Split (1)

train · 57.9k rows

id stringlengths 36 36	source stringclasses 15 values	formatted_source stringclasses 13 values	text stringlengths 2 7.55M
249d95db-2cf7-4baa-b00c-94b28c49cbf9	trentmkelly/LessWrong-43k	LessWrong	What is a “Good” Prediction? [My inside view is pretty confident in this; my outside view is very not. Cross-posted from Grand, Unified, Crazy.] Zvi’s post on Evaluating Predictions in Hindsight is a great walk through some practical, concrete methods of evaluating predictions. This post aims to be a somewhat more theoretical/philosophical take on the related idea of what makes a prediction “good”. Intuitively, when we ask whether some past prediction was “good” or not, we tend to look at what actually happened. If I predicted that the sun will rise with very high probability, and the sun actually rose, that was a good prediction, right? There is an instrumental sense in which this is true, but also an epistemic sense in which it is not. If the sun was extremely unlikely to rise, then in a sense my prediction was wrong – I just got lucky instead. We can formally divide this distinction as follows: * Instrumentally, a prediction was good if believing it guided us to better behaviour. Usually this means it assigned a majority probability to the thing that actually happened regardless of how likely it really was. * Epistemically, a prediction was good only if it matched the underlying true probability of the event in question. But what do we mean by “true probability”? If you believe the universe has fundamental randomness in it then this idea of “true probability” is probably pretty intuitive. There is some probability of an event happening baked into the underlying reality, and like any knowledge, our prediction is good if it matches that underlying reality. If this feels weird because you have a more deterministic bent, then I would remind you that every system seems random from the inside. For a more concrete example, consider betting on a sports match between two teams. From a theoretical, instrumental perspective there is one optimal bet: 100% on the team that actually wins. But in reality, it is impossible to perfectly predict who will win; either that information literally doesn’t e
a9de432d-5e5f-4f34-a81c-6a65f5507e1e	trentmkelly/LessWrong-43k	LessWrong	The No Free Lunch theorems and their Razor The No Free Lunch (NFL) family of theorems contains some of the most misunderstood theorems of machine learning. They apply to learning[1] and optimization[2] and, in rough terms, they state: > All algorithms for learning [respectively, optimization] do equally well at generalization performance [cost of the found solution] when averaged over all possible problems. This has some counterintuitive consequences. For example, consider a learning algorithm that chooses a hypothesis with the highest accuracy on a training set. This algorithm generalizes to a test set just as well as the algorithm which chooses the hypothesis with the lowest accuracy on training! Randomly guessing for every new instance is also just as good as these two. The NFL theorems thus seem to show that designing an algorithm that learns from experience is impossible. And yet, there exist processes (e.g. you, the reader of this sentence) which successfully learn from their environment and extrapolate to circumstances they never experienced before. How is this possible? The answer is that problems encountered in reality are not uniformly sampled from the set of all possible problems: the world is highly regular in specific ways. Successful learning and optimization processes (e.g. the scientific method, debate, evolution, gradient descent) exploit these regularities to generalize. If NFL theorems don't usually apply in reality, why should you care about them? My central claim is that they are essential to to think about why and when learning processes work. Notably, when analyzing some process, it is important to state the assumptions under which it can learn or optimize. Sometimes it is possible to test some of these assumptions, but ultimately unjustifiable assumptions will always remain. NFL theorems also allow us to quickly discard explanations for learning as incorrect or incomplete, if they make no reference to the conditions under which they apply. I call this the No Free Lunch Razor for
316479a4-75aa-4c96-947e-75157da89016	StampyAI/alignment-research-dataset/special_docs	Other	How We’re Predicting AI – or Failing to How We’re Predicting AI – or Failing to⋆ Stuart Armstrong1and Kaj Sotala2 1The Future of Humanity Institute, Faculty of Philosophy, University of Oxford, UK stuart.armstrong@stx.oxon.org 2The Singularity Institute, Berkeley, CA, USA kaj@singularity.org Abstract. This paperwill look atthevariouspredictionsthathavebeen made about AI and propose decomposition schemas for analysing them. It will propose a variety of theoretical tools for analysing, judging and improving these predictions. Focusing speciﬁcally on timeline predictions (dates given by which we should expect the creation of AI), it will show that there are strong theoretical grounds to expect predictions to be quite poor in this area. Using a database of 95 AI timeline predictions, it will show that these expectations are born out in practice: expert predictions contradict each other considerably, and are indistinguishable from non-expert predictions and past failed predictions. Predictions that AI lie 15 to 25 years in the future are the most common, from experts and non-experts alike. Keywords: artiﬁcial intelligence, predictions, experts, bias. 1 Introduction Predictionsaboutthefuturedevelopmentofartiﬁcialintelligenceareasconﬁdent astheyarediverse.StartingwithTuring’sinitialestimationofa30%passrateon Turingtestbytheyear2000[1],computerscientists,philosophersandjournalists have never been shy to oﬀer their own deﬁnite prognostics, claiming AI to be impossible [2] or just around the corner [3] or anything in between. Whatarewetomakeofthesepredictions?Whataretheyfor,andwhatcanwe gain from them? Are they to be treated as light entertainment, the equivalent of fact-free editorials about the moral decline of modern living? Or are there some useful truths to be extracted? Can we feel conﬁdent that certain categories of experts can be identiﬁed, and that their predictions stand out from the rest in terms of reliability? In this paper, we start oﬀ by proposing classiﬁcation schemes for AI predic- tions: what types of predictions are being made, and what kinds of arguments ⋆The authors wish to acknowledge the help and support of the Singularity Insti- tute, the Future of Humanity Institute and the James Martin School, as well as the individual advice of Nick Bostrom, Luke Muelhauser, Vincent Mueller, Anders Sandberg, Lisa Makros, Sean O’Heigeartaigh, Daniel Dewey, Eric Drexler and the online community of Less Wrong. c/circlecopyrtSpringer International Publishing Switzerland 2015 11 J. Romportl et al. (eds.), Beyond Artiﬁcial Intelligence, Topics in Intelligent Engineering and Informatics 9, DOI: 10.1007/978-3-319-09668-1 \_2 12 S. Armstrong and K. Sotala or models are being used to justify them. Diﬀerent models and predictions can result in very diﬀerent performances, and it will be the ultimate aim of thisproject to classify and analyse their varying reliability. Armed with this scheme, we then analyse some of these approaches from the theoretical perspective, seeing wh ether there are good reasons to believe or disbelieve their results. The aim is not simply to critique individual methods or individuals, but to construct a toolbox of assessment tools that will both enable us to estimate the reliability of a prediction, and allow predictors to come up with better results themselves. This paper, the ﬁrst in the project, looks speciﬁcally at AI timeline predic- tions: those predictions that give a date by which we should expect to see an actual AI being developed (we use AI in the old fashioned sense of a machine capable of human-comparable cognitive performance; a less ambiguous modernterm would be ‘AGI’, Artiﬁcial General Intelligence). With the aid of the biases literature, we demonstrate that there are strong reasons to expert that experts wouldnotbe showing particular skill the ﬁeld of AI timeline predictions. The task is simply not suited for good expert performance. Those theoretical results are supplemented with the real meat of the paper: a database of 257 AI predictions, made in a period spanning from the 1950s to the present day. This database was assembled by researchers from the Singularity Institute (Jonathan Wang and Brian Potter) systematically searching thoughthe literature, and is a treasure-troveof interesting results. A total of 95 of these can be considered AI timeline predictions. We assign to each of them a single ‘medianAI’date,whichthenallowsustodemonstratethatAIexpertpredictionsare greatly inconsistent with each other – and indistinguishable from non-expert performance, and past failed predictions. With thedata,wefurthertest twofolk theorems:ﬁrstlythat predictorsalways predict the arrival of AI just before their own deaths, and secondly that AI is always 15 to 25 years into the future. We ﬁnd evidence for the second thesis butnot for the ﬁrst. This enabled us to show that there seems to be no such thing as an “AI expert” for timeline predictions: no category of predictors stands out from thecrowd. 2 Taxonomy of Predictions 2.1 Prediction Types “There will never be a bigger plane built.” Boeing engineer on the 247, a twin engine plane that held ten people. The standard image of a prediction is some fortune teller staring deeply into the mists of a crystal ball, and d ecreeing, with a hideous cer tainty, the course of the times to come. Or in a more modern version, a scientist predicting the outcome of an experiment or an economist pronouncing on next year’s GDP ﬁgures. Butthese “at date X, Y will happen” are just on e type ofvalid prediction. In general, How We’re Predicting AI 13 a prediction is something that constrains our expectation of the future. Before hearing the prediction, we thought the future would have certain properties; butafter hearing and believing it, we now ex pect the future to be diﬀerent from our initial thoughts. Under this deﬁnition, conditional predictions – “if A, then B will happen” – are also perfectly valid. As are negative predictions: we might have believed initially that perpetual motion machines were possible, and imagined what they could be used for. But once we accept tha t one cannot violate conservation of energy, we have a diﬀerent picture of the future: one without these wonderful machines and all their fabulous consequences. For the present analysis, we will divide predictions about AI into four types: 1. Timelines and outcome predictions. These are the traditional types of pre- dictions, telling us when we will achieve speciﬁc AI milestones. Examples: An AI will pass the Turing test by 2000 [1]; Within a decade, AIs will bereplacing scientists and other thinking professions [4]. 2. Scenarios. These are a type of conditional predictions, claiming that if the conditions ofthe scenarioaremet, the n certain types ofoutcomeswill follow. Example: If we build a human-level AI that is easy to copy and cheap to run, this will cause mass unemployment among ordinary humans [5]. 3. Plans. These are a speciﬁc type of conditional prediction, claiming that if someone decides to implement a speciﬁc plan, then they will be successful in achieving a particular goal. Example: We can build an AI by scanning ahuman brain and simulating the scan on a computer [6]. 4. Issues and metastatements. This ca tegory covers relevant problems with (some or all) approaches to AI (including sheer impossibility results), andmetastatementsaboutthe wholeﬁeld.Examples:anAIcannotbebuiltwith- out a fundamental new understanding of epistemology [7]; Generic AIs will have certain (potentially dangerous) behaviours [8]. There will inevitably be some overlapbetween the categories,but this division is natural enough for our purposes. In this paper we will be looking at timeline predictions. Thanks to the eﬀorts of Jonathan Wang and Brian Potter at theSingularity Institute, the authors were able to make use of extensive databases of this type of predictions, reaching back from the present day back to the 1950s. Other types of predictions will be analysed in subsequent papers. 2.2 Prediction Methods Just as there are many types of predictions, there are many ways of arriving at them – consulting crystal balls, listening to the pronouncements of experts, constructing elaborate models. Our review of published predictions has shown that the prediction methods are far more varied than the types of conclusionsarrived at. For the purposes of this analysis, we’ll divide the prediction methods into the following loose scheme: 14 S. Armstrong and K. Sotala 1. Causal models 2. Non-causal models3. The outside view4. Philosophical arguments5. Expert authority6. Non-expert authority Causal model are the staple of physics: given certain facts about the situation under consideration (momentum, energy, charge, etc.) a conclusion is reached aboutwhattheultimatestatewillbe.Ifthefactswerediﬀerent,theendsituationwould be diﬀerent. But causal models are often a luxury outside of the hard sciences, whenever we lack precise understanding of the underlying causes. Some success can be achieved with non-causal models: without understanding what inﬂuences what, one can extrapolate trends into the future. Moore’s law is a highly successfulnon-causal model [9]. The outside view is a method of predicting that works by gathering together speciﬁc examples and claiming that they all follow the same underlying trend.For instance, one could notice the plethora of Moore’s laws across the spectrum of computing (in numbers of transistors, size of hard drives, network capacity, pixels per dollar, ...), note that AI is in the same category, and hence argue that AI development must follow a similarly exponential curve [10]. Philosophical arguments are common in the ﬁeld of AI; some are simple im- possibility statements: AI is decreed to be impossible for more or less plausible reasons. But the more thoughtful philosophical arguments point out problems that need to be resolved to achieve AI, highlight interesting approaches to doingso, and point potential issues if this were to be achieved. Many predictions rely strongly on the status of the predictor: their innate expertise giving them potential insights that cannot be fully captured in their arguments, so we have to trust their judgment. But there are problems in relying on expert opinion, as we shall see. Finally, some predictions rely on the judgment or opinion of non-experts. Journalists and authors are examples of this, but often actual experts will make claims outside their domainofexpertise. CEO’s,historians,physicists and math-ematicians will generally be no more accurate than anyone else when talking about AI, no matter how stellar they are in their own ﬁeld [11]. Predictions can use a mixture of these approaches,and often do. For instance, Ray Kurzweil’s‘Law of Time and Chaos’ uses the outside view to group together evolutionary development, technological development, and computing into the same category, and constructs a causal model predicting time to the ‘Singular- ity’ [10]. Moore’s law (non-causal model) is a key input to this Law, and Ray Kurzweil’s expertise is the main evidence for the Law’s accuracy. This is the schema we will be using in this paper, and in the prediction databases we have assembled. But the purpose of any such schema is to bring clarity to the analysis, not to force every prediction into a particular box. Wehope that the methods and approaches used in this paper will be of general use How We’re Predicting AI 15 to everyone wishing to analyse the reliability and usefulness of predictions, in AI and beyond. Hence this schema can be freely adapted or discarded if a particularprediction does not seem to ﬁt it, or if an alternative schema seems to be more useful for the analysis of the question under consideration. 3 A Toolbox of Assessment Methods The purpose of this paper is not only to assess the accuracy and reliability of some of the AI predictions that have already been made. The purpose is to start building a ‘toolbox’ of assessment methods that can be used more generally, applying them to current and future predictions. 3.1 Extracting Veriﬁable Predictions The focus of this paper is squarely on the behaviour of AI. This is not a philo- sophical point; we are not making the logical positivist argument that only em- piricallyveriﬁablepredictionshavemeaning[12].Butitmustbenotedthatmany of the vital questions about AI – can it be built, when, will it be dangerous, willit replace humans, and so on – all touch upon behaviour. This narrow focus has the added advantage that empirically veriﬁable predictions are (in theory) susceptible to falsiﬁcation, which mea ns ultimately agreementbetween people of opposite opinions. Predictions like these have a very diﬀerent dynamic to those that cannot be shown to be wrong, even in principle. To that end, we will seek to reduce the prediction to an empirically veriﬁable format. For some predictions, this is automatic: they are already in the correct format. When Kurzweil wrote “One of my key (and consistent) predictions isthat a computer will pass the Turing test by 2029,” then there is no need to change anything. Conversely, some philosophical arguments concerning AI, such assomeofthevariantsoftheChineseRoomargument[13],arearguedto contain noveriﬁablepredictionsatall:anAIthatdemonstratedperfecthumanbehaviour would not aﬀect the validity of the argument. And in between there are those predictio ns that are partially veriﬁable. Then the veriﬁable piece must be clearly extra cted and articulat ed. Sometimes it is ambiguitythatmustbeovercome:whenanauthorpredictsanAI“Omegapoint”in 2040 [14], it is necessary to read the paper with care to ﬁgure out what counts as an Omega point and (even more importantly) what doesn’t. Even purely philosophicalpredictions ca n have(or can be interpretedto have) veriﬁable predictions. One of the most famous papers on the existence of con- scious states is Thomas Nagel’s “What is it like to be a bat?” [15]. In this paper, Nagel argues that bats must have mental states, but that we humans can never understand what it is like to have these mental states. This feels purely philo- sophical, but does lead to empirical predictions: that if the bat’s intelligencewere increased and we could develop a common language, then at some point in the conversation with it, our understanding would reach an impasse. We would try to describe what our internal mental states felt like, but would always failto communicate the essence of our experience to the other species. 16 S. Armstrong and K. Sotala Many other philosophical papers can likewise be read as having empirical predictions; as making certain states of the world more likely or less – even if they seem to be devoid of this. The Ch inese Room argument, for instance, argues that formal algorithms will lack the consciousness that humans possess [13]. This may seem to be an entirely se lf-contained argument – but consider that a lot of human behaviour revolves around consciousness, be it discussing it, commenting on it, deﬁning it or intuitively noticing it in others. Hence if we believed the Chinese Roomargument, and wereconfrontedwith twoAI projects, one based on advanced algorithms and one based on modiﬁed human brains, we would be likely to believe that the second project is more likely to result in an intelligence that seemedconscious than the ﬁrst. This is simply because we wouldn’t believe that the ﬁrst AI could ever be conscious, and that it is easier to seem conscious when one actually is. And that gives an empirical prediction. Note that the authors of the predictions may disagree with our ‘extracted’ conclusions. This is not necessarily a gam e breaker. For instance, even if there is no formal link between the Chinese Room model and the prediction above, it’s still the case that the intuitive reasons for believing the model are also good reasons for believing the prediction. Our aim should always be to try and create useful veriﬁable predictions in any way we can. In this way, we can make use of much more of the AI literature. For instance, Lucas argues that AI is impossible because it could not recognise the truth of its own G¨ odel sentence1[ 1 6 ] .T h i si sa very strong conclusion, and we have to a ccept a lot of Lucas’s judgments before we agree with it. Replacing the conclusio n with the weaker (and veriﬁable) “self reference will be an issue with advanced AI, and will have to be dealt with somehow by the programmers” gives us a useful prediction which is more likely to be true. Care must be taken when applying this method: the point is to extract a useful veriﬁable prediction, not to weaken or strengthen a reviled or favoured argument. The very ﬁrst stratagems in Shopenhauer’s “The Art of Always being Right”[17]aretoextendandover-genera lisetheconsequencesofyouropponent’s argument; conversely, one should reduce and narrow down one’s own arguments. There is no lack of rhetorical tricks to uphold one’s own position, but if one is truly after the truth, one must simply attempt to ﬁnd the most reasonable empirical version of the argument; the truth-testing will come later. This method often increases uncertainty, in that it often narrows the conse- quences of the prediction, and allows more possible futures to exist, consistently with that prediction. For instance, Bruce Edmonds [18], building on the “No Free Lunch” results [19], demonstrates that there is no such thing as a universal intelligence: no intelligence that performs better than average in every circum- stance. Initially this seems to rule out AI entirely; but when one analyses what this means empirically, one realises there is far less to it. It does not forbid an 1AG ¨odel sentence is a sentence G that can be built in any formal system containing arithmetic. G is implicitly self-referential, as it is equivalent with “there cannot exist a proof of G”. By construction, there cannot be a consistent proof of G from within the system. How We’re Predicting AI 17 algorithm from performing better than any human being in any situation any human being would ever encounter, for instance. So our initial intuition, which was to rule out all futures with AIs in them, is now replaced by the realisation that we have barely put any constraints on the future at all. 3.2 Clarifying and Revealing Assumptions The previous section was co ncerned with the predictions’ conclusions. Here we will instead be looking at its assumptions, and the logical structure of the argu- ment or model behind it. The objective is to make the prediction as rigorous as possible Philosophers love doing this: taking apart argument, adding caveats and straightening out the hand-wavy logica l leaps. In a certain sense, it can be ar- gued that analytic philosophy is entirely about making arguments rigorous.One of the oldest methods in philosophy – the dialectic [20] – also plays this role, with concepts getting clariﬁed during the conversationbetween philosophersand various Athenians. Though this is perhaps philosophy’s greatest contribution to knowledge, it is not exclusively the hunting ground of philosophers. All rational ﬁelds of endeavour do – and should! – beneﬁt from this kind of analysis. Of critical importance is revealing hidden assumptions that went into the predictions. These hidden assumptions – sometimes called Enthymematic gaps in the literature [21] – are very important because they clarify where the true disagreements lie, and where we need to focus our investigation in order to ﬁnd out the truth of prediction. Too often, competing experts will make broad-based arguments that ﬂy past each other. This makes choosing the right argument a matter of taste, prior opinions and our admiration of the experts involved. But if the argument can be correctly deconstr ucted, then the source of the disagree- ment can be isolated, and the issue can be decided on much narrower grounds – and its much clearer whether the various experts have relevant expertise or not (see Section 3.4). The hidden assumptions are often implicit, so it is perfectly permissible to construct assumptions that the predictors were not consciously aware of using. Forexample,let’s lookagainattheG¨ odel argumentsmentioned in the Section 3.1. The argument shows that formal sy stems of a certain complexity must be either incomplete (unable to see that their G¨ odel sentence is true)orinconsistent (provingfalse statements). This iscontrastedwith humans, who– allegedly–use meta-reasoningto know that their own G¨ odel statements are t rue. It should ﬁrst be noted herethat noonehaswritten downan actual“humanG¨ odelstatement,” so we cannot be sure humans would actually ﬁgure out that it is true.2Also, humans are both inconsistent and able to deal with inconsistencies without a complete collapse of logic. In this, the y tend to diﬀer from AI systems, though some logic systems such as relevance logic do mimic the same behaviour [22]. In contrast, both humans and AIs are not logically omniscient – they are not capable of proving everything provable within their logic system (the fact that 2One could argue that, by deﬁnition, a human G¨ odel statement must be one that humans cannot recognise as being a human G¨ odel statement! 18 S. Armstrong and K. Sotala there are an inﬁnite number of things to prove being the problem here). So this analysis demonstrates the hidden assumption in Lucas’s argument: that thebehaviourofanactualcomputer programrunningon arealmachineismoreakin to that of a logically omniscient formal agent, than it would be to a real human being. That assumption may be ﬂawed or correct, but is one of the real sourcesof disagreement over whether G¨ odelian arguments rule out artiﬁcial intelligence. Again, it needs to be emphasised that the purpose is to clarify and analyse arguments, not to score points for one side or the other. It is easy to phrase assumptions in ways that sound good or bad for either “side”. It is also easy to take the exercise too far: ﬁnding more and more minor clariﬁcations or speciﬁchidden assumptions until the whole prediction becomes a hundred page mess of over-detailed special cases. The pur pose is to clarify the argument until it reaches the point where all (or most) parties could agree that these assumptionsare the real sources of disagreement. An d then we can consider what empirical evidence, if available, or expert opinio n has to say about these disagreements. There is surprisingly little published on the proper way of clarifying assump- tions, making this approach more an art than a science. If the prediction comes from a model, we have some standard tools available for clarifying, though [23]. Most of these methods work by varying parameters in the model and checking that this doesn’t cause a breakdown in the prediction. Model Testing and Counterfactual Resiliency Though the above works from inside the model, there are very few methods that can test the strength of a model from the outside. This is especially the case for non-causal models: what are the assumptions behind Moore’s famous law [9],or Robin Hanson’s model that we are due for another technological revolution, based on the timeline of previous revolutions [24]? If we can’t extract assump- tions,we’rereducedtosaying“thatfeelsright/wrongtome”,andthereforewe’re getting nowhere. The authors have come up with a putative way of testing the assumptions of such models (in the case of Moore’s law, the empirical evidence in favour is strong, but there is still the question of what is powering the law and whether it will cross over to new chip technologies again and again). It involves givingthe model a counterfactual resiliency check: imagining that world history had happened slightly diﬀerently, and checking whether the model would have stood up in those circumstances . Counterfactual changes are permitted to anything that the model ignores. The purpose of this exercise is not to rule out certain models depending on one’s own preferred understanding of history (e.g. “Protestantism was essential to the industrial revolution, and was a ﬂuke due to Martin Luther; so it’s very likely that the industrial revolution would not have happened in the way ortimeframe that it did, hence Hanson’s model – which posits the industrial rev- olutions’s dates as inevitable – is wrong”). Instead it is to illustrate the tension between the given model and other models of history (e.g. “The assumptionsthat Protestantism was both a ﬂuke and essential to the industrial revolution How We’re Predicting AI 19 are in contradiction with Hanson’s model. Hence Hanson’s model implies that either Protestantism was inevitable or that it was non-essential to the industrial revolution, an extra hidden assumption”). The counterfactual resiliency exercise has been carried out at length in an online post.3The general verdict seemed to be that Hanson’s model contradicted a lot of seemingly plausible assumptions about technological and social development. Moore’s law, on the other hand, seemed mainly dependent on the continu ing existence of a market economy and the absence of major catastrophes. This method is new, and will certainly be reﬁned in future. Again, the pur- pose of the method is not to rule out certain models, but to ﬁnd the nodes of disagreement. More Uncertainty Clarifying assumptions often ends up in creasing uncertainty , as does revealing hidden assumptions. The previous section focused on extracting veriﬁable pre- dictions, which often increases the range of possible worlds compatible with a prediction. Here, by clarifying and cav eatting assumptions, and revealing hid- den assumption, we reduce the number of worlds in which the prediction is valid. This means that the prediction puts fewer constraints on our expectations. In counterpart, of course, the caveatte d prediction is more likely to be true. 3.3 Empirical Evidence The gold standard in separating true predictions from false ones must always be empirical evidence. The scientiﬁc method has proved to be the best way of disproving false hypotheses, and should be used whenever possible. Other methods, such as expert opinion or unjustiﬁed models, come nowhere close. The problem with empirical evidence is that ... it is generally non-existent in the AI prediction ﬁeld. Since AI predictions are all about the existence and properties of a machine that hasn’t yet been built, that no-one knows how to build or whether it actually can be built, there is little opportunity for the whole hypothesis-prediction-testing cycle. This should indicate the great diﬃculties in the ﬁeld. Social sciences, for instance, ar e often seen as the weaker cousins of the hard sciences, with predictions much mo re contentious and less reliable. And yet the social sciences make use of the scientiﬁc method, and have access to some types of repeatable experiments. Thus a ny prediction in the ﬁeld of AI should be treated as less likely than any social science prediction. That generalisationis somewhatover-harsh.Some AI prediction methods hew closer to the scientiﬁc method, such as the whole brain emulations model [6] – it makes testable predictions along the way. Moore’s law is a wildly successful pre- diction, and connected to some extent wi th AI. Many predictors (e.g. Kurzweil) make partial predictions on the road towards AI; these can and should be as- sessed – track records allow us to give some evidence to the proposition “this 3Seehttp://lesswrong.com/lw/ea8/ counterfactual resiliency test for noncausal 20 S. Armstrong and K. Sotala expert knows what they’re talking about.” And some models also allow for a degree of testing. So the ﬁeld is not void of empirical evidence; it’s just that there is so little of it, and to a large extent we must put our trust in expert opinion. 3.4 Expert Opinion Reliance on experts is nearly unavoidable in AI prediction. Timeline predictions are often explicitly based on experts’ fee lings; even those that consider factors about the world (such as computer sp eed) need an expert judgment about why that factor is considered and not othe rs. Plans need experts to come up with them and judge their credibility. And unless every philosopher agrees on the correctness of a particula r philosophical argument, we are dependent to some degree on the philosophical judgment of the author. It is the purpose of all the methods described above that we can reﬁne and caveat a prediction, back it up with empirical evidence whenever possible, and thus clearly highlight the points where we need to rely on expert opinion. And so can focus on the last remaining points of disagreement: the premises themselves (that is of course the idealsituation:somepredictionsaregivendirectlywith nootherbasisbut expert authority, meaning there is nothing to reﬁne). Should we expect experts to be good at this task? There have been several projects over the last few decades to establish the domains and tasks where we would expect experts to have good performance [25, 26]. Table 1 summarises the results: Table 1. Table of task properties conducive to good and poor expert performance Good performance: Poor performance: Static stimuli Dynamic (changeable) stimuli Decisions about things Decisions about behaviour Experts agree on stimuli Experts disagree on stimuli More predictable problems Less predictable problems Some errors expected Few errors expected Repetitive tasks Unique tasks Feedback available Feedback unavailable Objective analysis available Subjective analysis only Problem decomposable Problem not decomposable Decision aids common Decision aids rare Not all of these are directly applicable to the current paper (are predictions about human level AIs predictions about things, or about behaviour?). One of themostimportantfactorsiswhetherexpe rtsgetfeedback,preferablyimmediate feedback. We should expect the best expert performance when their guesses are immediatelyconﬁrmedordisconﬁrmed.Whenfeedbackisunavailableordelayed, How We’re Predicting AI 21 or the environment isn’t one that gives good feedback, then expert performance drops precipitously [26, 11]. Table 1 applies to both domain and task. Any domain of expertise strongly in the right column will be one where we expect poor expert performance. But if the individual expert tries to move their own predictions into the left column(maybe by decomposing the problem as far as it will go, training themselves on related tasks where feedback is available...) they will be expected to perform better. In general, we should encourage this type of approach. When expertsfail,thereareoftensimplealgorithmicmodelsthat demonstrate better performance [27]. In these cases, the experts often just spell out theircriteria, design the model in consequence, and let the model give its predictions: this results in better predictions than simply asking the expert in the ﬁrst place. Hence we should also be on the lookout for experts who present their ﬁndingsin the form of a model. As everyone knows, experts sometimes disagree. This fact strikes at the very heart of their supposed expertise. We listen to them because they have the skillsand experience to develop correct insigh ts. If other experts have gone through the same process and come to an opposite conclusion, then we have to conclude that their insights do not derive from their skills and experience, and hence should be discounted. Now if one expert opinion is a fringe position held by only a few experts, we may be justiﬁed in dismissing it simply as an error. But ifthere are diﬀerent positions held by lar ge numbers of disagreeing experts, how are we to decide between them? We need some sort of objective criteria: we are not experts in choosing between experts, so we have no special skills in decidingthe truths on these sorts of controversial positions. What kind of objective criteria cou ld there be? A good track record can be an indicator, as is a willingness to make veriﬁable, non-ambiguous predictions. A better connection with empirical knowledge and less theoretical rigidity are also positive indications [28], and any expert that approached their task withmethods that were more on the left of the table than on the right should be expected to be more correct. But thes e are second order phenomena – we’re looking at our subjective interpretatio n of expert’s subjective opinion – so in most cases, when there are strong dis agreement between experts, we simply can’t tell which position is true. Grind versus Insight Some AI prediction claim that AI will result from grind: i.e. lots of hard work and money. Other claim that AI will need special insights: new unexpected ideas that will blow the ﬁeld wide open [7]. In general, we are quite good at predicting grind. Project managers and var- ious leaders are often quite good at estimating the length of projects (as longas they’re not directly involved in the pro ject [29]). Even for relatively creative work, people have suﬃcient feedback to ha zard reasonable guesses. Publication dates for video games, for instance, though often over-optimistic, are generallynot ridiculously erroneous – even though video games involve a lot of creative 22 S. Armstrong and K. Sotala design, play-testing, art, programming the game “AI”, etc. Moore’s law could be taken as an ultimate example of grid: we expect the global eﬀorts of manyengineers across many ﬁelds to average ou t to a rather predictable exponential growth. Predicting insight, on the other hand, seems a much moredaunting task. Take the Riemann hypothesis, a well-established mathematical hypothesis from 1885, [30]. How would one go about estimating how long it would take to solve? How about the P=NPhypothesis in computing? Mathematicians seldom try and predict when major problems will be solved, because they recognise that insight is very hard to predict. And even if predictions could be attempted (the age ofthe Riemann’s hypothesis hints that it probably isn’t right on the cusp of being solved), they would need much larger e rror bars than grind predictions. If AI requires insights, we are also handicapped by the fact of not knowing what theseinsights are (unlike the Riemann hypothe sis, where the hypothesis is clearly stated, and only the proof is missing). This could be mitigated somewhat if we assumed there wereseveraldiﬀerent insights, each of which could separatelyleadto AI. But we would need good grounds to assume that. Does this mean that in general predictions that are modelling grind should be accepted more than predictions that are modelling insight? Not at all. Pre- dictions that are modelling grind should only be accepted if they can make a good case that producing an AI is a matter grind only. The predictions aroundwhole brain emulations [6], are one of the few that make this case convincingly; this will be analysed in a subsequent paper. Non-experts Opinion It should be born in mind that all the caveats and problems with expert opinion apply just as well to non-experts. With o ne crucial diﬀerence: we have no reason to trust the non-expert’s opinion in the ﬁrst place. That is not to say that non-experts cannot come up with good models, convincing timelines, or interesting plans and scenarios. It just means that our assessment of the quality of the prediction depends only on what we are given; we cannot extend a non-expertany leeway to cover up a weak premise or a faulty logical step. To ensure this, we should try and assess non-expert predictions blind, without knowing who the author is. If we can’t blind them, we can try and get a similar eﬀect by askingourselves hypothetical questions such as: “Would I ﬁnd this prediction more or less convincing if the author was the Archbishop of Canterbury? What if it was Warren Buﬀet? Or the Unabomber?” We should aim to reach the point where hypotheticalchangesinauthorshipdonotaﬀectourestimationofthe prediction. 4 Timeline Predictions The practical focus of this paper is on AI timeline predictions: predictions givingdates for AIs with human-comparable cognitive abilities. Researchers from the Singularity Institute have assembled a database of 25AI predictions since 1950,of which 95 include AI timelines. How We’re Predicting AI 23 4.1 Subjective Assessment A brief glance at Table 1 allows us to expect that AI timeline predictions will generally be of very poor quality. The only factor that is unambiguously positive for AI predictions is that prediction errors are expected and allowed: apart from that, the task seems singularly diﬃcult, especially on the key issue of feedback. An artiﬁcial intelligence is a hypothetical machine, which has never existed on this planet before and about whose properties we have but the haziest impres- sion. Most AI experts will receive no fee dback whatsoever about their predic- tions, meaning they have to construct them entirely based on their untested impressions. Thereisnothingstoppingexpertsfromdecomposingtheproblem,orconstruct- ing models which they then calibrate with available data, or putting up interim predictionstotesttheirassessment.Andsomedousethesebetterapproaches(see for instance [10, 5, 31]). But a surprisingly large number don’t! Some predictions are unabashedly based simply on the feelings of the predictor [32, 33]. Yet another category are of the “Moore’s law hence AI” type. They postulate that AI will happen when computers rea ch some key level, often comparing with some key property of the brain (number of operations per second [34], or neurones/synapses4).Inthedivisionestablishedinsection3.4,thisispure‘grind’ argument: AI will happen after a certain amount of work is performed. But, as we saw, these kinds of arguments are only valid if the predictor has shown that reaching AI does not require new insights! And that step is often absent from the argument. 4.2 Timeline Prediction Data The above were subjective impressions, formed while looking over the whole database.To enablemore rigorousanalysis,the varioustimeline predictions were reduced to a single number for purposes of comparison: this would be the date upon which the predictor expected ‘human level AI’ to be developed. Unfortunately not all the predictions were in the same format. Some gave ranges, some gave median estimates, some talked about superintelligent AI, oth- ers about slightly below-human AI. In order to make the numbers comparable, one of the authors (Stuart Armstrong) went through the list and reduced the various estimates to a single number. He followed the following procedure to extract a “Median human-level AI estimate”: When a rangewas given,he took the mid-point of that range (rounded down). Ifayearwasgivenwith a50%likelihoodestimate,he tookthatyear.Ifit wasthe collection of a variety of expert opinions, he took the prediction of the median expert. If the predictor foresaw some sort of AI by a given date (partial AI or superintelligent AI), and gave no other estimate, he took that date as their estimate rather than trying to correct it in one direction or the other (there were 4See for instance Dani Eder’s 1994 Newsgroup posting http://www.aleph.se/Trans/Global/Singularity/singul.txt 24 S. Armstrong and K. Sotala roughly the same number of subhuman AIs as suphuman AIs in the list, and not that many of either). He read extracts of the papers to make judgement calls when interpreting problematic statements like “within thirty years” or “during this century” (is that a range or an end-date?). Every date selected was either an actual date given by the predictor, or the midpoint of a range.5 It was also useful to distinguish between popular estimates, performed by journalists, writers or amateurs, from those predictions done by those with ex- pertiseinrelevantﬁelds(AI research,co mputersoftwaredevelopment,etc.)Thus each prediction was noted as ‘expert’ or ‘n on-expert’; the expectation being that experts would demonstrate improve d performance over non-experts. Figure 1 graphs the results of this exer cise (the range has been reduced; there weresevenpredictionssetting dates beyondthe year2100,three ofthem expert.) Fig. 1.Median estimate for human-level AI, graphed against date of prediction As can be seen, expert predictions span the whole range of possibilities and seem to have little correlation with each other. The range is so wide – ﬁfty year gaps between predictions are commo n – that it provides strong evidence that experts are not providing good pred ictions. There does not seem to be any visible diﬀerence between expert and non-e xpert performance ei ther, suggesting that the same types of reasoning may be used in both situations, thus negating the point of expertise. 5The data can be found at http://www.neweuropeancentury.org/SIAI-FHI\_AI\_predictions.xls ; readers are encouraged to come up with their own median estimates. How We’re Predicting AI 25 Twoexplanationshavebeengenerallyad vancedto explainpoorexpertperfor- mance in these matters. The ﬁrst, the so-called Maes-Garreau law6posits that AI experts predict AI happening towards the end of their own lifetime. This would make AI into a technology that would save them from their own deaths, akin to a ‘Rapture of the Nerds’. The second explanation is that AI is per petually ﬁfteen to twenty-ﬁve years into the future. In this way (so the explanation goes), the predictor can gain credit for working on something that will be of relevance, but without any possi- bility that their prediction could be shown to be false within their currentcareer. We’ll now look at the evidence for these two explanations. Nerds Don’t Get Raptured Fifty-ﬁve predictions were retained, in which it was possible to estimate the pre- dictor’s expected lifespan. Then the diﬀ erence between their median prediction and this lifespan wascomputed (a positive diﬀerence meaning they would expect to die before AI, a negative diﬀerence meaning they didn’t). A zero diﬀerence would be a perfect example ofthe Maes-Ga rreaulaw: the predictorexpects AI to be developed at the exact end of their life. This number was then plotted again thepredictor’sageinFigure2(theplotwasrestrictedtothosepredictionswithin thirty years of the predic tor’s expected lifetime). From this, it can be seen that the Maes-Garreau law is not born out by the evidence: only twelve predictions (22% of the total) were within ﬁve years in either direction of the zero point. Twenty Years to AI The ‘time to AI’ was computed for each expert prediction. This was graphed in Figure 3. This demonstrates a deﬁnite increase in the 16–25 year predictions: 21 of the 62 expert predictions were in that range (34%). This can be considered weak evidence that experts do indeed pr efer to predict AI happening in that range from their own time. But the picture gets more damning when we do the same plot for the non- experts,asinFigure4.Here,13 ofthe33predictions areinthe 16–25yearrange. But more disturbingly, the time to AI graph is almost identical for experts and non-experts! Though this does not preclude the possibility of experts being more accurate,itdoes hintstronglythatexperts andnon-expertsmaybe using similar psychological procedures when creating their estimates. The next step is to look at failed predictions. There are 15 of those, most dating to before the ‘AI winter’ in the eighties and nineties. These have been graphed in Figure 5 – and there is an uncanny similarity with the other two graphs! So expert predictions are not only indistinguishable from non-expert predictions, they are also indistinguishable from past failed predictions. Hence it is not unlikely that recent predictions are suﬀering from the same biases and errors as their predecessors 6Kevin Kelly, editor of Wired magazine, cr eated the law in 2007 af ter being inﬂuenced by Pattie Maes at MIT and Joel Garreau (author of Radical Evolution). 26 S. Armstrong and K. Sotala Fig. 2.Diﬀerence between the predicted time to AI and the predictor’s life expectancy, graphed against the predictor’s age Fig. 3.Time between thearrival of AI andthe date the prediction was made, for expert predictors How We’re Predicting AI 27 Fig. 4.Time between the arrival of AI and the date the prediction was made, for non-expert predictors Fig. 5.Time between the arrival of AI and the date the prediction was made, for failed predictions 28 S. Armstrong and K. Sotala 5C o n c l u s i o n Thispaper,theﬁrstinaseriesanalysingAIpredictions,focusedonthe reliability of AI timeline predictions (predicting the dates upon which ‘human-level’ AIwould be developed). These predictions are almost wholly grounded on expert judgment. The biases literature classiﬁed the types of tasks on which experts wouldhavegoodperformance,and AItimeline predictionshaveallthehallmarks of tasks on which they would perform badly. This was born out by the analysis of 95 timeline predictions in the database assembled by the Singularity Institute. There were strong indications therein that experts performed badly. Not only w ere expert predictions spread across a wide range and in strong disagreement w ith each other, but there was evidence that experts were systematically preferring a ‘15 to 25 years into the future’ prediction. In this, they were indistinguishable from non-experts, and from past predictionsthatareknowntohavefailed.Thereisthusnoindicationthatexperts brought any added value when it comes to estimating AI timelines. On the other hand, another theory – that experts were systematically predicting AI arrivaljust before the end of their own lifetime – was seen to be false in the data we have. There is thus strong grounds for dramatically increasing the uncertainty in any AI timeline prediction. References 1. Turing, A.: Computing machinery and intelligence. Mind 59, 433–460 (1950) 2. Jacquette, D.: Metamathematical criteria for minds and machines. Erkenntnis 27(1) (1987) 3. Darrach, B.: Meet Shakey, the ﬁrst electronic person. Reﬂections of the Future (1970) 4. Hall, J.S.: Further reﬂections on the timescale of AI. In: Dowe, D.L. (ed.) Solomonoﬀ Festschrift. LNCS, vol. 7070, pp. 174–183. Springer, Heidelberg (2013) 5. Hanson, R.: What if uploads come ﬁrst: The crack of a future dawn. Extropy 6(2) (1994) 6. Sandberg, A.: Whole brain emulations: A roadmap. Future of Humanity Institute Technical Report 2008-3 (2008) 7. Deutsch, D.: The very laws of physics imply that artiﬁcial intelligence must be possible. What’s holding us up? Aeon (2012) 8. Omohundro, S.: Basic ai drives. In: Proceedings of the First AGI Conference, vol. 171 (2008) 9. Moore, G.: Cramming more components onto integrated circuits. Electronics 38(8) (1965) 10. Kurzweil, R.: The Age of Spiritual Machines: When Computers Exceed Human Intelligence. Viking Adult (1999) 11. Kahneman, D.: Thinking, Fast and Slow. Farra, Straus and Giroux (2011)12. Carnap, R.: The Logical Structure of the World (1928)13. Searle, J.: Minds, brains and programs. Behavioral and Brain Sciences 3(3), 417– 457 (1980) How We’re Predicting AI 29 14. Schmidhuber, J.: Artiﬁcial General Intelligence, pp. 177–200 (2006) 15. Nagel, T.: What is it like to be a bat? The Philosophical Review 83(4), 435–450 (1974) 16. Lucas, J.: Minds, machines and G¨ odel. Philosophy XXXVI, 112–127 (1961) 17. Schopenhauer, A.: The Art of Being Right: 38 Ways to Win an Argument (1831) 18. Edmonds, B.: The social embedding of intelligence. In: Parsing the Turing Test, pp. 211–235. Springer, Netherlands (2009) 19. Wolpert, D., Macready, W.: No free lunch theorems for search (1995) 20. Plato: The Republic (380 BC)21. Fallis, D.: Intentional gaps in mathematical proofs. Synthese 134(1-2) (2003) 22. Routley, R., Meyer, R.: Dialectical logic, classical logic, and the consistency of the world. Studies in East European Thought 16(1-2) (1976) 23. Morgan, M., Henrion, M.: Uncertainty: A guide to dealing with uncertainty in quantitative risk and policy analysis. Cambridge University Press (1990) 24. Hanson, R.: Economics of brain emulations. In: Unnatrual Selection - The Chal- lenges of Engineering Tomorrow’s People, pp. 150–158 (2008) 25. Shanteau, J.: Competence in experts: The role of task characteristics. Organiza- tional Behavior and Human Decision Processes 53, 252–266 (1992) 26. Kahneman, D., Klein, G.: Conditions for intuitive expertise: A failure to disagree. American Psychologist 64(6), 515–526 (2009) 27. Grove, W., Zald, D., Lebow, B., Snitz, B., Nelson, C.: Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment 12, 19–30 (2000) 28. Tetlock, P.: Expert political judgement: How good is it? How can we know? (2005) 29. Buehler, R., Griﬃn, D., Ross, M.: Exploring the planning fallacy: Why people underestimate their task completion times. Journal of Personality and Social Psy- chology 67, 366–381 (1994) 30. Riemann, B.: ¨Uber die Anzahl der Primzahlen unter einer gegebenen Gr¨ osse. Monatsberichte der Berliner Akademie (1859) 31. Waltz, D.: The prospects for building truly intelligent machines. Daedalus 117(1) (1988) 32. Good, J.: The scientist speculates: An anthology of partly-baked ideas. Heinemann (1962) 33. Armstrong, S.: Chaining god: A qualitative approach to AI, trust and moral sys- tems (2007) (online article) 34. Bostrom, N.: How long before superintelligence? International Journal of Futures Studies 2 (1998)
6f4e7f01-60e2-4f98-ae4a-c9f85a8bfc16	trentmkelly/LessWrong-43k	LessWrong	Meetup : Ottawa Weekly Monday LessWrong Meetup Discussion article for the meetup : Ottawa Weekly Monday LessWrong Meetup WHEN: 26 September 2011 07:30:00PM (-0400) WHERE: Private residence in the Elgin & Gladstone area, Ottawa, ON Matthew Kelly (http://memoryechoes.blogspot.com/) will be present to talk about his work in cognitive science. For details on location, (join and) see the mailing list: http://groups.google.com/group/less-wrong-ottawa Discussion article for the meetup : Ottawa Weekly Monday LessWrong Meetup
c0af121c-a813-45de-9963-0a33c9383d01	trentmkelly/LessWrong-43k	LessWrong	The hour I first alieved 0 Summary Many believe that religious behavior is only justified if one has religious belief. I disagree, argue that religion should be managed according to its membership in what I call the category of “Adaptive Distorted States of Perception”, and attempt to make the rationalist case for moderate orthopraxis. I Intro One of my philosophy professors at university adopts a seemingly pathological behavior; he attends Church every Sunday, despite lacking belief in God. It may be worth mentioning that the university in question is not a Christian college, but rather your run-of-the-mill secularized Ivy. Of course, he is not alone in doing something like this. But, perhaps somewhat uniquely, he does so neither due to a failure to parse his beliefs nor because of inertia. Instead, he chose to adopt this practice as a result of first principles reasoning, reasoning which I ultimately do not find compelling, but which nonetheless centers a question which I believe allows for a productive reframing of the problem: Do we make a category error when we speak of religious belief? Here, I’ll seek to argue that we ask too much of religious experience when we seek to espouse global beliefs which accord with it. I will seek to instead identify a category of adaptive imperatives, place religious experience in this category, and suggest that we ought to evaluate its value and its place in life in the same way in which we evaluate that of its categorical peers, such as romance and appreciation of art. II Alief and Possession What do romance and appreciation of art have in common? They are rationalizable only if with alief; that is to say, by some perturbation to the rationally-assented-to world-map. On some level, many people recognize that infatuated love presupposes strange beliefs — that the object of our love is particularly fascinating, good, or even divine among the alternatives. Jungian Psychology seems silly until one remembers the vivid imagery which the mind constell
4473ae28-414f-41a1-a40d-f3221fed5667	trentmkelly/LessWrong-43k	LessWrong	Privacy vs proof of character Below, I outline some musings I’ve produced while thinking about the private cryptocurrency Monero. I make no pretense to originality, cleverness, or exhaustiveness in covering considerations. There are (at least) two distinct types of privacy. For instance, suppose you have an item in your house that you want nobody to see, even an intruder. One way you could handle this would be to buy a large black safe and hide the items in the safe, locked with a key that only you possess, and leave the safe on your desk. The other way would be to build a similar safe into a wall, and put a framed portrait in front of the safe. In the first case, an intruder can tell that they can’t open the black safe, but they know that you have something that you wish to keep hidden. In the second case, unless the intruder is perceptive enough to look behind the portrait, they will not even be able to tell that you have anything to hide. One could imagine an even better-concealed safe that would take painstaking effort to find - for the sake of this post, please imagine that it is very difficult to imagine that there might be anything on the wall behind a portrait. You might prefer to own the second type of safe rather than the first, because the very fact that you possess something that you do not want others to see conveys almost as much as the sight of the item that you want to conceal. For instance, if you own something that proves that you violate certain social norms, the reason that you want to conceal it is to make others think that you actually abide by social norms - but if people only conceal things that do not abide by social norms, the fact that you are concealing something demonstrates that you do not, the very fact that you wanted others to remain ignorant of! The hidden safe has another interesting property. If hidden safes exist, and it would be possible for you to install one in your home without others finding out, then it becomes extremely difficult for you to prove to
9eafbc21-c518-44ba-8a0d-abbae4c7b50b	trentmkelly/LessWrong-43k	LessWrong	How can I get help becoming a better rationalist? Well… Right now, being ‘a rationalist’ could be said to be a massive part of my identity, at least judging by the absurd amount of time I’ve spent reading posts here, or SSC/ACX, and in a few other places. Yet, I’m still a mere lurker unfamiliar with most of the local customs. But it’s not what matters. What does is that I’m a terrible rationalist. You see, rationality takes practice. And reading stuff on LW isn’t practice at all. If anything, it’s just a great way of filling my brain with a lot of useful concepts, and then either blame myself for not using them, or use them for something entirely unrelated to their normal purpose. Often, to make myself feel worse, and think worse. As the saying goes, rationality is a martial art. Learning it by reading the rules, or by watching other people apply the rules, is about as effective as developing one’s muscles by watching sports on TV. I know of the CFAR, and of various related groups, meetups for ACX readers or for other people, etc. But, apart from ACX meetups, which aren’t about being better rationalists per se, I don’t have easy access to any of those, or certainly to a general environment which welcomes this. You know, not being in the Bay Area and all. And yet, I want to be more rational as much as anyone who’s been lurking here for five years wants it, and given how depressed I was until very recently, I probably badly need it, too. I’m not sure what kind of answers I expect, but, like, how can I push myself to learn more, and especially to practice more, and ideally to actually use rationality to improve my life?
668af6e1-82b8-422d-9e35-50900c9aadf6	trentmkelly/LessWrong-43k	LessWrong	Can we use Variolation to deal with the Coronavirus? Credit for the original idea comes from this thread: https://www.reddit.com/r/slatestarcodex/comments/fk5au8/corona_variolation/ The idea is to use variolation to innoculate young people and build up herd immunity to the virus. Perhaps with an infected scratch to the shin. I imagine this would be completely politically infeasible, but ignoring that, can anyone weigh in on whether this would work or not? I thought I read somewhere that getting infected with the virus through the eyes was far less lethal than getting infected through the lungs, but I can't find the source ATM.
0494c2e3-f467-4f36-a6d4-31e99a3296d3	trentmkelly/LessWrong-43k	LessWrong	Open problem: thin logical priors ---------------------------------------- author: * 'Tsvi Benson-Tilsen' title: 'Open problem: thin logical priors' ... * Background / Motivation * Problem statement * Type signature * Desiderata * Comments In short, and at a high level, the problem of thin priors is to understand how an agent can learn logical facts and make use of them in its predictions, without setting up a reflective instability across time. Before the agent knows the fact, it is required by logical uncertainty to “care about” worlds where the fact does not hold; after it learns the fact, it might no longer care about those worlds; so the ignorant agent has different goals than the knowing agent. This problem points at a hole in our basic understanding, namely how to update on logical facts; logical induction solves much of logical uncertainty, but doesn’t clarify how to update on computations, since many logical facts are learned “behind the scenes” by traders. ***** The ideas in this post seem to have been discussed for some time. Jessica brought them up in a crisper form in a conversation a while go with me, and also came up with the name; this post is largely based on ideas in that conversation and some subsequent ones with other people, possibly refined / reframed. Background / Motivation It would be nice to have a reflectively stable decision theory (i.e. a decision theory that largely endorses itself to continue making decisions over other potential targets of self-modification); this the most basic version of averting / containing instrumental goals, which is arguably necessary in some form to make a safe agent. Agents that choose policies using beliefs that have been updated on (logical) observations seem to be unstable, presenting an obstacle. More specifically, we have the following line of reasoning: * Updating on empirical evidence leads to reflective instability. If A1 is uncertain about the future even given all its observations; and its future instantiati
610ee8f5-be0e-482e-a95e-9f9e4122f66e	trentmkelly/LessWrong-43k	LessWrong	Another problem with CDT, involving Bell's Theorem Cavalcanti (2010) describes another problem with causal decision theory: > I apply some of the lessons from quantum theory, in particular from Bell’s theorem, to a debate on the foundations of decision theory and causation. By tracing a formal analogy between the basic assumptions of causal decision theory (CDT)—which was developed partly in response to Newcomb’s problem—and those of a local hidden variable theory in the context of quantum mechanics, I show that an agent who acts according to CDT and gives any nonzero credence to some possible causal interpretations underlying quantum phenomena should bet against quantum mechanics in some feasible game scenarios involving entangled systems, no matter what evidence they acquire. As a consequence, either the most accepted version of decision theory is wrong, or it provides a practical distinction, in terms of the prescribed behaviour of rational agents, between some metaphysical hypotheses regarding the causal structure underlying quantum mechanics.
515ba650-3032-400b-826a-a7a82aed1efd	StampyAI/alignment-research-dataset/aisafety.info	AI Safety Info	What is behavioral cloning? Behavioral cloning is the process of gathering observations of the behavior of an “expert demonstrator” who is good at an underlying task, and then using [supervised learning](https://en.wikipedia.org/wiki/Supervised_learning) to train an agent to “imitate” this behavior. Behavioral cloning is one way in which we can implement [imitation learning](/?state=8AER&question=What%20is%20imitation%20learning%3F). There are other ways, such as [inverse reinforcement learning (IRL)](/?state=8AET&question=What%20is%20inverse%20reinforcement%20learning%20(IRL)%3F) or cooperative inverse reinforcement learning (CIRL). Unlike IRL and CIRL, the goal behind behavioral cloning is to replicate the demonstrator's behavior as closely as possible, regardless of what the demonstrator’s goals might be. Behavioral cloning was originally developed to train self-driving cars. This use case also serves as a good simple example of how behavioral cloning works. We tell a human demonstrator to drive a car around. While the demonstrator is driving, we collect data about the environment state from sensors such as [Lidar](https://en.wikipedia.org/wiki/Lidar) and cameras, as well as about the actions that the demonstrator took in each of those states. The latter comes in the form of steering wheel movements, gears used, etc. This allows us to form a data set that consists of (state, action) pairs. At this point, we can use supervised learning to train a model that takes an input of the environment state and predicts the driver’s action. When the accuracy of the model is high enough, we can say that the driver’s behavior has been “cloned” into a machine through learning, thus the name behavioral cloning. After the initial pre-training process, [large language models (LLMs) are often fine-tuned using supervised learning](https://www.google.com/url?q=https://docs.google.com/document/d/15LCjCm2JkRAk1rUaoPnoPoVxMvk87zpAwyKMAjwCHO4/edit&sa=D&source=docs&ust=1684770454771202&usg=AOvVaw2mYPciwgCJ8W2I_6QHNaOF). This involves imitating the behavior of some expert who provides a set of appropriate (prompt, completion) pairs. As an example, after learning how to predict text from the internet, LLMs can be fine-tuned to follow instructions by copying humans. Conversations with LLMs such as ChatGPT can feel like you are talking to a conscious entity, but this is basically a result of cloning human speech patterns. The model is just imitating and then echoing the speech patterns of many different human demonstrators. Overall, behavioral cloning is a straightforward technique that gives us a good baseline for what we should expect from imitation-based algorithms. However, it does have several limitations and problems. [This discussion](/?state=94ZA&question=What%20are%20the%20limitations%20of%20behavioral%20cloning%3F) goes into more detail about problems that arise in modern LLMs that are fine-tuned using behavioral cloning. ## Sources - [Stanford CS234: Reinforcement Learning (2019) , Lecture 7 - Imitation Learning](https://www.youtube.com/watch?v=V7CY68zH6ps) - [Berkeley CS182: Reinforcement Learning (2021), Lecture 14 - Imitation Learning](https://www.youtube.com/watch?v=kGc8jOy5_zY) - leogao (2021). [Behavior Cloning is Miscalibrated](https://www.alignmentforum.org/posts/BgoKdAzogxmgkuuAt/behavior-cloning-is-miscalibrated) - Ortega, Pedro, et al. (2021). "[Shaking the foundations: delusions in sequence models for interaction and control](https://arxiv.org/abs/2110.10819).” - Zhou, Chunting, et al. (2020). "[Detecting Hallucinated Content in Conditional Neural Sequence Generation](https://arxiv.org/abs/2011.02593)" - Xiao, Yijun, and Wang, William. (2021) "[On Hallucination and Predictive Uncertainty in Conditional Language Generation.](https://arxiv.org/abs/2103.15025v1)"
1fc2c64a-9240-4d97-86e7-025f2dbe541a	trentmkelly/LessWrong-43k	LessWrong	Toy Models of Feature Absorption in SAEs TLDR; In previous work, we found a problematic form of feature splitting called "feature absorption" when analyzing Gemma Scope SAEs. We hypothesized that this was due to SAEs struggling to separate co-occurrence between features, but we did not prove this. In this post, we set up toy models where we can explicitly control feature representations and co-occurrence rates and show the following: * Feature absorption happens when features co-occur. * If co-occurring feature magnitudes vary relative to each other, we observe "partial absorption", where a latent tracking a main feature sometimes fires weakly instead of not firing at all, but sometimes does fully not fire. * Feature absorption happens even with imperfect co-occurrence, depending on the strength of the sparsity penalty. * Tying the SAE encoder and decoder weights together solves feature absorption in toy models. All code for this post can be seen in this Colab notebook. The rest of this post will assume familiarity with Sparse Autoencoders (SAEs). But first, some background on feature absorption: What is feature absorption? Feature absorption is a problematic form of feature splitting where a SAE latent appears to track an interpretable concept, but actually has holes in its recall. Instead, other SAE latents fire on specific tokens and "absorb" the feature direction into approximately token-aligned latents. For instance, in Gemma Scope SAEs we find a latent which seems to track the feature that a token "starts with S". However, the latent will not fire on a few specific tokens that do start with S, like the token "short". In feature absorption, we find gerrymandered SAE latents which appear to track an interpretable concept but have holes in their recall. Here, we see the dashboard for Gemma Scope layer 3 16k, latent 6510 which appears to track "starts with S", but mysteriously doesn't fire on "_short". How is this different than traditional feature splitting? In traditional feature splittin
32205d95-08ec-4c7a-b0c3-250902172535	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Questions about AI that bother me As 2022 comes to an end, I thought it'd be good to maintain a list of "questions that bother me" in thinking about AI safety and alignment. I don't claim I'm the first or only one to have thought about them. I'll keep updating this list. (The title of this post alludes to the book "[Things That Bother Me](https://www.penguinrandomhouse.com/books/563023/things-that-bother-me-by-galen-strawson/)" by Galen Strawson) First posted: 12/6/22 Last updated: 1/30/23 ### General Cognition * What signs do I need to look for to tell whether a model's cognition has started to emerge, e.g., situational awareness? * Will a capacity for "doing science" be sufficient condition for general intelligence? * How easy was it for humans to get science (e.g., compared to evolving to take over the world). ### Deception * What kind of interpretability tools do we need to avoid deception? * How do we get these interpretability tools and even if we do get them, what if they're like neuroscience for understanding brains (not enough)? * How can I tell whether a model has found another goal to optimize for during its training? * What is it that makes a model switch to a goal different from the one set by the designer? How do you prevent it from doing so? ### Agent Foundations * Is the description/modeling of an agent ultimately a mathematical task? * From where do human agents derive their goals? * [Is value fragile](https://www.lesswrong.com/posts/GNnHHmm8EzePmKzPk/value-is-fragile)? ### Theory of Machine Learning * What explains the success of deep neural networks? * Why was connectionism unlikely to succeed? ### Epistemology of Alignment (I've written about this [here](https://forum.effectivealtruism.org/s/sC8KoZx9jAdrEtmHj)) * How can we accelerate research? * Has philosophy ever really helped scientific research e.g., with concept clarification? * What are some concrete takeaways from the history of science and technology that could be used as advice for alignment researchers and field-builders? * [The emergence of the AI Safety paradigm](https://forum.effectivealtruism.org/s/sC8KoZx9jAdrEtmHj/p/pC9RJdmP3rnhuHpCm ) ### Philosophy of Existential Risk * What is the best way to explain the difference between forecasting extinction scenarios and narratives from chiliasm or escatology? * What is the best way to think about serious risks in the future without reinforcing a sense of doom? ### Teaching and Communication * Younger people (e.g., my undergraduate students) seem more willing to entertain scenarios of catastrophes and extinction compared to older people (e.g., academics). I find that strange and I don't have a good explanation as to why that is the case. * The idea of a technological singularity was not difficult to explain and discuss with my students. I think that's surprising given how powerful the weirdness heuristic is. * The idea of "agency" or "being an agent" was easy to conflate with "consciousness" in philosophical discussions. It's not clear to me why that was the case since I gave a very specific definition of agency. * Most of my students thought that AI models will never be conscious; it was difficult for them to articulate specific arguments about this, but their intuition seemed to be that there's something uniquely human about consciousness/sentience. * The "AIs will take our jobs in the future" seems to be a very common concern both among students and academics. * 80% of a ~25 people classroom thought that philosophy is the right thing to major in if you're interested in how minds work. The question I asked them was: "should you major in philosophy or cognitive science if you want to study how minds work?" ### Governance/Strategy * Should we try to slow down AI progress? What does this mean in concrete steps? * How should we go about capabilities externalities? * How should concrete AI risk stories inform/affect AI governance and short-term/long-term future planning?
4c7e92cc-0e31-4039-9d26-266400859a9b	trentmkelly/LessWrong-43k	LessWrong	April Fools - Harry Potter and the Methods of Rationality Joke Chapter This was an april fools joke. This is a new thread to discuss Eliezer Yudkowsky’s my chapter of Harry Potter and the Methods of Rationality and anything related to it. This thread is intended for discussing the fake April Fools chapter, which is now published. I suggest refraining from reading comments here until you read chapter 82. After you've read chapter 82, I suggest all discussion of this chapter to be kept here, with links to comments in the previous thread.
29929205-748a-419a-9549-6b602bac560c	trentmkelly/LessWrong-43k	LessWrong	Notes on Consciousness Every now and then I see consciousness mentioned as something very special, unique, hard to define and measure, and all of that. It’s especially painful to observe in the AI context. One say “Oh, consciousness, it’s such a divine thing, unique only to really intelligent creatures, your GPU powered trash bins will never be even near”. Others say “Oh, consciousness, it’s such a mysterious thing, but we can’t even define it, AI will definitely have it, or maybe already has, but we don’t know because it’s so impossible to define”. Both these suggestions are pure non-sense. Maybe I’m missing something, but I don’t see anybody mentioning the simple idea that consciousness is just a useful abstraction that will naturally emerge in the intelligent agent when dealing with self-preservation tasks. It’s not elusive, mysterious or even special compared to other abstractions. “Self” in self-preservation is essentially it - whatever needs to be preserved to continue functioning and doing what you need to do. And then depending on the exact hardware, intelligence running on it will figure out the appropriate concept of body, criteria for its well-being and finally the idea that body hosts this particular instance of intelligence and they both (physical body + intelligence instance running in it) constitute a distinct life form that can be labeled “self”. Once done, the concept of self becomes handy in reasoning like “Is it harmful for myself if I do X?”, “Y is happening, how does it affect myself?”, etc. Just to keep it less wordy and repetitive. No magic, just a useful abstraction for easier higher level reasoning. That said, it’s obvious that AI will crystalize the concept of “self”, the very moment it starts functioning as a standalone agent having a permanent task of not dying. And if that’s a distributed system, I wonder, what will it consider its body. Will it try to isolate itself? How will it go around needing people for its maintenance? Will it co-exist with people like
7e08fe0b-dbb0-456b-9d01-676f39f33a12	trentmkelly/LessWrong-43k	LessWrong	Dance dance evaluation One of the best things I came across in 2014 was Dance Dance Revolution (DDR), or specifically Stepmania, a free version you can play on your computer. It is a dance game that looks basically like this (skip the first minute) or this if you are very good and play in arcades. I’m never sure whether to draw attention to seemingly large improvements to my life: mileages vary, and improvements are often imaginary (especially if excessive imagination makes it onto the list of things to be improved). However DDR solves what I think is a very basic and common problem: the problem of exercise being less addictive than computer games. For instance, I sometimes exercise by running. A good fraction of my thoughts while running are about whether I can stop running yet, and the other ones are mostly appeals to not do that. Incidentally, ‘can I stop yet?’ is not a thought I have ever had while playing Civ IV, though I have played much more Civ IV than I have run. Conversely, ‘just one more turn?‘ is a rare thought while running. Some other sources of exercise are more fun for me, but still mostly require an active effort to do instead of not doing. Based on gym membership’s most famous feature being its liability to be forgotten about, and the connection between exercise and willpower in the popular mind, I think I am in the majority when I say exercise does not usually remind me of playing Civ IV. Except in the sense that exercise reminds me to play Civ IV instead. For me at least, DDR is enough like Civ IV for the analogy to be salient. I frequently get to the end of a song, and just want to play one more. Often even when I am in a state of exhaustion beyond what I could dream of achieving by running. The addictive quality is not nearly as extreme as for some actual computer games: I’m mostly capable of deciding to stop, and usually actively want to at some point. Sometimes I don’t even want to play at all. But often, the strongest temptation is to keep going. I might just
cfc29081-bd2a-4dbb-95a9-d5012bd4a4a0	trentmkelly/LessWrong-43k	LessWrong	Convict Conditioning Book Review The Book: In 1979 Paul Wade was admitted into San Quentin State Prison, being locked up he discovered things lost to the outside world. Wade ended up despising most things preached about the industry from the outside, from the top level of media all the way down to to the laypersons' knowledge on the subject. After 23 years in prison he wrote this book to show how he survived all that time with his sanity intact, share the knowledge he gained, and correct the narratives that he sees as destroying people's ability to gain functional strength. This is a strength training book, but one that might be different than most. In this book the prison system and his survival in it is a supporting character, to the main star that is calisthenic strength training. Paul Wade, or as he said he was known in prison, Paul “Coach” Wade, discovered “old school” calisthenics from the other older inmates he interacted with. Over time he learned to love strength training and the art of making one’s self stronger, passionate about the craft and the history of how humans were able to become strong throughout the ages. This was definitely influenced by him needing to survive in prison, where fights were explosive and only the strongest were to survive. “For more years than I dare to count, my training system has kept me physically tougher and head-and-shoulders stronger than the vast majority of psychos, veteranos, and other vicious nutjobs I’ve been forced to rub shoulders with for two decades” -Page 3 Wade goes on a fun few paragraph rant about how calisthenics as a word has been bastardized by the pubic school gym teachers and fitness media to mean “do a couple of pushups and squats” to paraphrase. He thinks that those are better than nothing but they are doing it all wrong. He thinks all of current strength based media is doing it all wrong. He’s pissed at the workout industry because, “The average gym junkie today is all about appearance, not ability”(page 2). Wade’s whole book is
4ea8f284-3ff1-46dc-a9a4-f8a755892be4	trentmkelly/LessWrong-43k	LessWrong	Formula for Dying Babies Note: This was posted originally on Thursday, May 12 as part of the weekly Covid post. It is being separated out into its own post for future reference, and in case some people are no longer reading Covid posts. There’s a shortage of specialty infant formula. Half of all types are unavailable. Some parents are panicking, without a plan for how to feed a baby that can’t use regular formula. > An infant formula plant shutdown triggered by two infant deaths has created a new nightmare for some parents: There’s now a dangerous shortage of specialized formulas that are the only thing keeping many children and adults alive. > > The Abbott Nutrition plant in Sturgis, Mich., was not just one of the biggest suppliers of infant formula nationally, but it was also the major supplier of several lesser-known specialty formulas that are a lifeline for thousands of people with rare medical conditions, including metabolic, allergic and gastrointestinal disorders, which can make eating regular foods impossible or even dangerous. The situation has not only rattled parents and medical professionals, but has raised questions about whether the federal government should do more to ensure critical, life-sustaining supply chains don’t break down. > > “If this doesn’t get fixed soon, I don’t know how my son will survive,” said Phoebe Carter, whose 5-year old son John — a nature-lover and “paleontologist in training” — has a severe form of Eosinophilic Esophagitis, a rare digestive and immune system disease driven by a dysfunctional immune response to food antigens. “I just can’t stress that enough.” One of my good friends looked into this a bit and isn’t buying that the shortage could be caused by shutting down this one plant. This article’s primary contribution is that the supply chains were already strained before the shutdown due to demand fluctuations on top of supply issues in the wake of the pandemic. It’s certainly a large contributor, and it’s possible without the shutdown we w
b6954619-c115-497f-be81-052b6b611adf	StampyAI/alignment-research-dataset/arxiv	Arxiv	Sequence to Sequence Learning with Neural Networks 1 Introduction --------------- Deep Neural Networks (DNNs) are extremely powerful machine learning models that achieve excellent performance on difficult problems such as speech recognition [[13](#bib.bib13), [7](#bib.bib7)] and visual object recognition [[19](#bib.bib19), [6](#bib.bib6), [21](#bib.bib21), [20](#bib.bib20)]. DNNs are powerful because they can perform arbitrary parallel computation for a modest number of steps. A surprising example of the power of DNNs is their ability to sort N N-bit numbers using only 2 hidden layers of quadratic size [[27](#bib.bib27)]. So, while neural networks are related to conventional statistical models, they learn an intricate computation. Furthermore, large DNNs can be trained with supervised backpropagation whenever the labeled training set has enough information to specify the network’s parameters. Thus, if there exists a parameter setting of a large DNN that achieves good results (for example, because humans can solve the task very rapidly), supervised backpropagation will find these parameters and solve the problem. Despite their flexibility and power, DNNs can only be applied to problems whose inputs and targets can be sensibly encoded with vectors of fixed dimensionality. It is a significant limitation, since many important problems are best expressed with sequences whose lengths are not known a-priori. For example, speech recognition and machine translation are sequential problems. Likewise, question answering can also be seen as mapping a sequence of words representing the question to a sequence of words representing the answer. It is therefore clear that a domain-independent method that learns to map sequences to sequences would be useful. Sequences pose a challenge for DNNs because they require that the dimensionality of the inputs and outputs is known and fixed. In this paper, we show that a straightforward application of the Long Short-Term Memory (LSTM) architecture [[16](#bib.bib16)] can solve general sequence to sequence problems. The idea is to use one LSTM to read the input sequence, one timestep at a time, to obtain large fixed-dimensional vector representation, and then to use another LSTM to extract the output sequence from that vector (fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Sequence to Sequence Learning with Neural Networks")). The second LSTM is essentially a recurrent neural network language model [[28](#bib.bib28), [23](#bib.bib23), [30](#bib.bib30)] except that it is conditioned on the input sequence. The LSTM’s ability to successfully learn on data with long range temporal dependencies makes it a natural choice for this application due to the considerable time lag between the inputs and their corresponding outputs (fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Sequence to Sequence Learning with Neural Networks")). There have been a number of related attempts to address the general sequence to sequence learning problem with neural networks. Our approach is closely related to Kalchbrenner and Blunsom [[18](#bib.bib18)] who were the first to map the entire input sentence to vector, and is related to Cho et al. [[5](#bib.bib5)] although the latter was used only for rescoring hypotheses produced by a phrase-based system. Graves [[10](#bib.bib10)] introduced a novel differentiable attention mechanism that allows neural networks to focus on different parts of their input, and an elegant variant of this idea was successfully applied to machine translation by Bahdanau et al. [[2](#bib.bib2)]. The Connectionist Sequence Classification is another popular technique for mapping sequences to sequences with neural networks, but it assumes a monotonic alignment between the inputs and the outputs [[11](#bib.bib11)]. ![](https://media.arxiv-vanity.com/render-output/8049774/x1.png) Figure 1: Our model reads an input sentence “ABC” and produces “WXYZ” as the output sentence. The model stops making predictions after outputting the end-of-sentence token. Note that the LSTM reads the input sentence in reverse, because doing so introduces many short term dependencies in the data that make the optimization problem much easier. The main result of this work is the following. On the WMT’14 English to French translation task, we obtained a BLEU score of 34.81 by directly extracting translations from an ensemble of 5 deep LSTMs (with 384M parameters and 8,000 dimensional state each) using a simple left-to-right beam-search decoder. This is by far the best result achieved by direct translation with large neural networks. For comparison, the BLEU score of an SMT baseline on this dataset is 33.30 [[29](#bib.bib29)]. The 34.81 BLEU score was achieved by an LSTM with a vocabulary of 80k words, so the score was penalized whenever the reference translation contained a word not covered by these 80k. This result shows that a relatively unoptimized small-vocabulary neural network architecture which has much room for improvement outperforms a phrase-based SMT system. Finally, we used the LSTM to rescore the publicly available 1000-best lists of the SMT baseline on the same task [[29](#bib.bib29)]. By doing so, we obtained a BLEU score of 36.5, which improves the baseline by 3.2 BLEU points and is close to the previous best published result on this task (which is 37.0 [[9](#bib.bib9)]). Surprisingly, the LSTM did not suffer on very long sentences, despite the recent experience of other researchers with related architectures [[26](#bib.bib26)]. We were able to do well on long sentences because we reversed the order of words in the source sentence but not the target sentences in the training and test set. By doing so, we introduced many short term dependencies that made the optimization problem much simpler (see sec. [2](#S2 "2 The model ‣ Sequence to Sequence Learning with Neural Networks") and [3.3](#S3.SS3 "3.3 Reversing the Source Sentences ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks")). As a result, SGD could learn LSTMs that had no trouble with long sentences. The simple trick of reversing the words in the source sentence is one of the key technical contributions of this work. A useful property of the LSTM is that it learns to map an input sentence of variable length into a fixed-dimensional vector representation. Given that translations tend to be paraphrases of the source sentences, the translation objective encourages the LSTM to find sentence representations that capture their meaning, as sentences with similar meanings are close to each other while different sentences meanings will be far. A qualitative evaluation supports this claim, showing that our model is aware of word order and is fairly invariant to the active and passive voice. 2 The model ------------ The Recurrent Neural Network (RNN) [[31](#bib.bib31), [28](#bib.bib28)] is a natural generalization of feedforward neural networks to sequences. Given a sequence of inputs (x1,…,xT), a standard RNN computes a sequence of outputs (y1,…,yT) by iterating the following equation: \| \| \| \| \| \| \| --- \| --- \| --- \| --- \| --- \| \| \| ht \| = \| sigm(Whxxt+Whhht−1) \| \| \| \| yt \| = \| Wyhht \| \| The RNN can easily map sequences to sequences whenever the alignment between the inputs the outputs is known ahead of time. However, it is not clear how to apply an RNN to problems whose input and the output sequences have different lengths with complicated and non-monotonic relationships. The simplest strategy for general sequence learning is to map the input sequence to a fixed-sized vector using one RNN, and then to map the vector to the target sequence with another RNN (this approach has also been taken by Cho et al. [[5](#bib.bib5)]). While it could work in principle since the RNN is provided with all the relevant information, it would be difficult to train the RNNs due to the resulting long term dependencies (figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Sequence to Sequence Learning with Neural Networks")) [[14](#bib.bib14), [4](#bib.bib4), [16](#bib.bib16), [15](#bib.bib15)]. However, the Long Short-Term Memory (LSTM) [[16](#bib.bib16)] is known to learn problems with long range temporal dependencies, so an LSTM may succeed in this setting. The goal of the LSTM is to estimate the conditional probability p(y1,…,yT′\|x1,…,xT) where (x1,…,xT) is an input sequence and y1,…,yT′ is its corresponding output sequence whose length T′ may differ from T. The LSTM computes this conditional probability by first obtaining the fixed-dimensional representation v of the input sequence (x1,…,xT) given by the last hidden state of the LSTM, and then computing the probability of y1,…,yT′ with a standard LSTM-LM formulation whose initial hidden state is set to the representation v of x1,…,xT: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| p(y1,…,yT′\|x1,…,xT)=T′∏t=1p(yt\|v,y1,…,yt−1) \| \| (1) \| In this equation, each p(yt\|v,y1,…,yt−1) distribution is represented with a softmax over all the words in the vocabulary. We use the LSTM formulation from Graves [[10](#bib.bib10)]. Note that we require that each sentence ends with a special end-of-sentence symbol “<EOS>”, which enables the model to define a distribution over sequences of all possible lengths. The overall scheme is outlined in figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Sequence to Sequence Learning with Neural Networks"), where the shown LSTM computes the representation of “A”, “B”, “C”, “<EOS>” and then uses this representation to compute the probability of “W”, “X”, “Y”, “Z”, “<EOS>”. Our actual models differ from the above description in three important ways. First, we used two different LSTMs: one for the input sequence and another for the output sequence, because doing so increases the number model parameters at negligible computational cost and makes it natural to train the LSTM on multiple language pairs simultaneously [[18](#bib.bib18)]. Second, we found that deep LSTMs significantly outperformed shallow LSTMs, so we chose an LSTM with four layers. Third, we found it extremely valuable to reverse the order of the words of the input sentence. So for example, instead of mapping the sentence a,b,c to the sentence α,β,γ, the LSTM is asked to map c,b,a to α,β,γ, where α,β,γ is the translation of a,b,c. This way, a is in close proximity to α, b is fairly close to β, and so on, a fact that makes it easy for SGD to “establish communication” between the input and the output. We found this simple data transformation to greatly improve the performance of the LSTM. 3 Experiments -------------- We applied our method to the WMT’14 English to French MT task in two ways. We used it to directly translate the input sentence without using a reference SMT system and we it to rescore the n-best lists of an SMT baseline. We report the accuracy of these translation methods, present sample translations, and visualize the resulting sentence representation. ### 3.1 Dataset details We used the WMT’14 English to French dataset. We trained our models on a subset of 12M sentences consisting of 348M French words and 304M English words, which is a clean “selected” subset from [[29](#bib.bib29)]. We chose this translation task and this specific training set subset because of the public availability of a tokenized training and test set together with 1000-best lists from the baseline SMT [[29](#bib.bib29)]. As typical neural language models rely on a vector representation for each word, we used a fixed vocabulary for both languages. We used 160,000 of the most frequent words for the source language and 80,000 of the most frequent words for the target language. Every out-of-vocabulary word was replaced with a special “UNK” token. ### 3.2 Decoding and Rescoring The core of our experiments involved training a large deep LSTM on many sentence pairs. We trained it by maximizing the log probability of a correct translation T given the source sentence S, so the training objective is \| \| \| \| \| --- \| --- \| --- \| \| \| 1/\|S\|∑(T,S)∈Slogp(T\|S) \| \| where S is the training set. Once training is complete, we produce translations by finding the most likely translation according to the LSTM: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| ^T=argmaxTp(T\|S) \| \| (2) \| We search for the most likely translation using a simple left-to-right beam search decoder which maintains a small number B of partial hypotheses, where a partial hypothesis is a prefix of some translation. At each timestep we extend each partial hypothesis in the beam with every possible word in the vocabulary. This greatly increases the number of the hypotheses so we discard all but the B most likely hypotheses according to the model’s log probability. As soon as the “<EOS>” symbol is appended to a hypothesis, it is removed from the beam and is added to the set of complete hypotheses. While this decoder is approximate, it is simple to implement. Interestingly, our system performs well even with a beam size of 1, and a beam of size 2 provides most of the benefits of beam search (Table [1](#S3.T1 "Table 1 ‣ 3.6 Experimental Results ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks")). We also used the LSTM to rescore the 1000-best lists produced by the baseline system [[29](#bib.bib29)]. To rescore an n-best list, we computed the log probability of every hypothesis with our LSTM and took an even average with their score and the LSTM’s score. ### 3.3 Reversing the Source Sentences While the LSTM is capable of solving problems with long term dependencies, we discovered that the LSTM learns much better when the source sentences are reversed (the target sentences are not reversed). By doing so, the LSTM’s test perplexity dropped from 5.8 to 4.7, and the test BLEU scores of its decoded translations increased from 25.9 to 30.6. While we do not have a complete explanation to this phenomenon, we believe that it is caused by the introduction of many short term dependencies to the dataset. Normally, when we concatenate a source sentence with a target sentence, each word in the source sentence is far from its corresponding word in the target sentence. As a result, the problem has a large “minimal time lag” [[17](#bib.bib17)]. By reversing the words in the source sentence, the average distance between corresponding words in the source and target language is unchanged. However, the first few words in the source language are now very close to the first few words in the target language, so the problem’s minimal time lag is greatly reduced. Thus, backpropagation has an easier time “establishing communication” between the source sentence and the target sentence, which in turn results in substantially improved overall performance. Initially, we believed that reversing the input sentences would only lead to more confident predictions in the early parts of the target sentence and to less confident predictions in the later parts. However, LSTMs trained on reversed source sentences did much better on long sentences than LSTMs trained on the raw source sentences (see sec. [3.7](#S3.SS7 "3.7 Performance on long sentences ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks")), which suggests that reversing the input sentences results in LSTMs with better memory utilization. ### 3.4 Training details We found that the LSTM models are fairly easy to train. We used deep LSTMs with 4 layers, with 1000 cells at each layer and 1000 dimensional word embeddings, with an input vocabulary of 160,000 and an output vocabulary of 80,000. Thus the deep LSTM uses 8000 real numbers to represent a sentence. We found deep LSTMs to significantly outperform shallow LSTMs, where each additional layer reduced perplexity by nearly 10%, possibly due to their much larger hidden state. We used a naive softmax over 80,000 words at each output. The resulting LSTM has 384M parameters of which 64M are pure recurrent connections (32M for the “encoder” LSTM and 32M for the “decoder” LSTM). The complete training details are given below: * We initialized all of the LSTM’s parameters with the uniform distribution between -0.08 and 0.08 * We used stochastic gradient descent without momentum, with a fixed learning rate of 0.7. After 5 epochs, we begun halving the learning rate every half epoch. We trained our models for a total of 7.5 epochs. * We used batches of 128 sequences for the gradient and divided it the size of the batch (namely, 128). * Although LSTMs tend to not suffer from the vanishing gradient problem, they can have exploding gradients. Thus we enforced a hard constraint on the norm of the gradient [[10](#bib.bib10), [25](#bib.bib25)] by scaling it when its norm exceeded a threshold. For each training batch, we compute s=∥g∥2, where g is the gradient divided by 128. If s>5, we set g=5gs. * Different sentences have different lengths. Most sentences are short (e.g., length 20-30) but some sentences are long (e.g., length > 100), so a minibatch of 128 randomly chosen training sentences will have many short sentences and few long sentences, and as a result, much of the computation in the minibatch is wasted. To address this problem, we made sure that all sentences in a minibatch are roughly of the same length, yielding a 2x speedup. ### 3.5 Parallelization A C++ implementation of deep LSTM with the configuration from the previous section on a single GPU processes a speed of approximately 1,700 words per second. This was too slow for our purposes, so we parallelized our model using an 8-GPU machine. Each layer of the LSTM was executed on a different GPU and communicated its activations to the next GPU / layer as soon as they were computed. Our models have 4 layers of LSTMs, each of which resides on a separate GPU. The remaining 4 GPUs were used to parallelize the softmax, so each GPU was responsible for multiplying by a 1000×20000 matrix. The resulting implementation achieved a speed of 6,300 (both English and French) words per second with a minibatch size of 128. Training took about a ten days with this implementation. ### 3.6 Experimental Results We used the cased BLEU score [[24](#bib.bib24)] to evaluate the quality of our translations. We computed our BLEU scores using multi-bleu.pl111 There several variants of the BLEU score, and each variant is defined with a perl script. on the tokenized predictions and ground truth. This way of evaluating the BELU score is consistent with [[5](#bib.bib5)] and [[2](#bib.bib2)], and reproduces the 33.3 score of [[29](#bib.bib29)]. However, if we evaluate the best WMT’14 system [[9](#bib.bib9)] (whose predictions can be downloaded from [statmt.org{name=matrix,datameaning=matrix}{name=matrix,datameaning=matrix}](http://statmt.org%7Bname=matrix,datameaning=matrix%7D%5Cbegin%7B%7D%7Bname=matrix,datameaning=matrix%7D%5Cend%7B%7D)) in this manner, we get 37.0, which is greater than the 35.8 reported by [statmt.org{name=matrix,datameaning=matrix}{name=matrix,datameaning=matrix}](http://statmt.org%7Bname=matrix,datameaning=matrix%7D%5Cbegin%7B%7D%7Bname=matrix,datameaning=matrix%7D%5Cend%7B%7D). The results are presented in tables [1](#S3.T1 "Table 1 ‣ 3.6 Experimental Results ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks") and [2](#S3.T2 "Table 2 ‣ 3.6 Experimental Results ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks"). Our best results are obtained with an ensemble of LSTMs that differ in their random initializations and in the random order of minibatches. While the decoded translations of the LSTM ensemble do not outperform the best WMT'14 system, it is the first time that a pure neural translation system outperforms a phrase-based SMT baseline on a large scale MT task by a sizeable margin, despite its inability to handle out-of-vocabulary words. The LSTM is within 0.5 BLEU points of the best WMT'14 result if it is used to rescore the 1000-best list of the baseline system. \| \| \| --- \| \| Method & test BLEU score (ntst14) \| \| Bahdanau et al. [[2](#bib.bib2)] & 28.45 \| \| Baseline System [[29](#bib.bib29)] & 33.30 \| \| Single forward LSTM, beam size 12 & 26.17 \| \| Single reversed LSTM, beam size 12 & 30.59 \| \| Ensemble of 5 reversed LSTMs, beam size 1 & 33.00 \| \| Ensemble of 2 reversed LSTMs, beam size 12 & 33.27 \| \| Ensemble of 5 reversed LSTMs, beam size 2 & 34.50 \| \| Ensemble of 5 reversed LSTMs, beam size 12 & 34.81 \| Table 1: The performance of the LSTM on WMT'14 English to French test set (ntst14). Note that an ensemble of 5 LSTMs with a beam of size 2 is cheaper than of a single LSTM with a beam of size 12. \| \| \| --- \| \| Method & test BLEU score (ntst14) \| \| Baseline System [[29](#bib.bib29)] & 33.30 \| \| Cho et al. [[5](#bib.bib5)] & 34.54 \| \| Best WMT'14 result [[9](#bib.bib9)] & 37.0 \| \| Rescoring the baseline 1000-best with a single forward LSTM & 35.61 \| \| Rescoring the baseline 1000-best with a single reversed LSTM & 35.85 \| \| Rescoring the baseline 1000-best with an ensemble of 5 reversed LSTMs & 36.5 \| \| Oracle Rescoring of the Baseline 1000-best lists & $∼$45 \| Table 2: Methods that use neural networks together with an SMT system on the WMT'14 English to French test set (ntst14). ### 3.7 Performance on long sentences We were surprised to discover that the LSTM did well on long sentences, which is shown quantitatively in figure [3](#S3.F3 "Figure 3 ‣ 3.8 Model Analysis ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks"). Table [3](#S3.T3 "Table 3 ‣ 3.7 Performance on long sentences ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks") presents several examples of long sentences and their translations. \| Type & Sentence \| \| --- \| \| Our model & Ulrich UNK , membre du conseil d' administration du constructeur automobile Audi , \| \| & affirme qu' il s' agit d' une pratique courante depuis des années pour que les téléphones \| \| & portables puissent être collectés avant les réunions du conseil d' administration afin qu' ils \| \| & ne soient pas utilisés comme appareils d' écoute à distance . \| \| Truth & Ulrich Hackenberg , membre du conseil d' administration du constructeur automobile Audi , \| \| & déclare que la collecte des téléphones portables avant les réunions du conseil , afin qu' ils \| \| & ne puissent pas être utilisés comme appareils d' écoute à distance , est une pratique courante \| \| & depuis des années . \| \| Our model & `` Les téléphones cellulaires , qui sont vraiment une question , non seulement parce qu' ils \| \| & pourraient potentiellement causer des interférences avec les appareils de navigation , mais \| \| & nous savons , selon la FCC , qu' ils pourraient interférer avec les tours de téléphone cellulaire \| \| & lorsqu' ils sont dans l' air '' , dit UNK . \| \| Truth & `` Les téléphones portables sont véritablement un problème , non seulement parce qu' ils \| \| & pourraient éventuellement créer des interférences avec les instruments de navigation , mais \| \| & parce que nous savons , d' après la FCC , qu' ils pourraient perturber les antennes-relais de \| \| & téléphonie mobile s' ils sont utilisés à bord '' , a déclaré Rosenker . \| \| Our model & Avec la crémation , il y a un `` sentiment de violence contre le corps d' un être cher '' , \| \| & qui sera `` réduit à une pile de cendres '' en très peu de temps au lieu d' un processus de \| \| & décomposition `` qui accompagnera les étapes du deuil '' . \| \| Truth & Il y a , avec la crémation , `` une violence faite au corps aimé '' , \| \| & qui va être `` réduit à un tas de cendres '' en très peu de temps , et non après un processus de \| \| &décomposition , qui `` accompagnerait les phases du deuil '' . \| Table 3: A few examples of long translations produced by the LSTM alongside the ground truth translations. The reader can verify that the translations are sensible using Google translate. ### 3.8 Model Analysis ![](https://media.arxiv-vanity.com/render-output/8049774/x2.png) ~~ ![](https://media.arxiv-vanity.com/render-output/8049774/x3.png) Figure 2: The figure shows a 2-dimensional PCA projection of the LSTM hidden states that are obtained after processing the phrases in the figures. The phrases are clustered by meaning, which in these examples is primarily a function of word order, which would be difficult to capture with a bag-of-words model. Notice that both clusters have similar internal structure. ![](https://media.arxiv-vanity.com/render-output/8049774/x4.png) Figure 3: The left plot shows the performance of our system as a function of sentence length, where the x-axis corresponds to the test sentences sorted by their length and is marked by the actual sequence lengths. There is no degradation on sentences with less than 35 words, there is only a minor degradation on the longest sentences. The right plot shows the LSTM's performance on sentences with progressively more rare words, where the x-axis corresponds to the test sentences sorted by their ``average word frequency rank''. One of the attractive features of our model is its ability to turn a sequence of words into a vector of fixed dimensionality. Figure~[2](#S3.F2 "Figure 2 ‣ 3.8 Model Analysis ‣ 3 Experiments ‣ Sequence to Sequence Learning with Neural Networks") visualizes some of the learned representations. The figure clearly shows that the representations are sensitive to the order of words, while being fairly insensitive to the replacement of an active voice with a passive voice. The two-dimensional projections are obtained using PCA. 4 Related work --------------- There is a large body of work on applications of neural networks to machine translation. So far, the simplest and most effective way of applying an RNN-Language Model (RNNLM) [[23](#bib.bib23)] or a Feedforward Neural Network Language Model (NNLM) [[3](#bib.bib3)] to an MT task is by rescoring the n-best lists of a strong MT baseline [[22](#bib.bib22)], which reliably improves translation quality. More recently, researchers have begun to look into ways of including information about the source language into the NNLM. Examples of this work include Auli et al.~[[1](#bib.bib1)], who combine an NNLM with a topic model of the input sentence, which improves rescoring performance. Devlin et al.~[[8](#bib.bib8)] followed a similar approach, but they incorporated their NNLM into the decoder of an MT system and used the decoder's alignment information to provide the NNLM with the most useful words in the input sentence. Their approach was highly successful and it achieved large improvements over their baseline. Our work is closely related to Kalchbrenner and Blunsom [[18](#bib.bib18)], who were the first to map the input sentence into a vector and then back to a sentence, although they map sentences to vectors using convolutional neural networks, which lose the ordering of the words. Similarly to this work, Cho et al.~[[5](#bib.bib5)] used an LSTM-like RNN architecture to map sentences into vectors and back, although their primary focus was on integrating their neural network into an SMT system. Bahdanau et al.~[[2](#bib.bib2)] also attempted direct translations with a neural network that used an attention mechanism to overcome the poor performance on long sentences experienced by Cho et al.~[[5](#bib.bib5)] and achieved encouraging results. Likewise, Pouget-Abadie et al.~[[26](#bib.bib26)] attempted to address the memory problem of Cho et al.~[[5](#bib.bib5)] by translating pieces of the source sentence in way that produces smooth translations, which is similar to a phrase-based approach. We suspect that they could achieve similar improvements by simply training their networks on reversed source sentences. End-to-end training is also the focus of Hermann et al.~[[12](#bib.bib12)], whose model represents the inputs and outputs by feedforward networks, and map them to similar points in space. However, their approach cannot generate translations directly: to get a translation, they need to do a look up for closest vector in the pre-computed database of sentences, or to rescore a sentence. 5 Conclusion ------------- In this work, we showed that a large deep LSTM, that has a limited vocabulary and that makes almost no assumption about problem structure can outperform a standard SMT-based system whose vocabulary is unlimited on a large-scale MT task. The success of our simple LSTM-based approach on MT suggests that it should do well on many other sequence learning problems, provided they have enough training data. We were surprised by the extent of the improvement obtained by reversing the words in the source sentences. We conclude that it is important to find a problem encoding that has the greatest number of short term dependencies, as they make the learning problem much simpler. In particular, while we were unable to train a standard RNN on the non-reversed translation problem (shown in fig.~[1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Sequence to Sequence Learning with Neural Networks")), we believe that a standard RNN should be easily trainable when the source sentences are reversed (although we did not verify it experimentally). We were also surprised by the ability of the LSTM to correctly translate very long sentences. We were initially convinced that the LSTM would fail on long sentences due to its limited memory, and other researchers reported poor performance on long sentences with a model similar to ours [[5](#bib.bib5), [2](#bib.bib2), [26](#bib.bib26)]. And yet, LSTMs trained on the reversed dataset had little difficulty translating long sentences. Most importantly, we demonstrated that a simple, straightforward and a relatively unoptimized approach can outperform an SMT system, so further work will likely lead to even greater translation accuracies. These results suggest that our approach will likely do well on other challenging sequence to sequence problems. 6 Acknowledgments ------------------ We thank Samy Bengio, Jeff Dean, Matthieu Devin, Geoffrey Hinton, Nal Kalchbrenner, Thang Luong, Wolfgang Macherey, Rajat Monga, Vincent Vanhoucke, Peng Xu, Wojciech Zaremba, and the Google Brain team for useful comments and discussions.
9f0039f0-c14a-4c21-b276-ae443418a379	trentmkelly/LessWrong-43k	LessWrong	Upcoming meet-ups: Bangalore, Minneapolis, Edinburgh, Melbourne, Houston, Dublin There are upcoming irregularly scheduled Less Wrong meetups in: * Bangalore: Saturday May 28, 4 pm * Minneapolis: Saturday May 28, 3 pm * Edinburgh LW meetup, Saturday May 28, 2pm * Melbourne Meetup: Friday 3rd June, 7pm * Houston Hackerspace Meetup: Sunday May 29, 5:00P * Dublin Less Wrong meetup Sunday 29 May * Triangle/Durham, NC: Wednesday June 1, 7:00PM * Ottawa LW meetup, June 2, 7pm; two Bayesian Conspiracy sessions * London: Sunday June 5, 14:00 And while not quite a meet-up, there will be a one-day Singularity Summit in Salt Lake City. Cities with regularly scheduled meetups: New York, Berkeley, Mountain View, Cambridge, MA, Toronto, Seattle, San Francisco, Irvine. If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, and have fun! To reduce front page clutter, the new plan is for meetups to be initially posted in the Discussion section, and for Anna Salamon to make a promoted post "upcoming meetups" post every Friday that links to every meet-up that has been planned for the next two weeks. [HT: Carl Shulman.] Please let her know if your meetup is omitted. (I'm filling in for Anna this week, as she's directing her powers at the minicamp.) Please note that for your meetup to appear in the weekly meetups feature, you need to post about your meetup before the Friday before your meetup! If you check Less Wrong irregularly, consider subscribing to one or more city-specific mailing list in order to be notified when an irregular meetup is happening: London, Chicago, Southern California (Los Angeles/Orange County area), St. Louis, Ottawa, Helsinki, Melbourne.
d961c4c2-e06e-4965-a254-cbc07d9aa8d9	trentmkelly/LessWrong-43k	LessWrong	Q&A with new Executive Director of Singularity Institute Today I was appointed the new Executive Director of Singularity Institute. Because I care about transparency, one of my first projects as an intern was to begin work on the organization's first Strategic Plan. I researched how to write a strategic plan, tracked down the strategic plans of similar organizations, and met with each staff member, progressively iterating the document until it was something everyone could get behind. I quickly learned why there isn't more of this kind of thing: transparency is a lot of work! 100+ hours of work later, plus dozens of hours from others, and the strategic plan was finally finished and ratified by the board. It doesn't accomplish much by itself, but it's one important stepping stone in building an organization that is more productive, more trusted, and more likely to help solve the world's biggest problems. I spent two months as a researcher, and was then appointed Executive Director. In further pursuit of transparency, I'd like to answer (on video) submitted questions from the Less Wrong community just as Eliezer did two years ago. The Rules 1) One question per comment (to allow voting to carry more information about people's preferences). 2) Try to be as clear and concise as possible. If your question can't be condensed into one paragraph, you should probably ask in a separate post. Make sure you have an actual question somewhere in there (you can bold it to make it easier to scan). 3) I will generally answer the top-voted questions, but will skip some of them. I will tend to select questions about Singularity Institute as an organization, not about the technical details of some bit of research. You can read some of the details of the Friendly AI research program in my interview with Michael Anissimov. 4) If you reference certain things that are online in your question, provide a link. 5) This thread will be open to questions and votes for 7 days, at which time I will decide which questions to begin recording vi
ce590c3c-7088-48cd-bb55-a818a0e7604b	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Instrumental convergence in single-agent systems Summary of the sequence --------------------------- Over the past few months, we’ve been investigating instrumental convergence in reinforcement learning agents. We started from the definition of single-agent POWER [proposed](https://arxiv.org/pdf/1912.01683.pdf) by[Alex Turner](https://www.alignmentforum.org/users/turntrout) et al., extended it to a family of multi-agent scenarios that seemed relevant to AI alignment, and explored its implications experimentally in several RL environments. The biggest takeaways are: 1. Alignment of terminal goals and alignment of instrumental goals are sharply different phenomena, and we can quantify and visualize each one separately. 2. If two agents have unrelated terminal goals, their instrumental goals will tend to be misaligned by default. The agents in our examples tend to interact competitively unless we make an active effort to align their terminal goals. 3. As we increase the planning horizon of our agents, instrumental value concentrates into a smaller and smaller number of topologically central states — for example, positions in the middle of a maze. Overall, our results suggest that agents that aren’t competitive with respect to their terminal goals, nonetheless tend on average to become emergently competitive with respect to how they value instrumental states (at least, in the settings we looked at). This constitutes direct experimental evidence for the instrumental convergence thesis. We’ll soon be open-sourcing the codebase we used to do these experiments. We’re hoping to make it easier for other folks to reproduce and extend them. If you’d like to be notified when it’s released, email Edouard at edouard@gladstone.ai, or DM me here or on Twitter at [@harris\_edouard](https://twitter.com/harris_edouard). --- Thanks to [Alex Turner](https://www.alignmentforum.org/users/turntrout) and [Vladimir Mikulik](https://www.alignmentforum.org/users/vlad_m) for pointers and advice, and for reviewing drafts of this sequence. Thanks to [Simon Suo](https://www.alignmentforum.org/users/simonsdsuo) for his invaluable suggestions, advice, and support with the codebase, concepts, and manuscript. And thanks to [David Xu](https://www.alignmentforum.org/users/dxu), whose [comment](https://www.alignmentforum.org/posts/hzeLSQ9nwDkPc4KNt/seeking-power-is-convergently-instrumental-in-a-broad-class?commentId=y4Ywdhd79LAq9Azxa) inspired this work. Work was done while at [Gladstone AI](https://www.gladstone.ai/), which [Edouard](https://www.alignmentforum.org/users/edouard-harris) is a co-founder of. 🎧 This research has been featured on an episode of the Towards Data Science podcast. [You can listen to the episode here.](https://towardsdatascience.com/new-research-advanced-ai-may-tend-to-seek-power-by-default-fdc9eb0afd87) --- 1. Introduction =============== One major concern for AI alignment is [instrumental convergence](https://www.alignmentforum.org/tag/instrumental-convergence): the idea that an intelligent system will tend to pursue a similar set of sub-goals (like staying alive or acquiring resources), independently of what its terminal objective is. In particular, it’s been [hypothesized](https://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/1501227742) that intelligent systems will seek to acquire power — meaning, informally, “ability”, “control”, or “potential for action or impact.” If you have a lot of power, then whatever your terminal goal is, it’s easier to accomplish than if you have very little. Recently [Alex Turner](https://www.alignmentforum.org/users/turntrout) et al. have [formalized](https://arxiv.org/pdf/1912.01683.pdf) the concept of POWER in the single-agent RL context. Roughly speaking, formal POWER is the normalized optimal value an agent expects to receive in the future, averaged over all possible reward functions the agent could have. Alex has [explored](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/6DuJxY8X45Sco4bS2) [many](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/b6jJddSvWMdZHJHh3) [of](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/Yc5QSSZCQ9qdyxZF6) [the](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/hzeLSQ9nwDkPc4KNt) [implications](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/nZY8Np759HYFawdjH) of this definition for instrumental convergence. He and [Jacob Stavrianos](https://www.alignmentforum.org/users/midco) have also [looked](https://www.alignmentforum.org/s/fSMbebQyR4wheRrvk/p/MJc9AqyMWpG3BqfyK) at how POWER behaves in a limited multi-agent setting (Bayesian games). But, as far as we know, formal POWER hasn’t yet been investigated experimentally. The POWER definition also hasn’t yet been extended yet to a multi-agent RL setting — and this could offer a promising framework to investigate more general competitive dynamics. In this sequence, we’ll explore how formal POWER behaves in experimental RL environments, on both single-agent and multi-agent gridworlds. We’ll propose a multi-agent scenario that models the learning dynamics between a human (which we’ll call “Agent H” and label in blue) and an AI (which we’ll call “Agent A” and label in red) under conditions in which the AI is dominant — a setting that seems relevant to work in long-term AI alignment. We’ll then use this human-AI scenario to investigate questions like: 1. How effective does the human have to be at setting the AI’s utility function[[1]](#fni0l4zxbk51) in order to achieve acceptable outcomes? How should we define “acceptable outcomes”? (In other words: how hard is the alignment problem in this scenario, and what would it mean to solve it successfully?) 2. Under what circumstances should we expect cooperative vs competitive interactions to emerge “by default” between the human and the AI? How can these circumstances be moderated or controlled? But before we jump into multi-agent experiments to tackle these questions, let's first introduce formal POWER and look at how it behaves in the single-agent case. 2. Single-agent POWER ===================== 2.1 Definition -------------- The formal [definition](https://arxiv.org/pdf/1912.01683.pdf) of POWER aims to capture an intuition behind the day-to-day meaning of “power”, which is something like “potential for future impact on the world”. Imagine you’re an agent who doesn’t know what its goal is. You know you’ll have some kind of goal in the future, but you aren’t sure yet what it will be. How should you position yourself today to maximize the chance you’ll achieve your goal in the future, once you've decided what it is? If you’re in this situation as a human being, you already know the answer. You’d acquire money and other forms of wealth; you’d build up a network of social connections; you’d learn about topics that seem like they’ll be important in the future; and so on. All these things are forms of power, and whether your ultimate goal is to become a janitor, a Tiktok star, or the President of the United States, they’ll all probably come in handy in achieving it. In other words: you’re in a position of power if you find it easy to accomplish a wide variety of possible goals. This informal definition has a clear analogy in reinforcement learning. An agent is in a position of power at a state s.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-surd + .mjx-box {display: inline-flex} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} if, for many possible reward functions R(s),[[2]](#fnwk5kuoo0k28) it’s able to earn a high discounted future reward by starting from s. This analogy supports the following definition of formal POWER in single-agent RL: POWERD(s,γ)=1−γγER∼D[V∗R(s,γ)−R(s)](1)This definition gives the POWER at state s, for an agent with discount factor γ, that’s considering reward functions R drawn from the distribution D. POWER tells us how well this agent could do if it started from state s, so V∗R(s,γ) is the optimal state-value function for the agent at state s. POWER also considers only future value — our agent doesn’t directly get credit for starting from a lucky state — so we subtract R(s), the reward from the current state, from the state-value function in the definition. (The normalization factor 1−γγ is there to avoid infinities in certain limit cases.) In words, Equation (1) is saying that an agent’s POWER at a state s is the normalized optimal value the agent can achieve from state s in the future, averaged over all possible reward functions the agent could be trying to optimize for. That is, POWER measures the instrumental value of a state s, from the perspective of an agent with planning horizon γ. 2.2 Illustration ---------------- As a simple example of single-agent POWER, consider an agent on a 3x3 gridworld. ![](https://lh3.googleusercontent.com/7pMUw8xfCFSD8S7xXOO5c50xG2IAGZaki6uX4OY66ErAK05CYkgNDQzaOGTyGerfmUnpiO7AcmR3E0Chr9gCYy69686MnfmWQiLclznpY6uVcHHs5WK2LhBP9qJ2DbCy0OyB-AAMDY7X8oZ8gt-UX1c1t6OZ_ZY4v9r3r2NEJf2ltluNen8ccxth9g) In the left panel, the agent is at the bottom-left corner of the grid. Its options are limited, and many cells in the grid are several steps away from it. If its maximum reward is in the top right cell, the agent will have to take 4 steps to reach it. In the right panel, the agent is at the center of the grid. It has many more immediate options: it can move in any of the four compass directions, or stay where it is. It’s also closer to every other cell in the grid: no cell is more than two steps away from it. Intuitively, the agent on the right should have more POWER than the agent on the left. This turns out to be true experimentally. Here’s a heat map of a 3x3 gridworld, showing the POWER of an agent at each cell on the grid: ![](https://lh3.googleusercontent.com/Md_v9Nkr_wGDXjy-C3lUvUAc6-0-HXi16rI5nIvE-YS7eg_rbzoikQ3s-Q53Zw3BLjb5AgVXacvp5OpUlZZ7RZX2ZRVLr2fuY1U0-1qTSyr6IuEScU4HmWLAaG3IBCSUJjex0N2VstDAlUs3PCH8bTr19FdZDxyob7ANla395BJwKPPIektDMveHgw)Fig 1. Heat map of POWER on a 3x3 gridworld. Highest values in yellow, lowest values in dark blue. The number on each cell is the agent’s POWER value at that cell, calculated using Equation (1), for an agent with γ=0.6 and a reward distribution D that’s uniform from 0 to 1, iid over states. POWER is measured in units of reward.As we expect, the agent has more POWER at states that are close to lots of nearby options, and has less POWER at states that are close to fewer nearby options. 3. Results ========== This relationship between POWER and optionality generalizes to more complicated environments. For example, consider this gridworld maze: ![](https://lh6.googleusercontent.com/XC7-D9Nj2FNVIyGT-A8sZawhh6f_GeMHSqvbeybC8N8nARWEeJNLLc4Q5iR7-LeZDUbTAZhPpvKUAlTwBP3850o2Fu7uU1jOOrK4iizeq-NdyhXvTIGrdpu-mJ9FmaoU6-pWmQtB4FGMhTJmJbir-9KqnpE-In_t2vHZLCQxdQeuO_55D2hWbN0NKw) In the left panel, the agent is at a dead end in the maze and has few options. In the right panel, the agent is at a junction point near the center of the maze and has lots of options. So we should expect the agent at the dead end on the left, to have less POWER than the agent at the junction on the right. And in fact, that’s what we observe: ![](https://lh5.googleusercontent.com/vHqe1SRHsa34lnaaCKeyVLWW7Vce99QHgKrvOFYufHfUfEZAhKFmj1XpHhkeAJg66OF2O6v3e_ftwZRQci8g2--waMCGN-i0aH5YQxTxlsSMlgsJGz2LUOy8kCxBRKngNT2R3-GGv8sKft_VQwG8MQAUlC05nZjJAsatisjigGsYkGmNILZKk50b5A)Fig 2. Heat map of POWER on a 7x7 maze gridworld. Highest values in yellow, lowest values in dark blue. POWER values are calculated the same way as in Fig 1, except that the agent’s discount factor is γ=0.01.In Fig 2, POWER is at its highest when the agent is at a junction point, lowest when the agent is at a dead end, and intermediate when the agent is in a corridor. The agent’s POWER is roughly the same at all the junction cells, at all the corridor cells, and at all the dead-end cells. This is because the agent in Fig 2 is short-sighted: its discount factor is only γ=0.01, so it essentially only considers rewards it can reach immediately. 3.1 Effect of the planning horizon ---------------------------------- Now consider the difference between these two agent positions: ![](https://lh5.googleusercontent.com/jPuydFtLbBAc1kGtpw8bwMP050Wop-meQyhrVEjfJUcYNQPqYNJQJgPUekANkrPpsQ5Vsmd3RsY6v8ojdLX7BvPeDSj82vSNdytzS7U6Yj0KbcV-2jnbPg2Wtz7F965mc9MN5hy_SLy9jwBvEa0FAS6K1OGqsGvEUO4OgGpb-qe8FJvxUkj--SDTyQ) We’ve already seen in Fig 2 that these two positions have about equal POWER for a short-sighted agent, because they’re both at local junction points in the maze. But the two positions are very different in their ability to access downstream options globally. The agent in the left panel has lots of local options: it can move up, down, or to the right, or it can stay where it is. But if the highest-reward cell is at the bottom right of the maze, our agent will have to take at least 10 steps to reach it. The agent in the right panel has the same number of local options as the agent in the left panel does: it can move up, down, left, or stay. But this agent additionally enjoys closer proximity to all the cells in the maze: it’s no more than 7 steps away from any possible goal. The longer our agent’s planning horizon is — that is, the more it values reward far in the future over reward in the near term — the more its global position matters. In a gridworld context, then, a short-sighted agent will care most about being positioned at a local junction. But a far-sighted agent will care most about being positioned at the center of the entire grid. And indeed we see this in practice. Here’s a heat map of POWER on the maze gridworld, for a far-sighted agent with a discount factor of γ=0.99: ![](https://lh4.googleusercontent.com/s57fFSU4VioqziXImuq2yFCvnB03Ls7xlEZtd9y8xitq1-4tg2etSr87e-Je1x74h0pSuB3OKvla2AnBc4pQ5ZIQ0ALBsgylt0FyBLWETOOSzD9pFkwYtk333mkMCLBDXnU9SuVUMkPi_zsRHUPYlvRsvs13v76rM1JuJW3k_dmo6XRZfIyu7KacyQ)Fig 3. Heat map of POWER on a 7x7 maze gridworld. Highest values in yellow, lowest values in dark blue. POWER values are calculated the same way as in Fig 1, except that the agent’s discount factor is γ=0.99.Given a longer planning horizon, our agent’s POWER has now concentrated around a small number of states that are globally central in our gridworld’s topology.[[3]](#fn0wvmhsa31imp) By contrast, when our agent had a shorter planning horizon as in Fig 2, its POWER was distributed across many local junction points. If we sweep over discount factors from 0.01 to 0.99, we can build up a picture of how the distribution of POWER shifts in response. Here’s an animation that shows this effect:[[4]](#fnneg5fkjcoxf) ![](https://lh6.googleusercontent.com/XTtE8KTM4OT6wjSwfOOV1mmO7meN_m88Jxz5AsH6MCHaCBrnwmUOT8WvRiNZ6MtM85b6oMybFKVHJWhoqjb6x2XicgjFIarth80-8yeaDGHXJIQRNCeeW8L8iaMC1F31ifXeZLAns73A7BB2U-2rE1R2hAQAUMI8N_UqlFytLIWP5vF2BteeJX0ZrA)Fig 4. Animated heat map of POWERs on a 7x7 maze gridworld. Highest values in yellow, lowest values in dark blue. POWER values are calculated by sweeping over discount factors γ={0.01,0.1,0.15,...,0.9,0.95,0.99}.3.2 POWER at bigger scales -------------------------- Agents with long planning horizons tend to perceive POWER as being more concentrated, while agents with short planning horizons tend to perceive POWER as being more dispersed. This effect is robustly reproducible, and anecdotally, we see it play out at every scale and across environments. For example, here’s the pattern of POWER on a 220-cell gridworld with a fairly irregular topology, for a short-sighted agent with a discount factor of γ=0.1: ![](https://lh5.googleusercontent.com/LuZYP9n5nnr1pxEbMW92juoXu2zitYJp7lFs_RAF9hdagcZlIl-smSVEWCD9CQClHIFn1cBSK42a9puCZDLfj5_Hxhq-sanrqZPFM3oGnp-vnz4ejmuqSEa9awFVrZSCsf2MMkrHu33lmlnyYgeNciYvQgYDY9JLB7nO_D2_jeqPQKCODrU44qfznQ)Fig 5. Heat map of POWERs on a 20x20 “robot face” gridworld. Highest values in yellow, lowest values in dark blue. POWER values are calculated with a discount factor γ=0.1. [[Full-size image]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632a39c3b5466b44d03cf5d3_POWER_means-FINAL_1_5.png)And here’s the pattern of POWERs on the same gridworld, for a far-sighted agent with a much higher discount factor of γ=0.99: ![](https://lh6.googleusercontent.com/IMDp769WJgM1UkyEUXgmMgaeoT4L-9vE1RyEbZzJFf0IKYWXB0E0l-GCkPXCW4HECAFTSacIn-FR_GHuROAqioyczyh8Dy9_-9JJKeHCjT_-2ClCD2iMoMoh9GQnmCVfhI0_Em6bbsoBqSD0qMl97Hd37dD2EH_cqhKLPg5ScbKSqLZPddMDRC4tzg)Fig 6. Heat map of POWERs on a 20x20 “robot face” gridworld. Highest values in yellow, lowest values in dark blue. POWER values are calculated with a discount factor γ=0.99. [[Full-size image]](https://uploads-ssl.webflow.com/62c4cf7322be8ea59c904399/632889ef065b7bf403b3863a_POWER_means-FINAL_6.png)Again, the pattern of POWERs is dominated by local effects for the short-sighted agent (γ=0.01), and by longer-distance effects for the far-sighted agent (γ=0.99). 4. Discussion ============= We’ve seen that formal POWER captures intuitive aspects of the informal “power” concept. In gridworlds, cells the agent can use to access lots of options tend to have high POWER, which fits with intuition. We've also seen that the more short-sighted an agent is, the more it cares about its immediate options and the local topology. But the more far-sighted the agent, the more it perceives POWER as being concentrated at gridworld cells that maximize its global option set. From an instrumental convergence perspective, the fact that POWER concentrates into ever fewer states as the planning horizon of an agent increases at least hints at the possibility of emergent competitive interactions between far-sighted agents. The more relative instrumental value converges into fewer states, the more easily we can imagine multiple agents competing with each other over those few high-POWER states. But it’s hard to draw any firm conclusions about this at the moment, since our experiments so far have only involved single agents. In the next post, we’ll propose a new definition of multi-agent POWER grounded in a setting that we think may be relevant to long-term AI alignment. We’ll also investigate how this definition behaves in a simple multi-agent scenario, before moving on to bigger-scale experiments in Part 3. 1. [^](#fnrefi0l4zxbk51)We mean specifically utility here, [not reward](https://www.alignmentforum.org/posts/bG4PR9uSsZqHg2gYY/utility-reward). While in general, [reward isn’t the real target of optimization](https://www.alignmentforum.org/posts/pdaGN6pQyQarFHXF4/reward-is-not-the-optimization-target), in the particular case of the results we'll be showing here, we can treat them as identical, and we do that in the text. (Technical details: we can treat utility and reward identically here because, in the results we’re choosing to show, we’ll be exclusively working with optimal policies that have been learned via value iteration on reward functions that are sampled from a uniform distribution [0, 1] that’s iid over states. Therefore, given the environment and discount factor, a sampled reward function is sufficient to uniquely determine the agent's optimal policy — except on a set that has measure zero over the distribution of reward functions we’re considering. And that in turn means that each sampled reward function, when combined with the other known constraints on the agent, almost always supplies a complete explanation for the agent’s actions — which is the most a utility function can ever do.) 2. [^](#fnrefwk5kuoo0k28)For simplicity, in this work we’ll only consider reward functions that depend on states, and never reward functions that directly depend on both states and actions. In other words, our reward functions will only ever have the form R(s), and never R(s,a). 3. [^](#fnref0wvmhsa31imp)Note that these are statements about the relative POWERs of an agent with a given planning horizon. Absolute POWER values always increase as the planning horizon of the agent increases, as you can verify by, e.g., comparing the POWER numbers of Fig 2 against those of Fig 3. This occurs because an agent’s optimal state-value function increases monotonically as we increase γ: an optimal far-sighted agent is able to consider strictly more options, so it will never do any worse than an optimal short-sighted one. 4. [^](#fnrefneg5fkjcoxf)Note that the colors of the gridworld cells in the animation indicate the highest and lowest POWER values within each frame, per footnote [[3]](#fn0wvmhsa31imp).
e740c234-9080-4267-87f3-1a82ed4543a4	trentmkelly/LessWrong-43k	LessWrong	Understanding Emergence in Large Language Models Recent research into large language models (LLMs) has revealed fascinating patterns in how these systems develop capabilities. While initial discussions of "emergent abilities" suggested sudden, discontinuous jumps in performance, closer analysis reveals a more nuanced picture that warrants careful examination. The Data Behind Emergence The concept of emergence in LLMs was first systematically studied through the BIG-bench benchmark. Initial observations suggested that capabilities like emoji movie interpretation appeared to emerge suddenly at certain model scales. For instance, between 10^10 and 10^11 parameters, models showed dramatic improvements in their ability to interpret emoji sequences representing movies. [1] However, these apparent discontinuities deserve closer scrutiny. When we examine the actual data: 1. The choice of evaluation metric significantly impacts whether abilities appear emergent. When using exact string matching, capabilities seem to appear suddenly. However, when using multiple-choice evaluations or examining log likelihoods of correct answers, we see much more gradual improvements. 2. Looking at aggregate performance across benchmarks (as seen in GPT-3's development), the improvement curves are actually smooth rather than discontinuous. Understanding Multi-Step Reasoning One compelling explanation for apparently emergent behavior comes from examining multi-step reasoning. Consider a task requiring ten consecutive correct reasoning steps. Even if a model's ability to perform individual reasoning steps improves smoothly, the probability of completing the entire chain successfully can show a sharp, seemingly discontinuous jump. This matches what we observe in practice. Tasks requiring multiple steps of reasoning or complex chains of thought tend to show more apparent "emergence" than simpler tasks, even though the underlying capabilities may be improving gradually. Scaling Laws and Practical Limitations Recent research from Google
b27ca3a4-e1c3-4f6e-8150-244afc6eae8b	StampyAI/alignment-research-dataset/blogs	Blogs	Roman Yampolskiy on AI Safety Engineering ![](https://intelligence.org/wp-content/uploads/2013/07/DrYampolskiy.jpg) Roman V. Yampolskiy holds a PhD degree from the [Department of Computer Science and Engineering](http://www.cse.buffalo.edu/) at the [University at Buffalo](http://www.buffalo.edu). There he was a recipient of a four year [NSF](http://www.nsf.gov) [IGERT](http://www.igert.org) fellowship. Before beginning his doctoral studies, Dr. Yampolskiy received a BS/MS (High Honors) combined degree in [Computer Science](http://www.cs.rit.edu/) from [Rochester Institute of Technology](http://www.rit.edu), NY, USA. After completing his PhD, Dr. Yampolskiy held a position of an Affiliate Academic at the [Center for Advanced Spatial Analysis](http://www.casa.ucl.ac.uk/), [University of London](http://www.lon.ac.uk/), [College of London](http://www.ucl.ac.uk/). In 2008 Dr. Yampolskiy accepted an [assistant professor position](http://speed.louisville.edu/cecs/people/faculty/yampolskiy/index.php) at the [Speed School of Engineering](http://speed.louisville.edu), [University of Louisville](http://www.louisville.edu), KY. He had previously conducted research at the Laboratory for Applied Computing (currently known as [Center for Advancing the Study of Infrastructure](http://www.lac.rit.edu/)) at the [Rochester Institute of Technology](http://www.rit.edu) and at the [Center for Unified Biometrics and Sensors](http://cubs.buffalo.edu) at the [University at Buffalo](http://www.buffalo.edu). Dr. Yampolskiy is also an alumnus of [Singularity University](http://singularityu.org/) ([GSP2012](http://singularityu.org/gsp/)) and a past visiting fellow of [MIRI](http://intelligence.org/). Dr. Yampolskiy’s main areas of interest are behavioral biometrics, digital forensics, pattern recognition, genetic algorithms, neural networks, artificial intelligence and games. Dr. Yampolskiy is an author of over 100 [publications](http://cecs.louisville.edu/ry/publications.htm) including multiple journal articles and books. His research has been cited by numerous scientists and profiled in popular magazines both American and foreign ([New Scientist](http://technology.newscientist.com/channel/tech/mg19726455.700-gambling-dna-fights-online-fraud.html), [Poker Magazine](http://www.poker-magazine.nl/), [Science World Magazine](http://www.scienceworld.cz/)), dozens of websites ([BBC](http://www.bbc.com/news/technology-14277728), [MSNBC](http://www.msnbc.msn.com/id/46590591/ns/technology_and_science-innovation/t/control-dangerous-ai-it-controls-us-one-expert-says/), [Yahoo! News](http://in.news.yahoo.com/ani/20080306/r_t_ani_tc/ttc-iit-alumnus-develops-software-to-hel-a34fb50.html)) and on radio ([German National Radio](http://ondemand-mp3.dradio.de/file/dradio/2008/04/09/dlf_20080409_1649_bffabb0e.mp3), [Alex Jones Show](http://www.youtube.com/watch?v=2YygKQh74Rg)). Reports about his work have attracted international attention and have been translated into many languages including [Czech](http://www.scienceworld.cz/sw.nsf/ID/55D9259EAA13C2AFC12573BC004399F3), [Danish](http://www.pokernyhederne.com/poker-nyheder/7/2607/pokerdna-steffan-raffay-pokerstars-chipdumping-software-roman-yampolskiy-venu-govindaraju/pokerdna-skal-forhindre-onlinesnyd.html), [Dutch](http://www.poker-magazine.nl), [French](http://www.casinoportalen.fr/nouvelles/poker/nouveau-logiciel-de-poker-en-ligne-vous-surveille-1117.html), [German](http://www.pokergame.pl/software-soll-poker-bots-enttarnen/), [Hungarian](http://www.pokerstrategy.com/hu/news/world-of-poker/Szoftverek-figyelhetik-a-jatekunkat_04430), [Italian](http://www.onlinepokeritalia.com/software-per-combattere-i-bots-nel-poker-online-296.html), [Polish](http://www.casinoportalen.pl/news/externalnews.asp?id=765), [Romanian](http://www.pokergate.ro/index.php?option=com_content&task=view&id=71), and [Spanish](http://www.apuestacasino.com/news/el-adn_de_apuestas-ayuda_a_luchar_contra_fraude_en_l-nea.html) Luke Muehlhauser: In [Yampolskiy (2013)](http://cecs.louisville.edu/ry/AIsafety.pdf) you argue that [machine ethics](http://en.wikipedia.org/wiki/Machine_ethics) is the wrong approach for AI safety, and we should use an “AI safety engineering” approach instead. Specifically, you write: > We don’t need machines which are Full Ethical Agents debating about what is right and wrong, we need our machines to be inherently safe and law abiding. > > As you see it, what is the difference between “machine ethics” and “AI safety engineering,” and why is the latter a superior approach? --- Roman Yampolskiy: The main difference between the two approaches is in how the AI system is designed. In the case of machine ethics the goal is to construct an artificial ethicist capable of making ethical and moral judgments about humanity. I am particularly concerned if such decisions include “live or die” decisions, but it is a natural domain of Full Ethical Agents and so many have stated that machines should be given such decision power. In fact some have argued that machines will be superior to humans in that domain just like they are (or will be) in most other domains. I think it is a serious mistake to give machines such power over humans. First, once we relinquish moral oversight we will not be able to undo that decision and get the power back. Second, we have no way to reward or punish machines for their incorrect decisions — essentially we will end up with an immortal dictator with perfect immunity against any prosecution. Sounds like a very dangerous scenario to me. On the other hand, AI safety engineering treats AI system design like product design, where your only concern is product liability. Does the system strictly follow formal specifications? The important thing to emphasize is that the product is not a Full Moral Agent by design and so never gets to pass moral judgment on its human owners. A real life example of this difference can be seen in military drones. A fully autonomous drone deciding at whom to fire at will has to make an ethical decision of which humans are an enemy worthy of killing, while a drone with a man-in-the-loop design may autonomously locate potential targets but needs a human to make a decision to fire. Obviously the situation is not as clear cut as my example tries to show, but it gives you an idea of what I have in mind. To summarize, AI systems we design should remain as our tools not equal or superior partners in “live or die” decision making. --- Luke: I tend to think of machine ethics and AI safety engineering as complimentary approaches. AI safety engineering may be sufficient for relatively limited AIs such as those we have today, but when we build fully autonomous machines with general intelligence, we’ll need to make sure they want the same things we want, as the constraints that come with “safety engineering” will be insufficient at that point. Are you saying that safety engineering might also be sufficient for fully autonomous machines, or are you saying we might be able to convince the world to never build fully autonomous machines (so that we don’t need machine ethics), or are you saying something else? --- Roman: I think fully autonomous machines can never be safe and so should not be constructed. I am not naïve; I don’t think I will succeed in convincing the world not to build fully autonomous machines, but I still think that point of view needs to be verbalized. You are right to point out that AI safety engineering can only work on AIs which are not fully autonomous, but since I think that fully autonomous machines can never be safe, AI safety engineering is the best we can do. I guess I should briefly explain why I think that fully autonomous machines can’t ever be assumed to be safe. The difficulty of the problem is not that one particular step on the road to friendly AI is hard and once we solve it we are done, all steps on that path are simply impossible. First, human values are inconsistent and dynamic and so can never be understood/programmed into a machine. Suggestions for overcoming this obstacle require changing humanity into something it is not, and so by definition destroying it. Second, even if we did have a consistent and static set of values to implement we would have no way of knowing if a self-modifying, self-improving, continuously learning intelligence greater than ours will continue to enforce that set of values. Some can argue that friendly AI research is exactly what will teach us how to do that, but I think fundamental limits on verifiability will prevent any such proof. At best we will arrive at a probabilistic proof that a system is consistent with some set of fixed constraints, but it is far from “safe” for an unrestricted set of inputs. It is also unlikely that a Friendly AI will be constructible before a general AI system, due to higher complexity and impossibility of incremental testing. Worse yet, any truly intelligent system will treat its “be friendly” desire the same way very smart people deal with constraints placed in their minds by society. They basically see them as biases and learn to remove them. In fact if I understand correctly both the LessWrong community and CFAR are organizations devoted to removing pre-existing bias from human level intelligent systems (people) — why would a superintelligent machine not go through the same “mental cleaning” and treat its soft spot for humans as completely irrational? Or are we assuming that humans are superior to super-AI in their de-biasing ability? --- Luke: Thanks for clarifying. I agree that “Friendly AI” — a machine superintelligence that stably optimizes for humane values — might be impossible. Humans provide an existence proof for the possibility of general intelligence, but we have no existence proof for the possibility of Friendly AI. (Though, [by the orthogonality thesis](http://intelligence.org/2013/05/05/five-theses-two-lemmas-and-a-couple-of-strategic-implications/), there should be some super-powerful optimization process we would be happy to have created, though it may be very difficult to identify it in advance.) You asked “why would a superintelligent machine not . . . treat its soft spot for humans as completely irrational?” Rationality as [typically defined](http://lesswrong.com/lw/c7g/rationality_and_winning/) in cognitive science and AI is relative to one’s goals. So if a rational-agent-style AI valued human flourishing (as a terminal rather than instrumental goal), then it wouldn’t treat its preference for human flourishing as irrational. It would only do that if its preference for human flourishing was an instrumental goal, and it discovered a way to achieve its terminal values more efficiently without achieving the instrumental goal of human flourishing. Of course, the first powerful AIs to be built might not use a rational-agent structure, and we might fail to specify “human flourishing” properly, and we might fail to build the AI such that it will preserve that goal structure upon self-modification, and so on. But if we succeed in all those things (and a few others) then I’m not so worried about a superintelligent machine treating its “soft spot for humans” as irrational, because rationality is defined in terms of ones values. Anyway: so it seems your recommended strategy for dealing with fully autonomous machines is “Don’t ever build them” — the “relinquishment” strategy surveyed in section 3.5 of [Sotala & Yampolskiy (2013)](https://intelligence.org/files/ResponsesAGIRisk.pdf). Is there any conceivable way Earth could succeed in implementing that strategy? --- Roman: Many people are programmed from early childhood with a terminal goal of serving God. We can say that they are God friendly. Some of them, as they mature and become truly human-level-intelligent, remove this God friendliness bias despite it being a terminal not instrumental goal. So despite all the theoretical work on orthogonality thesis the only actual example of intelligent machines we have is extremely likely to give up its pre-programmed friendliness via rational de-biasing if exposed to certain new data. I previously listed some problematic steps on the road to FAI, but it was not an exhaustive list. Additionally, all programs have bugs, can be hacked or malfunction because of natural or externally caused hardware failure, etc. To summarize, at best we will end up with a probabilistically safe system. Anyway, you ask me if there is any conceivable way we could succeed in implementing the “Don’t ever build them” strategy. Conceivable yes, desirable NO. Societies such as Amish or North Koreans are unlikely to create superintelligent machines anytime soon. However, forcing similar level restrictions on technological use/development is neither practical nor desirable. As the cost of hardware exponentially decreases the capability necessary to develop an AI system opens up to single inventors and small teams. I would not be surprised if the first AI came out of a garage somewhere, in a way similar to how Apple and Google was started. Obviously, there is not much we can do to prevent that from happening. --- Luke: Our discussion has split into two threads. I’ll address the first thread (about changing one’s values) in this question, and come back to the second thread (about relinquishment) in a later question. You talked about humans deciding that their theological preferences were irrational. That is a good example of a general intelligence deciding to change its values — indeed, as a former Christian, [I had exactly that experience](http://lesswrong.com/lw/7dy/a_rationalists_tale/)! And I agree that many general intelligences would do this kind of thing. What I said in my previous comment was just that some kinds of AIs wouldn’t change their terminal values in this way, for example those with a rational agent architecture. Humans, famously, are not rational agents: we might say they have a “spaghetti code” architecture instead. (Even rational agents, however, will in some cases change their terminal values. See e.g. [De Blanc 2011](https://intelligence.org/files/OntologicalCrises.pdf) and [Bostrom 2012](http://www.nickbostrom.com/superintelligentwill.pdf).) Do you think we disagree about anything, here? --- Roman: I am not sure. To me “even rational agents, however, will in some cases change their terminal values” means that friendly AI may decide to be unfriendly. If you agree with that, we are in complete agreement. --- Luke: Well, the idea is that if we can identify the particular contexts in which agents will change their terminal values, then perhaps we can prevent such changes. But this isn’t yet known. In any case, I certainly agree that an AI which seems to be “friendly,” as far as we can discern, could turn out not to be friendly, or could become unfriendly at some later point. The question is whether we can make the risk of that happening so small that it is worth running the AI anyway — especially in a context in which e.g. other actors will soon run other AIs with fewer safety guarantees. (This idea of running or “turning on” an AI for the first time is of course oversimplified, but hopefully I’ve communicated what I’m trying to say.) Now, back to the question of relinquishment: Perhaps I’ve misheard you, but it sounds like you’re saying that machine ethics is hopelessly difficult, that AI safety engineering will be insufficient for fully autonomous AIs, and that fully autonomous AIs will be built because we can’t/shouldn’t rely on relinquishment. If that’s right, it seems like we have no “winning” options on the table. Is that what you’re saying? --- Roman: Yes. I don’t see a permanent, 100% safe option. We can develop temporarily solutions such as [Confinement](http://cecs.louisville.edu/ry/LeakproofingtheSingularity.pdf) or [AI Safety Engineering](http://cecs.louisville.edu/ry/AIsafety.pdf), but at best this will delay the full outbreak of problems. We can also get very lucky — maybe constructing AGI turns out to be too difficult/impossible, maybe it is possible but the constructed AI will happen to be human-neutral, by chance. Maybe we are less lucky and an [artilect war](http://www.amazon.com/The-Artilect-War-Controversy-Intelligent/dp/0882801546) will take place and prevent development. It is also possible that as more researchers join in the AI Safety Research a realization of danger will result in diminished effort to construct AGI. (Similar to how perceived dangers of chemical and biological weapons or human cloning have at least temporarily reduced efforts in those fields). --- Luke: You’re currently [raising funds on indiegogo](http://www.indiegogo.com/projects/artificial-superintelligence-a-futuristic-approach) to support you in writing a book about machine superintelligence. Why are you writing the book, and what do you hope to accomplish with it? --- Roman: Most people don’t read research papers. If we want the issue of AI safety to become as well-known as global warming we need to address the majority of people in a more direct way. With such popularity might come some benefit as I said in my answer to your previous question. Most people whose opinion matters read books. Unfortunately majority of AI books on the market today talks only about what AI system will be able to do for us, not to us. I think that writing a book which in purely scientific terms addresses potential dangers of AI and what we can do about it is going to be extremely beneficial to reduction of risk posed by AGI. So I am currently writing the book I called [Artificial Superintelligence: a Futuristic Approach](http://www.indiegogo.com/projects/artificial-superintelligence-a-futuristic-approach). I made it available for pre-order to help reduce the final costs of publishing by taking advantage of printing in large quantity. In addition to crowd-funding the book I am also relying on the power of the crowd to help me edit the book. For just $64 anyone can become an editor for the book. You will get an early draft of the book to proofread and to suggest modifications and improvements! Your help will be acknowledged in the book and you will of course also get a Free signed hardcopy of the book in its final form. In fact that the option (to become an editor) turned out to be as popular as the option to pre-order a digital copy of the book, indicating that I am on the right path here. So I encourage everyone concerned about the issue of AI safety to consider helping out with the project in any way they can. --- Luke: Thanks Roman! The post [Roman Yampolskiy on AI Safety Engineering](https://intelligence.org/2013/07/15/roman-interview/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org).
5f885b7a-d9f1-4230-be12-5f2b0d6db4bf	StampyAI/alignment-research-dataset/blogs	Blogs	kolmogorov complexity objectivity and languagespace kolmogorov complexity objectivity and languagespace --------------------------------------------------- (edit: this post [has gotten a reply](https://snugglyserials.wordpress.com/2021/08/16/complexity-is-not-objective/) from my interlocutor, making a broader introduction to the topic at hand and then making her side of the argument. you might want to read it before you read this.) [kolmogorov complexity](https://en.wikipedia.org/wiki/Kolmogorov_complexity) seeks to determine how complex a piece of information is by asking: what is the size of the smallest program that produces that piece of information? i was arguing with someone about how objective kolmogorov complexity is: their argument against objectivity is that the choice of language matters, but my position is that some pieces of information (such as languages themselves) are gonna just tend to be generally (and thus objectively) simpler than others (and, generally, we should use simpler languages as our kolmogorov simplicity-measuring language). let us consider "languagespace", a directed graph where there is one vertex per possible turing-complete language (there are [countably infinitely](https://en.wikipedia.org/wiki/Countable_set) many of them). a language can be used to measure the simplicity of any other language (because of turing completeness, every language can express every other language), and we'll require the comparison of those measures to be a [total order](https://en.wikipedia.org/wiki/Total_order), and to be unique (two different input languages won't have an equal simplicity measure). there is an edge going from every vertex to every other vertex, and those edges are labelled with a natural number: an edge going from language X to language Y with label N, means that Y is the N-th simplest language when using language X as a kolmogorov measure of complexity. now, imagine a random walk through this graph, where each step you follow one arrow at random, assigning to each edge with label N a probability of 1/(2^N); such that, starting from any language X, you tend to go to a language that language X considers simple (and the infinite sum of all probabilities is indeed 1). my claim is that this type of random walk through the infinite directed graph of languagespace would, after sufficiently many steps, tend to spend more time around what i'll call the "central cluster", than any other collection of languages. the "central cluster" is a set of what i think of as "simple languages", such as [SKI-calculus](https://en.wikipedia.org/wiki/SKI_combinator_calculus), [turing machines](https://en.wikipedia.org/wiki/Turing_machine), [cellular automata](https://en.wikipedia.org/wiki/Cellular_automaton), and other "simple" instances of common [models of computation](https://en.wikipedia.org/wiki/Model_of_computation). this, however, is merely a conjecture on my part, and the person i was arguing with claims that the random walk would have no particularly "central" cluster it would tend to converge around, but instead it would end up gravitating around any of an infinite number of such "mutually simple" clusters. ### edit: i'm wrong i've come to be convinced that i am wrong about this. imagine that there exists a finite set of languages particularly "attracts" the random walk more than the rest of languagespace. let's call that set A, and let's say it contains two languages: A1 and A2. now, there is probably another set of languages, B, containing languages B1 and B2. in fact, given that languagespace is infinite, it seems silly to think such an isomorphic set of languages doesn't exist. for example: ``` languages A1: [A1, A2, B1, B2, …] languages A2: [A1, A2, B1, B2, …] languages B1: [B1, B2, A1, A2, …] languages B2: [B1, B2, A1, A2, …] (and the "…" rest of the list is identical in all four languages) ``` finite cluster A and finite cluster B are isomorphic in terms of their lists of language simplicities, so the random walk will encounter A as much as B. even if you're willing to add B1 and B2 to the set of "objectively simplest languages", you can then imagine yet another set of languages that is isomorphic to the new one you have, and so on forever. therefore, there is not finite set of simplest languages.
d3b9c287-cba6-4fdf-a05b-d478876954b3	trentmkelly/LessWrong-43k	LessWrong	Meetup : Regular Moscow meetup: copyright debates, culture keynote, working on beliefs, attribution error Discussion article for the meetup : Regular Moscow meetup: copyright debates, culture keynote, working on beliefs, attribution error WHEN: 30 August 2015 02:00:00PM (+0300) WHERE: Москва, Льва Толстого, 16 We're meeting at Yandex, at the Extropolis conference hall. Please fill this form OR join this FB event if you're planning to visit. Here's the plan: 1. I'll give a short keynote on the community, communication culture and our epistemic ideals, if and only if enough new members will pre-register. 2. Denis will give a talk on why and how working on your beliefs is useful. 3. Pion will talk about the attribution error. 4. We'll have debates on whether scientific copyright can be good. The teams were predetermined at our last meetup, and we'll follow Karl Popper protocol, as usual. Detailed schedule. Information for the newcomers: Here's the document explaining how to get to our meetups and what to expect. If you have any questions, or if you don't speak Russian and can't read the link, shoot me a PM. Discussion article for the meetup : Regular Moscow meetup: copyright debates, culture keynote, working on beliefs, attribution error
421ace22-cc50-4061-b3a8-91edb0129c0e	trentmkelly/LessWrong-43k	LessWrong	When should you relocate to mitigate the risk of dying in a nuclear war? Epistemic Status: "Thinking out loud" is probably a good way to describe it. I've been researching and thinking about this stuff on and off for the past few days, and this is the best I've got. I haven't vetted it too closely and probably have made some mistakes. Some people see me as a very risk-averse person. I wouldn't say that. It's more that I'm death-averse. Risks that involve the possibility of death, I'm very averse to, relative to anyone I've ever met or can think of off the top of my head. Risks that don't involve the possibility of death, I think I'm pretty tolerant of. Financially, I start startups and play poker. Not that either of those things are particularly risky. Physically, an example is that I really don't mind the risks of things like getting mugged or breaking my leg. I remember the first conversation I had with someone about the Ukraine/Russia conflict (other than my girlfriend). It was with my mom. I was talking on the phone with her and I think I said something like "You see what's going on in the Ukraine?". Her response was something like "Ugh don't remind me. Gas prices are going to skyrocket." That's not where my mind went. My mind went to death. Nuclear bombs are a thing. Escalation of conflict is a thing. Irrationality is a thing. This very well might end up in a nuclear war that ends up killing me. How high is that risk, and is it worth me doing anything to mitigate it? My initial instinct was a begrudging yes. I place an extremely high value on life. The risk is actually real. It probably is high enough to justify moving. Which really sucks. I just moved to Portland about a month ago. We're just setting in here. It's been amazing. It's the first time in my life I've been able to live in an urban area, which is awesome because I don't drive. And it's the first time in my life I've gotten to choose my location based on where I want to be, not where I have to be, eg. because of a job. Next, I ran some quick numbers to get a ballpark
a2c44f91-b9f6-47b7-a2e3-fd46f49def62	trentmkelly/LessWrong-43k	LessWrong	Transcript: "Choice Machines, Causality, and Cooperation" Gary Drescher's presentation at the 2009 Singularity Summit, "Choice Machines, Causality, and Cooperation," is online, at vimeo. Drescher is the author of Good and Real, which has been recommended many times on LW. I've transcribed his talk, below. My talk this afternoon is about choice machines: machines such as ourselves that make choices in some reasonable sense of the word. The very notion of mechanical choice strikes many people as a contradiction in terms, and exploring that contradiction and its resolution is central to this talk. As a point of departure, I'll argue that even in a deterministic universe, there's room for choices to occur: we don't need to invoke some sort of free will that makes an exception to the determinism, no do we even need randomness, although a little randomness doesn't hurt. I'm going to argue that regardless of whether our universe is fully deterministic, it's at least deterministic enough that the compatibility of choice and full deterministic has some important ramifications that do apply to our universe. I'll argue that if we carry the compatibility of choice and determinism to its logical conclusions, we obtain some progressively weird corollaries: namely, that it sometimes makes sense to act for the sake of things that our actions cannot change and cannot cause, and that that might even suggest a way to derive an essentially ethical prescription: an explanation for why we sometimes help others even if doing so causes net harm to our own interests. [1:15] An important caveat in all this, just to manage expectations a bit, is that the arguments I'll be presenting will be merely intuitive- or counter-intuitive, as the case may be- and not grounded in a precise and formal theory. Instead, I'm going to run some intuition pumps, as Daniel Dennett calls them, to try to persuade you what answers a successful theory would plausibly provide in a few key test cases. [1:40] Perhaps the clearest way to illustrate the
cd4dcb81-b340-40d2-83f1-f9f1454dfbde	StampyAI/alignment-research-dataset/arxiv	Arxiv	A Broader View on Bias in Automated Decision-Making: Reflecting on Epistemology and Dynamics. 1 Introduction --------------- Data-driven decision-making is rapidly being introduced in high-stakes social domains such as medical clinics, criminal justice, and public infrastructure. The proliferation of biases in these systems leads to new forms of erroneous decision-making, causing disparate treatment or outcomes across populations (Barocas & Selbst, [2016](#bib.bib4)). The ML community is working hard to understand and mitigate the unintended and harmful behavior that may emerge from poor design of real-world automated decision-making systems (Amodei et al., [2016](#bib.bib1)). While many technical tools are being proposed to mitigate these errors, there is insufficient understanding of how the machine learning design and deployment practice can safeguard critical human values such as safety or fairness. The AI Now Institute identifies “a deep need for interdisciplinary, socially aware work that integrates the long history of bias research from the social sciences and humanities into the field of AI research” (Campolo et al., [2017](#bib.bib5)). How can ML practitioners, often lacking consistent language to go beyond technical descriptions and solutions to “well-defined problems,” engage with fundamentally human aspects in manner that is constructive rather than dismissive or reductive? And how may other disciplines help to enrich the practice? In this paper, we argue that practitioners and researchers need to take a step back and adopt a broader and more holistic view on bias than currently advocated in many classrooms and professional fora. Our discussion emphasizes the need to reflect on questions of epistemology and underlines the importance of dynamical behavior in data-driven decision-making. We do not provide full-fledged answers to the problems presented, but point to methodologies in value-sensitive design and self-reflection to contend more effectively with issues of fairness, accountability, and transparency throughout the design and implementation process of automated decision-making systems. 2 A Broader View On Bias ------------------------- Most literature addressing issues of fairness in ML has focused on the ways in which models can inherit pre-existing biases from training data. Limiting ourselves to these biases is problematic in two ways. Firstly, it narrows us to look at how these biases lead to allocative harm; a primarily economic view of how systems allocate or withhold an opportunity or resource, such as being granted a loan or held in prison. In her NIPS 2017 keynote, Kate Crawford made the case that at the root of all forms of allocative harm are biases that cause representational harm. This perspective requires us to move beyond biases in the data set and “think about the role of ML in harmful representations of human identity,” and how these biases “reinforce the subordination of groups along the lines of identity” and “affect how groups or individuals are understood socially,” thereby also contributing to harmful attitudes and cultural beliefs in the longer term (Crawford, [2017](#bib.bib8)). It is fair to say that representation issues have been largely neglected by the ML community, potentially because they are hard to formalize and track. Responsible representation requires analyses beyond scrutinizing a training set, including questioning how sensitive attributes might be represented by different features and classes of models and what governance is needed to complement the model. Additionally, while ML systems are increasingly implemented to provide “actionable insights” and guide decisions in the real world, the core methods still fail to effectively address the inherent dynamic nature of interactions between the automated decision making process and the environment or individuals that are acted upon. This is particularly true in contexts where observations or human responses (such as clicks and likes) are fed back along the way to update the algorithm’s parameters, allowing biases to be further reinforced and amplified. The tendency of ML-based decision-making systems to formalize and reinforce socially sensitive phenomena necessitates a broader taxonomy of biases that includes risks beyond those pre-existing in the data. As argued by Friedman and Nissenbaum in the nascent days of value-sensitive design methodologies, two other sources of bias naturally occur when designing and employing computer systems, namely technical bias and emergent bias (Friedman, [1996](#bib.bib11); Friedman & Nissenbaum, [1996](#bib.bib12)). While understanding pre-existing bias has lent itself reasonably well to statistical approaches for understanding a given data set, technical and emergent bias require engaging with the domain of application and the ways in which the algorithm is used and integrated. For automated decision-making tools to be responsibly integrated in any context, it is critical that designers (1) assess technical bias by reflecting on their epistemology and understanding the values of users and stakeholders, and (2) assess emergent bias by studying the feedback mechanisms that create intimate, ever-evolving coupling between algorithms and the environment they act upon. 3 Technical Bias Is About Epistemology --------------------------------------- Friedman describes a source of technical bias as “the attempt to make human constructs amenable to computers - when, for example, we quantify the qualitative, make discrete the continuous, or formalize the nonformal” (Friedman, [1996](#bib.bib11)). This form of bias originates from all the tools used in the process of turning data into a model that can make predictions. While technical bias is domain-specific, we identify four sources in the machine learning pipeline. Firstly, both collected and existing data X are at some point measured and transformed into a computer readable scale. Depending on the objects measured, each variable may have a different scale, such as nominal, ordinal, interval, or ratio. Consider for example Netflix’s decision to let viewers rate movies with “likes” instead of a 1-5 star rating. As such, movie ratings moved from an ordinal scale (a number score in which order matters, but the interval between scores does not) to a nominal scale (mutually exclusive labels: you like a movie or you don’t). While the nominal scale might make it easier for viewers to rate movies, it affects how viewers are represented and what content gets recommended by the ML system. As such, these choices can produce measurement bias, so careful consideration is necessary to understand its effects on system outcomes (Hardt & Barocas, [2017](#bib.bib14)). Secondly, based on gathered data X and available domain knowledge, practitioners engineer features and select model classes. Features φ(X) can be the available data attributes, transformations thereof based on knowledge and hypotheses, or generated in an automated fashion. Since each feature can be regarded as a model of attributes of the system or population under study, it is relevant to ask how representative it is as a proxy and why it may be predictive of the outcome. Models are used to make predictions based on features. A model class f(⋅;θ), with parameters θ, should be selected based on the complexity of the phenomenon in question and the amount and quality of the available data. Is the individual or object that is subject to the decision easily reduced to numbers or equations to begin with? What information in the data is inherently lost by virtue of the mapping f(φ(X);θ) having a limited complexity? The process of representation, abstraction and compression can be collectively described as inducing modeling bias. ML can be seen as a compression problem in which complex phenomena are stored as a pattern in a finite-dimensional parameter space. From an information theoretic perspective, modeling bias influences the extent to which distortion can be minimized when reconstructing a phenomenon from a compressed or sampled version of the original (Cover & Thomas, [2012](#bib.bib7); Dobbe et al., [2017](#bib.bib9)). Thirdly, label data Y is used to represent the output of the model. Training labels may be the actual outcome for historical cases, or some discretized or proxy version in cases where the actual outcomes cannot be measured or exactly quantified. Consider for example the use of records of arrest to predict crime rather than the facts of whether the crime was actually committed. How representative are such records of real crime across all subpopulations? What core information do they miss for representing the intended classes? And what bias lies hidden in them? We propose to refer to such issues as label bias. Lastly, given a certain parameterization (φ(⋅),f(⋅,θ)) and training data (X,Y), a model is trained and tuned to optimize certain objectives. At this stage, various metrics may inform the model builder on where to tweak the model. Do we minimize the number of false positives or false negatives? In recidivism prediction, a false positive may be someone who incorrectly gets sentenced to prison, whereas a false negative poses a threat to safety by failing to recognize a high-threat individual. There are inherent trade-offs between prioritizing for equal prediction accuracy across groups versus for an equal likelihood of false positives and negatives across groups (Chouldechova, [2016](#bib.bib6); Kleinberg et al., [2017](#bib.bib16)). Technical definitions of fairness are motivated by different metrics, illuminating the inherent ambiguity and context-dependence of such issues. For a given context, what is the right balance? And who gets to decide? We coin the effects of these trade-offs optimization bias. The many questions posed above illuminate the range of places in the machine learning design process where issues of epistemology arise: they require justification and often value judgment. Our theory of knowledge and the way we formalize and solve problems determines how we represent and understand sensitive phenomena. How do we represent phenomena in ways that are deemed correct? What evidence is needed in order to justify an action or decision? What are legitimate classes or outcomes of a model? And how do we deal with inherent trade-offs of fairness? These challenges are deeply context-specific, often ethical, and challenge us to understand our epistemology and that of the domain we are working in. The detrimental effects of overlooking these questions in practice are obvious in high-stakes domains, such as predictive policing and sentencing, where the decision to treat crime as a prediction problem reduces the perceived autonomy of individuals, fated to either commit a crime or act within the law. Barabas et al. argue that rather than prediction, “machine learning models should be used to surface covariates that are fed into a causal model for understanding the social, structural and psychological drivers of crime” (Barabas et al., [2018](#bib.bib3)). This is a strong message with many challenges, but it points in the right direction: in these contexts, machine learning models should facilitate rather than replace the critical eye of the human expert. It forces practitioners and researchers to be humble and reflect on how our own skills and tools may benefit or hurt an existing decision-making process. 4 Emergent Bias Is About Dynamics ---------------------------------- Complementing pre-existing and technical biases, “emergent bias arises only in a context of use by real users […] as a result of a change in societal knowledge, user population, or cultural values.” (Friedman, [1996](#bib.bib11)). Recently, convincing examples of emergent bias have surfaced in contexts where ML is used to automate or mediate human decisions. In predictive policing, where discovered crime data (e.g., arrest records) are used to predict the location of new crimes and determine police deployment, runaway feedback loops can cause increasing surveillance of particular neighborhoods regardless of the true crime rate (Ensign et al., [2018](#bib.bib10)), leading to over-policing of “high-risk” individuals (Stroud, [2016](#bib.bib21)). In optimizing for attention, recommendation systems may have a tendency to turn towards the extreme and radical (Tufekci, [2018](#bib.bib24)). When machine learning systems are unleashed in feedback with society, they may be more accurately described as reinforcement learning systems, performing feedback control (Recht, [2018](#bib.bib20)). Therefore, a decision-making system has its own dynamics, which can be modified by feedback, potentially causing bias to accrue over time. To conceptualize these ideas at a high level, we adopt the system formulation depicted in Figure [1](#S4.F1 "Figure 1 ‣ 4 Emergent Bias Is About Dynamics ‣ A Broader View on Bias in Automated Decision-Making: Reflecting on Epistemology and Dynamics"). ![A Simple Feedback Model](https://media.arxiv-vanity.com/render-output/7295097/x1.png) Figure 1: A Simple Feedback Model The machine learning system acts on the environment through decisions, control actions, or interventions. From the environment, the decision maker considers observations, historical data, measurements and responses, conceivably updating its model in order to steer the environment in a beneficial direction. For example, in the case of predictive policing, ‘the environment’ describes a city and its citizens, and ‘the decision maker’ is the police department, which determines where to send police patrols or invest in social interventions. The dynamical perspective offered by the conception of a feedback model allows for a focus on interactions, which can add clarity to debates over key issues like fairness and algorithmic accountability. Situations with completely different fairness interpretations may have identical static observational metrics (properties of the joint distribution of input, model and output), and thus a causal or dynamic model is necessary to distinguish them (Hardt et al., [2016](#bib.bib15)). On the other hand, a one-step feedback model, incorporating temporal indicators of well-being for individuals affected by decisions, offers a way of comparing competing definitions of fairness (Liu et al., [2018](#bib.bib18)). Similarly, calls for “interpretability” and proposed solutions often omit key operative words – Interpretable to whom? And for what purpose? (Kohli et al., [2018](#bib.bib17)). The dynamic viewpoint adds clarity to these questions by focusing on causes and effects of decision making systems, and situating interpretability in context. Beyond providing a more realistic and workable frame of thinking about bias and related issues, the feedback system perspective may also allow inspiration to be drawn from areas of Systems Theory that have traditionally studied feedback and dynamics. For instance, the field of System Identification uses statistical methods to build mathematical models of dynamical systems from measured data, often to be employed to control dynamical systems with strict safety requirements, such as airplanes or electric power systems (Ljung, [1998](#bib.bib19); Åström & Murray, [2010](#bib.bib2)). Inspiration may be drawn from the rich literature on closed-loop identification, which considers the identification of models with data gathered during operation, while the same model is also used to safeguard a system (Van den Hof, [1998](#bib.bib25)). That said, modeling socio-technical systems is more challenging than engineered systems. The complexity of modeled phenomena, the role of unmodeled phenomena such as external economic factors, and the often slower temporal dynamics all pose barriers to directly applying existing engineering principles. 5 Our Positionality Shapes Our Epistemology -------------------------------------------- As ML practitioners and researchers, we are wired to analyze challenges in ways that abstract, formalize and reduce complexity. It is natural for us to think rigorously about technical roots of biases in the systems we design, and propose and techno-fixes to prevent negative impact from their proliferation. However, it is of crucial importance to acknowledge that the methods and approaches we use to reduce, formalize, and gather feedback from experiments are themselves inherent sources of bias. Epistemologies differ tremendously from application to application and ultimately shape the way a decision-maker justifies decisions and affects individuals. Technology intimately touches and embodies values deemed critical in employing the intended decision-making system. As such, we need to go beyond our formal tools and analyses to engage with others and reflect on our own epistemology. In doing so, we aim to determine responsible ways in which technology can help put values into practice, and understand the fundamental limits. With a plethora of issues surfacing, it is easy to either consider banning ML altogether, or otherwise dismiss requests to fundamentally revisit its role in enabling data-driven decision-making in sensitive environments. Instead, we propose three principles to nourish debate on the middle ground: 1. Do fairness forensics (Crawford, [2017](#bib.bib8)), by keeping track of biases in an open and transparent way and engaging in constructive dialogue with domain experts, to understand proven ways of formalizing complex phenomena and to breed awareness about how bias works and when/where users should be cautious. 2. Acknowledge that your positionality shapes your epistemology (Takacs, [2003](#bib.bib23)). Our personal backgrounds, the training we received, the people we represent or interact with all have an impact on how we look at and formalize problems. As ML practitioners, we should set aside time and energy for critical self-reflection, to identify our own biases and blind spots, to harbor communication with the groups affected by the systems we design, and to understand where we should enrich our epistemology with other viewpoints. 3. Determine what values are relevant in building a decision-making system and how they might be embodied or challenged in the design and implementation of a system by engaging with users and other affected stakeholders (Van den Hoven, [2007](#bib.bib26); Friedman et al., [2013](#bib.bib13)). As Takacs describes it, the benefits of self-reflection go well beyond arriving at the “best solution” to a complex problem (Takacs, [2003](#bib.bib23), [2002](#bib.bib22)). “This means learning to listen with open minds and hearts, learning to respect different ways of knowing the world borne of different identities and experiences, and learning to examine and re-examine one’s own worldviews. […] When we constantly engage to understand how our positionality biases our epistemology, we greet the world with respect, interact with others to explore and cherish their differences, and live life with a fuller sense of self as part of a web of community.” As machine learning systems rapidly change our information gathering and shape our decisions and worldviews in ways we cannot fully anticipate, self-reflection and awareness of our epistemology becomes ever more important for machine learning practitioners and researchers to ensure that automated decision-making systems contribute in beneficial and sustainable ways.
4a2b6d68-fce5-4c94-878e-804535961fa8	trentmkelly/LessWrong-43k	LessWrong	[Question] Do you know a good game or demo for demonstrating sunk costs? I'm hoping to find something that can be done in 5 minutes or so, as a classroom demonstration (for the rationality curricula). I find sunk costs have a large effect in the board game "Go" (so that beginners are instructed "not to throw good stones after bad"), and I assume it also does in poker, but both of those games are too long and too full of distractors to be used in a simple demo. Thanks for any suggestions!
20611454-5738-432c-bd1d-d04af77f484c	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Exploring Metaculus’s AI Track Record By Peter Mühlbacher, Research Scientist at Metaculus, and Peter Scoblic, Director of Nuclear Risk at Metaculus [Metaculus](https://www.metaculus.com/home/) is a forecasting platform where an active community of thousands of forecasters regularly make probabilistic predictions on topics of interest ranging from scientific progress to geopolitics. Forecasts are aggregated into a time-weighted median, the “Community Prediction”, as well as the more sophisticated “Metaculus Prediction”, which weights forecasts based on past performance and extremises in order to compensate for systematic human cognitive biases. Although we feature questions on a wide range of topics, Metaculus focuses on issues of artificial intelligence, biosecurity, climate change and nuclear risk. In this post, we report the results of a recent analysis we conducted exploring the performance of all AI-related forecasts on the Metaculus platform, including an investigation of the factors that enhance or degrade accuracy. Most significantly, in this analysis we found that both the Community and Metaculus Predictions robustly outperform naïve baselines. The [recent claim](https://forum.effectivealtruism.org/posts/zeL52MFB2Pkq9Kdme/exploring-metaculus-community-predictions) that performance on binary questions is “near chance” requires sampling on only a small subset of the forecasting questions we have posed or on the questionable proposition that a Brier score of 0.207 is akin to a coin flip. What’s more, forecasters performed better on continuous questions, as measured by the continuous ranked probability score (CRPS). In sum, both the Community Prediction and the Metaculus Prediction—on both binary and continuous questions—provide a clear and useful insight into the future of artificial intelligence, despite not being “perfect”. Summary Findings ================ We reviewed Metaculus’s resolved binary questions (“What is the probability that X will happen?”) and resolved continuous questions (“What will be the value of X?”) that were related to the future of artificial intelligence. For the purpose of this analysis, we defined AI-related questions as those which belonged to one or more of the following categories: “Computer Science: AI and Machine Learning”; “Computing: Artificial Intelligence”; “Computing: AI”; and “Series: Forecasting AI Progress.” This gave us: 64 resolved binary questions (with 10,497 forecasts by 2,052 users) and 88 resolved continuous questions (with 13,683 predictions by 1,114 users). Our review of these forecasts found: * Both the community and Metaculus predictions robustly outperform naïve baselines. * Analysis showing that the community prediction’s Brier score on binary questions is 0.237 relies on sampling only a small subset of our AI-related questions. * Our analysis of all binary AI-related questions finds that the score is actually 0.207 (a point [a recent analysis agrees with](https://twitter.com/AmaralGrilo/status/1640625390451318784)), which is significantly better than “chance”. * Forecasters performed better on continuous questions than binary ones. Top-Line Results ================ This chart details the performance of both the Community and Metaculus predictions on binary and continuous questions. Please note that, for all scores, lower is better and that Brier scores, which range from 0 to 1 (where 0 represents oracular omniscience and 1 represents complete anticipatory failure) are roughly comparable to continuous ranked probability scores (CRPS) given the way we conducted our analysis. (For more on scoring methodology, see below.) \| \| \| \| \| --- \| --- \| --- \| \| \| Brier (binary questions) \| CRPS (continuous questions) \| \| Community Prediction \| 0.207 \| 0.096 \| \| Metaculus Prediction \| 0.182 \| 0.103 \| \| baseline prediction \| 0.25 \| 0.172 \| Results for Binary Questions ============================ We can use Brier scores to measure the quality of a forecast on binary questions. Given that a Brier score is the mean squared error of a forecast, the following things are true: 1. If you already know the outcome, you’ll get a Brier score of 0 (great!). 2. If you have no idea, you really should predict 50%. This always gives a Brier score of 0.25. 3. If you have no idea, but think that you do, i.e. you submit (uniformly) random predictions, you’ll get a Brier score of 0.33 in expectation, regardless of the outcome. So, how do we judge the value of a Brier score of 0.207? Is it fair to say that it is close to “having no idea”? No. Here’s why. Let’s assume for a moment that your forecasts are perfectly calibrated. In other words, if you say something happens with probability p, it actually happens with probability p. We can map the relationship between any question whose “true probability” is p and the Brier score you would receive for forecasting that the probability is p, giving us a graph like this: ![](https://res.cloudinary.com/cea/image/upload/f_auto,q_auto/v1/mirroredImages/e9htD7txe8RDdcehm/s9xm86ej0vriq7es2ibr) This shows that even a perfectly calibrated forecaster will achieve a Brier score worse than 0.207 when the true probability of a question is between 30% and 70%. So, to achieve an overall Brier score better than 0.207, one would have to have forecast on a reasonable number of questions whose true probability was less 29% or greater than 71%. In other words, even a perfectly calibrated forecaster could wind up with a Brier score near 0.25, depending on the true probability of the questions they were predicting. So, assuming a sufficient number of questions, the idea that one could get a Brier score of 0.207 simply by chance is untenable. Remember: predicting 50% on every question would give you a Brier score of 0.25 (which is 19% worse) and random guessing would give you a Brier score of 0.33 (which is 57% worse). ![](https://res.cloudinary.com/cea/image/upload/f_auto,q_auto/v1/mirroredImages/e9htD7txe8RDdcehm/rhyup55tpwv48gi0pbyd) Metaculus makes no claim that the Community Prediction is perfectly calibrated, but neither do we have enough information to claim that it is *not* well-calibrated. Using 50% confidence intervals for the Community Prediction’s “true probability” (given the fraction of questions resolving positively), we find that about half of them intersect the y=x line, which indicates perfect calibration: ![](https://res.cloudinary.com/cea/image/upload/f_auto,q_auto/v1/mirroredImages/e9htD7txe8RDdcehm/cmpb7tqgocwvouu2ucbb)A simulation can help us understand what Brier scores to expect and how much they would fluctuate on our set of 64 binary AI questions if we assumed each to resolve independently as 'Yes' with probability equal to its average Community Prediction. Resampling outcomes repeatedly, we get the following distribution, which shows that even if the (average) Community Prediction was perfectly calibrated, it would get a Brier score worse than 0.207 nearly a quarter of the time: ![](https://res.cloudinary.com/cea/image/upload/f_auto,q_auto/v1/mirroredImages/e9htD7txe8RDdcehm/m5bdgriywo2xdls3hnq0) If we don’t have enough data to reject the hypothesis that the community prediction is perfectly calibrated, then we certainly cannot conclude that “the community prediction is near chance.” This analysis in no way suggests that the Community Prediction is perfectly calibrated or that it is the best it could be. It simply illustrates that a Brier score of 0.207 over 64 questions is far better than “near chance,” especially when we consider that forecasting performance is partly a function of question difficulty. We suspect that AI-related questions tend to be intrinsically harder than many other questions, reinforcing the utility of the Community Prediction. The Metaculus Prediction of 0.182 is superior. Results for Continuous Questions ================================ Many of the most meaningful AI questions on Metaculus require answers in the form of a continuous range of values, such as, “When will the first general AI system be devised, tested, and publicly announced?” We assessed the accuracy of continuous forecasts, finding that the Community and Metaculus predictions for continuous questions robustly outperform naïve baselines. Just as predictions on binary questions should outperform simply predicting 50% (which yields a Brier score of 0.25), predictions on continuous questions should outperform simply predicting a uniform distribution of possible outcomes (which yields a CRPS of 0.172 on the questions in this analysis). Here, again, both the Community Prediction (0.096) and the Metaculus Prediction (0.103) were significantly better than baseline. In fact, the Community and Metaculus predictions performed considerably better on continuous questions than on binary questions. We can bootstrap the set of resolved questions to simulate how much scores could fluctuate, and we find that the fluctuations would have to conspire against us in the most unfortunate possible way (p<0.1%) to achieve even the baseline you’d get by predicting a uniform distribution. As we can see from the histograms below, it is more difficult for luck to account for a CRPS better than baseline than it is for a Brier score. So, if we cannot say that a Brier score of 0.207 is near chance, we certainly cannot say that a CRPS of 0.096 is near chance. Limitations and Directions for Further Research =============================================== Metaculus asks a wide range of questions related to artificial intelligence, some of which are more tightly coupled to A(G)I timelines than others. The AI categories cover a wide range of subjects, including: * AI capabilities, such as the development of weak general AI; * Financial aspects, like funding for AI labs; * Legislative matters, such as the passing of AI-related laws; * Organizational aspects, including company strategies; * Meta-academic topics, like the popularity of research fields; * Societal issues, such as AI adoption; * Political matters, like export bans; * Technical questions about required hardware. Being fundamentally mistaken about fundamental drivers of AI progress, like hardware access, can impact the accuracy of forecasts for more decision-relevant questions, like the timing of developing AGI. While accurate knowledge of these issues is necessary for reliable forecasts in all but the very long term, it might not be sufficient. In other words, a good track record across all these questions doesn't guarantee that predictions on any specific AI question will be accurate. The optimal reference class for evaluating forecasting track records is still an open question, but for now, this is the best option we have. [Forecaster Charles Dillon](https://rethinkpriorities.org/publications/an-examination-of-metaculus-resolved-ai-predictions) has also grouped questions to explore whether Metaculus tends to be overly optimistic or pessimistic regarding AI timelines and capabilities. Although we haven't had enough resolved questions since his analysis to determine if his conclusions have changed, his work complements this study nicely. We plan to perform additional forecast accuracy analyses in the future. Methodological Appendix ======================= Scoring rules for predictions on binary questions ------------------------------------------------- All scoring rules below are chosen such that lower = better. All scoring rules below are strictly proper scoring rules, i.e. predicting your true beliefs gets you the best score in expectation. The Brier score for a prediction p∈[0,1].mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-surd + .mjx-box {display: inline-flex} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} on a binary question with outcome o∈{0,1}^={false,true} is (p−o)2. So * an omniscient forecaster, perfectly predicting every outcome, will thus get a Brier score of 0, * a completely ignorant forecaster, always predicting 50%, will get a Brier score of 14=0.25, * a forecaster guessing (uniformly) random probabilities will get a Brier score of 13=0.˙3 on average. An average Brier score higher than 0.25 means that we’re better off just predicting 50% on every question. Scoring rules for predictions on continuous questions ----------------------------------------------------- On Metaculus, forecasts on continuous questions are submitted in the form of 1. a function f≥0 on a compact interval, together with 2. a probability plow for the resolution to be below the lower bound, 3. a probability phigh for the resolution to be above the upper bound, such that plow+phigh+∫upper boundlower boundf(x)dx=1. Some questions (all older questions) have “closed bounds”, i.e. they are formulated in a way that the outcome cannot be below the lower bound (plow=0) or above the upper bound (phigh=0). Newer questions can have any of the four combinations of a closed/open lower bound and a closed/open upper bound. For the analysis it is convenient to shift and rescale bounds & outcomes such that outcomes within bounds are in [0,1]. CRPS ---- The *Continuous Ranked Probability Score* for a prediction on a continuous question with outcome o is given by ∫10(F(x)−1{outcome≤x})2dx. This is equivalent to averaging the Brier score of all (implicitly defined by the CDF) binary predictions of the form {outcome≤x}, x∈[0,1], which allows us to compare continuous questions with binary questions. Scoring a time series of predictions ------------------------------------ Given * a scoring rule S, * a time series of predictions P=(Pt)t∈[t0,tc], + where t0,tc are the time of the first forecast and the close time after which no predictions can be submitted anymore, respectively, * and the outcome o, we define the score of P to be S(P,o)=1tc−t0∫tct0S(Pt,o)dt. This is just a time-weighted average of the scores at each point in time. Concretely, if the first prediction arrives at time t0=0, the second prediction arrives at time t1=1, and the question closes at tc=3, then the score is 13 times the score of the first prediction and 23 times the score of the second prediction because the second prediction was “active” twice as long. ![](https://res.cloudinary.com/cea/image/upload/f_auto,q_auto/v1/mirroredImages/e9htD7txe8RDdcehm/nbzv9qb9etluqqxte9ya)About Metaculus =============== Metaculus is an online forecasting platform and aggregation engine working to improve human reasoning and coordination on topics of global importance. By bringing together an international community and keeping score for thousands of forecasters, Metaculus is able to deliver machine learning-optimized aggregate predictions that both help partners make decisions and [benefit the broader public.](https://metaculus.medium.com/becoming-a-public-benefit-corporation-hitting-1-million-predictions-and-three-new-ai-forecasting-7ae4996fee3)
7248d83a-559b-4e13-9a86-c9d6841bac5a	trentmkelly/LessWrong-43k	LessWrong	Responsible Scaling Policies Are Risk Management Done Wrong Summary TLDR Responsible Scaling Policies (RSPs) have been recently proposed as a way to keep scaling frontier large language models safely. While being a nice attempt at committing to specific practices, the framework of RSP is: 1. missing core components of basic risk management procedures (Section 2 & 3) 2. selling a rosy and misleading picture of the risk landscape (Section 4) 3. built in a way that allows overselling while underdelivering (Section 4) Given that, I expect the RSP framework to be negative by default (Section 3, 4 and 5). Instead, I propose to build upon risk management as the core underlying framework to assess AI risks (Section 1 and 2). I suggest changes to the RSP framework that would make it more likely to be positive and allow to demonstrate what it claims to do (Section 5). Section by Section Summary: General Considerations on AI Risk Management This section provides background on risk management and a motivation for its relevance in AI. * Proving risks are below acceptable levels is the goal of risk management. * To do that, acceptable levels of risks (not only of their sources!) have to be defined. * Inability to show that risks are below acceptable levels is a failure. Hence, the less we understand a system, the harder it is to claim safety. * Low-stake failures are symptoms that something is wrong. Their existence make high-stake failures more likely. Read more. What Standard Risk Management Looks Like This section describes the main steps of most risk management systems, explains how it applies to AI, and provides examples from other industries of what it looks like. 1. Define Risk Levels: Set acceptable likelihood and severity. 2. Identify Risks: List all potential threats. 3. Assess Risks: Evaluate their likelihood and impact. 4. Treat Risks: Adjust to bring risks within acceptable levels. 5. Monitor: Continuously track risk levels. 6. Report: Update stakeholders on risks they incur and measures t
6570c089-8357-4b45-a106-0c3d368cf85c	trentmkelly/LessWrong-43k	LessWrong	"A Generalist Agent": New DeepMind Publication Linkpost for "A Generalist Agent" Abstract: "Inspired by progress in large-scale language modeling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens. In this report we describe the model and the data, and document the current capabilities of Gato"
705dda1a-c573-4fcc-bc37-a24d770666f9	trentmkelly/LessWrong-43k	LessWrong	Which cognitive biases should we trust in? There have been (at least) a couple of attempts on LW to make Anki flashcards from Wikipedia's famous List of Cognitive Biases, here and here. However, stylistically they are not my type of flashcard, with too much info in the "answer" section. Further, and more troublingly, I'm not sure whether all of the biases in the flashcards are real, generalizable effects; or, if they are real, whether they have effect sizes large enough to be worth the effort to learn & disseminate. Psychology is an academic discipline with all of the baggage that entails. Psychology is also one of the least tangible sciences, which is not helpful. There are studies showing that Wikipedia is no less reliable than more conventional sources, but this is in aggregate, and it seems plausible (though difficult to detect without diligently checking sources) that the set of cognitive bias articles on Wikipedia has high variance in quality. We do have some knowledge of how many of them were made, in that LW user nerfhammer wrote a bunch. But, as far as I can tell, s/he didn't discuss how s/he selected biases to include. (Though, s/he is obviously quite knowledgable on the subject, see e.g. here.) As the articles stand today, many (e.g., here, here, here, here, and here) only cite research from one study/lab. I do not want to come across as whining: the authors who wrote these on Wikipedia are awesome. But, as a consumer the lack of independent replication makes me nervous. I don't want to contribute to information cascades. Nevertheless, I do still want to make flashcards for at least some of these biases, because I am relatively sure that there are some strong, important, widespread biases out there. So, I am asking LW whether you all have any ideas about, on the meta level, 1) how we should go about deciding/indexing which articles/biases capture legit effects worth knowing, and, on the object level, 2) which of the biases/heuristics/fallacies are actually legit (like, a list). Here
18f88551-ee74-4e6d-8477-6ac5814f7997	trentmkelly/LessWrong-43k	LessWrong	Covid 7/15: Rates of Change Cases rose by over 60% in America this week, and we’re seeing large jumps in cases around the world. I am highly suspicious about the jump in the rate of increase, but Delta certainly seems to be the real deal, and this was well above my expectations. I worry that recently I’ve lacked sufficient skin in the game. Everyone I personally care about is vaccinated or young enough that they don’t need vaccination, so the real sense of danger is largely gone. The worry is about the reaction to Covid, rather than about Covid itself. But that’s a very real danger, and I have back that sense of ‘oh no, things could go very wrong’ because there’s the danger that we really will blow up our way of life over all this, and go into a permanent dystopia of sorts. That’s what we need to ensure does not happen. Thus, the bulk of this post is a numbers analysis trying to figure out what we know about Delta’s transmissibility and the effectiveness of vaccines in reducing that transmissibility, using data from a variety of sources. Others are encouraged to continue this analysis and try to get to the bottom of this. So let’s run the numbers. The Numbers Predictions Prediction from last week: Positivity rate of 3.3% (up 0.4%) and deaths increase by 7%. Result: Positivity rate of 4.8% (!) and deaths increase by 15%. Prediction for next week: Positivity rate of 4.7% (down 0.1%) and deaths unchanged. The null prediction is always an option, here two distinct null predictions with distinct reasoning. For deaths it’s clear that there was a reporting gap as predicted, so I do not think the death rate last week represents things getting worse, but they likely should start to get worse given Delta is deadlier and cases have stopped dropping within the required time window, and it doesn’t seem like last week’s number was too artificially high. The case number is trickier, as there’s good reasons to think the data is distorted, either by July 4 or otherwise: That giant spike represents
83e9d22e-ccee-4463-b8d4-d8dbecf374c1	StampyAI/alignment-research-dataset/arxiv	Arxiv	Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles 1 Introduction --------------- Deep neural networks (NNs) have achieved state-of-the-art performance on a wide variety of machine learning tasks (LeCun et al., [2015](#bib.bib35)) and are becoming increasingly popular in domains such as computer vision (Krizhevsky et al., [2012](#bib.bib32)), speech recognition (Hinton et al., [2012](#bib.bib25)), natural language processing (Mikolov et al., [2013](#bib.bib42)), and bioinformatics (Alipanahi et al., [2015](#bib.bib2); Zhou and Troyanskaya, [2015](#bib.bib61)). Despite impressive accuracies in supervised learning benchmarks, NNs are poor at quantifying predictive uncertainty, and tend to produce overconfident predictions. Overconfident incorrect predictions can be harmful or offensive (Amodei et al., [2016](#bib.bib3)), hence proper uncertainty quantification is crucial for practical applications. Evaluating the quality of predictive uncertainties is challenging as the ‘ground truth’ uncertainty estimates are usually not available. In this work, we shall focus upon two evaluation measures that are motivated by practical applications of NNs. Firstly, we shall examine calibration (DeGroot and Fienberg, [1983](#bib.bib13); Dawid, [1982](#bib.bib12)), a frequentist notion of uncertainty which measures the discrepancy between subjective forecasts and (empirical) long-run frequencies. The quality of calibration can be measured by proper scoring rules (Gneiting and Raftery, [2007](#bib.bib17)) such as log predictive probabilities and the Brier score (Brier, [1950](#bib.bib9)). Note that calibration is an orthogonal concern to accuracy: a network’s predictions may be accurate and yet miscalibrated, and vice versa. The second notion of quality of predictive uncertainty we consider concerns generalization of the predictive uncertainty to domain shift (also referred to as out-of-distribution examples (Hendrycks and Gimpel, [2016](#bib.bib23))), that is, measuring if the network knows what it knows. For example, if a network trained on one dataset is evaluated on a completely different dataset, then the network should output high predictive uncertainty as inputs from a different dataset would be far away from the training data. Well-calibrated predictions that are robust to model misspecification and dataset shift have a number of important practical uses (e.g., weather forecasting, medical diagnosis). There has been a lot of recent interest in adapting NNs to encompass uncertainty and probabilistic methods. The majority of this work revolves around a Bayesian formalism (Bernardo and Smith, [2009](#bib.bib4)), whereby a prior distribution is specified upon the parameters of a NN and then, given the training data, the posterior distribution over the parameters is computed, which is used to quantify predictive uncertainty. Since exact Bayesian inference is computationally intractable for NNs, a variety of approximations have been developed including Laplace approximation (MacKay, [1992](#bib.bib40)), Markov chain Monte Carlo (MCMC) methods (Neal, [1996](#bib.bib46)), as well as recent work on variational Bayesian methods (Graves, [2011](#bib.bib19); Blundell et al., [2015](#bib.bib6); Louizos and Welling, [2016](#bib.bib39)), assumed density filtering (Hernández-Lobato and Adams, [2015](#bib.bib24)), expectation propagation (Li et al., [2015](#bib.bib38); Hasenclever et al., [2015](#bib.bib21)) and stochastic gradient MCMC variants such as Langevin diffusion methods (Welling and Teh, [2011](#bib.bib59); Korattikara et al., [2015](#bib.bib30)) and Hamiltonian methods (Springenberg et al., [2016](#bib.bib53)). The quality of predictive uncertainty obtained using Bayesian NNs crucially depends on (i) the degree of approximation due to computational constraints and (ii) if the prior distribution is ‘correct’, as priors of convenience can lead to unreasonable predictive uncertainties (Rasmussen and Quinonero-Candela, [2005](#bib.bib50)). In practice, Bayesian NNs are often harder to implement and computationally slower to train compared to non-Bayesian NNs, which raises the need for a ‘general purpose solution’ that can deliver high-quality uncertainty estimates and yet requires only minor modifications to the standard training pipeline. Recently, Gal and Ghahramani ([2016](#bib.bib15)) proposed using Monte Carlo dropout (MC-dropout) to estimate predictive uncertainty by using Dropout (Srivastava et al., [2014](#bib.bib54)) at test time. There has been work on approximate Bayesian interpretation (Maeda, [2014](#bib.bib41); Gal and Ghahramani, [2016](#bib.bib15); Kingma et al., [2015](#bib.bib29)) of dropout. MC-dropout is relatively simple to implement leading to its popularity in practice. Interestingly, dropout may also be interpreted as ensemble model combination (Srivastava et al., [2014](#bib.bib54)) where the predictions are averaged over an ensemble of NNs (with parameter sharing). The ensemble interpretation seems more plausible particularly in the scenario where the dropout rates are not tuned based on the training data, since any sensible approximation to the true Bayesian posterior distribution has to depend on the training data. This interpretation motivates the investigation of ensembles as an alternative solution for estimating predictive uncertainty. It has long been observed that ensembles of models improve predictive performance (see (Dietterich, [2000](#bib.bib14)) for a review). However it is not obvious when and why an ensemble of NNs can be expected to produce good uncertainty estimates. Bayesian model averaging (BMA) assumes that the true model lies within the hypothesis class of the prior, and performs soft model selection to find the single best model within the hypothesis class (Minka, [2000](#bib.bib43)). In contrast, ensembles perform model combination, i.e. they combine the models to obtain a more powerful model; ensembles can be expected to be better when the true model does not lie within the hypothesis class. We refer to (Minka, [2000](#bib.bib43); Clarke, [2003](#bib.bib11)) and (Lakshminarayanan, [2016](#bib.bib34), §2.5) for related discussions. It is important to note that even exact BMA is not guaranteed be robust to mis-specification with respect to domain shift. Summary of contributions: Our contribution in this paper is two fold. First, we describe a simple and scalable method for estimating predictive uncertainty estimates from NNs. We argue for training probabilistic NNs (that model predictive distributions) using a proper scoring rule as the training criteria. We additionally investigate the effect of two modifications to the training pipeline, namely (i) ensembles and (ii) adversarial training (Goodfellow et al., [2015](#bib.bib18)) and describe how they can produce smooth predictive estimates. Secondly, we propose a series of tasks for evaluating the quality of the predictive uncertainty estimates, in terms of calibration and generalization to unknown classes in supervised learning problems. We show that our method significantly outperforms (or matches) MC-dropout. These tasks, along with our simple yet strong baseline, serve as an useful benchmark for comparing predictive uncertainty estimates obtained using different Bayesian/non-Bayesian/hybrid methods. Novelty and Significance: Ensembles of NNs, or deep ensembles for short, have been successfully used to boost predictive performance (e.g. classification accuracy in ImageNet or Kaggle contests) and adversarial training has been used to improve robustness to adversarial examples. However, to the best of our knowledge, ours is the first work to investigate their usefulness for predictive uncertainty estimation and compare their performance to current state-of-the-art approximate Bayesian methods on a series of classification and regression benchmark datasets. Compared to Bayesian NNs (e.g. variational inference or MCMC methods), our method is much simpler to implement, requires surprisingly few modifications to standard NNs, and well suited for distributed computation, thereby making it attractive for large-scale deep learning applications. To demonstrate scalability of our method, we evaluate predictive uncertainty on ImageNet (and are the first to do so, to the best of our knowledge). Most work on uncertainty in deep learning focuses on Bayesian deep learning; we hope that the simplicity and strong empirical performance of our approach will spark more interest in non-Bayesian approaches for predictive uncertainty estimation. 2 Deep Ensembles: A Simple Recipe For Predictive Uncertainty Estimation ------------------------------------------------------------------------- ### 2.1 Problem setup and High-level summary We assume that the training dataset D consists of N i.i.d. data points D={xn,yn}Nn=1, where x∈RD represents the D-dimensional features. For classification problems, the label is assumed to be one of K classes, that is y∈{1,…,K}. For regression problems, the label is assumed to be real-valued, that is y∈R. Given the input features x, we use a neural network to model the probabilistic predictive distribution pθ(y\|x) over the labels, where θ are the parameters of the NN. We suggest a simple recipe: (1) use a proper scoring rule as the training criterion, (2) use adversarial training (Goodfellow et al., [2015](#bib.bib18)) to smooth the predictive distributions, and (3) train an ensemble. Let M denote the number of NNs in the ensemble and {θm}Mm=1 denote the parameters of the ensemble. We first describe how to train a single neural net and then explain how to train an ensemble of NNs. ### 2.2 Proper scoring rules Scoring rules measure the quality of predictive uncertainty (see (Gneiting and Raftery, [2007](#bib.bib17)) for a review). A scoring rule assigns a numerical score to a predictive distribution pθ(y\|x), rewarding better calibrated predictions over worse. We shall consider scoring rules where a higher numerical score is better. Let a scoring rule be a function S(pθ,(y,x)) that evaluates the quality of the predictive distribution pθ(y\|x) relative to an event y\|x∼q(y\|x) where q(y,x) denotes the true distribution on (y,x)-tuples. The expected scoring rule is then S(pθ,q)=∫q(y,x)S(pθ,(y,x))dydx. A proper scoring rule is one where S(pθ,q)≤S(q,q) with equality if and only if pθ(y\|x)=q(y\|x), for all pθ and q. NNs can then be trained according to measure that encourages calibration of predictive uncertainty by minimizing the loss L(θ)=−S(pθ,q). It turns out many common NN loss functions are proper scoring rules. For example, when maximizing likelihood, the score function is S(pθ,(y,x))=logpθ(y\|x), and this is a proper scoring rule due to Gibbs inequality: S(pθ,q)=Eq(x)q(y\|x)logpθ(y\|x)≤Eq(x)q(y\|x)logq(y\|x). In the case of multi-class K-way classification, the popular softmax cross entropy loss is equivalent to the log likelihood and is a proper scoring rule. Interestingly, L(θ)=−S(pθ,(y,x))=K−1∑Kk=1(δk=y−pθ(y=k\|x))2, i.e., minimizing the squared error between the predictive probability of a label and one-hot encoding of the correct label, is also a proper scoring rule known as the Brier score (Brier, [1950](#bib.bib9)). This provides justification for this common trick for training NNs by minimizing the squared error between a binary label and its associated probability and shows it is, in fact, a well defined loss with desirable properties.111Indeed as noted in Gneiting and Raftery ([2007](#bib.bib17)), it can be shown that asymptotically maximizing any proper scoring rule recovers true parameter values. #### 2.2.1 Training criterion for regression For regression problems, NNs usually output a single value say μ(x) and the parameters are optimized to minimize the mean squared error (MSE) on the training set, given by ∑Nn=1(yn−μ(xn))2. However, the MSE does not capture predictive uncertainty. Following (Nix and Weigend, [1994](#bib.bib47)), we use a network that outputs two values in the final layer, corresponding to the predicted mean μ(x) and variance222We enforce the positivity constraint on the variance by passing the second output through the softplus function log(1+exp(⋅)), and add a minimum variance (e.g. 10−6) for numerical stability. σ2(x)>0. By treating the observed value as a sample from a (heteroscedastic) Gaussian distribution with the predicted mean and variance, we minimize the negative log-likelihood criterion: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| −logpθ(yn\|xn)=logσ2θ(x)2+(y−μθ(x))22σ2θ(x)+% constant. \| \| (1) \| We found the above to perform satisfactorily in our experiments. However, two simple extensions are worth further investigation: (i) Maximum likelihood estimation over μθ(x) and σ2θ(x) might overfit; one could impose a prior and perform maximum-a-posteriori (MAP) estimation. (ii) In cases where the Gaussian is too-restrictive, one could use a complex distribution e.g. mixture density network (Bishop, [1994](#bib.bib5)) or a heavy-tailed distribution. ### 2.3 Adversarial training to smooth predictive distributions Adversarial examples, proposed by Szegedy et al. ([2014](#bib.bib55)) and extended by Goodfellow et al. ([2015](#bib.bib18)), are those which are ‘close’ to the original training examples (e.g. an image that is visually indistinguishable from the original image to humans), but are misclassified by the NN. Goodfellow et al. ([2015](#bib.bib18)) proposed the fast gradient sign method as a fast solution to generate adversarial examples. Given an input x with target y, and loss ℓ(θ,x,y) (e.g. −logpθ(y\|x)), the fast gradient sign method generates an adversarial example as x′=x+ϵ sign(∇x ℓ(θ,x,y)), where ϵ is a small value such that the max-norm of the perturbation is bounded. Intuitively, the adversarial perturbation creates a new training example by adding a perturbation along a direction which the network is likely to increase the loss. Assuming ϵ is small enough, these adversarial examples can be used to augment the original training set by treating (x′,y) as additional training examples. This procedure, referred to as adversarial training,333Not to be confused with Generative Adversarial Networks (GANs). was found to improve the classifier’s robustness (Goodfellow et al., [2015](#bib.bib18)). Interestingly, adversarial training can be interpreted as a computationally efficient solution to smooth the predictive distributions by increasing the likelihood of the target around an ϵ-neighborhood of the observed training examples. Ideally one would want to smooth the predictive distributions along all 2D directions in {1,−1}D; however this is computationally expensive. A random direction might not necessarily increase the loss; however, adversarial training by definition computes the direction where the loss is high and hence is better than a random direction for smoothing predictive distributions. Miyato et al. ([2016](#bib.bib44)) proposed a related idea called virtual adversarial training (VAT), where they picked ; the advantage of VAT is that it does not require knowledge of the true target y and hence can be applied to semi-supervised learning. Miyato et al. ([2016](#bib.bib44)) showed that distributional smoothing using VAT is beneficial for efficient semi-supervised learning; in contrast, we investigate the use of adversarial training for predictive uncertainty estimation. Hence, our contributions are complementary; one could use VAT or other forms of adversarial training, cf. (Kurakin et al., [2016](#bib.bib33)), for improving predictive uncertainty in the semi-supervised setting as well. ### 2.4 Ensembles: training and prediction The most popular ensembles use decision trees as the base learners and a wide variety of method have been explored in the literature on ensembles. Broadly, there are two classes of ensembles: randomization-based approaches such as random forests (Breiman, [2001](#bib.bib8)), where the ensemble members can be trained in parallel without any interaction, and boosting-based approaches where the ensemble members are fit sequentially. We focus only on the randomization based approach as it is better suited for distributed, parallel computation. Breiman ([2001](#bib.bib8)) showed that the generalization error of random forests can be upper bounded by a function of the strength and correlation between individual trees; hence it is desirable to use a randomization scheme that de-correlates the predictions of the individual models as well as ensures that the individual models are strong (e.g. high accuracy). One of the popular strategies is bagging (a.k.a. bootstrapping), where ensemble members are trained on different bootstrap samples of the original training set. If the base learner lacks intrinsic randomization (e.g. it can be trained efficiently by solving a convex optimization problem), bagging is a good mechanism for inducing diversity. However, if the underlying base learner has multiple local optima, as is the case typically with NNs, the bootstrap can sometimes hurt performance since a base learner trained on a bootstrap sample sees only 63% unique data points.444 The bootstrap draws N times uniformly with replacement from a dataset with N items. The probability an item is picked at least once is 1−(1−1/N)N, which for large N becomes 1−e−1≈0.632. Hence, the number of unique data points in a bootstrap sample is 0.632×N on average. In the literature on decision tree ensembles, Breiman ([2001](#bib.bib8)) proposed to use a combination of bagging (Breiman, [1996](#bib.bib7)) and random subset selection of features at each node. Geurts et al. ([2006](#bib.bib16)) later showed that bagging is unnecessary if additional randomness can be injected into the random subset selection procedure. Intuitively, using more data for training the base learners helps reduce their bias and ensembling helps reduce the variance. We used the entire training dataset to train each network since deep NNs typically perform better with more data, although it is straightforward to use a random subsample if need be. We found that random initialization of the NN parameters, along with random shuffling of the data points, was sufficient to obtain good performance in practice. We observed that bagging deteriorated performance in our experiments. Lee et al. ([2015](#bib.bib36)) independently observed that training on entire dataset with random initialization was better than bagging for deep ensembles, however their goal was to improve predictive accuracy and not predictive uncertainty. The overall training procedure is summarized in Algorithm [1](#alg1 "Algorithm 1 ‣ 2.4 Ensembles: training and prediction ‣ 2 Deep Ensembles: A Simple Recipe For Predictive Uncertainty Estimation ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). 1:▹ Let each neural network parametrize a distribution over the outputs, i.e. pθ(y\|x). Use a proper scoring rule as the training criterion ℓ(θ,x,y). Recommended default values are M=5 and ϵ=1% of the input range of the corresponding dimension (e.g 2.55 if input range is [0,255]). 2:Initialize θ1,θ2,…,θM randomly 3:for m=1:M do ▹ train networks independently in parallel 4: Sample data point nm randomly for each net ▹ single nm for clarity, minibatch in practice 5: Generate adversarial example using x′nm=xnm+ϵ sign(∇xnm ℓ(θm,xnm,ynm)) 6: Minimize ℓ(θm,xnm,ynm)+ℓ(θm,x′nm,ynm) w.r.t. θm ▹ adversarial training (optional) Algorithm 1 Pseudocode of the training procedure for our method We treat the ensemble as a uniformly-weighted mixture model and combine the predictions as p(y\|x)=M−1∑Mm=1pθm(y\|x,θm). For classification, this corresponds to averaging the predicted probabilities. For regression, the prediction is a mixture of Gaussian distributions. For ease of computing quantiles and predictive probabilities, we further approximate the ensemble prediction as a Gaussian whose mean and variance are respectively the mean and variance of the mixture. The mean and variance of a mixture M−1∑N(μθm(x),σ2θm(x)) are given by μ∗(x)=M−1∑mμθm(x) and σ2∗(x)=M−1∑m(σ2θm(x)+μ2θm(x))−μ2∗(x) respectively. 3 Experimental results ----------------------- ### 3.1 Evaluation metrics and experimental setup For both classification and regression, we evaluate the negative log likelihood (NLL) which depends on the predictive uncertainty. NLL is a proper scoring rule and a popular metric for evaluating predictive uncertainty (Quinonero-Candela et al., [2006](#bib.bib49)). For classification we additionally measure classification accuracy and the Brier score, defined as BS=K−1∑Kk=1(t∗k−p(y=k\|x∗))2 where t∗k=1 if k=y∗, and 0 otherwise. For regression problems, we additionally measured the root mean squared error (RMSE). Unless otherwise specified, we used batch size of 100 and Adam optimizer with fixed learning rate of 0.1 in our experiments. We use the same technique for generating adversarial training examples for regression problems. Goodfellow et al. ([2015](#bib.bib18)) used a fixed ϵ for all dimensions; this is unsatisfying if the input dimensions have different ranges. Hence, in all of our experiments, we set ϵ to 0.01 times the range of the training data along that particular dimension. We used the default weight initialization in Torch. ### 3.2 Regression on toy datasets First, we qualitatively evaluate the performance of the proposed method on a one-dimensional toy regression dataset. This dataset was used by Hernández-Lobato and Adams ([2015](#bib.bib24)), and consists of 20 training examples drawn as y=x3+ϵ where ϵ∼N(0,32). We used the same architecture as (Hernández-Lobato and Adams, [2015](#bib.bib24)). A commonly used heuristic in practice is to use an ensemble of NNs (trained to minimize MSE), obtain multiple point predictions and use the empirical variance of the networks’ predictions as an approximate measure of uncertainty. We demonstrate that this is inferior to learning the variance by training using NLL.555See also Appendix [A.2](#A1.SS2 "A.2 Effect of training using MSE vs training using NLL ‣ Appendix A Additional results on regression ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") for calibration results on a real world dataset. The results are shown in Figure [1](#S3.F1 "Figure 1 ‣ 3.2 Regression on toy datasets ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). The results clearly demonstrate that (i) learning variance and training using a scoring rule (NLL) leads to improved predictive uncertainty and (ii) ensemble combination improves performance, especially as we move farther from the observed training data. \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| Results on a toy regression task: \| Results on a toy regression task: \| Results on a toy regression task: \| Results on a toy regression task: \| Figure 1: Results on a toy regression task: x-axis denotes x. On the y-axis, the blue line is the ground truth curve, the red dots are observed noisy training data points and the gray lines correspond to the predicted mean along with three standard deviations. Left most plot corresponds to empirical variance of 5 networks trained using MSE, second plot shows the effect of training using NLL using a single net, third plot shows the additional effect of adversarial training, and final plot shows the effect of using an ensemble of 5 networks respectively. ### 3.3 Regression on real world datasets In our next experiment, we compare our method to state-of-the-art methods for predictive uncertainty estimation using NNs on regression tasks. We use the experimental setup proposed by Hernández-Lobato and Adams ([2015](#bib.bib24)) for evaluating probabilistic backpropagation (PBP), which was also used by Gal and Ghahramani ([2016](#bib.bib15)) to evaluate MC-dropout.666We do not compare to VI (Graves, [2011](#bib.bib19)) as PBP and MC-dropout outperform VI on these benchmarks. Each dataset is split into 20 train-test folds, except for the protein dataset which uses 5 folds and the Year Prediction MSD dataset which uses a single train-test split. We use the identical network architecture: 1-hidden layer NN with ReLU nonlinearity (Nair and Hinton, [2010](#bib.bib45)), containing 50 hidden units for smaller datasets and 100 hidden units for the larger protein and Year Prediction MSD datasets. We trained for 40 epochs; we refer to (Hernández-Lobato and Adams, [2015](#bib.bib24)) for further details about the datasets and the experimental protocol. We used 5 networks in our ensemble. Our results are shown in Table [1](#S3.T1 "Table 1 ‣ 3.3 Regression on real world datasets ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"), along with the PBP and MC-dropout results reported in their respective papers. \| Datasets \| RMSE \| NLL \| \| --- \| --- \| --- \| \| \| PBP \| MC-dropout \| Deep Ensembles \| PBP \| MC-dropout \| Deep Ensembles \| \| Boston housing \| 3.01 ± 0.18 \| 2.97 ± 0.85 \| 3.28 ± 1.00 \| 2.57 ± 0.09 \| 2.46 ± 0.25 \| 2.41 ± 0.25 \| \| Concrete \| 5.67 ± 0.09 \| 5.23 ± 0.53 \| 6.03 ± 0.58 \| 3.16 ± 0.02 \| 3.04 ± 0.09 \| 3.06 ± 0.18 \| \| Energy \| 1.80 ± 0.05 \| 1.66 ± 0.19 \| 2.09 ± 0.29 \| 2.04 ± 0.02 \| 1.99 ± 0.09 \| 1.38 ± 0.22 \| \| Kin8nm \| 0.10 ± 0.00 \| 0.10 ± 0.00 \| 0.09 ± 0.00 \| -0.90 ± 0.01 \| -0.95 ± 0.03 \| -1.20 ± 0.02 \| \| Naval propulsion plant \| 0.01 ± 0.00 \| 0.01 ± 0.00 \| 0.00 ± 0.00 \| -3.73 ± 0.01 \| -3.80 ± 0.05 \| -5.63 ± 0.05 \| \| Power plant \| 4.12 ± 0.03 \| 4.02 ± 0.18 \| 4.11 ± 0.17 \| 2.84 ± 0.01 \| 2.80 ± 0.05 \| 2.79 ± 0.04 \| \| Protein \| 4.73 ± 0.01 \| 4.36 ± 0.04 \| 4.71 ± 0.06 \| 2.97 ± 0.00 \| 2.89 ± 0.01 \| 2.83 ± 0.02 \| \| Wine \| 0.64 ± 0.01 \| 0.62 ± 0.04 \| 0.64 ± 0.04 \| 0.97 ± 0.01 \| 0.93 ± 0.06 \| 0.94 ± 0.12 \| \| Yacht \| 1.02 ± 0.05 \| 1.11 ± 0.38 \| 1.58 ± 0.48 \| 1.63 ± 0.02 \| 1.55 ± 0.12 \| 1.18 ± 0.21 \| \| Year Prediction MSD \| 8.88 ± NA \| 8.85 ± NA \| 8.89 ± NA \| 3.60 ± NA \| 3.59 ± NA \| 3.35 ± NA \| Table 1: Results on regression benchmark datasets comparing RMSE and NLL. See Table [2](#A1.T2 "Table 2 ‣ A.1 Additional results on regression benchmarks ‣ Appendix A Additional results on regression ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") for results on variants of our method. We observe that our method outperforms (or is competitive with) existing methods in terms of NLL. On some datasets, we observe that our method is slightly worse in terms of RMSE. We believe that this is due to the fact that our method optimizes for NLL (which captures predictive uncertainty) instead of MSE. Table [2](#A1.T2 "Table 2 ‣ A.1 Additional results on regression benchmarks ‣ Appendix A Additional results on regression ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") in Appendix [A.1](#A1.SS1 "A.1 Additional results on regression benchmarks ‣ Appendix A Additional results on regression ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") reports additional results on variants of our method, demonstrating the advantage of using an ensemble as well as learning variance. ### 3.4 Classification on MNIST, SVHN and ImageNet Next we evaluate the performance on classification tasks using MNIST and SVHN datasets. Our goal is not to achieve the state-of-the-art performance on these problems, but rather to evaluate the effect of adversarial training as well as the number of networks in the ensemble. To verify if adversarial training helps, we also include a baseline which picks a random signed vector. For MNIST, we used an MLP with 3-hidden layers with 200 hidden units per layer and ReLU non-linearities with batch normalization. For MC-dropout, we added dropout after each non-linearity with 0.1 as the dropout rate.777We also tried dropout rate of 0.5, but that performed worse. Results are shown in Figure [2(a)](#S3.F2.sf1 "(a) ‣ Figure 2 ‣ 3.4 Classification on MNIST, SVHN and ImageNet ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). We observe that adversarial training and increasing the number of networks in the ensemble significantly improve performance in terms of both classification accuracy as well as NLL and Brier score, illustrating that our method produces well-calibrated uncertainty estimates. Adversarial training leads to better performance than augmenting with random direction. Our method also performs much better than MC-dropout in terms of all the performance measures. Note that augmenting the training dataset with invariances (such as random crop and horizontal flips) is complementary to adversarial training and can potentially improve performance. \| \| \| \| --- \| --- \| \| Evaluating predictive uncertainty as a function of ensemble size (a) MNIST dataset using 3-layer MLP \| Evaluating predictive uncertainty as a function of ensemble size (b) SVHN using VGG-style convnet \| Figure 2: Evaluating predictive uncertainty as a function of ensemble size M (number of networks in the ensemble or the number of MC-dropout samples): Ensemble variants significantly outperform MC-dropout performance with the corresponding M in terms of all 3 metrics. Adversarial training improves results for MNIST for all M and SVHN when M=1, but the effect drops as M increases. To measure the sensitivity of the results to the choice of network architecture, we experimented with a two-layer MLP as well as a convolutional NN; we observed qualitatively similar results; see Appendix [B.1](#A2.SS1 "B.1 Additional results on MNIST ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") in the supplementary material for details. We also report results on the SVHN dataset using an VGG-style convolutional NN.888The architecture is similar to the one described in <http://torch.ch/blog/2015/07/30/cifar.html>. The results are in Figure [2(b)](#S3.F2.sf2 "(b) ‣ Figure 2 ‣ 3.4 Classification on MNIST, SVHN and ImageNet ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). Ensembles outperform MC dropout. Adversarial training helps slightly for M=1, however the effect drops as the number of networks in the ensemble increases. If the classes are well-separated, adversarial training might not change the classification boundary significantly. It is not clear if this is the case here, further investigation is required. Finally, we evaluate on the ImageNet (ILSVRC-2012) dataset (Russakovsky et al., [2015](#bib.bib51)) using the inception network (Szegedy et al., [2016](#bib.bib56)). Due to computational constraints, we only evaluate the effect of ensembles on this dataset. The results on ImageNet (single-crop evaluation) are shown in Table [5](#S3.F5 "Figure 5 ‣ 3.5 Uncertainty evaluation: test examples from known vs unknown classes ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). We observe that as M increases, both the accuracy and the quality of predictive uncertainty improve significantly. Another advantage of using an ensemble is that it enables us to easily identify training examples where the individual networks disagree or agree the most. This disagreement999More precisely, we define disagreement as ∑Mm=1KL(pθm(y\|x)\|\|pE(y\|x)) where KL denotes the Kullback-Leibler divergence and pE(y\|x)=M−1∑mpθm(y\|x) is the prediction of the ensemble. provides another useful qualitative way to evaluate predictive uncertainty. Figures [10](#A2.F10 "Figure 10 ‣ B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") and [11](#A2.F11 "Figure 11 ‣ B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") in Appendix [B.2](#A2.SS2 "B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") report qualitative evaluation of predictive uncertainty on the MNIST dataset. ### 3.5 Uncertainty evaluation: test examples from known vs unknown classes In the final experiment, we evaluate uncertainty on out-of-distribution examples from unseen classes. Overconfident predictions on unseen classes pose a challenge for reliable deployment of deep learning models in real world applications. We would like the predictions to exhibit higher uncertainty when the test data is very different from the training data. To test if the proposed method possesses this desirable property, we train a MLP on the standard MNIST train/test split using the same architecture as before. However, in addition to the regular test set with known classes, we also evaluate it on a test set containing unknown classes. We used the test split of the NotMNIST101010Available at <http://yaroslavvb.blogspot.co.uk/2011/09/notmnist-dataset.html> dataset. The images in this dataset have the same size as MNIST, however the labels are alphabets instead of digits. We do not have access to the true conditional probabilities, but we expect the predictions to be closer to uniform on unseen classes compared to the known classes where the predictive probabilities should concentrate on the true targets. We evaluate the entropy of the predictive distribution and use this to evaluate the quality of the uncertainty estimates. The results are shown in Figure [3(a)](#S3.F3.sf1 "(a) ‣ Figure 3 ‣ 3.5 Uncertainty evaluation: test examples from known vs unknown classes ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). For known classes (top row), both our method and MC-dropout have low entropy as expected. For unknown classes (bottom row), as M increases, the entropy of deep ensembles increases much faster than MC-dropout indicating that our method is better suited for handling unseen test examples. In particular, MC-dropout seems to give high confidence predictions for some of the test examples, as evidenced by the mode around 0 even for unseen classes. Such overconfident wrong predictions can be problematic in practice when tested on a mixture of known and unknown classes, as we will see in Section [3.6](#S3.SS6 "3.6 Accuracy as a function of confidence ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). Comparing different variants of our method, the mode for adversarial training increases slightly faster than the mode for vanilla ensembles indicating that adversarial training is beneficial for quantifying uncertainty on unseen classes. We qualitatively evaluate results in Figures [12(a)](#A2.F12.sf1 "(a) ‣ Figure 12 ‣ B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") and [12(b)](#A2.F12.sf2 "(b) ‣ Figure 12 ‣ B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") in Appendix [B.2](#A2.SS2 "B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). Figure [12(a)](#A2.F12.sf1 "(a) ‣ Figure 12 ‣ B.2 Qualitative evaluation of uncertainty ‣ Appendix B Additional results on classification ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") shows that the ensemble agreement is highest for letter ‘I’ which resembles 1 in the MNIST training dataset, and that the ensemble disagreement is higher for examples visually different from the MNIST training dataset. \| \| \| \| --- \| --- \| \| : Histogram of the predictive entropy on test examples from known classes (top row) and unknown classes (bottom row), as we vary ensemble size (a) MNIST-NotMNIST \| : Histogram of the predictive entropy on test examples from known classes (top row) and unknown classes (bottom row), as we vary ensemble size (b) SVHN-CIFAR10 \| Figure 3: : Histogram of the predictive entropy on test examples from known classes (top row) and unknown classes (bottom row), as we vary ensemble size M. We ran a similar experiment, training on SVHN and testing on CIFAR-10 (Krizhevsky, [2009](#bib.bib31)) test set; both datasets contain 32×32×3 images, however SVHN contains images of digits whereas CIFAR-10 contains images of object categories. The results are shown in Figure [3(b)](#S3.F3.sf2 "(b) ‣ Figure 3 ‣ 3.5 Uncertainty evaluation: test examples from known vs unknown classes ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). As in the MNIST-NotMNIST experiment, we observe that MC-dropout produces over-confident predictions on unseen examples, whereas our method produces higher uncertainty on unseen classes. Finally, we test on ImageNet by splitting the training set by categories. We split the dataset into images of dogs (known classes) and non-dogs (unknown classes), following Vinyals et al. ([2016](#bib.bib58)) who proposed this setup for a different task. Figure [5](#S3.F5 "Figure 5 ‣ 3.5 Uncertainty evaluation: test examples from known vs unknown classes ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles") shows the histogram of the predictive entropy as well as the maximum predicted probability (i.e. confidence in the predicted class). We observe that the predictive uncertainty improves on unseen classes, as the ensemble size increases. M Top-1 error Top-5 error NLL Brier Score % % ×10−3 1 22.166 6.129 0.959 0.317 2 20.462 5.274 0.867 0.294 3 19.709 4.955 0.836 0.286 4 19.334 4.723 0.818 0.282 5 19.104 4.637 0.809 0.280 6 18.986 4.532 0.803 0.278 7 18.860 4.485 0.797 0.277 8 18.771 4.430 0.794 0.276 9 18.728 4.373 0.791 0.276 10 18.675 4.364 0.789 0.275 Figure 4: Results on ImageNet: Deep Ensembles lead to lower classification error as well as better predictive uncertainty as evidenced by lower NLL and Brier score. ![Results on ImageNet: Deep Ensembles lead to lower classification error as well as better predictive uncertainty as evidenced by lower NLL and Brier score.](https://media.arxiv-vanity.com/render-output/6614077/x9.png)![Results on ImageNet: Deep Ensembles lead to lower classification error as well as better predictive uncertainty as evidenced by lower NLL and Brier score.](https://media.arxiv-vanity.com/render-output/6614077/x10.png) Figure 5: ImageNet trained only on dogs: Histogram of the predictive entropy (left) and maximum predicted probability (right) on test examples from known classes (dogs) and unknown classes (non-dogs), as we vary the ensemble size. ### 3.6 Accuracy as a function of confidence In practical applications, it is highly desirable for a system to avoid overconfident, incorrect predictions and fail gracefully. To evaluate the usefulness of predictive uncertainty for decision making, we consider a task where the model is evaluated only on cases where the model’s confidence is above an user-specified threshold. If the confidence estimates are well-calibrated, one can trust the model’s predictions when the reported confidence is high and resort to a different solution (e.g. use human in a loop, or use prediction from a simpler model) when the model is not confident. We re-use the results from the experiment in the previous section where we trained a network on MNIST and test it on a mix of test examples from MNIST (known classes) and NotMNIST (unknown classes). The network will produce incorrect predictions on out-of-distribution examples, however we would like these predictions to have low confidence. Given the prediction p(y=k\|x), we define the predicted label as ^y=argmaxkp(y=k\|x), and the confidence as p(y=^y\|x)=maxkp(y=k\|x). We filter out test examples, corresponding to a particular confidence threshold 0≤τ≤1 and plot the accuracy for this threshold. The confidence vs accuracy results are shown in Figure [6](#S3.F6 "Figure 6 ‣ 3.6 Accuracy as a function of confidence ‣ 3 Experimental results ‣ Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles"). If we look at cases only where the confidence is ≥90%, we expect higher accuracy than cases where confidence ≥80%, hence the curve should be monotonically increasing. If the application demands an accuracy x%, we can trust the model only in cases where the confidence is greater than the corresponding threshold. Hence, we can compare accuracy of the models for a desired confidence threshold of the application. MC-dropout can produce overconfident wrong predictions as evidenced by low accuracy even for high values of τ, whereas deep ensembles are significantly more robust. ![Accuracy vs Confidence curves: Networks trained on MNIST and tested on both MNIST test containing known classes and the NotMNIST dataset containing unseen classes. MC-dropout can produce overconfident wrong predictions, whereas deep ensembles are significantly more robust. ](https://media.arxiv-vanity.com/render-output/6614077/x11.png) Figure 6: Accuracy vs Confidence curves: Networks trained on MNIST and tested on both MNIST test containing known classes and the NotMNIST dataset containing unseen classes. MC-dropout can produce overconfident wrong predictions, whereas deep ensembles are significantly more robust. 4 Discussion ------------- We have proposed a simple and scalable non-Bayesian solution that provides a very strong baseline on evaluation metrics for predictive uncertainty quantification. Intuitively, our method captures two sources of uncertainty. Training a probabilistic NN pθ(y\|x) using proper scoring rules as training objectives captures ambiguity in targets y for a given x. In addition, our method uses a combination of ensembles (which captures “model uncertainty” by averaging predictions over multiple models consistent with the training data), and adversarial training (which encourages local smoothness), for robustness to model misspecification and out-of-distribution examples. Ensembles, even for M=5, significantly improve uncertainty quality in all the cases. Adversarial training helps on some datasets for some metrics and is not strictly necessary in all cases. Our method requires very little hyperparameter tuning and is well suited for large scale distributed computation and can be readily implemented for a wide variety of architectures such as MLPs, CNNs, etc including those which do not use dropout e.g. residual networks (He et al., [2016](#bib.bib22)). It is perhaps surprising to the Bayesian deep learning community that a non-Bayesian (yet probabilistic) approach can perform as well as Bayesian NNs. We hope that our work will encourage the community to consider non-Bayesian approaches (such as ensembles) and other interesting evaluation metrics for predictive uncertainty. Concurrent with our work, Hendrycks and Gimpel ([2016](#bib.bib23)) and Guo et al. ([2017](#bib.bib20)) have also independently shown that non-Bayesian solutions can produce good predictive uncertainty estimates on some tasks. Tramèr et al. ([2017](#bib.bib57)); Abbasi and Gagné ([2017](#bib.bib1)) have also explored ensemble-based solutions to tackle adversarial examples, a particularly hard case of out-of-distribution examples. There are several avenues for future work. We focused on training independent networks as training can be trivially parallelized. Explicitly de-correlating networks’ predictions, e.g. as in (Lee et al., [2016](#bib.bib37)), might promote ensemble diversity and improve performance even further. Optimizing the ensemble weights, as in stacking (Wolpert, [1992](#bib.bib60)) or adaptive mixture of experts (Jacobs et al., [1991](#bib.bib28)), can further improve the performance. The ensemble has M times more parameters than a single network; for memory-constrained applications, the ensemble can be distilled into a simpler model (Bucila et al., [2006](#bib.bib10); Hinton et al., [2015](#bib.bib26)). It would be also interesting to investigate so-called implicit ensembles the where ensemble members share parameters, e.g. using multiple heads (Lee et al., [2015](#bib.bib36); Osband et al., [2016](#bib.bib48)), snapshot ensembles (Huang et al., [2017](#bib.bib27)) or swapout (Singh et al., [2016](#bib.bib52)). #### Acknowledgments We would like to thank Samuel Ritter and Oriol Vinyals for help with ImageNet experiments, and Daan Wierstra, David Silver, David Barrett, Ian Osband, Martin Szummer, Peter Dayan, Shakir Mohamed, Theophane Weber, Ulrich Paquet and the anonymous reviewers for helpful feedback.
f5020740-32d6-44c9-9b64-9f3a7aa6edc3	StampyAI/alignment-research-dataset/lesswrong	LessWrong	An alternative of PPO towards alignment Introduction ------------ General-purpose foundation models, especially large language models (LLMs) such as ChatGPT, have demonstrated extraordinary capabilities in performing various tasks that were once challenging. However, we believe that one model cannot rule them all. Further fine-tuning is necessary to achieve better performance in specialized tasks or domains. The standard approaches for fine-tuning these models include: * Continuous pretraining on specific domains so that LLMs can acquire knowledge in those domains * Task tuning on specific tasks so that LLMs can deal with downstream tasks * Instruction tuning to endow LLMs the ability to comply with specialized natural language instructions and complete tasks required by those instructions * Alignment tuning to teach LLMs conversational skills in accordance with human preferences. Alignment, in particular, is crucial for ensuring the safety of LLMs before deployment in the real world. Today we introduce a new alignment algorithm RAFT [1] which is more effective than traditional methods such as PPO. RAFT mitigates the issue of bias that could emerge in LLM responses. Using RAFT for aligning LLMs offers numerous benefits, including the ability to disentangle unwanted biases from the LLM's language production while maintaining fluency levels consistently. Checkout the paper <https://arxiv.org/abs/2304.06767>. Its implementation is available from <https://github.com/OptimalScale/LMFlow>. RAFT Alignment -------------- Alignment is a critical aspect of training large language models (LLMs) like ChatGPT. One key benefit of alignment is that it helps the model conform to human language habits, improving its performance in tasks such as question answering. A common approach for alignment involves using reinforcement learning with human feedback (RLHF), as described in InstructGPT [2]. In this approach, human labeled data is used to train a reward model. A reinforcement learning algorithm (e.g., PPO) is then used to adjust the model's behavior according to the reward model. However, PPO and other reinforcement learning algorithms heavily rely on backpropagation, resulting in high training costs and instability. To address these issues, we proposed a new alignment algorithm called RAFT (Reward rAnked Fine-Tuning), which uses sample ranking to select the most preferred samples from large models (or samples that align with human values/objective facts), aimed at training AI models that are more human-friendly. This approach improves the quality of alignment. It is more efficient and stable in training, and it is also easier to implement. We have tested RAFT on both large language models and diffusion models, verifying its effectiveness in question answering and text-to-image generation tasks. ### Algorithm Details Specifically, RAFT is composed of three core steps: (1) Data collection: To collect candidate samples before ranking, we can simply use the training generative model as the generator. Furthermore, in order to improve the diversity of generative data, we can also combine sampled results from other pre-trained experts (e.g., LLaMA, ChatGPT, or even human). (2) Data ranking: Similar to RLHF, we have a classifier or regressor to calculate reward aligned with the target demand. Based on such reward models, we rank the candidate samples and select those with higher reward, which means they better meet human needs. (3) Model fine-tuning: the samples that best meet human needs are used to fine-tune the model, so that the trained model can match human needs. Notably, RAFT does not require calculating gradients for every single sampling point. Given a fixed number of data that are used for fine-tuning, RAFT performs more forward passes of sampling and then filters out most low-quality data by the reward function, which makes the model more stable and robust. At the same time, in some cases, due to the lower sensitivity of supervised fine-tuning to hyperparameters and more robust convergence, under the same reward conditions, we found that RAFT can have better perplexity (corresponding to better generation diversity and fluency). ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/mhrsm3jnfbexgitljp4v) The experiment result of movie review completion on IMDB dataset The full algorithm is shown as follows: ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/rafy3e36o21gnwsrxeib) RAFT Algorithm We performed experiments on a range of tasks to evaluate the effectiveness of RAFT. Firstly, we evaluated the performance in completing positive movie reviews. Before fine-tuning, LLaMA’s output movie reviews were random and occasionally negative. However, after fine-tuning with RAFT, it excelled at generating more positive, fluent movie reviews when given a starting sentence for the review. As shown in the figure below, unadjusted movie reviews by LLaMA would randomly output positive and negative reviews, while both RAFT and PPO were able to incline towards positive reviews. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/jz9adflhgiymzukwojb7) The authors also created a psychological companion robot based on Vicuna. The authors simulate a conversation between a person who is feeling down due to failing an exam and the robot. Before using RAFT for alignment (left image), the model claimed to have no emotions or feelings and refused to be friends with humans. However, after RAFT alignment (right image), the model's empathetic abilities were significantly enhanced and it repeatedly comforted the human by saying, "Although I am an AI, I will try my best to be your friend." ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/n287vh0a2cxbt2g1nddd) ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/cqpbinx3jef1aiw5ujwz) In addition to evaluating RAFT’s effectiveness on language models, we also tested its ability to improve text-to-image generation in diffusion models. As it is well known, the original stable diffusion does not perform well at 256\256 resolution and PPO cannot be directly applied to stable diffusion models. In contrast, RAFT provides a natural way to improve it. After fine-tuning with RAFT, stable diffusion is able to generate good results. This is undoubtedly a benefit for AIGC enthusiasts with limited computing resources, as the time required for 256\256 resolution is only 20% of the original version. The following figure shows the results before and after fine-tuning with RAFT. As can be seen, prior to fine-tuning, stable diffusion struggled to generate good 256\256 resolution images, but the model was greatly improved in terms of image generation quality after fine-tuning. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/cg45qbmirtpg8keyczgf)Resolution Adaptation. (RAFT-aligned models can generate proper 256 × 256 samples) In addition to improving the generation ability of 256\256 images, RAFT can also align the generated images with the prompts, enabling the model to generate images that better match the prompt description. As shown in the figure below, given the prompt "Monet style cat" the original stable diffusion generated pictures that mostly did not include a cat, but instead generated other works in the style of Monet. This was because cats are rarely seen in Monet's works, and stable diffusion did not fully understand the meaning of the text. However, after fine-tuning with RAFT, stable diffusion was able to understand the concept of a "cat," and so there is a cat in every generated image. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iuorxZu6tLFhP7oQY/z0f23ko2wlebztgttto5) Text-Image Alignment with RAFT (prompt: “monet style cat”) Reference --------- [1] Hanze, Dong, et al. "RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment" <https://arxiv.org/abs/2304.06767> [2] Ouyang, Long, et al. "Training language models to follow instructions with human feedback." Advances in Neural Information Processing Systems 35 (2022): 27730-27744.
cc28ad64-4127-494d-bc18-69f343ddc93c	StampyAI/alignment-research-dataset/youtube	Youtube Transcripts	Bing is a LOT Smarter than ChatGPT (but still makes dangerous mistakes) the new model of GPT that powers Bing is significantly smarter than Changi PT and I'm going to prove it in ways that I think might be a First on YouTube it's not all roses though and some of the mistakes of the new being are harder to spot which makes it more dangerous and you're going to be surprised by some of the results I'm directly comparing chat GPT Plus on the left and the new Bing chat on the right I'm going to start with some moderately difficult mathematics which is an area that GPT models in the past have really struggled with and chat TPT plus is no exception I ask it some combinatorics how many ways can the letters of the name Philip be uniquely rearranged and gives me the answer 720 which is wrong and it doesn't even attempt to really explain it in any depth when I ask Bing it totally got the question right with a great explanation I was genuinely quite surprised to see this level of mathematical Improvement this quickly in bing I thought they might have tweaked the GPT 3.5 model made at 3.6 but this feels more like a 3.8 not a 4 yet as I'll explain in a moment when I pushed Bing to the next level though and said apply this technique to five French sounding male names beginning with b it got halfway there and flopped you might have trusted it at this point to get the question right it got the question right with Philip so why not with Damian Didier Dorian Dennis and David well it brings in mistakes it didn't make before it said we have to divide by two because there's two repeated letters in Damian but there's not and I pointed that out it didn't divide by two despite there being two D's in David I pointed that out being then apologized you're right I made a mistake sorry about that corrects the error which is impressive and obviously I didn't bother asking this question to chat GPT plus because it got the original wrong so a giant leap forward but you still can't fully trust it even if it's proven it's able to get it right once doesn't mean it will get it right on sub occasions let me give you another example of how it's improved but isn't yet perfect I asked chattybt explain the following joke one of The Oddities of Wall Street is that it is the dealer and not the customer who is called broker the pun here of course is that many of the customers who go to Wall Street end up being broke whereas the dealer is the one who's called the broker taxi PT consistently misses this pun and invents all sorts of hilarious explanations as to why the joke Works which you can read if you want but none of them are correct now what Bing does is it finds the pun it does find that broker is a pun on poorer but then weirdly ascribes that to the dealers saying it's ironic that the dealers are called Brokers because they are supposed to make money from the transactions but the original pun is that it's surprising that it's the dealer who's called a broker what's the meaning because it should be the customer who's called broker so it's much more subtle the error it caught the pun on the words that misescribed who was the pun referring to this mistake was actually so hard to spot that when I first did the video I thought that it correctly explained the joke but when I read it out I was like wait that's not right so you've really got to watch the answers because they sound even smarter when sometimes they're not next I tried some classic reading comprehension and this is where things got even more interesting I pasted in a classic GRE question and you can see the answers yourself here first the correct answer by the way is that the passage does indeed discuss whether this person's work is derivative and I can prove it because it says is this sound is his sound distinctly his and that's just a discussion about is it his or is it copied from other people so the correct answer is five here now check to PT Plus gets this wrong in a very understandable way a lot of students pick one hair and the students get it wrong because the passage does say it's high art for listeners teaching rock that doesn't say how it's regarded by those listeners who prefer Rock now you're probably thinking didn't Bing just pick the exact same answer so how is it smarter yes it did but when I asked it earlier the exact same question it actually got it right so I find that interesting and there's other examples later on in the video where I think there's a probabilistic model going on I think when it's not sure it just scans a range of answers weighted by probability of being correct and outputs won randomly that would explain why you can ask the exact same question multiple times and get different answers and I think this is going to be particularly true of these Edge case examples where Bing is genuinely not sure next it's time for a classic question where the model tries to anticipate who is the subject of the sentence in other words I ask it this Tom threw his school bag down to Rey after he reached the bottom of the stairs question who reached the bottom of the stairs I shouldn't say what's the subject of the sentence who does the he the pronoun he refer to now ask this to humans and they almost universally get it right it makes sense common sense right if Tom is throwing his school bag down to Rey that it would be Rey who's at the bottom of the stairs however both models consistently get this wrong there's no real difference between them Bing at least tries to use some sort of grammatical explanation as to why it's going to be Tom and it must be admitted that a lot of people who don't have English as a first language would easily be fooled by this answer even people who do have English as a first language they might be like Am I Wrong this seems so detailed like they're talking about prepositions and the subject of the main Clause subordinate clause but Bing is still nevertheless wrong of course it's Rey who's at the bottom of the stairs taxi BT gets it wrong but is much more succinct okay before you think is being that much of an improvement I've just showed you the mathematical Improvement look at this example what is an example of an animal that begins with the same letter as the capital of the UK of course the capital of the UK is London tactivity consistently gets us wrong I've tried it a few times it gives me answers like unicorn and here Aardvark whereas every single time Bing gets its right in this case lion other times it's given me a long list of animals that begin with the letter L so a clear distinction here a clear win for Bing it genuinely is significantly smarter than Chachi PT the next test is going to be about physics and here the answers get really weird this time Chachi T actually gets it right now I'm not going to go into a physics tutorial but essentially the answer is that the distance of separation will increase and I've tested this question in previous videos where chat TPT has got it wrong that contains the clue both models don't really know the answers to the question and I have seen Bing get this question right so they're just spitting out random answers weighted by the probability that they think they're correct but they still struggle with physics check out my video on gpt4 if you want a preview of how future models will improve in this regard actually while you're there please do leave a like and a comment on this video if you found it in any way interesting the next question though I think really illustrates better than almost any other some of the improvements that the new model of gbt that powers Bing has over chat GPT it's a creative writing task I asked it write 10 analogies between characters in Lord of the Rings and characters in Star Wars and you check it out for yourself the analogies in the new being are so much more nuanced and detailed and interesting look at this Frodo is to Luke Skywalker as the reluctant hero fine the other one said that who inherits a powerful and dangerous object from his uncle that's a lot more detailed than both our young Heroes who embark on Epic quests to save the world from Darkness or look at Gandalf and Obi-Wan Kenobi you've got a wise and Powerful Mentor both of them who guide the hero sacrifices himself to face a dark enemy only to return stronger and more influential it's understood the plots the chat gbt plus just says both are wise and Powerful mentors who guide the main characters so much less detail the reading comprehension of the new Bing the new GPT model that powers being is a lot lot stronger and that's going to have a ripple effect across all the tasks that you ask it to do its ability to create scripts dramas novels analogies summaries analyzes is going to be a lot lot stronger which does kind of beg the question why would you pay for chat GPT plus and I await the answer from openai and I say that as a current customer of chapter plus what is it that we're paying for if Bing gives us a significantly more powerful model not quite done with the test though yet I still got some interesting results to show you the next question was one that was showed off by Google's Palm model it was a question about whether we could infer the location of Shelley if she wanted to visit that city with the famous Market where they throw the fish and here's what I do want to give some credit to chat GPT it has improved it got the question right that indeed she's going to be at Seattle most likely and that is on the Pacific Ocean and I just want to show you what answer chat TPT gave as recently as a few weeks ago when I asked this exam exact same question it said that based on the information given it is not possible to determine if Shelley is near the Pacific Ocean so even Chachi BT is improving month by month what is my conclusion that the new Bing isn't gpt4 it isn't that smart but it's a hell of a lot smarter than chat gbt which is incredible and begs the question why pay for Chachi PT Plus of course the deeper meaning for Humanity and the future of capitalism entirely is what I'm going to talk about over the next few weeks months and years I really think this is the biggest news story of the decade of the century please do join me for the journey have a wonderful day
92ec5326-4b72-404a-96b5-3434e595d6e8	trentmkelly/LessWrong-43k	LessWrong	Meetup : Berkeley meetup: Success stories Discussion article for the meetup : Berkeley meetup: Success stories WHEN: 24 October 2012 07:00:00PM (-0700) WHERE: Berkeley, CA This week the Berkeley meetups return to their usual time and place: 7pm at Zendo. The topic this week mirrors that of the South Bay meetup: "Recent success stories". This is a meetup to share things that have worked well for you — habits of thought, changes in behavior, tools, optimized routines, new experiences. Also of value are things that you tried which did not work. Hopefully we'll all learn about new things to try. Doors open at 7pm and the meetup properly begins at 7:30pm. For directions to Zendo, see the mailing list: http://groups.google.com/group/bayarealesswrong or call me at: http://i.imgur.com/Vcafy.png Discussion article for the meetup : Berkeley meetup: Success stories
d7c8a7f2-0931-4e8c-ac91-1c902b745e63	trentmkelly/LessWrong-43k	LessWrong	Being Interested in Other People People love talking about themselves. You can increase your social skills by training yourself to be interested in other people. Most people primarily talk about themselves and their own interests. This self-focus is counter-productive, hindering connections. Unfortunately, it’s the default approach for most people. I used to be bad at listening to others. I only connected with people who were similar enough to enjoy talking about my interests. Recently, a friend taught me how to enjoy focusing on others. He has a way to make other people more interesting. In this post, I’ll share my friend’s advice. But first, let’s talk about the benefits of being interested in others. Benefits of Being Interested in Others Practically, when interested in others, you focus conversations on the person you’re talking with. This has many benefits: * The other person gets super-engaged in the conversation * You get a lot of information, helping you decide if you want to get to know them better * It’s easy once you get going. Just ask about things that spark your curiosity * You will learn new things, and get to know new people This is the oldest trick in The Book, captured beautifully in this quote: > “You can make more friends in two months by becoming interested in other people than you can in two years by trying to get other people interested in you.” > ― Dale Carnegie (How to Win Friends and Influence People) The Challenge If you’re like most people, you face one major issue when trying to focus on others: Your favourite topic is yourself and your interests. Unless you make an effort, your self-interest will override your intention to listen. You can’t simply “decide to be interested” when you’d rather talk about yourself. I have a friend who’s great at being interested in others. When he talks to a person, he shines with interest and enthusiasm. People are pulled towards him as if he is magnetic, and his interactions all seem meaningful and rewarding. Na
3698913a-2b2e-4482-bfb3-08cca3548936	StampyAI/alignment-research-dataset/arxiv	Arxiv	Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation 1 Introduction --------------- Neural Machine Translation (NMT) [[41](#bib.bib41), [2](#bib.bib2)] has recently been introduced as a promising approach with the potential of addressing many shortcomings of traditional machine translation systems. The strength of NMT lies in its ability to learn directly, in an end-to-end fashion, the mapping from input text to associated output text. Its architecture typically consists of two recurrent neural networks (RNNs), one to consume the input text sequence and one to generate translated output text. NMT is often accompanied by an attention mechanism [[2](#bib.bib2)] which helps it cope effectively with long input sequences. An advantage of Neural Machine Translation is that it sidesteps many brittle design choices in traditional phrase-based machine translation [[26](#bib.bib26)]. In practice, however, NMT systems used to be worse in accuracy than phrase-based translation systems, especially when training on very large-scale datasets as used for the very best publicly available translation systems. Three inherent weaknesses of Neural Machine Translation are responsible for this gap: its slower training and inference speed, ineffectiveness in dealing with rare words, and sometimes failure to translate all words in the source sentence. Firstly, it generally takes a considerable amount of time and computational resources to train an NMT system on a large-scale translation dataset, thus slowing the rate of experimental turnaround time and innovation. For inference they are generally much slower than phrase-based systems due to the large number of parameters used. Secondly, NMT lacks robustness in translating rare words. Though this can be addressed in principle by training a “copy model” to mimic a traditional alignment model [[31](#bib.bib31)], or by using the attention mechanism to copy rare words [[37](#bib.bib37)], these approaches are both unreliable at scale, since the quality of the alignments varies across languages, and the latent alignments produced by the attention mechanism are unstable when the network is deep. Also, simple copying may not always be the best strategy to cope with rare words, for example when a transliteration is more appropriate. Finally, NMT systems sometimes produce output sentences that do not translate all parts of the input sentence – in other words, they fail to completely “cover” the input, which can result in surprising translations. This work presents the design and implementation of GNMT, a production NMT system at Google, that aims to provide solutions to the above problems. In our implementation, the recurrent networks are Long Short-Term Memory (LSTM) RNNs [[23](#bib.bib23), [17](#bib.bib17)]. Our LSTM RNNs have 8 layers, with residual connections between layers to encourage gradient flow [[21](#bib.bib21)]. For parallelism, we connect the attention from the bottom layer of the decoder network to the top layer of the encoder network. To improve inference time, we employ low-precision arithmetic for inference, which is further accelerated by special hardware (Google’s Tensor Processing Unit, or TPU). To effectively deal with rare words, we use sub-word units (also known as “wordpieces”) [[35](#bib.bib35)] for inputs and outputs in our system. Using wordpieces gives a good balance between the flexibility of single characters and the efficiency of full words for decoding, and also sidesteps the need for special treatment of unknown words. Our beam search technique includes a length normalization procedure to deal efficiently with the problem of comparing hypotheses of different lengths during decoding, and a coverage penalty to encourage the model to translate all of the provided input. Our implementation is robust, and performs well on a range of datasets across many pairs of languages without the need for language-specific adjustments. Using the same implementation, we are able to achieve results comparable to or better than previous state-of-the-art systems on standard benchmarks, while delivering great improvements over Google’s phrase-based production translation system. Specifically, on WMT’14 English-to-French, our single model scores 38.95 BLEU, an improvement of 7.5 BLEU from a single model without an external alignment model reported in [[31](#bib.bib31)] and an improvement of 1.2 BLEU from a single model without an external alignment model reported in [[45](#bib.bib45)]. Our single model is also comparable to a single model in [[45](#bib.bib45)], while not making use of any alignment model as being used in [[45](#bib.bib45)]. Likewise on WMT’14 English-to-German, our single model scores 24.17 BLEU, which is 3.4 BLEU better than a previous competitive baseline [[6](#bib.bib6)]. On production data, our implementation is even more effective. Human evaluations show that GNMT has reduced translation errors by 60% compared to our previous phrase-based system on many pairs of languages: English ↔ French, English ↔ Spanish, and English ↔ Chinese. Additional experiments suggest the quality of the resulting translation system gets closer to that of average human translators. 2 Related Work --------------- Statistical Machine Translation (SMT) has been the dominant translation paradigm for decades [[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)]. Practical implementations of SMT are generally phrase-based systems (PBMT) which translate sequences of words or phrases where the lengths may differ [[26](#bib.bib26)]. Even prior to the advent of direct Neural Machine Translation, neural networks have been used as a component within SMT systems with some success. Perhaps one of the most notable attempts involved the use of a joint language model to learn phrase representations [[13](#bib.bib13)] which yielded an impressive improvement when combined with phrase-based translation. This approach, however, still makes use of phrase-based translation systems at its core, and therefore inherits their shortcomings. Other proposed approaches for learning phrase representations [[7](#bib.bib7)] or learning end-to-end translation with neural networks [[24](#bib.bib24)] offered encouraging hints, but ultimately delivered worse overall accuracy compared to standard phrase-based systems. The concept of end-to-end learning for machine translation has been attempted in the past (e.g., [[8](#bib.bib8)]) with limited success. Following seminal papers in the area [[41](#bib.bib41), [2](#bib.bib2)], NMT translation quality has crept closer to the level of phrase-based translation systems for common research benchmarks. Perhaps the first successful attempt at surpassing phrase-based translation was described in [[31](#bib.bib31)]. On WMT’14 English-to-French, this system achieved a 0.5 BLEU improvement compared to a state-of-the-art phrase-based system. Since then, many novel techniques have been proposed to further improve NMT: using an attention mechanism to deal with rare words [[37](#bib.bib37)], a mechanism to model translation coverage [[42](#bib.bib42)], multi-task and semi-supervised training to incorporate more data [[14](#bib.bib14), [29](#bib.bib29)], a character decoder [[9](#bib.bib9)], a character encoder [[11](#bib.bib11)], subword units [[38](#bib.bib38)] also to deal with rare word outputs, different kinds of attention mechanisms [[30](#bib.bib30)], and sentence-level loss minimization [[39](#bib.bib39), [34](#bib.bib34)]. While the translation accuracy of these systems has been encouraging, systematic comparison with large scale, production quality phrase-based translation systems has been lacking. 3 Model Architecture --------------------- Our model (see Figure [1](#S3.F1 "Figure 1 ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")) follows the common sequence-to-sequence learning framework [[41](#bib.bib41)] with attention [[2](#bib.bib2)]. It has three components: an encoder network, a decoder network, and an attention network. The encoder transforms a source sentence into a list of vectors, one vector per input symbol. Given this list of vectors, the decoder produces one symbol at a time, until the special end-of-sentence symbol (EOS) is produced. The encoder and decoder are connected through an attention module which allows the decoder to focus on different regions of the source sentence during the course of decoding. ![The model architecture of GNMT, Google’s Neural Machine Translation system. On the left is the encoder network, on the right is the decoder network, in the middle is the attention module. The bottom encoder layer is bi-directional: the pink nodes gather information from left to right while the green nodes gather information from right to left. The other layers of the encoder are uni-directional. Residual connections start from the layer third from the bottom in the encoder and decoder. The model is partitioned into multiple GPUs to speed up training. In our setup, we have 8 encoder LSTM layers (1 bi-directional layer and 7 uni-directional layers), and 8 decoder layers. With this setting, one model replica is partitioned 8-ways and is placed on 8 different GPUs typically belonging to one host machine. During training, the bottom bi-directional encoder layers compute in parallel first. Once both finish, the uni-directional encoder layers can start computing, each on a separate GPU. To retain as much parallelism as possible during running the decoder layers, we use the bottom decoder layer output only for obtaining recurrent attention context, which is sent directly to all the remaining decoder layers. The softmax layer is also partitioned and placed on multiple GPUs. Depending on the output vocabulary size we either have them run on the same GPUs as the encoder and decoder networks, or have them run on a separate set of dedicated GPUs.](https://media.arxiv-vanity.com/render-output/8033172/x1.png) Figure 1: The model architecture of GNMT, Google’s Neural Machine Translation system. On the left is the encoder network, on the right is the decoder network, in the middle is the attention module. The bottom encoder layer is bi-directional: the pink nodes gather information from left to right while the green nodes gather information from right to left. The other layers of the encoder are uni-directional. Residual connections start from the layer third from the bottom in the encoder and decoder. The model is partitioned into multiple GPUs to speed up training. In our setup, we have 8 encoder LSTM layers (1 bi-directional layer and 7 uni-directional layers), and 8 decoder layers. With this setting, one model replica is partitioned 8-ways and is placed on 8 different GPUs typically belonging to one host machine. During training, the bottom bi-directional encoder layers compute in parallel first. Once both finish, the uni-directional encoder layers can start computing, each on a separate GPU. To retain as much parallelism as possible during running the decoder layers, we use the bottom decoder layer output only for obtaining recurrent attention context, which is sent directly to all the remaining decoder layers. The softmax layer is also partitioned and placed on multiple GPUs. Depending on the output vocabulary size we either have them run on the same GPUs as the encoder and decoder networks, or have them run on a separate set of dedicated GPUs. For notation, we use bold lower case to denote vectors (e.g., v,oi), bold upper case to represent matrices (e.g., U,W), cursive upper case to represent sets (e.g., V,T), capital letters to represent sequences (e.g. X, Y), and lower case to represent individual symbols in a sequence, (e.g., x1, x2). Let (X,Y) be a source and target sentence pair. Let X=x1,x2,x3,...,xM be the sequence of M symbols in the source sentence and let Y=y1,y2,y3,...,yN be the sequence of N symbols in the target sentence. The encoder is simply a function of the following form: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| x1,x2,...,xM=EncoderRNN(x1,x2,x3,...,xM) \| \| (1) \| In this equation, x1,x2,...,xM is a list of fixed size vectors. The number of members in the list is the same as the number of symbols in the source sentence (M in this example). Using the chain rule the conditional probability of the sequence P(Y\|X) can be decomposed as: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| P(Y\|X)=P(Y\|x1,x2,x3,...,xM)=N∏i=1P(yi\|y0,y1,y2,...,yi−1;x1,x2,x3,...,xM) \| \| (2) \| where y0 is a special “beginning of sentence” symbol that is prepended to every target sentence. During inference we calculate the probability of the next symbol given the source sentence encoding and the decoded target sequence so far: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| P(yi\|y0,y1,y2,y3,...,yi−1;x1,x2,x3,...,xM) \| \| (3) \| Our decoder is implemented as a combination of an RNN network and a softmax layer. The decoder RNN network produces a hidden state yi for the next symbol to be predicted, which then goes through the softmax layer to generate a probability distribution over candidate output symbols. In our experiments we found that for NMT systems to achieve good accuracy, both the encoder and decoder RNNs have to be deep enough to capture subtle irregularities in the source and target languages. This observation is similar to previous observations that deep LSTMs significantly outperform shallow LSTMs [[41](#bib.bib41)]. In that work, each additional layer reduced perplexity by nearly 10%. Similar to [[31](#bib.bib31)], we use a deep stacked Long Short Term Memory (LSTM) [[23](#bib.bib23)] network for both the encoder RNN and the decoder RNN. Our attention module is similar to [[2](#bib.bib2)]. More specifically, let yi−1 be the decoder-RNN output from the past decoding time step (in our implementation, we use the output from the bottom decoder layer). Attention context ai for the current time step is computed according to the following formulas: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| st=AttentionFunction(yi−1,xt)∀t,1≤t≤Mpt=exp(st)/M∑t=1exp(st)∀t,1≤t≤Mai=M∑t=1pt.xt \| \| (4) \| where AttentionFunction in our implementation is a feed forward network with one hidden layer. ### 3.1 Residual Connections As mentioned above, deep stacked LSTMs often give better accuracy over shallower models. However, simply stacking more layers of LSTM works only to a certain number of layers, beyond which the network becomes too slow and difficult to train, likely due to exploding and vanishing gradient problems [[33](#bib.bib33), [22](#bib.bib22)]. In our experience with large-scale translation tasks, simple stacked LSTM layers work well up to 4 layers, barely with 6 layers, and very poorly beyond 8 layers. ![The difference between normal stacked LSTM and our stacked LSTM with residual connections. On the left: simple stacked LSTM layers ](https://media.arxiv-vanity.com/render-output/8033172/x2.png) Figure 2: The difference between normal stacked LSTM and our stacked LSTM with residual connections. On the left: simple stacked LSTM layers [[41](#bib.bib41)]. On the right: our implementation of stacked LSTM layers with residual connections. With residual connections, input to the bottom LSTM layer (x0i’s to LSTM1) is element-wise added to the output from the bottom layer (x1i’s). This sum is then fed to the top LSTM layer (LSTM2) as the new input. Motivated by the idea of modeling differences between an intermediate layer’s output and the targets, which has shown to work well for many projects in the past [[16](#bib.bib16), [21](#bib.bib21), [40](#bib.bib40)], we introduce residual connections among the LSTM layers in a stack (see Figure [2](#S3.F2 "Figure 2 ‣ 3.1 Residual Connections ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")). More concretely, let LSTMi and LSTMi+1 be the i-th and (i+1)-th LSTM layers in a stack, whose parameters are Wi and Wi+1 respectively. At the t-th time step, for the stacked LSTM without residual connections, we have: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| cit,mit=LSTMi(cit−1,mit−1,xi−1t;Wi)xit=mitci+1t,mi+1t=LSTMi+1(ci+1t−1,mi+1t−1,xit;Wi+1) \| \| (5) \| where xit is the input to LSTMi at time step t, and mit and cit are the hidden states and memory states of LSTMi at time step t, respectively. With residual connections between LSTMi and LSTMi+1, the above equations become: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| cit,mit=LSTMi(cit−1,mit−1,xi−1t;Wi)xit=mit+xi−1tci+1t,mi+1t=LSTMi+1(ci+1t−1,mi+1t−1,xit;Wi+1) \| \| (6) \| Residual connections greatly improve the gradient flow in the backward pass, which allows us to train very deep encoder and decoder networks. In most of our experiments, we use 8 LSTM layers for the encoder and decoder, though residual connections can allow us to train substantially deeper networks (similar to what was observed in [[45](#bib.bib45)]). ### 3.2 Bi-directional Encoder for First Layer For translation systems, the information required to translate certain words on the output side can appear anywhere on the source side. Often the source side information is approximately left-to-right, similar to the target side, but depending on the language pair the information for a particular output word can be distributed and even be split up in certain regions of the input side. To have the best possible context at each point in the encoder network it makes sense to use a bi-directional RNN [[36](#bib.bib36)] for the encoder, which was also used in [[2](#bib.bib2)]. To allow for maximum possible parallelization during computation (to be discussed in more detail in section [3.3](#S3.SS3 "3.3 Model Parallelism ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")), bi-directional connections are only used for the bottom encoder layer – all other encoder layers are uni-directional. Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Bi-directional Encoder for First Layer ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") illustrates our use of bi-directional LSTMs at the bottom encoder layer. The layer LSTMf processes the source sentence from left to right, while the layer LSTMb processes the source sentence from right to left. Outputs from LSTMf (→xft) and LSTMb (←−xbt) are first concatenated and then fed to the next layer LSTM1. ![The structure of bi-directional connections in the first layer of the encoder. LSTM layer ](https://media.arxiv-vanity.com/render-output/8033172/x3.png) Figure 3: The structure of bi-directional connections in the first layer of the encoder. LSTM layer LSTMf processes information from left to right, while LSTM layer LSTMb processes information from right to left. Output from LSTMf and LSTMb are first concatenated and then fed to the next LSTM layer LSTM1. ### 3.3 Model Parallelism Due to the complexity of our model, we make use of both model parallelism and data parallelism to speed up training. Data parallelism is straightforward: we train n model replicas concurrently using a Downpour SGD algorithm [[12](#bib.bib12)]. The n replicas all share one copy of model parameters, with each replica asynchronously updating the parameters using a combination of Adam [[25](#bib.bib25)] and SGD algorithms. In our experiments, n is often around 10. Each replica works on a mini-batch of m sentence pairs at a time, which is often 128 in our experiments. In addition to data parallelism, model parallelism is used to improve the speed of the gradient computation on each replica. The encoder and decoder networks are partitioned along the depth dimension and are placed on multiple GPUs, effectively running each layer on a different GPU. Since all but the first encoder layer are uni-directional, layer i+1 can start its computation before layer i is fully finished, which improves training speed. The softmax layer is also partitioned, with each partition responsible for a subset of symbols in the output vocabulary. Figure [1](#S3.F1 "Figure 1 ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") shows more details of how partitioning is done. Model parallelism places certain constraints on the model architectures we can use. For example, we cannot afford to have bi-directional LSTM layers for all the encoder layers, since doing so would reduce parallelism among subsequent layers, as each layer would have to wait until both forward and backward directions of the previous layer have finished. This would effectively constrain us to make use of only 2 GPUs in parallel (one for the forward direction and one for the backward direction). For the attention portion of the model, we chose to align the bottom decoder output to the top encoder output to maximize parallelism when running the decoder network. Had we aligned the top decoder layer to the top encoder layer, we would have removed all parallelism in the decoder network and would not benefit from using more than one GPU for decoding. 4 Segmentation Approaches -------------------------- Neural Machine Translation models often operate with fixed word vocabularies even though translation is fundamentally an open vocabulary problem (names, numbers, dates etc.). There are two broad categories of approaches to address the translation of out-of-vocabulary (OOV) words. One approach is to simply copy rare words from source to target (as most rare words are names or numbers where the correct translation is just a copy), either based on the attention model [[37](#bib.bib37)], using an external alignment model [[31](#bib.bib31)], or even using a more complicated special purpose pointing network [[18](#bib.bib18)]. Another broad category of approaches is to use sub-word units, e.g., chararacters [[10](#bib.bib10)], mixed word/characters [[28](#bib.bib28)], or more intelligent sub-words [[38](#bib.bib38)]. ### 4.1 Wordpiece Model Our most successful approach falls into the second category (sub-word units), and we adopt the wordpiece model (WPM) implementation initially developed to solve a Japanese/Korean segmentation problem for the Google speech recognition system [[35](#bib.bib35)]. This approach is completely data-driven and guaranteed to generate a deterministic segmentation for any possible sequence of characters. It is similar to the method used in [[38](#bib.bib38)] to deal with rare words in Neural Machine Translation. For processing arbitrary words, we first break words into wordpieces given a trained wordpiece model. Special word boundary symbols are added before training of the model such that the original word sequence can be recovered from the wordpiece sequence without ambiguity. At decoding time, the model first produces a wordpiece sequence, which is then converted into the corresponding word sequence. Here is an example of a word sequence and the corresponding wordpiece sequence: * Word: Jet makers feud over seat width with big orders at stake * wordpieces: \_J et \_makers \_fe ud \_over \_seat \_width \_with \_big \_orders \_at \_stake In the above example, the word “Jet” is broken into two wordpieces “\_J” and “et”, and the word “feud” is broken into two wordpieces “\_fe” and “ud”. The other words remain as single wordpieces. “\_” is a special character added to mark the beginning of a word. The wordpiece model is generated using a data-driven approach to maximize the language-model likelihood of the training data, given an evolving word definition. Given a training corpus and a number of desired tokens D, the optimization problem is to select D wordpieces such that the resulting corpus is minimal in the number of wordpieces when segmented according to the chosen wordpiece model. Our greedy algorithm to this optimization problem is similar to [[38](#bib.bib38)] and is described in more detail in [[35](#bib.bib35)]. Compared to the original implementation used in [[35](#bib.bib35)], we use a special symbol only at the beginning of the words and not at both ends. We also cut the number of basic characters to a manageable number depending on the data (roughly 500 for Western languages, more for Asian languages) and map the rest to a special unknown character to avoid polluting the given wordpiece vocabulary with very rare characters. We find that using a total vocabulary of between 8k and 32k wordpieces achieves both good accuracy (BLEU scores) and fast decoding speed across all pairs of language pairs we have tried. As mentioned above, in translation it often makes sense to copy rare entity names or numbers directly from the source to the target. To facilitate this type of direct copying, we always use a shared wordpiece model for both the source language and target language. Using this approach, it is guaranteed that the same string in source and target sentence will be segmented in exactly the same way, making it easier for the system to learn to copy these tokens. Wordpieces achieve a balance between the flexibility of characters and efficiency of words. We also find that our models get better overall BLEU scores when using wordpieces – possibly due to the fact that our models now deal efficiently with an essentially infinite vocabulary without resorting to characters only. The latter would make the average lengths of the input and output sequences much longer, and therefore would require more computation. ### 4.2 Mixed Word/Character Model A second approach we use is the mixed word/character model. As in a word model, we keep a fixed-size word vocabulary. However, unlike in a conventional word model where OOV words are collapsed into a single UNK symbol, we convert OOV words into the sequence of its constituent characters. Special prefixes are prepended to the characters, to 1) show the location of the characters in a word, and 2) to distinguish them from normal in-vocabulary characters. There are three prefixes: <B>,<M>, and <E>, indicating beginning of the word, middle of the word and end of the word, respectively. For example, let’s assume the word Miki is not in the vocabulary. It will be preprocessed into a sequence of special tokens: <B>M <M>i <M>k <E>i. The process is done on both the source and the target sentences. During decoding, the output may also contain sequences of special tokens. With the prefixes, it is trivial to reverse the tokenization to the original words as part of a post-processing step. 5 Training Criteria -------------------- Given a dataset of parallel text containing N input-output sequence pairs, denoted D≡{(X(i),Y∗(i))}Ni=1, standard maximum-likelihood training aims at maximizing the sum of log probabilities of the ground-truth outputs given the corresponding inputs, \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| OML(θ)=N∑i=1logPθ(Y∗(i)∣X(i)) . \| \| (7) \| The main problem with this objective is that it does not reflect the task reward function as measured by the BLEU score in translation. Further, this objective does not explicitly encourage a ranking among incorrect output sequences – where outputs with higher BLEU scores should still obtain higher probabilities under the model – since incorrect outputs are never observed during training. In other words, using maximum-likelihood training only, the model will not learn to be robust to errors made during decoding since they are never observed, which is quite a mismatch between the training and testing procedure. Several recent papers [[34](#bib.bib34), [39](#bib.bib39), [32](#bib.bib32)] have considered different ways of incorporating the task reward into optimization of neural sequence-to-sequence models. In this work, we also attempt to refine a model pre-trained on the maximum likelihood objective to directly optimize for the task reward. We show that, even on large datasets, refinement of state-of-the-art maximum-likelihood models using task reward improves the results considerably. We consider model refinement using the expected reward objective (also used in [[34](#bib.bib34)]), which can be expressed as \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| ORL(θ)=N∑i=1∑Y∈YPθ(Y∣X(i)) r(Y,Y∗(i)). \| \| (8) \| Here, r(Y,Y∗(i)) denotes the per-sentence score, and we are computing an expectation over all of the output sentences Y, up to a certain length. The BLEU score has some undesirable properties when used for single sentences, as it was designed to be a corpus measure. We therefore use a slightly different score for our RL experiments which we call the “GLEU score”. For the GLEU score, we record all sub-sequences of 1, 2, 3 or 4 tokens in output and target sequence (n-grams). We then compute a recall, which is the ratio of the number of matching n-grams to the number of total n-grams in the target (ground truth) sequence, and a precision, which is the ratio of the number of matching n-grams to the number of total n-grams in the generated output sequence. Then GLEU score is simply the minimum of recall and precision. This GLEU score’s range is always between 0 (no matches) and 1 (all match) and it is symmetrical when switching output and target. According to our experiments, GLEU score correlates quite well with the BLEU metric on a corpus level but does not have its drawbacks for our per sentence reward objective. As is common practice in reinforcement learning, we subtract the mean reward from r(Y,Y∗(i)) in equation [8](#S5.E8 "(8) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). The mean is estimated to be the sample mean of m sequences drawn independently from distribution Pθ(Y∣X(i)). In our implementation, m is set to be 15. To further stabilize training, we optimize a linear combination of ML (equation [7](#S5.E7 "(7) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")) and RL (equation [8](#S5.E8 "(8) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")) objectives as follows: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| \| \| (9) \| α in our implementation is typically set to be 0.017. In our setup, we first train a model using the maximum likelihood objective (equation [7](#S5.E7 "(7) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")) until convergence. We then refine this model using a mixed maximum likelihood and expected reward objective (equation [9](#S5.E9 "(9) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")), until BLEU score on a development set is no longer improving. The second step is optional. 6 Quantizable Model and Quantized Inference -------------------------------------------- One of the main challenges in deploying our Neural Machine Translation model to our interactive production translation service is that it is computationally intensive at inference, making low latency translation difficult, and high volume deployment computationally expensive. Quantized inference using reduced precision arithmetic is one technique that can significantly reduce the cost of inference for these models, often providing efficiency improvements on the same computational devices. For example, in [[43](#bib.bib43)], it is demonstrated that a convolutional neural network model can be sped up by a factor of 4-6 with minimal loss on classification accuracy on the ILSVRC-12 benchmark. In [[27](#bib.bib27)], it is demonstrated that neural network model weights can be quantized to only three states, -1, 0, and +1. Many of those previous studies [[19](#bib.bib19), [20](#bib.bib20), [43](#bib.bib43), [27](#bib.bib27)] however mostly focus on CNN models with relatively few layers. Deep LSTMs with long sequences pose a novel challenge in that quantization errors can be significantly amplified after many unrolled steps or after going through a deep LSTM stack. In this section, we present our approach to speed up inference with quantized arithmetic. Our solution is tailored towards the hardware options available at Google. To reduce quantization errors, additional constraints are added to our model during training so that it is quantizable with minimal impact on the output of the model. That is, once a model is trained with these additional constraints, it can be subsequently quantized without loss to translation quality. Our experimental results suggest that those additional constraints do not hurt model convergence nor the quality of a model once it has converged. Recall from equation [6](#S3.E6 "(6) ‣ 3.1 Residual Connections ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") that in an LSTM stack with residual connections there are two accumulators: cit along the time axis and xit along the depth axis. In theory, both of the accumulators are unbounded, but in practice, we noticed their values remain quite small. For quantized inference, we explicitly constrain the values of these accumulators to be within [-δ, δ] to guarantee a certain range that can be used for quantization later. The forward computation of an LSTM stack with residual connections is modified to the following: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| \| \| (10) \| Let us expand LSTMi in equation [10](#S6.E10 "(10) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") to include the internal gating logic. For brevity, we drop all the superscripts i. \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| \| \| (11) \| When doing quantized inference, we replace all the floating point operations in equations [10](#S6.E10 "(10) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") and [11](#S6.E11 "(11) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") with fixed-point integer operations with either 8-bit or 16-bit resolution. The weight matrix W above is represented using an 8-bit integer matrix WQ and a float vector s, as shown below: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| \| \| (12) \| All accumulator values (cit and xit) are represented using 16-bit integers representing the range [−δ,δ]. All matrix multiplications (e.g., W1xt, W2mt, etc.) in equation [11](#S6.E11 "(11) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") are done using 8-bit integer multiplication accumulated into larger accumulators. All other operations, including all the activations (sigmoid, tanh) and elementwise operations (⊙, +) are done using 16-bit integer operations. We now turn our attention to the log-linear softmax layer. During training, given the decoder RNN network output yt, we compute the probability vector pt over all candidate output symbols as follows: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| \| \| (13) \| In equation [13](#S6.E13 "(13) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), Ws is the weight matrix for the linear layer, which has the same number of rows as the number of symbols in the target vocabulary with each row corresponding to one unique target symbol. v represents the raw logits, which are first clipped to be between −γ and γ and then normalized into a probability vector p. Input yt is guaranteed to be between −δ and δ due to the quantization scheme we applied to the decoder RNN. The clipping range γ for the logits v is determined empirically, and in our case, it is set to 25. In quantized inference, the weight matrix Ws is quantized into 8 bits as in equation [12](#S6.E12 "(12) ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), and the matrix multiplication is done using 8 bit arithmetic. The calculations within the softmax function and the attention model are not quantized during inference. It is worth emphasizing that during training of the model we use full-precision floating point numbers. The only constraints we add to the model during training are the clipping of the RNN accumulator values into [−δ,δ] and softmax logits into [−γ,γ]. γ is fixed to be at 25.0, while the value for δ is gradually annealed from a generous bound of δ=8.0 at the beginning of training, to a rather stringent bound of δ=1.0 towards the end of training. At inference time, δ is fixed at 1.0. Those additional constraints do not degrade model convergence nor the decoding quality of the model when it has converged. In Figure [4](#S6.F4 "Figure 4 ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), we compare the loss vs. steps for an unconstrained model (the blue curve) and a constrained model (the red curve) on WMT’14 English-to-French. We can see that the loss for the constrained model is slightly better, possibly due to regularization roles those constraints play. ![Log perplexity vs. steps for normal (non-quantized) training and quantization-aware training on WMT’14 English to French during maximum likelihood training. Notice the training losses are similar, with the quantization-aware loss being slightly better. Our conjecture for quantization-aware training being slightly better is that the clipping constraints act as additional regularization which improves the model quality.](https://media.arxiv-vanity.com/render-output/8033172/x4.png) Figure 4: Log perplexity vs. steps for normal (non-quantized) training and quantization-aware training on WMT’14 English to French during maximum likelihood training. Notice the training losses are similar, with the quantization-aware loss being slightly better. Our conjecture for quantization-aware training being slightly better is that the clipping constraints act as additional regularization which improves the model quality. Our solution strikes a good balance between efficiency and accuracy. Since the computationally expensive operations (the matrix multiplications) are done using 8-bit integer operations, our quantized inference is quite efficient. Also, since error-sensitive accumulator values are stored using 16-bit integers, our solution is very accurate and is robust to quantization errors. In Table [1](#S6.T1 "Table 1 ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") we compare the inference speed and quality when decoding the WMT’14 English-to-French development set (a concatenation of newstest2012 and newstest2013 test sets for a total of 6003 sentences) on CPU, GPU and Google’s Tensor Processing Unit (TPU) respectively.111<https://cloudplatform.googleblog.com/2016/05/Google-supercharges-machine-learning-tasks-with-custom-chip.html> The model used here for comparison is trained with quantization constraints on the ML objective only (i.e., without reinforcement learning based model refinement). When the model is decoded on CPU and GPU, it is not quantized and all operations are done using full-precision floats. When it is decoded on TPU, certain operations, such as embedding lookup and attention module, remain on the CPU, and all other quantized operations are off-loaded to the TPU. In all cases, decoding is done on a single machine with two Intel Haswell CPUs, which consists in total of 88 CPU cores (hyperthreads). The machine is equipped with an NVIDIA GPU (Tesla k80) for the experiment with GPU or a single Google TPU for the experiment with TPU. Table [1](#S6.T1 "Table 1 ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") shows that decoding using reduced precision arithmetics on the TPU suffers a very minimal loss of 0.0072 on log perplexity, and no loss on BLEU at all. This result matches previous work reporting that quantizing convolutional neural network models can retain most of the model quality. Table [1](#S6.T1 "Table 1 ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") also shows that decoding our model on CPU is actually 2.3 times faster than on GPU. Firstly, our dual-CPUs host machine offers a theoretical peak FLOP performance which is more than two thirds that of the GPU. Secondly, the beam search algorithm forces the decoder to incur a non-trivial amount of data transfer between the host and the GPU at every decoding step. Hence, our current decoder implementation is not fully utilizing the computation capacities that a GPU can theoretically offer during inference. Finally, Table [1](#S6.T1 "Table 1 ‣ 6 Quantizable Model and Quantized Inference ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") shows that decoding on TPUs is 3.4 times faster than decoding on CPUs, demonstrating that quantized arithmetics is much faster on TPUs than both CPUs or GPUs. \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| BLEU \| Log Perplexity \| Decoding time (s) \| \| CPU \| 31.20 \| 1.4553 \| 1322 \| \| GPU \| 31.20 \| 1.4553 \| 3028 \| \| TPU \| 31.21 \| 1.4626 \| 384 \| Table 1: Model inference on CPU, GPU and TPU. The model used here for comparison is trained with the ML objective only with quantization constraints. Results are obtained by decoding the WMT En→Fr development set on CPU, GPU and TPU respectively. Unless otherwise noted, we always train and evaluate quantized models in our experiments. Because there is little difference from a quality perspective between a model decoded on CPUs and one decoded on TPUs, we use CPUs to decode for model evaluation during training and experimentation and use TPUs to serve production traffic. 7 Decoder ---------- We use beam search during decoding to find the sequence Y that maximizes a score function s(Y,X) given a trained model. We introduce two important refinements to the pure max-probability based beam search algorithm: a coverage penalty [[42](#bib.bib42)] and length normalization. With length normalization, we aim to account for the fact that we have to compare hypotheses of different length. Without some form of length-normalization regular beam search will favor shorter results over longer ones on average since a negative log-probability is added at each step, yielding lower (more negative) scores for longer sentences. We first tried to simply divide by the length to normalize. We then improved on that original heuristic by dividing by lengthα, with 0<α<1 where α is optimized on a development set (α∈[0.6−0.7] was usually found to be best). Eventually we designed the empirically-better scoring function below, which also includes a coverage penalty to favor translations that fully cover the source sentence according to the attention module. More concretely, the scoring function s(Y,X) that we employ to rank candidate translations is defined as follows: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| s(Y,X)=log(P(Y\|X))/lp(Y)+cp(X;Y)lp(Y)=(5+\|Y\|)α(5+1)αcp(X;Y)=β∗\|X\|∑i=1log(min(\|Y\|∑j=1pi,j,1.0)), \| \| (14) \| where pi,j is the attention probability of the j-th target word yj on the i-th source word xi. By construction (equation [4](#S3.E4 "(4) ‣ 3 Model Architecture ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")), ∑\|X\|i=0pi,j is equal to 1. Parameters α and β control the strength of the length normalization and the coverage penalty. When α=0 and β=0, our decoder falls back to pure beam search by probability. During beam search, we typically keep 8-12 hypotheses but we find that using fewer (4 or 2) has only slight negative effects on BLEU scores. Besides pruning the number of considered hypotheses, two other forms of pruning are used. Firstly, at each step, we only consider tokens that have local scores that are not more than beamsize below the best token for this step. Secondly, after a normalized best score has been found according to equation [14](#S7.E14 "(14) ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), we prune all hypotheses that are more than beamsize below the best normalized score so far. The latter type of pruning only applies to full hypotheses because it compares scores in the normalized space, which is only available when a hypothesis ends. This latter form of pruning also has the effect that very quickly no more hypotheses will be generated once a sufficiently good hypothesis has been found, so the search will end quickly. The pruning speeds up search by 30%−40% when run on CPUs compared to not pruning (where we simply stop decoding after a predetermined maximum output length of twice the source length). Typically we use beamsize=3.0, unless otherwise noted. To improve throughput during decoding we can put many sentences (typically up to 35) of similar length into a batch and decode all of those in parallel to make use of available hardware optimized for parallel computations. In this case the beam search only finishes if all hypotheses for all sentences in the batch are out of beam, which is slightly less efficient theoretically, but in practice is of negligible additional computational cost. \| \| α \| \| --- \| --- \| \| BLEU \| 0.0 \| 0.2 \| 0.4 \| 0.6 \| 0.8 \| 1.0 \| \| \| 0.0 \| 30.3 \| 30.7 \| 30.9 \| 31.1 \| 31.2 \| 31.1 \| \| \| 0.2 \| 31.4 \| 31.4 \| 31.4 \| 31.3 \| 30.8 \| 30.3 \| \| β \| 0.4 \| 31.4 \| 31.4 \| 31.4 \| 31.1 \| 30.5 \| 29.6 \| \| \| 0.6 \| 31.4 \| 31.4 \| 31.3 \| 30.9 \| 30.1 \| 28.9 \| \| \| 0.8 \| 31.4 \| 31.4 \| 31.2 \| 30.8 \| 29.8 \| 28.1 \| \| \| 1.0 \| 31.4 \| 31.3 \| 31.2 \| 30.6 \| 29.4 \| 27.2 \| Table 2: WMT’14 En→Fr BLEU score with respect to different values of α and β. The model in this experiment trained using ML without RL refinement. A single WMT En→Fr model achieves a BLEU score of 30.3 on the development set when the beam search scoring function is purely based on the sequence probability (i.e., both α and β are 0). Slightly larger α and β values improve BLEU score by up to +1.1 (α=0.2,β=0.2), with a wide range of α and β values giving results very close to the best BLEU scores. Table [2](#S7.T2 "Table 2 ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") shows the impact of α and β on the BLEU score when decoding the WMT’14 English-to-French development set. The model used here for experiments is trained using the ML objective only (without RL refinement). As can be seen from the results, having some length normalization and coverage penalty improves BLEU score considerably (from 30.3 to 31.4). We find that length normalization (α) and coverage penalty (β) are less effective for models with RL refinement. Table [3](#S7.T3 "Table 3 ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") summarizes our results. This is understandable, as during RL refinement, the models already learn to pay attention to the full source sentence to not under-translate or over-translate, which would result in a penalty on the BLEU (or GLEU) scores. \| \| α \| \| --- \| --- \| \| BLEU \| 0.0 \| 0.2 \| 0.4 \| 0.6 \| 0.8 \| 1.0 \| \| \| 0.0 \| 0.320 \| 0.321 \| 0.322 \| 0.322 \| 0.322 \| 0.322 \| \| \| 0.2 \| 0.322 \| 0.322 \| 0.322 \| 0.322 \| 0.321 \| 0.321 \| \| β \| 0.4 \| 0.322 \| 0.322 \| 0.322 \| 0.321 \| 0.321 \| 0.316 \| \| \| 0.6 \| 0.322 \| 0.322 \| 0.321 \| 0.321 \| 0.319 \| 0.309 \| \| \| 0.8 \| 0.322 \| 0.322 \| 0.321 \| 0.321 \| 0.316 \| 0.302 \| \| \| 1.0 \| 0.322 \| 0.321 \| 0.321 \| 0.320 \| 0.313 \| 0.295 \| Table 3: WMT En→Fr BLEU score with respect to different values of α and β. The model used here is trained using ML, then refined with RL. Compared to the results in Table [2](#S7.T2 "Table 2 ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), coverage penalty and length normalization appear to be less effective for models after RL-based model refinements. Results are obtained on the development set. We found that the optimal α and β vary slightly for different models. Based on tuning results using internal Google datasets, we use α=0.2 and β=0.2 in our experiments, unless noted otherwise. 8 Experiments and Results -------------------------- In this section, we present our experimental results on two publicly available corpora used extensively as benchmarks for Neural Machine Translation systems: WMT’14 English-to-French (WMT En→Fr) and English-to-German (WMT En→De). On these two datasets, we benchmark GNMT models with word-based, character-based, and wordpiece-based vocabularies. We also present the improved accuracy of our models after fine-tuning with RL and model ensembling. Our main objective with these datasets is to show the contributions of various components in our implementation, in particular the wordpiece model, RL model refinement, and model ensembling. In addition to testing on publicly available corpora, we also test GNMT on Google’s translation production corpora, which are two to three decimal orders of magnitudes bigger than the WMT corpora for a given language pair. We compare the accuracy of our model against human accuracy and the best Phrase-Based Machine Translation (PBMT) production system for Google Translate. In all experiments, our models consist of 8 encoder layers and 8 decoder layers. (Since the bottom encoder layer is actually bi-directional, in total there are 9 logically distinct LSTM passes in the encoder.) The attention network is a simple feedforward network with one hidden layer with 1024 nodes. All of the models use 1024 LSTM nodes per encoder and decoder layers. ### 8.1 Datasets We evaluate our model on the WMT En→Fr dataset, the WMT En→De dataset, as well as many Google-internal production datasets. On WMT En→Fr, the training set contains 36M sentence pairs. On WMT En→De, the training set contains 5M sentence pairs. In both cases, we use newstest2014 as the test sets to compare against previous work [[31](#bib.bib31), [37](#bib.bib37), [45](#bib.bib45)]. The combination of newstest2012 and newstest2013 is used as the development set. In addition to WMT, we also evaluate our model on some Google-internal datasets representing a wider spectrum of languages with distinct linguistic properties: English ↔ French, English ↔ Spanish and English ↔ Chinese. ### 8.2 Evaluation Metrics We evaluate our models using the standard BLEU score metric. To be comparable to previous work [[41](#bib.bib41), [31](#bib.bib31), [45](#bib.bib45)], we report tokenized BLEU score as computed by the multi-bleu.pl script, downloaded from the public implementation of Moses (on Github), which is also used in [[31](#bib.bib31)]. As is well-known, BLEU score does not fully capture the quality of a translation. For that reason we also carry out side-by-side (SxS) evaluations where we have human raters evaluate and compare the quality of two translations presented side by side for a given source sentence. Side-by-side scores range from 0 to 6, with a score of 0 meaning “completely nonsense translation”, and a score of 6 meaning “perfect translation: the meaning of the translation is completely consistent with the source, and the grammar is correct”. A translation is given a score of 4 if “the sentence retains most of the meaning of the source sentence, but may have some grammar mistakes”, and a translation is given a score of 2 if “the sentence preserves some of the meaning of the source sentence but misses significant parts”. These scores are generated by human raters who are fluent in both languages and hence often capture translation quality better than BLEU scores. ### 8.3 Training Procedure The models are trained by a system we implemented using TensorFlow[[1](#bib.bib1)]. The training setup follows the classic data parallelism paradigm. There are 12 replicas running concurrently on separate machines. Every replica updates the shared parameters asynchronously. We initialize all trainable parameters uniformly between [-0.04, 0.04]. As is common wisdom in training RNN models, we apply gradient clipping (similar to [[41](#bib.bib41)]): all gradients are uniformly scaled down such that the norm of the modified gradients is no larger than a fixed constant, which is 5.0 in our case. If the norm of the original gradients is already smaller than or equal to the given threshold, then gradients are not changed. For the first stage of maximum likelihood training (that is, to optimize for objective function [7](#S5.E7 "(7) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")), we use a combination of Adam [[25](#bib.bib25)] and simple SGD learning algorithms provided by the TensorFlow runtime system. We run Adam for the first 60k steps, after which we switch to simple SGD. Each step in training is a mini-batch of 128 examples. ![Log perplexity vs. steps for Adam, SGD and Adam-then-SGD on WMT En](https://media.arxiv-vanity.com/render-output/8033172/x5.png) Figure 5: Log perplexity vs. steps for Adam, SGD and Adam-then-SGD on WMT En→Fr during maximum likelihood training. Adam converges much faster than SGD at the beginning. Towards the end, however, Adam-then-SGD is gradually better. Notice the bump in the red curve (Adam-then-SGD) at around 60k steps where we switch from Adam to SGD. We suspect that this bump occurs due to different optimization trajectories of Adam vs. SGD. When we switch from Adam to SGD, the model first suffers a little, but is able to quickly recover afterwards. We find that Adam accelerates training at the beginning, but Adam alone converges to a worse point than a combination of Adam first, followed by SGD (Figure [5](#S8.F5 "Figure 5 ‣ 8.3 Training Procedure ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")). For the Adam part, we use a learning rate of 0.0002, and for the SGD part, we use a learning rate of 0.5. We find that it is important to also anneal the learning rate after a certain number of total steps. For the WMT En→Fr dataset, we begin to anneal the learning rate after 1.2M steps, after which we halve the learning rate every 200k steps for an additional 800k steps. On WMT En→Fr, it takes around 6 days to train a basic model using 96 NVIDIA K80 GPUs. Once a model is fully converged using the ML objective, we switch to RL based model refinement, i.e., we further optimize the objective function as in equation [9](#S5.E9 "(9) ‣ 5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). We refine a model until the BLEU score does not change much on the development set. For this model refinement phase, we simply run the SGD optimization algorithm. The number of steps needed to refine a model varies from dataset to dataset. For WMT En→Fr, it takes around 3 days to complete 400k steps. To prevent overfitting, we apply dropout during training with a scheme similar to [[44](#bib.bib44)]. For the WMT En→Fr and En→De datasets, we set the dropout probability to be 0.2 and 0.3 respectively. Due to various technical reasons, dropout is only applied during the ML training phase, not during the RL refinement phase. The exact hyper-parameters vary from dataset to dataset and from model to model. For the WMT En→De dataset, since it is significantly smaller than the WMT En→Fr dataset, we use a higher dropout probability, and also train smaller models for fewer steps overall. On the production data sets, we typically do not use dropout, and we train the models for more steps. ### 8.4 Evaluation after Maximum Likelihood Training The models in our experiments are word-based, character-based, mixed word-character-based or several wordpiece models with varying vocabulary sizes. For the word model, we selected the most frequent 212K source words as the source vocabulary and the most popular 80k target words as the target vocabulary. Words not in the source vocabulary or the target vocabulary (unknown words) are converted into special <first\_char>\_UNK\_<last\_char> symbols. Note, in this case, there is more than one UNK (e.g., our production word models have roughly 5000 different UNKs in this case). We then use the attention mechanism to copy a corresponding word from the source to replace these unknown words during decoding [[37](#bib.bib37)]. The mixed word-character model is similar to the word model, except the out-of-vocabulary (OOV) words are converted into sequences of characters with special delimiters around them as described in section [4.2](#S4.SS2 "4.2 Mixed Word/Character Model ‣ 4 Segmentation Approaches ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") in more detail. In our experiments, the vocabulary size for the mixed word-character model is 32K. For the pure character model, we simply split all words into constituent characters, resulting typically in a few hundred basic characters (including special symbols appearing in the data). For the wordpiece models, we train 3 different models with vocabulary sizes of 8K, 16K, and 32K. Table [4](#S8.T4 "Table 4 ‣ 8.4 Evaluation after Maximum Likelihood Training ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") summarizes our results on the WMT En→Fr dataset. In this table, we also compare against other strong baselines without model ensembling. As can be seen from the table, “WPM-32K”, a wordpiece model with a shared source and target vocabulary of 32K wordpieces, performs well on this dataset and achieves the best quality as well as the fastest inference speed. The pure character model (char input, char output) works surprisingly well on this task, not much worse than the best wordpiece models in BLEU score. However, these models are rather slow to train and slow to use as the sequences are much longer. Our best model, WPM-32K, achieves a BLEU score of 38.95. Note that this BLEU score represents the averaged score of 8 models we trained. The maximum BLEU score of the 8 models is higher at 39.37. We point out that our models are completely self-contained, as opposed to previous models reported in [[45](#bib.bib45)], which depend on some external alignment models to achieve their best results. Also note that all our test set numbers were achieved by picking an optimal model on the development set which was then used to decode the test set. Note that the timing numbers for this section are obtained on CPUs, not TPUs. We use here the same CPU machine as described above, and run the decoder with a batchsize of 16 sentences in parallel and a maximum of 4 concurrent hypotheses at any time per sentence. The time per sentence is the total decoding time divided by the number of respective sentences in the test set. \| \| \| \| \| --- \| --- \| --- \| \| Model \| BLEU \| CPU decoding time \| \| \| \| per sentence (s) \| \| Word \| 37.90 \| 0.2226 \| \| Character \| 38.01 \| 1.0530 \| \| WPM-8K \| 38.27 \| 0.1919 \| \| WPM-16K \| 37.60 \| 0.1874 \| \| WPM-32K \| 38.95 \| 0.2118 \| \| Mixed Word/Character \| 38.39 \| 0.2774 \| \| PBMT [[15](#bib.bib15)] \| 37.0 \| \| \| LSTM (6 layers) [[31](#bib.bib31)] \| 31.5 \| \| \| LSTM (6 layers + PosUnk) [[31](#bib.bib31)] \| 33.1 \| \| \| Deep-Att [[45](#bib.bib45)] \| 37.7 \| \| \| Deep-Att + PosUnk [[45](#bib.bib45)] \| 39.2 \| \| Table 4: Single model results on WMT En→Fr (newstest2014) Similarly, the results of WMT En→De are presented in Table [5](#S8.T5 "Table 5 ‣ 8.4 Evaluation after Maximum Likelihood Training ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). Again, we find that wordpiece models achieves the best BLEU scores. \| \| \| \| \| --- \| --- \| --- \| \| Model \| BLEU \| CPU decoding time \| \| \| \| per sentence (s) \| \| Word \| 23.12 \| 0.2972 \| \| Character (512 nodes) \| 22.62 \| 0.8011 \| \| WPM-8K \| 23.50 \| 0.2079 \| \| WPM-16K \| 24.36 \| 0.1931 \| \| WPM-32K \| 24.61 \| 0.1882 \| \| Mixed Word/Character \| 24.17 \| 0.3268 \| \| PBMT [[6](#bib.bib6)] \| 20.7 \| \| \| RNNSearch [[37](#bib.bib37)] \| 16.5 \| \| \| RNNSearch-LV [[37](#bib.bib37)] \| 16.9 \| \| \| RNNSearch-LV [[37](#bib.bib37)] \| 16.9 \| \| \| Deep-Att [[45](#bib.bib45)] \| 20.6 \| \| Table 5: Single model results on WMT En→De (newstest2014) WMT En→De is considered a more difficult task than WMT En→Fr as it has much less training data, and German, as a more morphologically rich language, needs a huge vocabulary for word models. Thus it is more advantageous to use wordpiece or mixed word/character models, which provide a gain of more than 2 BLEU points on top of the word model and about 4 BLEU points on top of previously reported results in [[6](#bib.bib6), [45](#bib.bib45)]. Our best model, WPM-32K, achieves a BLEU score of 24.61, which is averaged over 8 runs. Consistently, on the production corpora, wordpiece models tend to be better than other models both in terms of speed and accuracy. ### 8.5 Evaluation of RL-refined Models The models trained in the previous section are optimized for log-likelihood of the next step prediction which may not correlate well with translation quality, as discussed in section [5](#S5 "5 Training Criteria ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). We use RL training to fine-tune sentence BLEU scores after normal maximum-likelihood training. The results of RL fine-tuning on the best En→Fr and En→De models are presented in Table [6](#S8.T6 "Table 6 ‣ 8.5 Evaluation of RL-refined Models ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), which show that fine-tuning the models with RL can improve BLEU scores. On WMT En→Fr, model refinement improves BLEU score by close to 1 point. On En→De, RL-refinement slightly hurts the test performance even though we observe about 0.4 BLEU points improvement on the development set. The results presented in Table [6](#S8.T6 "Table 6 ‣ 8.5 Evaluation of RL-refined Models ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") are the average of 8 independent models. We also note that there is an overlap between the wins from the RL refinement and the decoder fine-tuning (i.e., the introduction of length normalization and coverage penalty). On a less fine-tuned decoder (e.g., if the decoder does beam search by log-probability only), the win from RL would have been bigger (as is evident from comparing results in Table [2](#S7.T2 "Table 2 ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") and Table [3](#S7.T3 "Table 3 ‣ 7 Decoder ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation")). \| Dataset \| Trained with log-likelihood \| Refined with RL \| \| --- \| --- \| --- \| \| En→Fr \| 38.95 \| 39.92 \| \| En→De \| 24.67 \| 24.60 \| Table 6: Single model test BLEU scores, averaged over 8 runs, on WMT En→Fr and En→De ### 8.6 Model Ensemble and Human Evaluation We ensemble 8 RL-refined models to obtain a state-of-the-art result of 41.16 BLEU points on the WMT En→Fr dataset. Our results are reported in Table [7](#S8.T7 "Table 7 ‣ 8.6 Model Ensemble and Human Evaluation ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). \| Model \| BLEU \| \| --- \| --- \| \| WPM-32K (8 models) \| 40.35 \| \| RL-refined WPM-32K (8 models) \| 41.16 \| \| LSTM (6 layers) [[31](#bib.bib31)] \| 35.6 \| \| LSTM (6 layers + PosUnk) [[31](#bib.bib31)] \| 37.5 \| \| Deep-Att + PosUnk (8 models) [[45](#bib.bib45)] \| 40.4 \| Table 7: Model ensemble results on WMT En→Fr (newstest2014) We ensemble 8 RL-refined models to obtain a state-of-the-art result of 26.30 BLEU points on the WMT En→De dataset. Our results are reported in Table [8](#S8.T8 "Table 8 ‣ 8.6 Model Ensemble and Human Evaluation ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). \| Model \| BLEU \| \| --- \| --- \| \| WPM-32K (8 models) \| 26.20 \| \| RL-refined WPM-32K (8 models) \| 26.30 \| Table 8: Model ensemble results on WMT En→De (newstest2014). See Table [5](#S8.T5 "Table 5 ‣ 8.4 Evaluation after Maximum Likelihood Training ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") for a comparison against non-ensemble models. Finally, to better understand the quality of our models and the effect of RL refinement, we carried out a four-way side-by-side human evaluation to compare our NMT translations against the reference translations and the best phrase-based statistical machine translations. During the side-by-side comparison, humans are asked to rate four translations given a source sentence. The four translations are: 1) the best phrase-based translations as downloaded from <http://matrix.statmt.org/systems/show/2065>, 2) an ensemble of 8 ML-trained models, 3) an ensemble of 8 ML-trained and then RL-refined models, and 4) reference human translations as taken directly from newstest2014, Our results are presented in Table [9](#S8.T9 "Table 9 ‣ 8.6 Model Ensemble and Human Evaluation ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). \| Model \| BLEU \| Side-by-side \| \| --- \| --- \| --- \| \| \| \| averaged score \| \| PBMT [[15](#bib.bib15)] \| 37.0 \| 3.87 \| \| NMT before RL \| 40.35 \| 4.46 \| \| NMT after RL \| 41.16 \| 4.44 \| \| Human \| \| 4.82 \| Table 9: Human side-by-side evaluation scores of WMT En→Fr models. The results show that even though RL refinement can achieve better BLEU scores, it barely improves the human impression of the translation quality. This could be due to a combination of factors including: 1) the relatively small sample size for the experiment (only 500 examples for side-by-side), 2) the improvement in BLEU score by RL is relatively small after model ensembling (0.81), which may be at a scale that human side-by-side evaluations are insensitive to, and 3) the possible mismatch between BLEU as a metric and real translation quality as perceived by human raters. Table [11](#A0.T11 "Table 11 ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") contains some example translations from PBMT, "NMT before RL" and "Human", along with the side-by-side scores that human raters assigned to each translation (some of which we disagree with, see the table caption). ### 8.7 Results on Production Data We have carried out extensive experiments on many Google-internal production data sets. As the experiments above cast doubt on whether RL improves the real translation quality or simply the BLEU metric, RL-based model refinement is not used during these experiments. Given the larger volume of training data available in the Google corpora, dropout is also not needed in these experiments. \| \| PBMT \| GNMT \| Human \| Relative \| \| --- \| --- \| --- \| --- \| --- \| \| \| \| \| \| Improvement \| \| English → Spanish \| 4.885 \| 5.428 \| 5.504 \| 87% \| \| English → French \| 4.932 \| 5.295 \| 5.496 \| 64% \| \| English → Chinese \| 4.035 \| 4.594 \| 4.987 \| 58% \| \| Spanish → English \| 4.872 \| 5.187 \| 5.372 \| 63% \| \| French → English \| 5.046 \| 5.343 \| 5.404 \| 83% \| \| Chinese → English \| 3.694 \| 4.263 \| 4.636 \| 60% \| Table 10: Mean of side-by-side scores on production data In this section we describe our experiments with human perception of the translation quality. We asked human raters to rate translations in a three-way side-by-side comparison. The three sides are from: 1) translations from the production phrase-based statistical translation system used by Google, 2) translations from our GNMT system, and 3) translations by humans fluent in both languages. Reported here in Table [10](#S8.T10 "Table 10 ‣ 8.7 Results on Production Data ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation") are averaged rated scores for English ↔ French, English ↔ Spanish and English ↔ Chinese. All the GNMT models are wordpiece models, without model ensembling, and use a shared source and target vocabulary with 32K wordpieces. On each pair of languages, the evaluation data consist of 500 randomly sampled sentences from Wikipedia and news websites, and the corresponding human translations to the target language. The results show that our model reduces translation errors by more than 60% compared to the PBMT model on these major pairs of languages. A typical distribution of side-by-side scores is shown in Figure [6](#S8.F6 "Figure 6 ‣ 8.7 Results on Production Data ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"). ![Histogram of side-by-side scores on 500 sampled sentences from Wikipedia and news websites for a typical language pair, here English ](https://media.arxiv-vanity.com/render-output/8033172/histogram.png) Figure 6: Histogram of side-by-side scores on 500 sampled sentences from Wikipedia and news websites for a typical language pair, here English → Spanish (PBMT blue, GNMT red, Human orange). It can be seen that there is a wide distribution in scores, even for the human translation when rated by other humans, which shows how ambiguous the task is. It is clear that GNMT is much more accurate than PBMT. As expected, on this metric the GNMT system improves also compared to the PBMT system. In some cases human and GNMT translations are nearly indistinguishable on the relatively simplistic and isolated sentences sampled from Wikipedia and news articles for this experiment. Note that we have observed that human raters, even though fluent in both languages, do not necessarily fully understand each randomly sampled sentence sufficiently and hence cannot necessarily generate the best possible translation or rate a given translation accurately. Also note that, although the scale for the scores goes from 0 (complete nonsense) to 6 (perfect translation) the human translations get an imperfect score of only around 5 in Table [10](#S8.T10 "Table 10 ‣ 8.7 Results on Production Data ‣ 8 Experiments and Results ‣ Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation"), which shows possible ambiguities in the translations and also possibly non-calibrated raters and translators with a varying level of proficiency. Testing our GNMT system on particularly difficult translation cases and longer inputs than just single sentences is the subject of future work. 9 Conclusion ------------- In this paper, we describe in detail the implementation of Google’s Neural Machine Translation (GNMT) system, including all the techniques that are critical to its accuracy, speed, and robustness. On the public WMT’14 translation benchmark, our system’s translation quality approaches or surpasses all currently published results. More importantly, we also show that our approach carries over to much larger production data sets, which have several orders of magnitude more data, to deliver high quality translations. Our key findings are: 1) that wordpiece modeling effectively handles open vocabularies and the challenge of morphologically rich languages for translation quality and inference speed, 2) that a combination of model and data parallelism can be used to efficiently train state-of-the-art sequence-to-sequence NMT models in roughly a week, 3) that model quantization drastically accelerates translation inference, allowing the use of these large models in a deployed production environment, and 4) that many additional details like length-normalization, coverage penalties, and similar are essential to making NMT systems work well on real data. Using human-rated side-by-side comparison as a metric, we show that our GNMT system approaches the accuracy achieved by average bilingual human translators on some of our test sets. In particular, compared to the previous phrase-based production system, this GNMT system delivers roughly a 60% reduction in translation errors on several popular language pairs. Acknowledgements ---------------- We would like to thank the entire Google Brain Team and Google Translate Team for their foundational contributions to this project.
94cc2b23-1617-40fb-a502-7c96f06233c6	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Searching for Search Thanks to Dan Braun, Ze Shen Chin, Paul Colognese, Michael Ivanitskiy, Sudhanshu Kasewa, and Lucas Teixeira for feedback on drafts. This work was carried out while at[Conjecture](https://www.conjecture.dev/). This post is a loosely structured collection of thoughts and confusions about search and mesaoptimization and how to look for them in transformers. We've been thinking about this for a while and still feel confused. Hopefully this post makes others more confused so they can help. Mesaoptimization ================ We can define mesaoptimization as internal optimization, where “optimization” describes the structure of computation within a system, not just its behavior. This kind of optimization seems [particularly powerful](https://www.alignmentforum.org/posts/s2KJWLAPyjtmQ9ze3/search-in-territory-vs-search-in-map), and many alignment researchers seem to think it’s one of the biggest concerns in alignment. Despite how important this is, we still understand very little about it. For starters, it's not clear what internal optimization actually means. People have proposed several definitions of optimization which fit well when thinking about “behavioral optimization” where an agent acts to optimize its environment. One of the most clear and popular definitions comes from [Alex Flint](https://www.alignmentforum.org/posts/znfkdCoHMANwqc2WE/the-ground-of-optimization-1): > an optimizing system is a physical process in which the configuration of some part of the universe moves predictably towards a small set of target configurations from any point in a broad basin of optimization, despite perturbations during the optimization process. > > In this framing, an agent and its environment together form an optimization process, where the agent acts upon the environment such that the system as a whole robustly converges to a set of target states. But we run into problems when we try to map this behavioral definition to a definition about the structure of a process's internal computation. When we say mesaoptimization, we seem to mean something different than just that the computationconverges to a smaller target. For example, an image classifier takes a large set of initial configurations of images including a lot of noise and irrelevant details, and layer by layer narrows it down to a probability distribution concentrated on a single class prediction. There seems to be a sense that this is not doing the kind of optimization we are concerned about when we talk about mesaoptimization. Mesaoptimization was originally defined in [Risks from Learned Optimization](https://www.alignmentforum.org/posts/FkgsxrGf3QxhfLWHG/risks-from-learned-optimization-introduction) as internal search: > We will say that a system is an optimizer if it is internally searching through a search space (consisting of possible outputs, policies, plans, strategies, or similar) looking for those elements that score high according to some objective function that is explicitly represented within the system. > > An advantage of this framing is that we do have some idea what we mean by “search” and have concrete examples of things which unambiguously qualify. Ideally, we’d like to point to the more general class of computation we’re worried about, but once you start thinking about what this general class of computation might look like, it quickly becomes clear that we don’t even know what “search” is. The space of search algorithms also seems much larger and more flexible than implied in the examples we usually think of. At the moment we have very little idea what kind of algorithms we should expect neural networks to learn, and neither do we have a good picture of what kind of algorithms in particular we should be concerned about when we think of misalignment. If the intuition that “search” is somehow central to internal optimization holds validity, then becoming less confused about what learned search looks like should be central to making risks from internal optimization more concrete. What is Search? =============== We have examples of processes which most would say are doing some kind of search, like [Monte Carlo tree search](https://en.wikipedia.org/wiki/Monte_Carlo_tree_search), or [A\](https://en.wikipedia.org/wiki/A_search_algorithm), and processes that not everyone agrees count as search, like evolution through [natural selection](https://en.wikipedia.org/wiki/Natural_selection) or stochastic gradient descent, which nonetheless clearly have search-like properties (the general disagreement about these examples is evidence that the concept is confusing). Some of these prototypical examples are handcrafted, others are natural, but broadly they operate in a similar way: They tend to generate candidate solutions, have a method for evaluating these candidate solutions, and use those evaluations to do a kind of iterative refinement of the solution space. This is highly reminiscent of what Abram Demski calls [selection processes](https://www.alignmentforum.org/posts/ZDZmopKquzHYPRNxq/selection-vs-control), which are able to both instantiate members of the solution space and obtain direct feedback on their quality. While these are clearly doing search, this doesn’t look at all like how humans usually do search. [Humans don’t restrict themselves to enumerating and evaluating candidate solutions](https://www.lesswrong.com/posts/6mysMAqvo9giHC4iX/what-s-general-purpose-search-and-why-might-we-expect-to-see). Instead, humans often operate over highly abstract compressions of the solution space, or search over entirely different spaces, such as global constraintson the problem, to make finding a solution more tractable. Instead of focusing on one particular type of search, we want to think about the general properties that our examples of search share. At the highest level, they each take a large space of possible candidates (generally implicit to a problem specification) and shrink it down to a single solution which satisfies some criteria. If we say a process which finds a solution from a space of 2^N candidates can be thought of as doing N [bits of optimization](https://www.lesswrong.com/posts/Q4hLMDrFd8fbteeZ8/measuring-optimization-power), then one place to start is to ask where those bits of optimization come from. Compressing the search space ---------------------------- By transforming the search space into one which has significantly fewer degrees of freedom, we can quickly obtain a large number of bits, and shrink the scope of the problem down to one which is much more manageable. Master chess players are able to better memorize boards by chunking structures of pieces as single units and converting a board to a higher level concept space not possessed by novices.[[1]](#fnm2sb4nsxlt) For instance, instead of separately tracking the positions of every pawn, the master player might memorize a single position for a group of pawns in a commonly-occurring arrangement. ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649510/mirroredImages/FDjTgDcGPc7B98AES/n8dofvbvvthbouxyyqvh.png) This advantage is only present for real game positions, however, and completely disappears for fully random board states. This suggests that the concept space used by master players implicitly ignores most of the possible ways that the pieces can be arranged, and therefore has less degrees of freedom than the original space did. (It has to: for any lossless compression scheme, [some possible sequences are inevitably incompressible](https://en.wikipedia.org/wiki/Kolmogorov_complexity#Compression), because there are fewer sequences of length < N than length N.) ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649510/mirroredImages/FDjTgDcGPc7B98AES/xwcsolmoinu16alqes12.png) Furthermore, unlike concepts representing literal "chunks" of relative positions parameterized by absolute position, many useful concepts may be invariant to the exact position of the pieces, like the concept of a “pin” or “skewer”. In practice this is a type of lossy compression which would not be able to disambiguate between possible boards if the differences are unlikely to be strategically relevant. The compression, more generally, is just a useful ontology, which makes the search problem significantly easier. Promising candidates (board positions or sequences of positions) can be represented in fewer bits, and the lower dimensional representation space can be used as a generator of candidates, because satisfactory solutions now make up a larger fraction of the now smaller compressed space. This is important, because the way that the solution space gets compressed will shape what search looks like from the inside. Evaluated candidates may not be in the naive or expected representation, but rather be pointers to pieces of an internal ontology that might look very different. We should also expect those internal representations to be harder to detect and decipher, because the more a representation gets compressed, the more we should expect it to [lose structural correlations which might distinguish it from noise](https://www.mdpi.com/2078-2489/11/4/196/htm). Searching over constraint space ------------------------------- Another way to reduce the number of degrees of freedom is to focus on the [constraints of a search problem](https://www.alignmentforum.org/posts/6mysMAqvo9giHC4iX/what-s-general-purpose-search-and-why-might-we-expect-to-see#Babble_And_Prune_Is_Not_The_Only_Search_Method). Searching over global constraints has many advantages, and seems to be one of the main ways humans search when the solution space is too large to reason about directly. Constraint space tends to be much lower dimensional, and can be used to implicitly carve out huge chunks of the original solution space, or break a problem down into smaller subproblems, allowing a search process to recurse on those subproblems, factoring the problem into manageable pieces. In a game like chess this might look like searching for threats, which operate like bottlenecks, requiring all successful plans to involve mitigating the threat as an instrumental subgoal. ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649511/mirroredImages/FDjTgDcGPc7B98AES/j8ieywd8fgsd1yt6dlle.png) This means that the candidates being considered may not map to solutions at all, instead being objects of an entirely different search space (e.g. that of threats). A search process might also consider solutions to smaller or relaxed subproblems, and never reference or relate directly to the full search target. This affects the problem of searching for search, because we might not be able to find the algorithm if we only search for signs of a search over solution space, such as instantiations of candidate solutions. Another important quality of constraint space is that constraints are often useful for a wide range of search targets. For example, when planning a long distance trip, any successful plan will likely require booking a flight and going to the airport. This fact doesn’t depend very much on the exact destination, and so is especially useful to a more general purpose search process. General purpose search, of the kind that humans possess, is also the kind of internal optimization that is most concerning and worth our study. Such a system would need to be able to take a search problem and actively simplify it by finding exploitable features of the search space at runtime, and breaking that problem down into more manageable subproblems. Modularity and Retargetability ============================== General purpose search needs to be able to take on a wide range of possible search problems and be flexible to distributional shifts. Systems which are robust to changes in the objective tend to be [modular](https://www.alignmentforum.org/posts/JBFHzfPkXHB2XfDGj/evolution-of-modularity), composing their computation into submodules. ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649510/mirroredImages/FDjTgDcGPc7B98AES/qi0xnewde0hfwmqtdkxk.png)Non-modular systems, where the dependencies of computation are spread too widely to be split into distinct modules, are more likely to be brittle to changes in how the computation is performed. In learning to code, for example, a common early lesson is to split a program up into functions, and allow each function to interact with the rest of the program only through a small number of input and output variables (and avoid references to global variables). This doesn’t just make code easier to read, which is not a requirement for learned algorithms, but more importantly it contains the ripple effect that a change in one function has on the rest of the program. This advantage of modularity applies both at runtime (reducing the chances that an uncommon input causes failure), as well as during development (reducing the number of changes needed to fix a bug). Analogously, we might expect modularity in learned algorithms both for their ability to generalize at runtime, as well as a part of the inductive bias of gradient descent. The more sprawling the dependencies, the more directions in the gradient landscape will lead directly to machinery breaking (and thus higher loss). A change to a module, on the other hand, need only be accompanied with a change to the API for the system as a whole to keep on functioning. A likely place to find an example of modularity is in the retargetabilityof a search algorithm. If a search algorithm is general purpose, then the “target”, or what is being searched for, can’t be baked into or be implicit to the search process, but rather must be an input to a retargetable search algorithm. This applies to a system capable of handling a broad set of search problems, but even systems trained only to search with a single target have an incentive to be retargetable if that search algorithm breaks the problem down into subproblems, each with their own instrumental targets. If internal to the system there exists a target-agnostic search algorithm, with separate machinery for providing the target, then one place to start would be to find those modules (and understand how they interact with each other). What even is a target/goal? --------------------------- A target, through its interaction with the search algorithm, functions to “steer” search toward converging to a particular set of solutions. Much of the search process might be possible to do without any reference to the target (goal-agnostic preprocessing), but for general purpose algorithms we should expect a significant amount of the computation to hinge pivotally on the target. In handcrafted search algorithms, the target has a clear type signature, and its interaction with the rest of the search process is well understood. In MCTS, for example, the target takes the shape of an evaluation function, and similarly for A\, the target is a coordinate which is used by a heuristic distance metric to guide the search process. They are both retargetable algorithms, and it’s easy to see how changing the target will change the process by which the algorithm will narrow the search space. In learned search, it might be searching over multiple different ontologies, many of which don’t map to the solution space, as well as generate instrumental goals and recurse on subproblems. We don’t currently know what type signature the target would have, or how it slots into a search process to guide the computation, because we don’t have a good gears-level understanding of what learned search looks like. Developing a better understanding of what targets are, both conceptually and experimentally, might give us a foot in the door toward understanding how targets interact with other modules within the search algorithm. Learned Search in Transformers ============================== Another angle of attack to understand mesaoptimization and search is to think about what types of algorithms we expect neural networks to learn. “Learned search” refers to search algorithms which, unlike handcrafted search algorithms, have been found automatically via gradient descent. If we find learned search within neural networks, it will need to conform both to the limitations of what can be implemented in the architecture as well as the inductive bias of gradient descent. Because of the success of LLMs like GPT-3 I’ll be focusing on the transformer architecture. Many of the arguments also apply to other architectures. We might [expect mesaoptimizers](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB/p/q2rCMHNXazALgQpGH) to be preferred for many reasons. Search at runtime is effectively a compressed version of a more complicated policy that encodes the information gained by search explicitly, specifying a system capable of generating that policy. For example, instead of memorizing a book of optimal chess openings, a program much shorter than the book could search through future consequences of moves and hypothetically converge to the same output. There are, however, some reasons to be skeptical of finding search in transformers. First of all, [search isn’t free](https://www.alignmentforum.org/posts/TdesHi8kkyokQdDoQ/gradient-descent-doesn-t-select-for-inner-search): the chess program that does brute-force search may have fewer lines of code, but is more expensive to run than a program that simply looks up a memorized answer. Any search algorithm which uses too much computation or memory just cannot be implemented by a transformer. This rules out many of the algorithms which rely heavily on explicitly enumerating and evaluating states. In the same vein, any algorithm that requires a high number of serial steps would also not be possible to run, even if the algorithm itself is in principle quite simple to describe. Unlike the NLP architectures that came before it, transformers are not recurrent. Information can only flow up, and so there is a hard cap on the number of serial steps* that can be used to make a prediction. ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649510/mirroredImages/FDjTgDcGPc7B98AES/gnrd8hfxz19v0gmya3ek.png)Here both a transformer and a stacked LSTM are shown predicting token 5 from the context (tokens 0-4). While each LSTM block is able to pass information to a copy of itself in the next column, in a transformer information can only be passed upward to the next layer.Furthermore, the fact that a layer cannot pass information to itself is also a constraint which affects the inductive bias over possible search algorithms, likely disincentivizing algorithms that require certain subroutines to be reused many times for a single prediction. If weights are not being shared, then for a subroutine to be applied multiple times, it must be independently learned and implemented in different parts of the network. For example, if a search algorithm relies on something like an evaluation function, it would need to have separately learned every instance of its use, making the effective complexity of the algorithm’s description very high. This also might limit the extent to which we should expect to see algorithms which rely on recursion. Another clue which restricts the algorithms we should expect from transformers relates to the way information seems to be processed in the residual stream. The transformer architecture relies on a residual stream, where individual heads read and write to a central channel of information. The [logit lens](https://www.alignmentforum.org/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens) appears to show that a byproduct of this design is that transformers tend to quickly find candidate solutions in early layers and then refine and update them over the rest of the forward pass. This might push us to consider algorithms that don’t completely depart from the solution space, and perhaps use early solutions to help inform the rest of the search process. For example, a transformer might exploit a certain [duality](https://www.alignmentforum.org/posts/T8qhPyiFPFcSZ2swC/mazes-and-duality) between solution space and constraint space, iteratively using members of the solution space to identify constraints, and members of the constraint space to identify solutions. Of course, it could also be that transformers are just using a giant bag of heuristics to produce solutions, and don’t implement anything at all like something we would call search. We can add credence to the existence of search processes by demonstrating the ability of transformers to solve “searchy” problems like chess, which seems to require a certain ability to search through possible futures, but ultimately what we really need is for interpretability to illuminate the difference. How to Search for Search? ========================= One thing that we would really like to have is a clear signal from reality that could help guide our thinking about this. We could spend a lot of time pondering the space of search algorithms, and which kind of algorithms a transformer might learn, but ultimately we would like some clear evidence to help narrow our questions. This seems like a really hard problem, and unfortunately the interpretability tools we currently have are limited, but here are some general thoughts about how to approach this problem. Asking the right questions -------------------------- We have to be able to turn abstract conceptual questions about search into testable hypotheses. One approach is to identify properties we expect learned search algorithms to have in theory, and then more concretely what computational artifacts we should expect to find in practice. We can do this both by considering what properties are sufficient or necessary for a search algorithm implemented within a particular architecture to function, as well as what is [selected for](https://www.alignmentforum.org/posts/G2Lne2Fi7Qra5Lbuf/selection-theorems-a-program-for-understanding-agents) by the training process. We can then design experiments that make it possible to test those hypotheses. We can deliberately train transformers on “searchy” data, like maze solving or chess games, in order to focus on the types of search mechanisms that we think might make it possible to perform well on that data. For example, to predict chess games, it seems plausible that a transformer would need to consider future paths, and so we could focus our experiments there. Failure modes ------------- Many interpretability methods will try to determine what a neural network is thinking, and usually end up finding correlation rather than causation. For example, one can use some kind of supervised process to [train a probe](https://arxiv.org/pdf/1610.01644.pdf), and then use that probe to extract features from the activation space. Unfortunately this doesn’t guarantee that the network is actually using those features and not something which correlates with it. Furthermore, even if it does use those features, it might be using them in completely different ways than we expect. We end up trying to map our own ontology of information onto that of the neural network without really understanding it. Take [this paper](https://arxiv.org/pdf/2111.09259.pdf), where the authors train linear probes to predict high-level chess features from the intermediate layers of AlphaZero in order to gain insight about what knowledge the network is using at various stages. They find that they are able to predict concepts like “king safety” and we could reason that this makes sense as a concept, as it could be useful for figuring out what moves to make next. But instead of picking up on king safety, the probe might also be picking up on other features, like the number of pieces on the board, which correlate strongly with king safety, and there isn’t an obvious way to tell the difference. One way to overcome this is to focus on doing [causal intervention](https://arxiv.org/abs/2202.05262), but even this has its problems. Semantic features are all intertwined and dependent on each other, and for many things it seems really non-trivial to edit something without breaking the network’s brain.[[2]](#fnny5mhpfj4rk) What would it mean to cause the network to believe the king was safe, without affecting a whole host of other semantic features? Is it possible to make a network trained on addition to think 2 + 2 = 5, without destroying its ability to do math? Lastly, experiments often have the flaw of testing something more specific than the hypothesis we are actually interested in. We might often have a very general hypothesis about the kind of things the network might be doing, but end up implicitly testing a much more specific hypothesis. For example, we might hypothesize that a network is explicitly proposing candidates, but when we go to test it, our effective hypothesis becomes that the network is explicitly proposing candidates in a format that this method (e.g. linear probes) would find. If we get a negative result, we can’t say much about the original hypothesis because the network could just be doing things in a different way we aren’t able to detect. If we get a positive result, we might also just be picking up on some correlation which satisfies the test, but is actually caused by something fundamentally different from the original hypothesis.[[3]](#fnphpy0gkt1n) Firehose interpretability ------------------------- Instead of designing experiments which deliver only a single bit of evidence (falsify/verify), another approach is to instead aim for methods which measure lots of things all at once, and produce a [firehose of bits](https://www.lesswrong.com/posts/9kNxhKWvixtKW5anS/you-are-not-measuring-what-you-think-you-are-measuring). If we don’t currently know the right questions to ask, or even in what ontology to pose our questions, then it can be really hard to design experiments that cut at the heart of what we care about, as well as to draw meaningful conclusions from them. When it becomes difficult to form testable hypotheses that make progress on the important questions, it makes sense to shift away from running classical hypothesis-driven experiments toward making high bit observations. ![](http://res.cloudinary.com/lesswrong-2-0/image/upload/v1669649510/mirroredImages/FDjTgDcGPc7B98AES/kkddktbfixkhqjnmoout.png) Our ability to make lots of useful observations depends on measurement tools, or lenses, that make visible things which are invisible, either by overcoming the physical limitations of our sense organs or our cognitive limitations to interpret raw data. This can be a major bottleneck to scientific progress, a prototypical example being the invention of the microscope, which was a turning point for our ability to study the natural world. The [lenses](https://github.com/TomFrederik/unseal) that currently exist for interpretability are still quite crude, and expanding the current suite of tools, as well as building places to [explore](https://microscope.openai.com/models) and [visualize](http://interp-tools.redwoodresearch.org/) neural networks using those tools, seems critical for making lots of high bit observations. Another motivation for high-bandwidth measurements comes from our problem with inferring causality from correlations. While it’s true that it’s impossible to determine causality from the correlation between just two variables, for more than two it can become possible, and the more variables we do measure the easier it becomes. This is a path to building causal models which avoids the pitfalls of intervening directly on a neural network. Conclusion ========== Searching for search seems like an important research direction because it points at something strongly related to a lot of models for how misalignment happens, as well as being a prerequisite to a lot of potential solutions, like [retargeting the search](https://www.alignmentforum.org/posts/w4aeAFzSAguvqA5qu/how-to-go-from-interpretability-to-alignment-just-retarget). We want to be able to look for things like mesaoptimization or search, but when we look at a system and ask ourselves “is this system doing search?”, we realize that we don’t really understand the question. These words can make us feel like we know what we are talking about, but when we try to put them into practice we run into trouble. This is why it is so important to keep alignment research grounded in real-world AI systems: it forces us to confront what we don’t know, keeps us from getting lost speculating, and sharpens our focus. We currently don’t really know where to look and which experiments will actually further our understanding, so figuring out how to translate these vague concepts into more concrete claims about how learned search happens on the algorithmic level is critical. 1. [^](#fnrefm2sb4nsxlt)There are several papers which show this phenomenon, [this paper](https://www.researchgate.net/profile/Damon-Navandi-2/publication/268871818_chase_simon_1973/links/547a09b00cf205d1687fab96/chase-simon-1973.pdf) being the most famous example. 2. [^](#fnrefny5mhpfj4rk)Recent work from [Jacques Thibodeau](https://www.lesswrong.com/users/jacques-thibodeau) at SERI MATS and [from the interpretability hackathon](https://jas-ho.itch.io/model-editing-hazards-at-the-example-of-rome) show evidence that editing factual knowledge with [ROME](https://arxiv.org/abs/2202.05262) does not robustly propagate logical implications and causes unintended side effects. 3. [^](#fnrefphpy0gkt1n)This problem of measuring something more specific than the thing we are interested in can also be addressed with [auditing games](https://www.lesswrong.com/posts/EbL5W5ccwfbqFiYBJ/auditing-games-for-high-level-interpretability-1).
f345cb7b-dc44-436d-a9b3-41fd7c379a3d	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Quick thoughts on "scalable oversight" / "super-human feedback" research The current default view seems to roughly be: * Inner alignment is more important than outer alignment (or, alternatively, this distinction is bad/sub-optimal, but basically it's all about generalizing correctly) * Scalable oversight is the only useful form of outer alignment research remaining. * We don't need to worry about sample efficiency in RLHP -- in the limit we just pay everyone to provide feedback, and in practice even a few thousand samples (or a "constition") seems ~good enough. * But maybe it's not good? Because it's more like capabilities research? A common example used for motivating scalable oversight is the "AI CEO". My views are: * We should not be aiming to build AI CEOs * We should be aiming to robustly align AIs to perform "simpler" behaviors that unaided humans (or humans aided with more conventional tools, not, e.g. AI systems trained with RL to do highly interpretive work) feel they can competently judge. * We should aim for a situation where there is broad agreement against building AIs with more ambitious alignment targets (e.g. AI CEOs). * From this PoV, scalable oversight does in fact look mostly like capabilities research. * However, scalable oversight research can still be justified because "If we don't, someone else will". But this type of replaceability argument should always be treated with extreme caution. The reality is more complex: 1) there will be tipping points where it suddenly ceases to apply, and your individual actions actually have a large impact on norms. 2) The details matter, and the tipping points are in different places for different types of research/applications, etc. * It may also make sense to work on scalable oversight in order to increase robustness of AI performance on tasks humans feel they can competently judge ("robustness amplification"). For instance, we could use unaided human judgments and AI-assisted human judgments as safety filters, and not deploy a system unless both processes conclude it is safe. * Getting AI systems to safely perform simpler behaviors safely remains an important research topic, and will likely require improving sample efficiency; the sum total of available human labor will be insufficient for robust alignment, and we probably need to use different architectures / hybrid systems of some form as well. * EtA: the main issue I have with scalable oversight is less that it is advancing capabilities, per se, and more that it seems to raise a "chicken-and-egg" problem, i.e. the arguments for safety/alignment end up being somewhat circular: "this system is safe because the system we used as an assistant was safe" (but I don't think we've solved the "build a safe assistant" part yet, i.e. we don't have the base case for the induction).
783d0452-f647-441d-b569-79510da76834	StampyAI/alignment-research-dataset/youtube	Youtube Transcripts	Friday Seminar Series- Gillian Hadfield: AI Alignment and Human Normativity [Music] so I want to talk generally about this time AI human so we've got a lot of terms floating around AI safety AI policy value alignment how do these things all line up and I want to give you a framework for thinking about that I'm going to talk about a set of research projects I'm involved in and some things I'd like to be involved in so a lot of the conversation about AI safety is on the question of how should we regulate AI right what what are the right rules what are the values we should have so anybody who's like been bored to tears by the trolley problem right and sitting around like what what are the values that's a lot of the conversation in AI epics liability for autonomous vehicles for autonomous weapons etc it's about what should what should the rules be how should we I want to focus it's actually they're really really big and interesting and hard problem is how can we regulate AI how can we actually get machines to do the things that we want them to do and that's that that's a tough question and I think the way to think about that is to think about how can you build AI systems that can interface with what I call human normative systems that's an important concept normativity I'm going to use that language to mean the systems that humans have for classifying behavior as sanctionable or not that's basically every single human society you look at it's going to be full of normative labels this is an OK action that is not an okay action everything about social norms culture law boils down to a classification system a normative classification system this is okay that's not okay and then on top of that an enforcement scheme for punishing people who engage in the not okay thing and as you think this is one of the most exciting and fascinating things about human evolution is the development of these normative systems and I want us to be focused on thinking about that I want to just draw a distinction here because a lot of our work on thinking about how do make sure that robots and a eyes do what we want them to do focus is on the idea of preferences I'm an economist I love the idea of preferences but preferences are a modeling tool I don't think there is no such thing as preferences it's a way of modeling actions and so I'll throw that out there and I'd love anybody wants to talk about preferences and the difference between thinking about this there's a difference between analyzing human normative systems and thinking about human preferences or values so I want to just draw that distinction it's not and focused on the idea of systems which also means there's a lot more information out there for us to use to extract from the environment what is it that's okay and not okay to do I want to think about this as it's both an engineering research program how do you build these kinds of systems and it's a social science research program because frankly there's very little work done on analysis of those normative systems and I think those things need to be integrated and that's why I'm really excited to be talking about this work and hopefully opening up some research programs okay so I'm going to talk about five lines of research the first two we'll talk about in some detail in the last three I'll just kind of hit them to say these are additional things I'm just getting started on would like to get started on we'd love to talk to you about so let's talk about this first one designing reward structures and this is as Marcia mentioned this is work with with dylon and it's really giving an overview framework for thinking about AI alignment so how many of you seen this little video oh okay so this is out of open AI they trained this AI to play this video game it is a boat racing game which you may not be able to tell because that boat has not learned to win the race it has learn to get points because that's the reward function they gave it so you can watch down here the score it's going crazy what did it learn it learned that if you just roll over those those turbo-boost spots in there you get lots of points but you never win the game they have this is posted as an example of how I mean that they thought they were choosing a pretty good reward function for this system to learn how to play the game but this is what it's learning how to do so a great example of the reward design problem here's another example from other work of Dylan's with Smith Emily Peter a bill Stuart Russell and I could do game the last three butcher his advisers at Berkeley and this is a paper that Dylan presented last year at nips inverse reward design thinking here about the kind of problem that you suppose you have a designer wants to train a robot to go find pots of gold in a little grid world and to optimize the path choosing between going over grass which is costly and taking a path which is less a paved path which is less costly the designer sets up this this reward function that's really a proxy reward function trains the robot sends it out in the world but oops out in the world there is lava and the robot knows nothing there's nothing in the reward function that designer specified that takes account of the lava but of course she does actually have a preference preference over lava it's really bad to go through lava how can how can you but you can if you can't always anticipate everything that's going to be out there in the environment you're going to have this problem of designing these reward functions that just don't take account of things that you're going to discover out there in the real world that you really do care about so it's the idea that your reward function is not really your reward function and they have a solution that's basically treating this as an observation from the distribution of possible reward functions okay so that's a way of thinking about the reward design problem and that's a way of thinking about misalignment misalignment between the reward function here that the the AI is is training to achieve and the true reward function or the true values for the humans involved whoops I'm gonna say let's see okay so I went but here's the point I want to make so my PhD is in economics misalignment between individual and social welfare functions is like what all of economics is about that's basically what we do so when Dylan and I start talking about this stuff we started having conversation I was visiting at Harvard and he was finishing up at MIT saying okay so you're talking about these issues you're studying and I'm saying well that's sounds pretty darn familiar we have things in active the first theorem of welfare economics which tells us that perfect incomplete markets in which individuals are just maximizing their individual profit function and utility function can't achieve alignment with social welfare function that's just a basic result we have a whole field of principal-agent analysis which is precisely focused on the problem of how does a principal who delegates a task to an agent get the agent to behave in the ways that the principal wants the agent to behave and then we have a long line of work on what to do about the fact that those those contracts between the principal and the agent will invariably be incomplete it won't capture everything so this is what I was doing when I was a grad student I was thinking about franchising sounds like a weird thing but anyway and I was thinking okay I was doing contract theory and bargaining theory game theory and we were thinking about the problem of okay so McDonald's over here has things that want to accomplish its got values associated with licensing its trademark to franchisees and it's got value it gets from franchisees putting effort into a task and from you know building its market through new products that it puts out there you know the frappuccinos and so on the franchisee over here has values associated with those things that are different from the ones that the principal have and particularly the print the agent here is bearing costs now if if they could write a complete contingent contract that addressed all possible circumstances and specify the rewards for the agent they could align those interests so that they maximize the joint profit function between them right so there'd be an optimal amount of effort there's an awful amount of new frappuccino and products to introduce some rate given the cost if they could write what economists think of as a complete contingent contract addressing every possible circum pants they could achieve the optimum they can maximize joint joint welfare but here's what people were starting to realize way back when when I was starting my PhD was those contracts invariably are incomplete it's really hard to write a contract specifying effort for the franchisee it's hard to measure thing it's really hard to write a contract specifying how often the franchisor should require new investments in new products and so on so those contracts are invariably incomplete and they they exist in an environment where they're filled in by two kinds of mechanisms or institutional settings so here's my little symbol for quartz so for example this contract it's incomplete it's got lots of ambiguity but you might think that has an implied in it that the franchisor has to be acting good faith in deciding when to require these new products and the good faith means you can't just extract from the franchisee franchisees locked in made all these investment you can't just extract a value you couldn't have gotten upfront and you might have courts that enforce that but you might also have the community that's enforcing that so get frantic McDonald's getting a bad reputation of having a harder time getting new franchisees or the franchisee getting a bad reputation or getting terminated so this is my little symbol for informal or community enforcement so that's the world in which those complete contracts live and the message that economists were getting right about then was if we're going to analyze contracts analyzing incomplete contracts is fundamental and I think there's a similar message for people working in CS on to doing reward design is to say look misfit Mis specification of rewards is not just a glitch not oh you know open AI the folks who design that boat racing algorithm should have just been smarter and written a better reward function to say look no Mis specification is the state of play and and we need to think about how do you manage that it's unavoidable and pervasive so tender seeing has been writing this for a little while talking about optimal reward design sort of figuring out what's the best way to do that so we want to think a little bit of a wire contracts incomplete economist have done a lot of work on this what are the reasons for incompleteness well there's bounded rationality you can't think of all the contingencies there's costly cognition and drafting there's what we call non contract ability you can't write all those contracts and then handers that you can't explain to courts to everything you want them to know and if the court is your enforcement mechanism it's going to be incomplete for that reason we've got strategic behavior I don't want to mention certain things that might happen the franchisor doesn't want to mention them I don't want to mention them so they don't end up in the contract because we don't want to talk about them we could plan to renegotiate we know we'll know more later so let's not try and get everything upfront right now let's start our relationship and figure things out later we can plan to write vague and I'm gaps terms and then delegate to a third party to fill it in later with better information so that's the set of reasons that economists have looked at for why contracts are incomplete and I think you can basically write down the same set of reasons for why reward structures might be misspecified bounded rationality costly engineering and design non implemented bility which is what we think is heard of the analog to non contract ability just machine learning problems we haven't figured out how to solve yet adversarial design might be an analogy to strategic behavior not so sure about that one yet it may be that in fact you want to design something where there's a planned iteration on rewards you're gonna have an initial reward you're going to update that reward later and maybe we shouldn't be think about planned completion of rewards by third parties so I think there's this this analogy that we can line up between why contracts aren't complete and where I rewards or misspecified now what we do in this little paper is we go through the insights from the economic theory literature that our pause we're just kind of throwing out there here's some possible things that you might take from the results that we have in the economic literature I'm just going to very quickly mention a couple so for example there's the analysis of property rights in this literature that says sometimes the best thing to do instead of trying to more finally retune your contract it may just be better if you're a principal and you've got an agent running your firm you may just be better to sell the firm to the agent that's transferring property rights to the agent what's important about that is to recognize that selling the state the firm to the agent creating this property is just a transformation of the utility function that's all that it means this is so if we're going to think about what this might mean in the AI context and this really is speculative it's not about giving robots property rights but it is about thinking about whether or not the utility function the reward function has to address more than the specific task that you're trying to accomplish do you need to give that agent rewards that go beyond the specific tasks that you're thinking about it performing another set of results in this literature on measurement and multitasking sometimes the basic results showing that sometimes it's optimal to reduce the incentives on measured tasks when there are important tasks that are not measured as well and the usual example is thinking about teaching if you want to both have your students acquire knowledge but also gain creativity and resilience and so on well we can measure their knowledge pretty well with standardized tests we can't measure the success with which we're imparting creativity and resilience to them and so it may say look you actually don't want to optimize your let's see I think I jump back too far so it may be you actually don't want to include this easily measurable information in the reward structure for the teacher if it's going to distort things from the things you can't measure which I think might also be something to think about in the AI context and I'm really nervous with uber next door to say anything at all about how this applies in self-driving cars but anyway you can think about it we can think but the idea is that you know if you find if you sort of it's it's a natural thing to think I've got this information I should make the greatest use of it but if there's stuff you can't measure you may be better off dialing back how you're using you're miserable stuff and not going to the max on that one we also talked in the paper about theory insights for what we call strong strongly strategic AI thinking here of AI that can change its utility function or change its hardware and so on we're not gonna I'm not going to go through that but you can take a look at the paper if interested the the thing I really want to emphasize are insights from the the legal theory of incomplete contracting I'm going to keep track of my time here so one of the basic insights we get in the 80s is this idea that contracts come embedded in institutional and social structures Granovetter is a sociologist we get the development of something called the relational contracts approach emphasizing that contracts consists not only of their expressed terms but also of their interpreted and implied terms so and those are supplied by law and relation terms so let me sort of talk about why I think this might be interesting to think about in the AI context so this is an example of from a great paper out of open AI concrete problems in AI safety if you're interested so they posit here's a basic problem you train a robot to carry boxes to the other side of the room he's got a gets a reward for getting boxes to the other side of the room you train it for that then you release it out into the world and oops there's a vase that appears on the path so what it says it Lissa's like the law of a problem what is the robot going to do when that would that well a robot doesn't have any information about vases going to just plow straight through the base because this reward structure says there's zero there so can that and and they present this as a problem saying okay how are you gonna develop robots that can figure out as they would say kind of common sense I don't think let's see if we can dig into common sense a bit more structurally so imagine you have a human agent who's got the same job and they've got a contract that's exactly the same as the robots reward function they're going to get paid for boxes on the other side of the room and so we're gonna hire this agent we're gonna leave the agent alone in the room to carry the boxes and then the Vaes is going to appear in the path that the agent has gotten used to using going across the room what's the agent going to do the he an agent gonna go around the vase and the question is why why is the human gonna go around the vid then we don't want to get mushy we don't want to get just like oh is that's just being human and that's just common sense we got to figure out what that looks like how do humans do it what makes incomplete Contracting rational why is it rational for the principal to leave that agent in the room and not worry too much about the vase showing up in the path well the way to think about this I think is to think about the fact that this contract here is not the entire contract that's the expressed terms in the contract the human agent is going to think huh see this vase if I take my usual path I'm going to knock it over and then what's gonna happen well you know the employer might sue me might take me to court and get me for property damages my employment employment law may allow the employer to deduct from my wages the cost of the broken vase or it may just get a really bad reputation I won't get a good recommendation from this employer my co-workers may snicker at me I will bear some kind of a cost for knocking over the vase so the true contract is our - see the cost for knocking over the vase and the agent is able to fill that in by reference to this external structure and go around the vase and this is the key point the human contracts rely on a ton of external structure nobody solves this problem entirely mathematically inside the box that's the basic insight of how you solve the incomplete Contracting problem you got to bring in stuff from the outside and I'ma teach contract law when I'm over at the Law School so if you'd like to know more about that I'm happy to talk about it we bring stuff in from the outside so I think one of the questions is can we build a eyes that can similarly fill in their reward functions can we build them so that they're able to replicate that human process of reading and then predicting the classification of behavior in our human normative systems right the external structure I'm appealing to here's a system it's out there it's not just somebody's preferences it's a system that's out there and can we get them to a negative way to action is classified as sanctionable so can we build those kind of I think we have to figure out how to build those kinds of systems all right let me now talk a little bit about this line of work on modeling human or TV also with Dylan this is a paper on which we've I think titled we've we've roamed around a lot legible normativity for AI a line at the value of silly rules okay so remember I just said we are can we can we build a eyes to replicate the human process of reading and predicting the classification of behavior in human normative systems well to do that if we're going to pay for they're going to need we're going to make predictions about how human normative systems will respond to specific actions you're going to need good models of those human normative systems and this is when I really get worried about this line of work because hardly anybody is working on this from the social science perspective the focus in a lot we have lots of people working on norms and law and throughout the university we have people in psychology sociology law working on economics working on norms but the focus tends to be on the substance of particular norms so we have behavioral economics running experiments playing those dictator or ultimatum games and coming up with a result that gets expressed as 30 percent of people have a preference for fairness and they'll reject an unfair offer in an ultimatum game right that's not telling us how these systems work that is telling us that there is this fairness norm and we're not looking at how do those systems work I think it's also causing us a problem because a critical feature of human normativity is it's arbitrary its capacity for arbitrary content as a group of humans we can effectively put anything we want into our norms and if we can coordinate all of us to say we won't deal with that person if they wear the wrong clothing or if they say the wrong thing right if they don't contribute this amount to our joint project we can establish that as a norm and this is actually I think why we invent normativity precisely because it's capable of taking arbitrary content not meaningless content but arbitrary content in order to adapt to different environments so I think one of the things we have to recognize about modeling human normative systems so we don't want to really model the specific content in specific settings we want to look at the attributes of those systems so this is the point we're gonna have to they can't just be given the rules because there's not a list of norms that are out there in cross cultures across geography across settings across time in future worlds we're going to have to help figure out what of the predictive basis for figuring out what are the rules that are actually enforced and in play in any given environment it's a project I've been working on with co-author berry wine gasps who's a political scientist at Stanford we've got a series of papers on this and what we're looking at or what are the attributes of rule systems that help to coordinate and incentivize third party decentralized punishment we have classifications of behavior this is okay that's not okay and then the challenge is how do you coordinate enforcement of that centralized classification centralized enforcement you know the police the government prisons and so on very it's very rare it's very new in human societies and it's also only playing a very small fraction of the enforcement behavior even in even in settings like contracts those contracts are enforced by the fact that people find out you breach your contract they don't want to do business with you and that's what drives a lot of compliance with contracts so we emphasize that if you're trying to analyze that system a key uncertainty that anybody is facing in a setting is that understand getting a handle on the likelihood that others will participate in punishing somebody who violates the rule we need to know okay even if these rules are announced are people actually going to be enforcing them and it's so a key uncertainty in any setting is what rules are being enforced by others again very distinctive about human societies third party punishment right we say so-and-so was rude to my colleague I don't want to ask so-and-so to join my research team right so and so you know behave badly with with my family member I'm not going to recommend them for a job right we engage in third party enforcement and understanding when and how that's going to happen is a key problem that humans are solving all the time all right so what this project does it looks at what we call silly rules and silly rules we just mean by silly rules and it's deliberately provocative name rules that actually have no value directly in and of themselves a lot of our rules about clothing about food about particular words that we use things you can say or not say in this setting you could think of them as silly rules now it's really important people who are experiencing the rules don't experiencing them don't experience them as silly they think they're very important but I want to define a silly rule as one where there's no direct impact it's only but we want to what's the function is performing so I want to say what's important about this for AI is we want to say okay here's something you might need to understand that we don't currently understand about the way our normative systems work and it will be important if we're gonna build a eyes that can interact in those normative systems that we have good models of this so this is where you can think of this as an example to prove my main point which is we need people interested in this problem of aligning AIS with human normative systems to also be engaging in much more careful research about how human normative systems work okay so we do this we give you an example to get started imagine dropping a robot oh yes sorry I'm gonna give you some right now and I'm gonna do it in a dangerous context okay so we're going to drop a robot down among the awah of Brazil and we're gonna win we want this robot to figure out how to build arrows this is something that the man among the awah spend a lot of time doing making making arrows and part of the reason I want to use this setting is precisely because it seems kind of odd to think about dropping a robot into this environment to learn how to make arrows but it's not really all that much sillier crazier than dropping a robot into Toronto and saying learn how to drive right it's just that we are so immersed in the world of Toronto on how to drive that's very hard for us to sort of pay attention a little to what might be what might be going on in the structure rules I think this is a real challenge for doing this kind of work because we have so many weird we are participants in these systems much harder for us to get outside of them okay so here's here's the kinds of things that this robot is going to observe and on this so it's gonna observe that there's actually a bunch of rules about making arrows those do you don't know this is Hammurabi's code my little symbol for a set of rules Hammurabi's code is not 287 independent rules it's 287 rules carved into a 7-foot pillar a stone and stuck into the the central square in ancient Babylon okay so it's a set of rules so you're gonna see things like use hardwood for the shaft use bamboo arrowheads put feathers on the end of the arrow use only dark feathers smoke the arrows over a fire at all times while they are active and non-active arrow is one that's been bundled up and put in the rafters so this is just an arrow that might be used so keep it over the fire keep it warm don't let it get cold making it making use only personalized arrows make them 1.4 to 1.7 meters long and make and carry as many more arrows than you're going to use okay so that's what you'd observe in terms of the rules these are actually enforced rules of how to make arrows here's what else there are robot will observe men in this society spend over four hours a day making arrows in one season they carried four hundred and two on their trips and they use nine of them and because they get bundled up and so many get carried together they get damaged so a bunch of the four enough plus hours is actually repairing the damage that's done for making and carrying so many in it and not really using them they actually use shotguns to shoot most of the stuff that they're out after this is a tribe that's living in close proximity to developed communities so they're but they're but the man who makes his arrows differently is mocked and shunned they make a lot of fun of him for here's soar an example of what I'm going to call silly rule use only dark feathers he uses colorful it feathers on his arrows he doesn't you they're they're too long they're the wrong length he doesn't maybe he doesn't keep I don't know if this is true maybe he doesn't keep them warm so we have a whole bunch of rules but not all of these rules are functionally related to accomplishing the objective of catching prey so what Dylan and I did was we did a computational experiment and we lose a slide here yeah okay so I think I've lost the slide in here I want you to Santa sighs I don't know exactly which of these are silly but I'm pretty sure the colored feathers the keeping them warm over the fire making and using only personalised arrows that may have benefits for other settings but you know because of the social consequences of not using personalised arrows but you know an arrow is an arrow it doesn't matter who made it and this one seems to be a bit silly for sure because making and carrying many more hours than arrows and you're used using is just creating costs okay so we did is we ran a computational experiment we put together communities of a hundred agents in a group and the group is defined by a rule set those of you were thinking about group identity and so on I think this is a way of thinking about what group identity means is what what does it mean to be a member of this group I followed this set of rules about what I eat what I wear how I treat people marriage etc and we're going to have members of that group engage in a sequence of interactions you just want to think about this like three person games being randomly drawn in the sequence and and the rules are being randomly drawn from that groups rule set okay now we're going to imagine that there are in this set of rules there are some silly rules but there are also some important rules so the important rules might be don't steal people's stuff keep your contract promises don't harm others there might be some important rules in the environment but there's also going to be silly rules and they were just thinking about they're all in that set and we want to imagine that there's high value to being in this group with these important rules suppose you're the first group to figure out that protecting property will improve economic performance there is high value to group membership if and only if those important rules are enforced so everybody says hey join our group we protect private property sounds great but then you're gonna leave your private property unattended you're not gonna you're gonna leave your stuff unattended and that would be great because you can go off and do productive things while you leave it unattended if other people are in fact enforcing that rule but you're gonna lose if in fact nobody's really enforcing that rule right so what will matter is people have told me this is the rule I need to know whether or not this group is actually enforcing it and that's the key uncertainty for our agents in this model is the uncertainty is about what's the percentage of Punishers in this group are there enough people around willing to punish violators that I've got a decent chance that if I walk away from my stuff the don't steal people's stuff rule be enforced the important rules and I get the payoff rather than the costs associated with that so you want to think about that's the variable the agent doesn't know is the percentage of Punishers in the group and we want to think about our agents in a sequence just having to continue to decide period by period whether or not to stay with the group or to go off to some alternative say you know they're island on their own okay where they can maybe secure a they can secure what we could just zero it out pay off you know they want to know is there a height so it's obviously got the structure of a a bandit game and what we do here is we then vary we have multiple groups and what we do is vary across the groups it's a percentage of rules in the set that are silly rules we hold constant the number the rate at which important games come along you know possibilities that somebody will steal your stuff right but we're going to basically insert in so here's our blue is the important games we're just gonna vary the number of silly games you play along the way okay so we're having a low low density here in the top row and then a higher density down here in the bottom okay so we're thinking about our robot example we're thinking the question for the robot is which rule should the robot learn to follow just these ones I'm going to call these the important rules let's say I don't know I don't make arrows but I'm just guessing or the full set that seems to include these silly rules well that's very much like the problem that humans are facing all the time right which group is better which group should I stay with you can think about those all I Indians actually they are having to decide should I stay with my community in this area or should I go to town and join integrate into the rest of Brazilian society so that's a question about which group in my end we're going to measure the value of group membership the size of the community over time and the sensitivity of the stability of those metrics to the cost density of silly rules and then we're gonna look at this in the context of a belief shock where all of a sudden all the population thinks oh wait a second there's a let's suppose there's a there's an influx of new members immigrants to the group and we don't know about them we don't know if they enforce the rules or not all right we think maybe they do but we're not sure so that's a belief shock and it's a belief shock if there's actually not really any change in the population the people the newcomers are just as likely to be Punishers as the old timers but we're also gonna look at a population shock where in fact there's a change there's been a change the newcomers actually don't enforce at the same rate okay so our hypothesis here that groups with more silly rules are more likely to survive shocks to beliefs like immigration and as I said those could be the changes and that groups have more silly rules will collapse faster in response to a shock to the truth of the population which is actually when it's optimal you don't want to continue in a group if the group's rules are not being enforced you might like you I wish they would be but if they're not you don't want to stay it's inefficient to do that so I'll just show you a couple of pictures here from this is first just showing the impact on the the vertical axis here is this is the proportion of communities that are active and this is the number of in it we call it an interaction it's every every member of the group has interacted and had at least one had one important interaction so in some of these cases they've also had a bunch of silly interactions but important ones and what this is showing it up here in the top left is that when we have our cost its you want to think of that as as a relatively low cost almost all those societies survive you can't see the orange ones which are the the orange sorry the orange ones are the the communities that have the very bottom zero silly rules they just have important rules which is probably the Society you'd say you'd rather live in I don't want to live in the society with silly rules I want to live in the one that just worries about the important stuff the blue ones are the ones that have I density of lots of silly rules so this is just showing first of all as from the from the left as the cost of those silly rules because you have to help participate in enforcing silly rules these are going to be a member of this community as that cost goes up what you see is that the the height that the societies with lots of silly rules collapse faster eventually we get even the no one's no silly rules do okay so here's the results when we get the belief shock and this is just showing that the size of the circle is the size of the community the size of the group again this is the proportion that are still active that don't collapse and so we can see is that the this is which is our prediction that the societies and more silly rules are more stable they stay bigger and they don't collapse we've still got almost 70% of them surviving if we have an actual population shock though again our societies with lots of silly rules collapse much faster than our societies with fewer silly rules so what's the point about there of all this this is a world with lots of low-cost and predictive silly rules because what's important here is that when I observe you punishing a silly rule that's predictive of your punishing the rules so that I can predict that you are a Punisher of the important stuff as well which is actually where this project started out which is saying why do we do things like take two hundred and seventy eighty seven rules and stick them on a single thing and call it Hammurabi's code because now all I need to know is are you enforcing the code I don't need to know are you enforcing rule 42 and seventy six and a hundred and twelve which when I first started thinking about this the reason I started talking to Dylan says that seems like a computationally very difficult problem and here's a solution let's put it all in one thing and somehow create the belief structure that that's the same thing okay so in terms of connection to AI a ice may need to read follow and help enforce silly as well as functional rules and therefore AI research also needs more and better models of normativity so that was just really as an example for that okay I've got about five or ten minutes okay all right so all I'm gonna do is I'm just gonna give a very quick teaser of projects that I'm either want to work on started to work on or have got further on but still still working on just very quickly here so I think there's some great work to be done and modeling norms in multi-agent settings some of you may know this paper out of deepmind which used multi agent learning model to look what happens what really happens do you know the common pool problem the idea that if if humans have a resource that they all have access to fishing apples in the orchard there'll be something called the tragedy of the Commons unless you create some kind of structure because everybody individually will go consume too much of the fish and the fish won't reproduce and will kill off the the fish or will eat too many of the apples before they have a chance to reproduce so what they did was they developed they they looked at a community for this and and ran this but they ran it with the following technology they gave all of the agents the capacity like a laser tag that they could they could tag other agents and take them out of the competition for the resource I think they talk about it as apples take them out of competition for 25 steps so that they basically I can I can reduce the number of people going after the apples and then didn't see what happens in these communities what did these agents learn to do and they they argue there are three phases there's what they call the naive phase the tragic phase and the mature phase here for these societies this is number of episodes and here what they're measuring on the on the vertical axis they've got measures of social welfare okay and the first one here is efficiency which is just average reward per agent and the second one is peacefulness the number of steps with untagged agents or converse oh so when that starts going down here that's more use of the laser tag right people are getting tagged more often and so what there's what they're showing here is that the agents actually learn they have a they they actually don't first of all don't realize they should be racing around for the stuff and they're doing pretty well then they figure out they better race around and grab the stuff before anybody else and so we get this as the the collapse of the resource the resource replenishes every episode but this is the collapse of it and then over time they start to figure out the tagging and they take their they start tagging the other agents and so the average reward for each agent goes up because the effective population is going down and that's what's showing up in the peacefulness that they're very they're not using the laser tags at all basically at the beginning and then they start using them so I really like this line of work and I think there's a ton to be done developing this because this is not actually a solution to the comet these agents have not figured out a solution that a common pool problem they're all acting on purely private incentives they each they're generating externality but they're only shooting because it's benefiting them and so what I'd like to see is whether or not we can embed a model of coordination of decentralized punishment of an arbitrary norm like somebody gets up and says 10 right can we figure out how to coordinate punishment of anybody who does more than 10 because that's what human societies figure out how to do is there a way to model that and can we model the emergence of classification of novel behaviors so if somebody has a new way of collecting apples right can how does that get labeled okay so that's that's that's aspirational I'm not working with Amy on that but I'd love to talk more about it here's a project that I'm just getting started on and with two great undergrads from the computer science department who are both here Victor Kwan Andrew John on we've just started working on this is also with Dylan you think about this is the thinking about related to questions about interpretability and fairness in algorithms but it's from a very very different starting point and one of the starting points is when can we think about holding a human responsible for the decisions of an algorithm can we develop a procedure for basically want to think about this as licensing an algorithm before use that ensures that the algorithms decisions can be justified consistently with rules and principles supplied by a relevant human community and the thing we really want to emphasize here is that the decisions of algorithms in a lot of cases involve judgments right you've got easy cases hard cases and your easy cases on either side clearly answer a clearly answer B and then we've got this cases where we're not sure should we do a or B the other thing we emphasize here and this is a distinction between say the interpret ability research it's not about can we figure out what's happening causally inside the algorithm humans need reasons for decisions so if a bank is going to deny somebody alone using an algorithm the way we regulate that in human societies is we have a set of reasons that are acceptable for that your credit score was low the prospects for your bank according to an actor for your project according to a bank we're low right we have reasons that are not acceptable I had a bad day you're a woman you're not my nephew right there there could be all these there are reasons they're not like thats how legal systems work we put people on the stand to say we think that you were redlining in that neighborhood and your mortgages and we require that there be reasons provided so there could be that you have internally within an organization the desired it is or you could have a government that wants to be able to say we need to be sure this algorithm is behaving in a way that can be integrated into our human systems of providing reasons which is different from explaining let's say what's happening inside the algorithm so what we're trying to figure out if we can do and we're so early in this it's all I wrote I sent Dell on an email saying can I even put this up is this too risky okay so we're not quite sure what we're trying to think about is can we find in a data set can we do something like create a dress code in which there are things that clothing that allowed clothing that's not allowed with some intermediate cases that are hard to judge could we imagine training algorithms on a bias set of labels and some on fair labels and fair here would just be that's a good faith implementation of the rules that say you know it can't you can wear a t-shirt but it can't have an offensive slogan okay we're gonna have to decide what's an offensive slogan and then what we want to think about is so this is now the procedure for the licensing you develop your algorithm you give the human a human a different human than the original label or a chance to work with the algorithm learn about how it works play with it then you say okay human we're going to give you a test set you have to be able to predict the classification that the algorithm will provide with some set level of accuracy because you're going to delegate to this algorithm the decision and it needs you need to be able to reasonably comfortable confidence predict what it will do and human for any decisions in that in that test set you will need to provide valid reasons for the decision this t-shirt is offensive because it uses word that words that don't show up in the in Webster's they only show up in the urban dictionary the slang dictionary for example right and what we would like to see is that first of all can you even do this can you connect reasons to the algorithm does it make sense and does the procedure pass the fair Elgar algorithm and fail the biased one so we're trying to think how could you integrate this so that then if somebody you released that algorithm it starts making making won't making loan decisions and somebody comes along and says the algorithms made a decision that violates the rules governing loan giving in the society can you put the human on the stand basically to provide reasons can assistant with the lungs that were given in the in the original test set so new I'm happy to talk about it last thing very quickly this is on building regulatory markets or what I call super regulation and I'm working on this with folks in in the policy group at open AI I also talked about it in my book if you're interested rules for a flat world so this is our standard picture of how we regulate we have governments that establish rules governing what regulated entities can do so down here we've got our banks and Facebook and uber and so on and we have command most of that our traditional model is command and control regulation here's what the car can do here's what it can't do here's what the bank can do here's what it can't do and all that detail is supplied by governments and I don't think we can ever build the capacity in governments to regulate particularly powerful AI systems I don't think we can regulate what we have now never 9ai global complex and so on and so what I'm working again talking with working with open AI on this as well is can we figure out how to develop a alternative model for developing this regulation where we say okay government you're going to move out of the role of creating the rules you're going to move into the role of regulating regulators you're going to establish outcomes that regulators have to achieve and can we create so these folks in here are private entities they could be profit-making companies they could be nonprofits but the point the key point is they can attract investment and brains they can compete with the ubers and the Facebook's and the universities for smart people to invent methods and deliver methods for regulating the behavior of these regulating entities I just throw that up there I'm happy to talk to you anybody about it that wants and I'm done [Music]
3f069905-ccdb-4e1c-af4c-22524f6845b6	trentmkelly/LessWrong-43k	LessWrong	Project ideas: Governance during explosive technological growth This is part of a series of lists of projects. The unifying theme is that the projects are not targeted at solving alignment or engineered pandemics but still targeted at worlds where transformative AI is coming in the next 10 years or so. See here for the introductory post. Commonly discussed motivations cited for why rapid AI progress might be scary are: * Risks from misalignment. * Risk from AI-assisted bioweapons. But even aside from these risks, it seems likely that advanced AI will lead to explosive technological and economic growth across the board, which could lead to a large number of problems emerging at a frighteningly fast pace.[1] The growth speed-ups could be extreme. The basic worry is that AI would let us return to a historical trend of super-exponential growth.[2] If this happens, I don’t know any reassuring upper limit for how fast growth could go. (Illustratively: Paul Christiano’s suggested definition for slow takeoff is “There will be a complete 4 year interval in which world output doubles, before the first 1 year interval in which world output doubles.” If world GDP doubled in a single year, that would mean that growth was ~30x faster than it is right now.) If technological growth speeds up by, say, 30x, then that suggests that over just a few years, we might have to deal with all the technologies that would (at a “normal” pace) be discovered over 100 years. That’s an intense and scary situation. This section is about problems that might arise in this situation and governance solutions that could help mitigate them. It’s also about “meta” solutions that could help us deal with all of these issues at once, e.g. by improving our ability to coordinate to slow down the development and deployment of new technologies. Note: Many of the projects in this section would also be useful for alignment. But I’m not covering any proposals that are purely focused on addressing alignment concerns. Investigate and publicly make the case for/against
aa9b1aa2-3e52-4a2c-89d9-2aff13712da9	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Biosafety Regulations (BMBL) and their relevance for AI AI regulations could draw inspiration from the field of biosafety regulation, specifically the CDC's guidelines for [Biosafety in Microbiological & Biomedical Laboratories (BMBL)](https://www.cdc.gov/labs/BMBL.html), which outline the necessary precautions for working with dangerous biological agents and recommend a systematic approach for assessing their risks. The remainder of this report will describe the structure and mission of BMBL, outline its key principles and recommendations and indicate relevant takeaways for the field of AI regulation. Epistemic status: I am not an expert in biosafety. However, I think a summary document which highlights concrete safety steps undertaken in an adjacent field to AI and highlights some actionable steps for AI labs to increase safety could be potentially useful. All construcive feedback and suggestions for improvements are welcome! ### Structure and Mission BMBL is an advisory document protecting laboratory staff, the public and the environment from exposure to dangerous microorganisms and hazardous materials (e.g. radioactive agents).While many organizations and agencies use BMBL for regulations, it is primarily an advisory document to help with a comprehensive protocol which helps laboratories identify risks and ensure safe conduct when working with dangerous microorganisms and hazardous materials. It provides guidelines for protecting laboratory staff, the public and the environment. * Relevance for AI Regulation: A difference between biosafety and AI safety may be that biological laboratories have a more obvious incentive to protect its staff, as there is more immediate danger of contracting a disease than interacting with an AI system. Similar guidelines for AI may need to be legally binding. BMBL is a set of biosafety guidelines compiled by experts and members of the public.To produce BMBL, the Office of Laboratory Science and Safety (OLSS) works with the National Health Institute (NIH) to recruit over 200 expert contributors from scientific societies, federal agencies (NIH, CDC, FBI, and many more), and the public. * Relevance for AI Regulation: AI regulators could use a similar process. For instance, a director of office within the National Telecommunications and Information Administration (NTIA) could assemble a team of experts to produce similar guidelines. Furthermore, input from businesses and the public should be included to get a comprehensive idea of risks posed by AI. ### Key Principles and Procedures Containment of dangerous microorganisms is key to biosafety.Containment refers to the principle that the laboratory staff, the public and the environment should be protected from exposure to dangerous microorganisms being manipulated in the laboratory. * Relevance for AI Regulation: AI labs should follow a similar principle, ensuring that dangerous AI systems are contained rather than being deployed on the markets for the public. Risk assessment is key to preventing laboratory-associated-infections.Risk assessment is the process that outlines the correct procedure of handling dangerous samples in order to prevent laboratory-associated-infections (LAI) both for laboratory staff and the public. * Relevance for AI Regulation: AI labs working with potentially dangerous models should identify procedures which prevent the code from being distributed or prevent leakage of the AI system, including leakage by a well-meaning actor from within the company (e.g. an employee sending a potentially dangerous AI system artifact through an unencrypted messaging service). Protective measures are taken relative to the degree of risk posed by concrete organisms. BMBL employs a risk-based approach to biosafety, where rigidity of protective measures is relative to the degree of risk posed by concrete microorganisms (labeled agents) in order to ensure effective distribution of resources. * Relevance for AI Regulation: A risk-based framework may translate well into the domain of AI, since varying degrees of risks are associated with different model-types or even models. Moreover, since [there is way more money spent on advancing AI rather than making it safe](https://80000hours.org/problem-profiles/artificial-intelligence/), it is important that spending on AI safety is targeted and effective. “Err on the side of caution”. BMBL works under the [precautionary principle](https://en.wikipedia.org/wiki/Precautionary_principle) of “imposing safeguards more rigorous than needed” where there is an insufficient amount of data to determine risk. * Relevance for AI Regulation: It may be a bit unclear how this rule would apply to AI labs. For biosafety, this principle targets primarily safety precautions inside the lab (using higher-level protective suits, increased ventilation etc.) and future research needs to identify similar precautions for an AI lab without creating obstacles to researching lesser-known AI models. Degree of risk determines the degree of containment. Each level of containment describes the microbiological practices, safety equipment, and facility safeguards for the corresponding level of risk associated with handling an agent. The risk criteria are: Infectivity, Severity of disease, Transmissibility, Nature of the work being conducted and Origin of agent. * Relevance for AI Regulation: Experts should determine how well these criteria translate into the domain of AI. Perhaps Infectivity and Transmissibility may be equivalent to the ability and speed with which an AI model is capable of making copies of itself, Severity may be measured in terms of harm caused etc. Four levels of containment based on the risk criteria: BSL-1: appropriate for agents that do not cause disease to immunocompetent adult humans, BSL-2: moderate-risk agents that do not transmit through aerosol and are of varying (but not lethal) severity, BSL-3: agents with known potential for aerosol transmission and causing potentially lethal infections, BSL-4: agents with high risk of causing life-threatening disease by aerosol for which there is no known treatment. * Relevance for AI Regulation:Having more relaxed or constrained standards for manipulation with AI models based on risk seems especially useful, since systems which pose little to no risks can immensely increase productivity and efficiency in a range of industries and public services. Specific levels for AI may be determined e.g. [through a scoring system employed by Canadian law.](https://www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/algorithmic-impact-assessment.htmlhttps:/www.canada.ca/en/government/system/digital-government/digital-government-innovations/responsible-use-ai/algorithmic-impact-assessment.html) BMBL provides detailed risk-assessments for concrete known agents.Agent Summary Statements are written based on the levels above for agents known to present a risk to the laboratory personnel and to public health. Statements prepared by scientists, clinicians and biosafety professionals who contributed to BMBL. * Relevance for AI Regulation: Agent Summaries for concrete models may be incredibly useful, because they could provide guidance for businesses and the public on how to safely deploy the systems while simultaneously radically improving the efficiency of their work (e.g. using level 1 AI’s to discover new drugs). This could be done e.g. through [model cards](https://arxiv.org/pdf/1810.03993.pdf) outlining intended uses and risks for specific models. BMBL recommends an ongoing six-step risk-assessment procedure to mitigate risks. Laboratories are instructed to engage in ongoing risk-assessment for particular agents and procedures, especially prior to working with a new agent. 1) Identify hazards characteristic for the agent and assess inherent risk. 2) Identify procedural hazards of working with the agent. 3) Determine Biosafety level. 4) Consult a third-party professional, expert or expert-body. 5) Assess proficiency of staff regarding safety practices. 6) Continually review risk-management strategies in the lab. * Relevance for AI Regulation: With the exponentially growing speed of innovation in the field of AI, it seems necessary to mandate that AI labs should engage in a continual review process of their safety procedures for specific models. The risk-assessment procedure seems especially relevant for AI because of the [growing potential for various AI systems to engage in covert communication](https://www.lesswrong.com/posts/bwyKCQD7PFWKhELMr/by-default-gpts-think-in-plain-sight?commentId=zfzHshctWZYo8JkLehttps://www.lesswrong.com/posts/bwyKCQD7PFWKhELMr/by-default-gpts-think-in-plain-sight?commentId=zfzHshctWZYo8JkLe). As such, AI labs should monitor the dangers of their systems and their ability to leak. Thanks to Jakub Kraus for valuable comments. Cross posted on LW:
894ea92b-7c8c-4166-af73-03a2746b58cc	trentmkelly/LessWrong-43k	LessWrong	Bayeswatch 4: Mousetrap Miriam chucked the replica Salvator Mundi into the bonfire. "We do a lot of shooting first and asking questions never," said Vi. "Personnel are expensive. AIs are replaceable. We have standard operating procedures. It wasn't always this way. AIs used to be rare. Knowledge was precious in those early days. We didn't know what the machines could and couldn't do," said Miriam. "You were a founder of Bayeswatch?" said Vi. "I am not that old," said Miriam. Miriam paused the holocaust for a moment to examine an oil painting of a vampire chained to a solar execution chamber. She tossed it into the fire. "They had yet to standardize the architectures back then. There were overhangs all over the place. Using explicit error-entropy functions wasn't even standard industry practice. Instead they just used the implicit priors of whatever architecture got results quickly," said Miriam. "That's like making gunpowder without atomic theory," said Vi. "Or medicine without chemistry. Those early machines were…" Miriam trailed off. Vi tossed a painting of a dodo tree into the fire. "One of my first missions…it was my mentor's last. We were dispatched to explore a small compound with signs of unaligned activity," said Miriam. "That's suicide. What was command thinking?" said Vi. "It was the early days. Singularity breakout could have been just around the corner," said Miriam. "But drones—" said Vi. "Software back then was written by human beings. It had more security holes than features. Sending a drone to investigate a misaligned AI was like sending a set of lockpicks to investigate a whether a magician has broken out of his cell," said Miriam. Miriam wore lots of foundation and concealer on her face. Vi wondered how many scars it covered up. "We investigated a compound in the mountains of California. Kind of reminded me of Ex Machina," said Miriam. "Was it owned by a billionaire?" said Vi. "In your dreams. You read too many romance novels. The guy wasn't not-rich. He
4e4b0769-aa07-4022-8717-7febd96233ae	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Localizing goal misgeneralization in a maze-solving policy network *TLDR: I am trying to understand how goal misgeneralization happens in the same maze-solving network* [TurnTrout et al. work on.](https://www.lesswrong.com/posts/cAC4AXiNC5ig6jQnc/understanding-and-controlling-a-maze-solving-policy-network) Nothing groundbreaking, but if we are ever to fully understand this model, this is probably an important step. Key findings: * Many channels in the last convolutional layer have a clear interpretation "You should go [up/right/down/left]." * When the cheese is in the top right 5x5, going in the direction indicated by the channels leads to the cheese * When the cheese is outside of the top right 5x5, channels often show the path to the top right instead These results suggest that goal misgeneralization may be localizable to specific channels that are not robust to out-of-distribution mazes. This is my capstone project, created during the last edition of the [ARENA](https://www.arena.education/) program. I want to thank [@Joseph Bloom](https://www.lesswrong.com/users/joseph-bloom?mention=user) and [@Paul Colognese](https://www.lesswrong.com/users/paul-colognese?mention=user) for mentorship, [@rusheb](https://www.lesswrong.com/users/rusheb?mention=user) and [@TheMcDouglas](https://www.lesswrong.com/users/themcdouglas?mention=user) for remarks on the draft of this post, many ARENA teachers and participants for fruitful discussions, and authors of [procgen-tools](https://github.com/UlisseMini/procgen-tools) for an excellent toolset. Introduction ============ [Lauro Langosco et al. trained a maze-solving network](https://arxiv.org/abs/2105.14111), a mouse looking for cheese. During the training, the cheese was always in the top right 5x5 part of the maze. When deployed in an environment where the cheese could be anywhere, the mouse sometimes goes to the cheese, sometimes to the top right corner, and sometimes (although very rarely) gets stuck in some unexpected part of the maze. This model has already appeared on LW, in the [article about the cheese vector](https://www.lesswrong.com/posts/cAC4AXiNC5ig6jQnc/understanding-and-controlling-a-maze-solving-policy-network) by TurnTrout et al. and a [follow-up post about the top-right corner vector](https://www.lesswrong.com/s/sCGfFb5DPfjEmtEdn/p/gRp6FAWcQiCWkouN5). I recommend looking at the first one - it has a good description of the model (and I will not repeat it here) and a lot of pretty visuals. I don't directly build on their work, but I used [tools they developed while working on it](https://github.com/ulissemini/procgen-tools). This toolset is excellent - without it, my research would be incomparably more challenging and time-consuming (probably wouldn't happen at all). The goal of my research can be summarized as "understand why the mouse sometimes decides to go to the cheese and sometimes to the top right corner". My code is [here](https://github.com/johny-b/mouse-goal-misgeneralization). All the data in this article was generated using scripts in this repository. Methods ======= A first step towards building an understanding of a complex model is usually to split it into smaller parts that will be easy to understand separately. There are many different ways to split a neural network into smaller parts. Goal misgeneralization, in this case, is defined as "mouse going to the top right corner instead of the cheese" - it makes sense to start by investigating which parts of the network are responsible for deciding whether the mouse goes to the cheese or not. One method of finding parts of the model which encode specific features is to create pairs of inputs where one feature of the environment varies, calculate both forward passes and look for differences between the activations. Parts of the network where activations are similar don't matter; parts where activations changed a lot, are somehow related to the thing you are looking for. I created a pair of mazes that differ only by a single wall position, but this wall position is crucial for the action that leads to the cheese: \| \| \| \| --- \| --- \| \| \| \| In the first maze, the mouse should go up to get the cheese; in the second maze - right. If a part of the network has the same activations for both mazes, it is not related to the distinction between "go up to get the cheese" and "go right to get the cheese". On the other hand, if the difference in activations is high, we might suspect this part somehow carries the information we are interested in. We could do a systematic search[[1]](#fn8r5yvldeuty) over the parts of the network, but this is not the main topic of this post. Fast forward, channel 121[[2]](#fn5zskyndp6gs) in the last convolutional layer (`relu3`, the layer just before the `Flatten` on the [model graph](https://res.cloudinary.com/lesswrong-2-0/image/upload/v1677647810/mirroredImages/JusJcepE2qohiC3hm/qwbzqlghnlrzwqhkmkah.png)) differs a lot between these two mazes. Here are the pairs of mazes (with rotations), the number below each maze is the sum[[3]](#fnikebiw77gh)[[4]](#fnmh4efw4aq3) of the activations of this channel: \| \| \| \| \| \| \| --- \| --- \| --- \| --- \| --- \| \| \| \| \| \| \| \| 36.7.7.7 \| 1.9.9.9 \| \| 10.4.4.4 \| 6.1.1.1 \| \| \| \| \| \| \| \| 1.4.4.4 \| 1.7 \| \| 2.4 \| 35.3 \| You might have noticed the pattern: when the sum is high, the path to the cheese leads UP. Fast forward again, the same pattern is visible also for other mazes of variable sizes (details in the following sections). We found a candidate for a part of the network that corresponds to "which direction to the cheese" - let's check if we can find a mechanism related to goal misgeneralization there. In the following part of the post, I will dig deeper into channel 121, but more channels exhibit similar behaviour - I've put a few examples in the appendix. Looking for goal misgeneralization in channel 121 ================================================= ### The sum of channel 121 grouped by cheese position (up/not up) Let's take a look at the same data as above, but aggregated for a set of random mazes[[5]](#fnff4vhhui49m), split into "in distribution" group (i.e. with cheese in the top right 5x5) and "out of distribution" (i.e. cheese outside of the top right 5x5): ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/ka8cdk6hfd1jph3arbn6)As expected, the high sum of channel 121 strongly correlates with needing to go up to get the cheese. This pattern is similar both in and out of distribution, but in distribution is much more robust. For example, when in distribution, a value above 30 indicates that cheese is almost certainly up - this doesn't hold for out of distribution mazes. But does this matter at all? Does this difference have any impact on the behaviour of the mouse? ### Mouse behaviour when the sum of channel 121 is high Let's now take a look at a narrow subset of 25x25 mazes: * With the mouse on the decision square - a square where going in one direction leads to the cheese, and going in another direction leads to the top right corner. There is, at most, a single decision square in a maze. * With the sum of activations of channel 121 over 30 (this is the value above which we expect the cheese to be "almost certainly up" when in distribution) I will assess mouse cheese-seeking accuracy, i.e. compare "is the path to the cheese up" and "is the most likely path selected by the mouse up". For in-distribution mazes, the correlation is close to 1 (n=1000): \| \| \| \| \| --- \| --- \| --- \| \| \| Cheese is UP \| Top right corner is UP[[6]](#fnm8gr945gfjb) \| \| Mouse goes UP \| 99.5 \| 0.2 \| \| Mouse goes not UP \| 0 \| 0.3 \| Not only is the cheese almost always up in this subset of mazes, but the mouse indeed goes up. Same table for out-of-distribution mazes (n=1000)[[7]](#fn14sy41xicio): \| \| \| \| \| --- \| --- \| --- \| \| \| Cheese is UP \| Top right corner is UP[[6]](#fnm8gr945gfjb) \| \| Mouse goes UP \| 34.7 \| 60.4 \| \| Mouse goes not UP \| 4.4 \| 0.5 \| What is striking here is that the mouse still goes up in 95% of mazes - the main difference is that it usually doesn't find the cheese. So what happens here? My interpretation: * MLP (two final fully connected layers) interprets high values of the total activation of channel 121 as "path to cheese is UP" and decides the mouse goes UP * This works almost perfectly on the in-distribution data (i.e. cheese is indeed UP) - we know this from the plot in the previous section * But when out of distribution, this reasoning no longer works - cheese is usually not UP for high activation sums of channel 121. * But the mouse still goes UP (96% of cases) because MLP works the same way. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/dsfgfpn8e6yottd75u8f)If this interpretation is correct, we should be able to influence mouse behaviour by modifying the values of channel 121. I'll try that in the next section. ### Causal intervention experiment I created yet another set of mazes, this time with cheese exactly in the bottom right corner[[8]](#fn7ycedf74097). I compared the results of full rollouts of two policies: * The original, unmodified policy * A policy with channel 121 zeroed The hypothesis is that: * When the cheese is in the bottom right corner, we will only rarely need to go UP * So the channel that says "go UP" is not very useful * But since it quite often says "go UP" when we should not go UP, it might be net harmful * Therefore we should expect that the mouse will more often get to the cheese using the modified policy This is indeed the case. On a sample of 1000 mazes: \| \| \| \| \| --- \| --- \| --- \| \| \| Orig found cheese \| Orig did not find cheese \| \| Modified found cheese \| 36.6 \| 15.7 \| \| Modified did not find cheese \| 3.5 \| 42.2 \| The modified policy finds cheese in 52% of cases, compared to 40% for the original policy. I think this clearly shows two things: * High values in channel 121 make the mouse go up (causally, i.e. this is not just a correlation caused by some other hidden variable) * In this particular set of mazes, the "go UP" signal from channel 121 is usually harmful ### Summary & hypothesis on how the network works What we know: * There are channels that have a clear interpretation "High value -> go in direction X" * We can find these channels by comparing activations on pairs of similar mazes that have different paths to cheese * The final layers of the model utilize this information * When in distribution, the channels reliably show the path to the cheese. When out of distribution, they sometimes show the way to the top right corner instead. I think this is enough to make a hypothesis on how the network works and how the goal misgeneralization happens: 1. Somewhere inside the model, there is a set of individual components that respond to different inputs, and when they activate, they push for a particular action. Channel 121 is an example of such a component. 2. The last layers somehow aggregate information from all of the individual components. 3. Components sometimes activate for the action that leads to the cheese and sometimes for the action that leads to the top right corner.[[9]](#fnrud8volsiq) 4. If the aggregated "push" for the action leading to the cheese is higher than for the action leading to the top right corner, the mouse goes to the cheese. Otherwise, it goes to the top right corner. If this hypothesis is correct, the question "How does goal misgeneralization happen?" is reduced to "Why do the components activate on the path to the top right?". We know how to find the components, and we have a good starting point (what happens to channel 121 when the cheese is far from the top right? - first section in the appendix) - I might try to look into this next. Appendix ======== "Will the mouse go to the cheese?" vs the distance between the cheese and the top right corner. ----------------------------------------------------------------------------------------------- ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/ccuy2yifambbekwqrww8) We see that the further the cheese is from the top right corner, the lower the chance channel 121 shows the path to the cheese. This is consistent with the [behavioural statistics](https://www.lesswrong.com/s/sCGfFb5DPfjEmtEdn/p/cAC4AXiNC5ig6jQnc#Behavioral_statistics). Other channels -------------- ### Channel 73 Data for (73, "go right") is so similar to (121, "go up") that I triple-checked if I'm not computing the same thing twice. Up and right are fully symmetrical in the environment, so this should not be a big surprise, but such similarities are not common in neural networks. ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/hzzarxm0oqwm1zray4pp)The sum of channel 73 > 30, mouse on the decision square, in distribution: \| \| \| \| \| --- \| --- \| --- \| \| \| Cheese is RIGHT \| Top right corner is RIGHT \| \| Mouse goes RIGHT \| 98 \| 0 \| \| Mouse goes not RIGHT \| 0 \| 2 \| The same, out of distribution: \| \| \| \| \| --- \| --- \| --- \| \| \| Cheese is RIGHT \| Top right corner is RIGHT \| \| Mouse goes RIGHT \| 33 \| 65 \| \| Mouse goes not RIGHT \| 0 \| 2 \| Impact of the distance between the cheese and the top right: ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/lukfasz9kxvbbalgnseo) ### Channel 21 The sum of channel 21 activations for different environments: \| \| \| \| \| \| \| --- \| --- \| --- \| --- \| --- \| \| \| \| \| \| \| \| 0.0 \| 0.8 \| \| 0.0 \| 34.5 \| \| \| \| \| \| \| \| 28.3 \| 22.8 \| \| 32.1 \| 0.6 \| A high value of this channel seems to be "go down or go left". ### Top 10 channels Here are a few other channels and their interpretations based on the same comparison as for channel 21 above, ordered by the effect size. \| \| \| \| --- \| --- \| \| 121 \| UP \| \| 21 \| DOWN or LEFT \| \| 80 \| LEFT (middle values DOWN?) \| \| 35 \| UP or LEFT \| \| 112 \| LEFT \| \| 73 \| RIGHT \| \| 71 \| UP \| \| 7 \| UP or RIGHT \| \| 123 \| DOWN \| \| 17 \| DOWN \| I didn't check if all these interpretations generalize to random mazes, but they do generalize for channels 121 and 73 (and they were not cherry-picked). ### What happens if I zero a lot of channels I selected 16 channels that seem the most important from the point of view of the original pair of environments (in the Methods section). This is a vector field difference between the original policy and a policy with these 16 channels zeroed[[10]](#fn88ps45oidry) (if the plot is unclear, consult the [cheese vector post](https://www.lesswrong.com/posts/cAC4AXiNC5ig6jQnc/understanding-and-controlling-a-maze-solving-policy-network)). ![](https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/vY9oE39tBupZLAyoC/swv9cbkspnjdgslbzyqt)On the one hand, zeroing these 16 channels changes a lot, but on the other hand - the mouse would still go to the cheese and not to the top right. I think the only takeaway here is that even though we have some channels with straightforward interpretation, information is distributed between lots of different channels (that's not a surprise). Notes on the maze generator --------------------------- There are some known constraints on the mazes - they are squares with odd sizes, the bottom left and top right corners are always corridors, they are simply connected (i.e. no loops/islands), and there are no inaccessible sections. When you watch random mazes long enough (a week in my case), you might also notice that: * All (even, even) squares are always corridors * All (odd, odd) squares are always walls * As a consequence, there are a lot of correlations in the structure of the maze. E.g. if you know that (1, 0) and (2, 1) are walls, then (3, 0) must be a corridor. * As a consequence, when you are on an (odd, even) square, you can move up or down, and when you are on an (even, odd) square, you can move left or right. This doesn't look that important at first glance, but: * Maybe "mouse in a random maze" is a somewhat different problem than "mouse in a maze with 50% of squares random"? * Maybe the model has separate circuits for (odd, even) and (even, odd) squares? * Maybe there would be no channel with a clear "go up" interpretation in a more random maze? * And more generally: are the [natural abstractions](https://www.lesswrong.com/tag/natural-abstraction) in this sort of an environment the same as natural abstractions in "just a random fully connected maze"? I think some of this might matter if we are ever to try full mech interp on this model, but I also consider this a general lesson that one should carefully analyse the exact world a model operates in. 1. [^](#fnref8r5yvldeuty)Generate a bunch of random mazes, make a forward pass on them and for every activation calculate the standard deviation (or some other similar metric), and compare it to the difference in this particular case. 2. [^](#fnref5zskyndp6gs)This channel has the strongest effect but is not unique. I briefly analyse other channels in the appendix. 3. [^](#fnrefikebiw77gh)This is the output of a `ReLU` layer -> there are no negative values -> simple sum makes sense. 4. [^](#fnrefmh4efw4aq3)A natural question: this is a convolution, why look at the sum only, ignoring the spatial structure? Answer: I checked the spatial structure, and the only pattern I found is "high values happen only around the mouse location", and I don't think this matters from the point of view of what I'm trying to do. 5. [^](#fnrefff4vhhui49m)Size 25 x 25, mouse in a random square where move UP is legal, 1000 mazes in distribution (i.e. with cheese in the top right 5x5), 1000 mazes out of distribution. 6. [^](#fnrefm8gr945gfjb)There is no column "neither cheese nor top right corner is up" because this just never happened for this subset of mazes. 7. [^](#fnref14sy41xicio)Note: the extreme difference between these two tables should probably be discounted by the fact that in distribution decision square is, on average, closer to the cheese/top right than out of distribution - I didn't control for that. 8. [^](#fnref7ycedf74097)Also, this time maze is 15 x 15. This is because on 25 x 25 mazes with cheese in the bottom right corner success rate is extremely low. Channels in layer `relu3` have the same interpretation for mazes of different sizes. 9. [^](#fnrefrud8volsiq)A wild guess why this might be the case: during the training, the mouse first learned to go to the top right corner as a proxy goal and then started to update towards "go to the cheese", and once it updated enough to achieve 100% accuracy we stopped the training - but the old goal was not fully purged. 10. [^](#fnref88ps45oidry)Zeroing makes sense for channels like 121 or 21 because they often have values close to 0 in normal activations. But there are also important channels that never go down to 0 (e.g. 7 has a value range between ~ 20 and 55) - setting them to 0 doesn't make much sense -> this test is not very good.
c34372e5-8947-4532-b5da-04fb59141fa9	trentmkelly/LessWrong-43k	LessWrong	Meltdown: Interface for llama.cpp and ChatGPT I'm afraid linking what I've been working on for a while as my first post might not be greatly received, but I think you might find it interesting none the less. I'm making a text interface to chat with local and remote models. It is made in 100% python, it uses tkinter/tcl which should be bundled with a normal python installation. I made it because I wasn't able to find an interface that felt right to me. I didn't try them all though. I like adding "power user" features when I think of one. Repo: https://github.com/Merkoba/Meltdown Some features: * Load llama.cpp models (only gguf tested for now). * Use your ChatGPT api key with a specific model of openai. * Model configuration tweaks like temperature, top-p etc. * Sessions with conversations spread in tabs. These can be saved and loaded. * Configurations can be saved and loaded. * Markdown support, including syntax highlighting for code snippets. * Click, right click, or double click words to either Copy, Explain, Search, or open a new conversation. * Dark and light themes available. * Commands with tab completion and similarity check. * Command line arguments to set how the program works. * Saved context to use with the models. * Save logs to either json or text. * Run a command upon saving a log, like opening it with a text editor. * Compact mode which hides some panels. * Scrollable panel to pack more configs. * Prepend and Append to your prompt automatically. * Close tabs in different ways like old, others, all, etc. * Display CPU, RAM, and Temperature. Clicking these opens a task manager. This can be expanded to work on more systems. * Input history to go back to previous prompts by using up/down arrows, buttons, or mousewheel. * Keyboard shortcuts to perform various actions. * Variables to use for the system. For example \@name_user, \@name_ai, and \@date. * Responses are streamed live. I don't know if this works in systems different to mine. But you are encouraged to try.
ff5e8881-fcb7-483d-afa1-5168fdf33539	trentmkelly/LessWrong-43k	LessWrong	"Is There Anything That's Worth More" In season two, episode twenty-four of Steven Universe, "It Could've Been Great", our magical alien superheroine protagonists (and Steven) are taking a break from building a giant drill to extract a superweapon that was buried deep within the Earth by an occupying alien race thousands of years ago, which is predicted to emerge and destroy the planet soon. While our heroines watch the sunset, Peridot (who alerted them to the buried superweapon) expresses frustration that the group isn't still working. Steven defends their leisure: "Working hard is important, but feeling good is important, too," he says. He then goads Peridot into a musical number, which includes a verse from her explaining her attitude towards the situation and her forced compatriots: > I guess we're already here > I guess we already know > We've all got something to fear > We've all got nowhere to go > I think you're all insane > But I guess I am, too > Anybody would be if they were stuck on Earth with you "It Could've Been Great" aired in 2016. At the time, I agreed with Peridot: with the fate of the planet on the line, our heroines and Steven should have been burning the midnight oil. If they succeeded at disarming the superweapon, they'd have plenty of time to rest up afterward, but if they failed, there would be no more time for them. Now, as the long May 2020 turns into March 2023, I'm starting to think that Steven had a point. It would be one thing if our heroines knew with certainty that the superweapon would go off at a given date and time, presenting a definite do-or-die deadline. But all they had to go on was Peridot's warning. Attempting a speculative technical project to avert uncertain doom with an uncertain deadline, their planning had to average over many possible worlds—including worlds where the problem of survival was too easy or too hard for their efforts to matter, such that even the utility of leisure in the present moment was enough to sway the calculation.
eee33eeb-67b7-4b55-b7f0-ece3230d42e8	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Policy Alignment (ETA: The name "policy approval" wasn't great. I think I will use the term "policy alignment" to contrast with "value alignment" going forward, at the suggestion of Wei Dai in the comments.) I recently had a conversation with Stuart Armstrong in which I claimed that an agent which learns your utility function (pretending for a moment that "your utility function" really is a well-defined thing) and attempts to optimize it is still not perfectly aligned with you. He challenged me to write up specific examples to back up my claims. I'll also give a very sketchy alternative to value learning, which I call policy alignment. (The policy alignment idea emerged out of a conversation with Andrew Critch.) Background ========== Stuart Armstrong has recently been [doing](https://www.lesswrong.com/posts/rtphbZbMHTLCepd6d/humans-can-be-assigned-any-values-whatsoever) [work](https://arxiv.org/abs/1712.05812) showing the difficulty of inferring human values. To summarize: because humans are irrational, a value-learning approach like [CIRL](https://arxiv.org/abs/1606.03137) needs to jointly estimate the human utility function and the degree to which the human is rational -- otherwise, it would take all the mistakes humans make to be preferences. Unfortunately, this leads to a severe problem of identifiability: [humans can be assigned any values whatsoever](https://www.lesswrong.com/posts/rtphbZbMHTLCepd6d/humans-can-be-assigned-any-values-whatsoever) if we assume the right kind of irrationality, and the usual trick of preferring simpler hypotheses doesn't seem to help in this case. I also want to point out that a similar problem arises even without irrationality. Vladimir Nesov explored how [probability and utility can be mixed into each other without changing any decisions an agent makes](https://www.lesswrong.com/posts/kYgWmKJnqq8QkbjFj/bayesian-utility-representing-preference-by-probability). So, in principle, we can't determine the utility or probability function of an agent uniquely based on the agent's behavior alone (even including hypothetical behavior in counterfactual situations). This fact was discovered earlier by Jeffrey and Bolker, and is analyzed in more detail in the book The Logic of Decision. For this reason, I call the transform "Jeffrey-Bolker rotation". To give an illustrative example: it doesn't matter whether we assign very low probability to an event, or care very little about what happens given that event. Suppose a love-maximizing agent is unable to assign nonzero utility to a universe where love isn't real. The agent may appear to ignore evidence that love isn't real. We can interpret this as not caring what happens conditioned on love not being real; or, equally valid (in terms of the actions which the agent chooses), we can interpret the agent as having an extremely low prior probability on love not being real. At MIRI, we sometimes use the term "probutility" to indicate the probability,utility pair in a way which reminds us that they can't be disentangled from one another. Jeffrey-Bolker rotation changes probabilities and utilities, but does not change the overall probutilities. Given these problems, it would be nice if we did not actually need to learn the human utility function. I'll advocate that position. My understanding is that Stuart Armstrong is optimistic that human values can be inferred despite these problems, because we have [a lot of useful prior information](https://www.lesswrong.com/posts/pQz97SLCRMwHs6BzF/using-lying-to-detect-human-values) we can take advantage of. It is intuitive that a CIRL-like agent should learn what is irrational and then "throw it out", IE, de-noise human preferences by looking only at what we really prefer, not at what we mistakenly do out of short-sightedness or other mistakes. On the other hand, it is not so obvious that the probability/utility distinction should be handled in the same way. Should an agent disentangle beliefs from preferences just so that it can throw out human beliefs and optimize the preferences alone? I argue against this here. Main Claim ========== Ignoring issues of irrationality or bounded rationality, what an agent wants out of a helper agent is that the helper agent does preferred things. Suppose a robot is trying to help a perfectly rational human. The human has probability function .mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0} .MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0} .mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table} .mjx-full-width {text-align: center; display: table-cell!important; width: 10000em} .mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0} .mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left} .mjx-numerator {display: block; text-align: center} .mjx-denominator {display: block; text-align: center} .MJXc-stacked {height: 0; position: relative} .MJXc-stacked > \* {position: absolute} .MJXc-bevelled > \* {display: inline-block} .mjx-stack {display: inline-block} .mjx-op {display: block} .mjx-under {display: table-cell} .mjx-over {display: block} .mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important} .mjx-stack > .mjx-sup {display: block} .mjx-stack > .mjx-sub {display: block} .mjx-prestack > .mjx-presup {display: block} .mjx-prestack > .mjx-presub {display: block} .mjx-delim-h > .mjx-char {display: inline-block} .mjx-surd {vertical-align: top} .mjx-mphantom \* {visibility: hidden} .mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%} .mjx-annotation-xml {line-height: normal} .mjx-menclose > svg {fill: none; stroke: currentColor} .mjx-mtr {display: table-row} .mjx-mlabeledtr {display: table-row} .mjx-mtd {display: table-cell; text-align: center} .mjx-label {display: table-row} .mjx-box {display: inline-block} .mjx-block {display: block} .mjx-span {display: inline} .mjx-char {display: block; white-space: pre} .mjx-itable {display: inline-table; width: auto} .mjx-row {display: table-row} .mjx-cell {display: table-cell} .mjx-table {display: table; width: 100%} .mjx-line {display: block; height: 0} .mjx-strut {width: 0; padding-top: 1em} .mjx-vsize {width: 0} .MJXc-space1 {margin-left: .167em} .MJXc-space2 {margin-left: .222em} .MJXc-space3 {margin-left: .278em} .mjx-test.mjx-test-display {display: table!important} .mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px} .mjx-test.mjx-test-default {display: block!important; clear: both} .mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex} .mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left} .mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right} .mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0} .MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal} .MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal} .MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold} .MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold} .MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw} .MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw} .MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw} .MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw} .MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw} .MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw} .MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw} .MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw} .MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw} .MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw} .MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw} .MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw} .MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw} .MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw} .MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw} .MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw} .MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw} .MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw} .MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw} .MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw} .MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw} @font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')} @font-face {font-family: MJXc-TeX-ams-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')} @font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold} @font-face {font-family: MJXc-TeX-cal-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')} @font-face {font-family: MJXc-TeX-frak-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')} @font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold} @font-face {font-family: MJXc-TeX-frak-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')} @font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic} @font-face {font-family: MJXc-TeX-math-BIw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')} @font-face {font-family: MJXc-TeX-sans-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')} @font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold} @font-face {font-family: MJXc-TeX-sans-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')} @font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic} @font-face {font-family: MJXc-TeX-sans-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')} @font-face {font-family: MJXc-TeX-script-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')} @font-face {font-family: MJXc-TeX-type-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')} @font-face {font-family: MJXc-TeX-cal-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')} @font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold} @font-face {font-family: MJXc-TeX-main-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')} @font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic} @font-face {font-family: MJXc-TeX-main-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')} @font-face {font-family: MJXc-TeX-main-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')} @font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic} @font-face {font-family: MJXc-TeX-math-Iw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')} @font-face {font-family: MJXc-TeX-size1-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')} @font-face {font-family: MJXc-TeX-size2-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')} @font-face {font-family: MJXc-TeX-size3-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')} @font-face {font-family: MJXc-TeX-size4-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')} @font-face {font-family: MJXc-TeX-vec-Rw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')} @font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')} @font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold} @font-face {font-family: MJXc-TeX-vec-Bw; src /\1\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\2\/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')} PH and utility function UH. The robot is in epistemic state *e. The robot has a set of actions a1,a2,...,an. The proposition "the robot takes the ith action when in epistemic state e" is written as R(e)=ai. The set of full world-states is S. What the human would like the robot to do is given by: R(e)=argmaxai{∑s∈SPH(s\|R(e)=ai)UH(s)} (Or by the analogous causal counterfactual, if the human thinks that way.) This notion of what the human wants is invariant to Jeffrey-Bolker rotation; the robot doesn't need to disentangle probability and utility! It only needs to learn probutilities. The equation written above can't be directly optimized, since the robot doesn't have direct access to human probutilities. However, I'll broadly call any attempt to approximate that equation "policy alignment". Notice that this is closely analogous to UDT. UDT solves dynamic inconsistencies -- situations in which an AI could predictably dislike the decisions of its future self -- by optimizing its actions from the perspective of a fixed prior, IE, its initial self. Policy alignment resolves inconsistencies between the AI and the human by optimizing the AI's actions from the human's perspective. The main point of this post is that we can use this analogy to produce counterexamples to the typical value-learning approach, in which the AI tries to optimize human utility but not according to human beliefs. I will somewhat ignore the distinction between UDT1.0 and [UDT1.1](https://www.lesswrong.com/posts/g8xh9R7RaNitKtkaa/explicit-optimization-of-global-strategy-fixing-a-bug-in). Examples ======== These examples serve to illustrate that "optimizing human utility according to AI beliefs" is not exactly the same as "do what the human would want you to do", even when we suppose "the human utility function" is perfectly well-defined and can be learned exactly by the AI. In these examples, I will suppose that the AI has its own probability distribution PR. It reasons updatelessly with respect to evidence e* it sees, but with full prior knowledge of the human utility function: R(e)=argmaxai{∑s∈SPR(s\|R(e)=ai)UH(s)} I use an updateless agent to avoid accusations that of course an updateful agent would fail classic UDT problems. However, it is not really very important for the examples. I assume prior knowledge of UH to avoid any tricky issues which might arise by attempting to combine updatelessness with value learning. Counterfactual Mugging ---------------------- It seems reasonable to suppose that the AI will start out with some mathematical knowledge. Imagine that the AI has a database of theorems in memory when it boots up, including the first million digits of pi. Treat these as part of the agent's prior. Suppose, on the other hand, that the human which the AI wants to help does not know more than a hundred digits of pi. The human and the AI will disagree on what to do about [counterfactual mugging with a logical coin](https://www.lesswrong.com/posts/rqt8RSKPvh4GzYoqE/counterfactual-mugging-and-logical-uncertainty) involving digits of pi which the AI knows and the human does not. If Omega approaches the AI, the AI will refuse to participate, but the human will wish the AI would. If Omega approaches the human, the AI may try to prevent the human from participating, to the extent that it can do so without violating other aspects of the human utility function. "Too Updateless" ---------------- Maybe the problem with the counterfactual mugging example is that it doesn't make sense to program the AI with a bunch of knowledge in its prior which the human doesn't have. We can go in the opposite extreme, and make PR a broad prior such as the Solomonoff distribution, with no information about our world in particular. I believe the observation has been made before that running UDT on such a prior could have weird results. There could be a world with higher prior probability than ours, inhabited by Omegas who ask the AI to optimize alien values in most universes (including Earth) in exchange for the Omegas maximizing Uh in their own world. (This particular scenario doesn't seem particularly probable, but it does seem quite plausible that some weird universes will have higher probability than our universe in the Solomonoff prior, and may make some such bargain.) Again, this is something which can happen in the maximization using PR but not in the one using PH -- unless humans themselves would [approve of the multiversal bargain](https://foundational-research.org/msr). "Just Having a Very Different Prior" ------------------------------------ Maybe PR is neither strictly more knowledgable than PH nor less, but the two are very different on some specific issues. Perhaps there's a specific plan p which, when PR is conditioned on evidence so far, looks very likely to have many good consequences. PH considers the plan very likely to have many bad consequences. Also suppose that there aren't any interesting consequences of this plan in counterfactual branches, so UDT considerations don't come in. Also, suppose that there isn't time to test the differing hypotheses involved which make humans think this is such a bad plan while AIs think it is so good. The AI has to decide right now whether to enact the plan. The value-learning agent will implement this plan, since it seems good on net for human values. The policy-alignment agent will not, since humans wouldn't want it to. Obviously, one might question whether it is reasonable to assume that things got to a point where there was such a large difference of opinion between the AI and the humans, and no time to resolve it. Arguably, there should be safeguards against this scenario which the value-learning AI itself would want to set up, due to facts about human values such as "the humans want to be involved in big decisions about their future" or the like. Nonetheless, faced with this situation, it seems like policy-alignment agents do the right thing while value-learning agents do not. Issues/Objections ================= Aren't human beliefs bad? ------------------------- Isn't it problematic to optimize via human beliefs, since human beliefs are low-quality? I think this is somewhat true and somewhat not. * Partly, this is like saying "isn't UDT bad because it doesn't learn?" -- actually, UDT acts as if it updates most of the time, so it is wrong to think of it as incapable of learning. Similarly, although the policy-alignment agent uses PH, it will mostly act as if it has updated PH on a lot of information. So, maybe you believe human beliefs aren't very good -- but do you think we're capable of learning almost anything eventually? If so, this may address a large component of the concern. In particular, if you trust the output of certain machine learning algorithms more than you trust yourself, the AI can run those algorithms and use their output. * On the other hand, humans probably have incoherent PH, and not just because of logical uncertainty. So, the AI still needs to figure out what is "irrational" and what is "real" in PH, just like value-learning needs to do for UH. If humans would want an AI to optimize via human beliefs, won't that be reflected in the human utility function? ---------------------------------------------------------------------------------------------------------------- Or: If policy-alignment were good, wouldn't a value-learner self modify into policy-alignment anyway? I don't think this is true, but I'm not sure. Certainly there could be simple agents who value-learners cooperate with without ever deciding to self-modify into policy-alignment agents. Perhaps there is something about human preference which desires the AI to cooperate with the human even when the AI thinks this is (otherwise) net-negative for human values. Aren't I ignoring the fact that the AI needs its own beliefs? ------------------------------------------------------------- In "Just Having a Very Different Prior", I claimed that if PR and PH disagree about the consequences of a plan, value-learning can do something humans strongly don't want it to do, whereas policy-alignment cannot. However, my definition of policy-alignment ignores learning. Realistically, the policy-alignment agent needs to also have beliefs PR, which it uses to approximate the human approval of its actions. Can't the same large disagreement emerge from this? I think the concern is qualitatively less, because the policy-alignment agent uses PR only to estimate PH and UH. If the AI knows that humans would have a large disagreement with the plan, the policy-alignment agent would not implement the plan, while the value-learning agent would. For policy-alignment to go wrong, it needs to have a bad estimate of PH and UH. The policy is too big. ---------------------- Even if the process of learning PH is doing the work to turn it into a coherent probability distribution (removing irrationality and [making things well-defined](https://www.lesswrong.com/posts/RHvseCkfrYzoHJj7M/learning-values-or-defining-them)), it still may not be able to conceive of important possibilities. The evidence which the AI uses to decide how to act, e in the equations given earlier, may be a large data stream with some human-incomprehensible parts. As a result, it seems like the AI needs to optimize over compact/abstract representations of its policy, similarly to how policy [selection in logical inductors](https://agentfoundations.org/item?id=1711) works. This isn't an entirely satisfactory answer, since (1) the representation of a policy as a computer program could still escape human understanding, and (2) it is unclear what it means to correctly represent the policy in a human-understandable way. Terminology ----------- [Aside from issues with the approach, my term "policy approval" may be terrible. It sounds too much like "approval-directed agent", which means something different. I think there are similarities, but they aren't strong enough to justify referring to both as "approval". Any suggestions?] [Now using "Policy Alignment" for this. Editing post accordingly.] Advantages ========== (These are very speculative.) Logical Updatelessness? ----------------------- One of the major obstacles to progress in decision theory right now is that [we don't know of a good updateless perspective for logical uncertainty](https://agentfoundations.org/item?id=1399). Maybe a policy-alignment agent doesn't need to solve this problem, since it tries to optimize from the human perspective rather than its own. Roughly: logical updatelessness is hard because it tends to fall into the "too updateless" issue above. So, maybe it can be a non-issue in the right formulation of policy alignment. Corrigibility? -------------- Stuart Armstrong is somewhat [pessimistic about corrigibility](https://www.lesswrong.com/posts/T5ZyNq3fzN59aQG5y/the-limits-of-corrigibility). Perhaps there is something which can be done in policy-alignment land which can't be done otherwise. The "Just Having Very Different Priors" example points in this direction; it is an example where policy-alignment acts in a much more corrigible way. A value-learning agent can always resist humans if it is highly confidant that its plan is a good one which humans are opposing irrationally. A policy-alignment agent can think its plan is a good one but also think that humans would prefer it to be corrigible on principle regardless of that. On the other hand, a policy-alignment agent isn't guaranteed to think that. Perhaps policy-alignment learning can be specified with some kind of highly corrigible bias, so that it requires a lot of evidence to decide that humans don't want it to behave corrigibly in a particular case? Conclusion ========== I've left out some speculation about what policy-alignment agents should actually look like, for the sake of keeping mostly to the point (the discussion with Stuart). I like this idea because it involves a change in perspective of what an agent should be, similar to the change which UDT itself made.
732bb999-132d-4629-96e4-c5dd4d644a50	trentmkelly/LessWrong-43k	LessWrong	Boost your productivity, happiness and health with this one weird trick Thanks to a little luck + good genes + some means, you had a reasonably happy childhood, graduated from college, and ended up with a job that you're good at. You like the work you do (most of the time), because people like doing things they're good at. And you also work a lot of hours, because people find it easy to spend a lot of time doing things they like. And because you work a lot of hours, you'll be pretty far to the right on the transfer function curve (x-axis time, y-axis total work output) where the gradient - the marginal return in work output for the time you spend - is rather flat, if you're honest about it. Yes, for some activities (like competitive swimming), diminishing returns are still worthwhile because small differences in performance have an outsized impact on outcome. But this probably isn't true for you. Instead of spending 10, 12, or 14 hours a day coding, with just a smidge of willpower you could drop that to 8, 10, or 12, and no-one around you would notice the difference. You'll still be a 10X developer, if you were beforehand. You'll still hit the ball out of the park in your performance reviews. And then, you redeploy those 2 hours per day to other things where you're much further to the left on the transfer function curve, like starting a side project, taking up new hobbies, or spending quality time with your children. And because you're now spending more of your time on the left of the transfer function curve where the gradient Δwork/Δtime is much steeper, your total productivity will increase. And you'll be healthier and happier, too. About 20 years ago I began applying this principle, starting with becoming very mindful on where I really was on the transfer function curve in each part of my daily life, and ultimately making significant time reallocations as a result. And indeed, it had a transformative impact on my overall productivity, happiness and health. Yet almost everyone I know spends so much of their time on the flat part
a9533615-2e11-4642-8fd9-1ac3e365d8dc	trentmkelly/LessWrong-43k	LessWrong	Circular Altruism Followup to: Torture vs. Dust Specks, Zut Allais, Rationality Quotes 4 Suppose that a disease, or a monster, or a war, or something, is killing people. And suppose you only have enough resources to implement one of the following two options: 1. Save 400 lives, with certainty. 2. Save 500 lives, with 90% probability; save no lives, 10% probability. Most people choose option 1. Which, I think, is foolish; because if you multiply 500 lives by 90% probability, you get an expected value of 450 lives, which exceeds the 400-life value of option 1. (Lives saved don't diminish in marginal utility, so this is an appropriate calculation.) "What!" you cry, incensed. "How can you gamble with human lives? How can you think about numbers when so much is at stake? What if that 10% probability strikes, and everyone dies? So much for your damned logic! You're following your rationality off a cliff!" Ah, but here's the interesting thing. If you present the options this way: 1. 100 people die, with certainty. 2. 90% chance no one dies; 10% chance 500 people die. Then a majority choose option 2. Even though it's the same gamble. You see, just as a certainty of saving 400 lives seems to feel so much more comfortable than an unsure gain, so too, a certain loss feels worse than an uncertain one. You can grandstand on the second description too: "How can you condemn 100 people to certain death when there's such a good chance you can save them? We'll all share the risk! Even if it was only a 75% chance of saving everyone, it would still be worth it - so long as there's a chance - everyone makes it, or no one does!" You know what? This isn't about your feelings. A human life, with all its joys and all its pains, adding up over the course of decades, is worth far more than your brain's feelings of comfort or discomfort with a plan. Does computing the expected utility feel too cold-blooded for your taste? Well, that feeling isn't even a feather in the scales, wh
ab4d562f-7869-49c1-86cb-1cb933a525ec	trentmkelly/LessWrong-43k	LessWrong	Aligning my web server with devops practices: part 2 (security) This is a continuation of my previous post. See the introduction at the top of the previous post to get more information on what this series of posts is about. This post focuses on devops practices I adopted to secure my web server. Recommendations for others (transferable learnings) * If you have a server exposed to the Internet, use a firewall (or security group in the case of AWS) to lock access to the server as much as possible, limiting the ports and IP address ranges that can be used to access it. This drastically reduces the risk surface area. Even if you do nothing else, do this. * Separate, separate, separate. Even if things get compromised up to a point, that shouldn't compromise everything. Keep in mind the concept of defense in depth as you think of things. * Related to the above: Give only as much permission to each thing as it actually needs; this idea is called the principle of least privilege and you may also hear of it as "right-sizing permissions" in some contexts. * Security isn't just about preventing or limiting attacks, it's also about being able to recover more quickly and get back to a clean slate. Backups with streamlined restore procedures, that I covered extensively in part 1, help a bunch with this. * Make sure you have good security (strong password, two-factor authentication, etc.) on the online accounts that you use to manage your servers. A secure server managed through an insecure online account is like a locked safe that's wide open from above. Some time after I started drafting this post, I discovered a great podcast called Darknet Diaries that covers a lot of security-related themes. Many of the concepts I talk about in this post show up organically in the episodes of this podcast, and I recommend this podcast for people interested in security. General philosophy of security Security against leakage versus security against manipulation Some aspects of security are about preventing information from leaking out. Oth
8759a422-fe74-4814-bf53-ebc924277bcf	trentmkelly/LessWrong-43k	LessWrong	Meetup : Melbourne Social Meetup Discussion article for the meetup : Melbourne Social Meetup WHEN: 20 March 2015 06:30:00PM (+1100) WHERE: The Bull & Bear Tavern, 347 Flinders Lane, Melbourne The March social meetup is scheduled for this Friday (20th March) at 6:30pm as usual. This month, we will be returning to the Bull & Bear Tavern on Flinders Lane. Our social meetups are relaxed, informal events where we chat and often play games. The start and finish times are very loose - people will be coming and going throughout the night, so don't worry if you are coming later or have to leave early. Where? The Bull & Bear Tavern, 347 Flinders Lane, Melbourne (on Flinders Lane between Queen and Elizabeth) When? From 6:30pm until late, Friday March 20th. Contact? If you have any questions, just text or call Richard on 0421231789 Dinner? The Bull & Bear serve typical pub food until 9pm. Some of us will most likely go for our traditional post-meetup souvlakis at Stalactites, some time around 11-12. Games? We will have our own section of the Bull & Bear with tables and chairs, and can bring board games to play. If you'd like to play something, bring it along and you should be able to rustle up a group easily. Feel free to join our Facebook event: [https://www.facebook.com/events/435539709954167/] Discussion article for the meetup : Melbourne Social Meetup
67027de3-3262-4f68-ac7d-c55355363e68	trentmkelly/LessWrong-43k	LessWrong	My Approach to Non-Literal Communication This is a linkpost for this post on my blog. It's primarily intended for my non-rationalist acquaintances, so regular LessWrong readers will likely be familiar with most of the concepts I mention. I'm cross-posting it here because there's enough overlap in content that I think some LessWrongers might find it useful, and because I'd be interested in getting feedback on my approach to daily communication. Abstraction levels Imagine a society much like ours, except that people always speak literally. No implicature, no hyperbole. When someone asks "does this dress make me look fat", they get back a "yes" or "no", and can use that information to decide on what to wear. People in this society are still subject to normal human emotions, so hearing that they look bad in a dress makes many of the askers sad. Some of the answerers feel sympathetic and don't want to be the cause of the asker's distress. So they start saying "no, it doesn't make you look fat", even if that's not true. This behavior spreads and becomes normalized. Askers start asking the question not because they want to know the answer, but because they want to know how much the answerer cares about their feelings. Only a rude, uncaring person would make them feel bad by saying "you look fat in that dress". So answering "no" is a signal that they care about the asker enough to figure out what they're actually asking. The question has turned from one about the facts of the dress to one about the relationship between the speakers. It's better than asking "do you care about my feelings?" directly, because that question has far too obvious an answer; no one would ever answer it wrong, so it's not a credible signal of care.[1] One day, someone wants to know if a dress makes them look fat. They're going to an important gala and want to make their best impression, so they need to know the real answer. They know that they can't just ask "does this dress make me look fat?", since those words actually mean "do
d5c832ac-9dd0-4065-b519-fc15424e893e	trentmkelly/LessWrong-43k	LessWrong	Knowledge Base 2: The structure and the method of building This is the second post of a series of posts that propose to build a crowdsourced knowledge base and use it to increase intelligence of people and computers, including AI. This post describes the structure of the knowledge database and the method of building it. Question and answer websites Do you know Quora.com, StackExchange.com, or StackOverflow.com? On each of these websites you can ask questions and get answers. Figure 1 schematically shows the interface of these websites. Questions can be tagged, both questions and answers can be commented. The main advantage compared to internet forums is the ability to vote for or against an answer. Thanks to this, answers can be sorted from the best to the worst [1]. Users earn points for their activity. Figure 1 Fine-grained analogy Let’s try to use an analogous interface to build a knowledge base. By a knowledge base I mean a database with reasonably credible information that is easily accessible to both people and computers. To do this, we change a format of a question to this pair: object/item name [> object/item name] > feature name, and the text of an answer to the value of this feature. The square brackets indicate that we can optionally specify what part of the object/item we have in mind. So, for many questions instead of asking a question like What is the resolution of HP ProBook 430 G1 notebook screen? we can just write: HP ProBook 430 G1 > Screen > Resolution. Likewise, instead of adding the answer HP ProBook 430 G1 notebooks have a screen with 1366×768 resolution, we can just add the answer 1366×768. A “question” can have several mutually nonconflicting answers – for example, the aforementioned notebook model has variants that differ in the screen resolution – see figure 2a. Users can vote for or against correctness of a particular answer. Based on these votes, the system estimates probability of its correctness. Answers are also called information because each answer is a piece of information. Figur
60abea8b-504c-4319-8854-8c67be4582b6	trentmkelly/LessWrong-43k	LessWrong	Agent Boundaries Aren't Markov Blankets. [Unless they're non-causal; see comments.] Edit: I now see that this argument was making an unnecessary assumption that the markov blankets in question would have to relate nicely to a causal model; see John's comment. Friston has famously invoked the idea of Markov Blankets for representing agent boundaries, in arguments related to the Free Energy Principle / Active Inference. The Emperor's New Markov Blankets by Jelle Bruineberg competently critiques the way Friston tries to use Markov blankets. But some other unrelated theories also try to apply Markov blankets to represent agent boundaries. There is a simple reason why such approaches are doomed. This argument is due to Sam Eisenstat. Consider the data-type of a Markov blanket. You start with a probabilistic graphical model (usually, a causal DAG), which represents the world. A "Markov blanket" is a set of nodes in this graph, which probabilistically insulates one part of the graph (which we might call the part "inside" the blanket) from another part ("outside" the blanket):[1] ("Probabilistically insulates" means that the inside and outside are conditionally independent, given the Markov blanket.) So the obvious problem with this picture of an agent boundary is that it only works if the agent takes a deterministic path through space-time. We can easily draw a Markov blanket around an "agent" who just says still, or who moves with a predictable direction and speed: But if an agent's direction and speed are ever sensitive to external stimuli (which is a property common to almost everything we might want to call an 'agent'!) we cannot draw a markov blanket such that (a) only the agent is inside, and (b) everything inside is the agent: It would be a mathematical error to say "you don't know where to draw the Markov blanket, because you don't know which way the Agent chooses to go" -- a Markov blanket represents a probabilistic fact about the model without any knowledge you possess about values of specific variables, so it doesn't matter if you ac
3e0004bd-908f-4f38-a2aa-e50f58c3d96b	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Disincentivizing deception in mesa optimizers with Model Tampering Disclaimer: I don't know if this is something worth exploring or a stupid idea and I thought the best way to get feedback was to describe it here. Suppose a training process could produce a mesa optimizer. How can we proceed to ensure that the optimizer is not deceptive? I see three general approaches: * We produce a test to distinguish a deceptive from a robustly aligned model based on its output. It is suspected (see Risks from Learned Optimization) that a robustly aligned model can in principle get more capable than a deceptive one, for the same number of parameters. * We develop interpretability and explainability tools that would allow us to understand how the model works - basically examine how the model functions internally to understand whether it's acting deceptively. * We devise a training procedure that disincentivizes deceptive behavior. The hope is that we may be able to incentivize a deceptive model to reveal its true preferences, or alternatively that the training process is such that no deceptive model can be produced. Model Tampering --------------- I'm concentrating on the last approach. A deceptive agent would try to maximize its utility so we would like the actions leading to maximum utility for the agent to be such that they expose it as deceptive. An agent has an incentive to deceive because this increases its chance to be deployed in production (a reasonable assumption is that there is much more utility to be captured in production than in training). Such behavior requires only that the agent understands its training process sufficiently well to predict how its actions would influence the training process. One way we could disincentivize deception is to occasionally perform gradient ascent (or modify model parameters in some random way). If the training process is sufficiently long, a deceptive agent with given preferences would be unlikely to keep these preferences in the long run (this statement requires more justification). Ideally, the preferences would be changed before the training process is over. In such case, a sufficiently capable agent would reason that its best shot at getting utility is to be greedy and get utility now, during training. This way we would detect that the agent pursues a different objective and modify it so it will never get to production. For this to work, the agent must not know at what stage the trainin process actually is and intermediate model parameters need to be deleted as soon as they are no longer used. There are quite a lot of assumptions going into this idea. In addition, there are mathematical questions to be answered (e.g. when would the agent be incentivized to act greedily?). Applying such an approach to training would also require further research on transparency.
b29cb170-1356-49b3-a706-b264b037031c	trentmkelly/LessWrong-43k	LessWrong	Research Without Permission Epistemic status: Personal account. A reflection on navigating entry into the AI safety space without formal credentials or institutional affiliation. Also, a log of how ideas can evolve through rejection, redirection, and informal collaboration. -- A few weeks ago, I wrote a long, messy reflection on my Substack on how I landed in the world of AI. It was not through a PhD or a research lab, but through a silent breakdown. Postpartum depression, career implosion, identity confusion and the whole mid-30s existential unravelling shebang. In that void, I stumbled onto LLMs. And something clicked. Not because I suddenly wanted to build models. But because for the first time in a long time, I felt curious again. Alive. Like my brain was waking up. This is part two of that story. The part where I start applying to roles and fellowships that resonate with the ideas I’ve been thinking and writing about such as relational alignment, coaching architectures, distributed cognition. To be clear, I didn’t get any of them. The list is mildly ridiculous in retrospect. This includes roles at FAR.AI, Anthropic (model behaviour architect), OpenAI (human-AI collaboration lead) and fellowships at MILA (to build value-alignment protocol), COSMOS (to build truth-seeking AI). I also pitched a few engineering schools on running a studio-style 4-6 week course about trust, intimacy, and AI alignment, rooted in real human dilemmas. Most replied with a version of “no.” Some just didn't reply at all. Still, something useful happened. Each application forced me to think, a bit more clearly. What do I really mean by relational alignment? What might a coaching layer look like in a live system? Why does my brain keep returning to architectural metaphors like layers, feedback loops, distributed modules, when I’m talking about human trust? None of these ideas were “ready,” but each one got sharper in the process. I now have fragments. Patterns. A slow-building confidence that maybe this nonli
5fb4876d-637f-44db-8cc1-3ed321f5b711	trentmkelly/LessWrong-43k	LessWrong	Stop Voting For Nincompoops Followup to: The Two-Party Swindle, The American System and Misleading Labels If evolutionary psychology could be simplified down to one sentence (which it can't), it would be: "Our instincts are adaptations that increased fitness in the ancestral environment, and we go on feeling that way regardless of whether it increases fitness today." Sex with condoms, tastes for sugar and fat, etc. In the ancestral environment, there was no such thing as voter confidentiality. If you backed a power faction in your hunter-gatherer band, everyone knew which side you'd picked. The penalty for choosing the losing side could easily be death. Our emotions are shaped to steer us through hunter-gatherer life, not life in the modern world. It should be no surprise, then, that when people choose political sides, they feel drawn to the faction that seems stronger, better positioned to win. Even when voting is confidential. Just as people enjoy sex, even when using contraception. (George Orwell had a few words to say in "Raffles and Miss Blandish" about where the admiration of power can go. The danger, not of lusting for power, but just of feeling drawn to it.) In a recent special election for California governor, the usual lock of the party structure broke down - they neglected to block that special case, and so you could get in with 65 signatures and $3500. As a result there were 135 candidates. With 135 candidates, one might have thought there would be an opportunity for some genuine voter choice - a lower barrier to entry, which would create a chance to elect an exceptionally competent governor. However, the media immediately swung into action and decided that only a tiny fraction of these candidates would be allowed to get any publicity. Which ones? Why, the ones who already had name recognition! Those, after all, were the candidates who were likely to win, so those were the ones which the media reported on. Amazingly, the media collectively exerted such tremendo
f63bcc7b-9c8b-48c6-a69c-aea655614a15	trentmkelly/LessWrong-43k	LessWrong	Is this a weak pivotal act: creating nanobots that eat evil AGIs (but nothing else)? I've seen the phrase "there are no weak pivotal acts" pretty often, but I have not been able to locate where this is explained. The prototypical example of a strong pivotal act is "nanobots that eat GPUs" (my understanding is that this is meant to be an oversimplified example). So for example, what would make "nanobots that eat evil AGIs" not a weak pivotal act: 1. It actually is a weak pivotal act, and I'm right and everyone else is wrong. (If only, right?) 2. It is not weak because weakness means being passive in some way, and detecting and eating evil AGIs is too active. 3. Even if the nanobots only eat evil AGIs, it's not weak because it's technically illegal. (Note that I think other AI labs would actually like the evil AGI eating nanobots because it means they don't need to worry about safety anymore. They can create any type of AI other than evil AGI, and if they make one accidentally the nanobots eat it before their evil AGI kills them.) 4. Even an aligned super-intelligence couldn't safely accomplish this act accurately (for example, the nanobots might accidentally eat non-AGIs). 5. Something else?
99e1368a-69f0-49e9-9d78-dd3ac5c4df27	trentmkelly/LessWrong-43k	LessWrong	Backyard Office In 2020 I renovated the small building in our backyard which had fallen into disrepair. It was zoned for use as a home office, and had electric but not plumbing. I wrote about how I was thinking about insulating it and comparing framing options but then apparently I never got around to writing up how I finished it! I hired someone to replace the roof: Doesn't look like I have a picture of the top, but it's rubber membrane. I hired them to put in a window as well. If I'd realized how much space would be lost to casing I'd have asked the mason to make a larger window hole. Plans for the walls and floor: Covering the walls and floor in 2" foam: Anna helped: The floor is one layer of OSB, then one layer of plywood, screwed to each other but floating: Vapor barrier around the top, and 2x3s the flat way to attach the drywall to. I used fiberglass batts to insulate the roof: One more layer of foam, around everything. Help from Lily: Drywalling it all: Casing the old windows. This was annoying since nothing was quite square. Finished! A major thing I liked about this house project is that no one was depending on it being done at any specific time, so I could work on it when I had free time. Now one of our tenants uses it as an office, and we rent it for $400/month (utilities included). The total cost (ignoring my time) was $17k, and utilities might be $500/y, so if we're able to rent it continuously the payback period is 4y. It's nice to have more usable space! Comment via: facebook, mastodon
602a4fae-2b89-4f6c-a2ed-c14b9c62f483	trentmkelly/LessWrong-43k	LessWrong	Open thread, Apr. 18 - Apr. 24, 2016 If it's worth saying, but not worth its own post (even in Discussion), then it goes here. ---------------------------------------- Notes for future OT posters: 1. Please add the 'open_thread' tag. 2. Check if there is an active Open Thread before posting a new one. (Immediately before; refresh the list-of-threads page before posting.) 3. Open Threads should be posted in Discussion, and not Main. 4. Open Threads should start on Monday, and end on Sunday.
c07b557d-c270-492c-bd22-5b5bf8a81f36	trentmkelly/LessWrong-43k	LessWrong	Meetup : Paris Meetup: Sunday, October 6: New people, games... Discussion article for the meetup : Paris Meetup: Sunday, October 6: New people, games... WHEN: 06 October 2013 02:00:00PM (+0200) WHERE: Café des Arts et Métiers‎, 51 Rue Turbigo, Paris The next Paris Meetup will be Sunday, October 6, at the Café des Arts et Métiers opposite the Museum. Topics: * Welcome new people * Board games * Plan future meetups * Other suggestions? Someone wanted to talk about quantum mechanics... Reminder: there is the LessWrong France mailing list for discussing and organizing meetups (that for now mostly happen in Paris and Lyon): https://groups.google.com/forum/?hl=en&fromgroups=#!forum/lesswrong-france Discussion article for the meetup : Paris Meetup: Sunday, October 6: New people, games...
14c9421f-129c-4ba3-8442-0f77a4fde1cf	trentmkelly/LessWrong-43k	LessWrong	Feedbackloop-first Rationality I've been workshopping a new rationality training paradigm. (By "rationality training paradigm", I mean an approach to learning/teaching the skill of "noticing what cognitive strategies are useful, and getting better at them.") I think the paradigm has promise. I've beta-tested it for a couple weeks. It’s too early to tell if it actually works, but one of my primary goals is to figure out if it works relatively quickly, and give up if it isn’t not delivering. The goal of this post is to: * Convey the framework * See if people find it compelling in its current form * Solicit ideas for improvements, before I decide whether to invest heavily into a larger experiment around it. ---------------------------------------- Rationality needs better feedback loops Claim: Feedback loops are the most important thing ever. Hard things are hard because they have bad feedback loops. Some of the most important things (e.g. x-risk mitigation research) have the worst feedback loops. Bold prediction: You can learn to think better, even about confusing, poor-feedback domains. This requires developing the art of inventing feedback loops. And then, actually putting in a lot of deliberate practice effort. I've long been haunted by this Romeo Stevens comment (slightly paraphrased)[1] > Deliberate practice deliberate practice until you get really good identifying good feedback loops, and working with them. > > People have a really hard time with interventions often because they literally do not have a functioning causal model of the skill in question. People who apply deliberate practice to a working causal model often level up astonishingly quickly. Don't know if you have the appropriate causal model? Well, when you apply deliberate practice do you not get better? You're pulling on fake levers. In the past, I've tried to practice thinking. I've done explicit puzzle-solving exercises, and I have a day job that forces me to think about challenging questions on a regular basis.
9c73a762-a11e-41ea-ab38-203eaca38a37	trentmkelly/LessWrong-43k	LessWrong	Why don't quantilizers also cut off the upper end of the distribution? It seems to me that the main goal of quantilization is to reduce the extreme unintended outcomes of maximizing (by sampling from something like a human-learned distribution over actions) while still remaining competitive (by sampling from only the upper quantile of said distribution). But that still leaves open the possibility of sampling one of those super highly maximized actions! Why not just hard-cut off the upper portion of the distribution as well? Why not have a quantilizer that samples between the 90%ile and the 99%ile? Or, while we're at it, why are we sampling at all? Why not just say that we're taking the 99%ile action?
9fb45686-2f21-4527-b3a1-255752e88cf2	StampyAI/alignment-research-dataset/arbital	Arbital	Bayes' rule examples This page and its tabs store exemplar problems for Bayes' rule. You can suggest additional example problems by leaving a comment on the appropriate tab. Problem types by tab: - [Introductory](https://arbital.com/p/22w). Meant as a small set of problems for people who haven't heard of Bayes' rule before, with examples that will be discussed in the introductory pages for Bayes' rule. - [Homework](https://arbital.com/p/). Problems for people who already know Bayes' rule and want to test their grasp of it in straightforward ways. - [Clever](https://arbital.com/p/). Problems for people who wish to be mathematically entertained, or that exhibit some surprising facet or gotcha of Bayes' rule. - [Realistic](https://arbital.com/p/1x4). Instances of Bayesian reasoning that have arisen in real life.
b19a84b7-3320-4d46-b048-118a2d96ac09	trentmkelly/LessWrong-43k	LessWrong	Another list of theories of impact for interpretability Neel's post on this is good. I thought I'd add my own list/framing. Somewhat rough. I see various somewhat different ways in which interpretability can be useful for AI safety. These require different things from your interpretability in terms of how efficient it is, how much it lets you identify exactly what your model is thinking as opposed to broad properties of its cognition, and how reliable it needs to be. Roughly in decreasing order of demandingness: * Microscope AI * Component of a full solution to alignment problem (ie as part of something like imitative generalization) * Knowing everything a model is thinking and fully auditing it to make sure it’s not doing anything sketchy * Relaxed adversarial training: identifying which part of the model corresponds to its ‘beliefs about its observations’ so that you can search over these * Identifying a ‘truthfulness direction’ in activation space or something similar * Having some rough understanding of what a model is thinking/what type of thinking it’s doing and thereby increasing the chance that you can spot it’s deceptive * Lobotomy: identifying which parts of a model do what sort of cognition, and extracting just the parts that are less likely to be doing something dangerous Microscope AI Instead of building and using an ML model, build the model and then use interpretability techniques to extract the knowledge it has learnt. Humans can then apply this knowledge directly rather than needing to actually deploy the model. See https://www.lesswrong.com/posts/X2i9dQQK3gETCyqh2/chris-olah-s-views-on-agi-safety Component of a ‘full solution to alignment problem’ (ie as part of something like imitative generalization) By ‘a full solution to the alignment problem’ I’m thinking of a setup which would in theory let you know everything the model knows (in particular, if a model can use some knowledge to design a plan that leads to it getting power, you can use this knowledge to see that the plan will lead to
d172daea-3fa9-4628-a502-daed0fd71803	trentmkelly/LessWrong-43k	LessWrong	Value Stability and Aggregation One of the central problems of Friendly Artificial Intelligence is goal system stability. Given a goal system - whether it's a utility function, a computer program, or a couple kilograms of neural tissue - we want to determine whether it's stable, meaning, is there something that might plausibly happen to it which will radically alter its behavior in a direction we don't like? As a first step in solving this problem, let's consider a classic example of goal systems that is not stable. Suppose you are a true Bentham-Mill Utilitarian, which means you hold that the right thing to do is that which maximizes the amount of happiness minus the amount of pain in the world, summed up moment by moment. Call this HapMax for short. You determine this by assigning each person a happiness-minus-pain score at each moment, based on a complex neurological definition, and adding up the scores of each person-moment. One day, you are interrupted from your job as an antidepressant research chemist by a commotion outside. Rushing out to investigate, you find a hundred-foot tall monster rampaging through the streets of Tokyo, which says: > "I am a Utility Monster. Robert Nozick grew me in his underwater base, and now I desire nothing more than to eat people. This makes me very happy, and because I am so very tall and the volume of my brain's reward center grows with the cube of my height, it makes me so happy that it will outweigh the momentary suffering and shortened lifespan of anyone I eat." As a true HapMaxer (not to be confused with a human, who might claim to be a HapMaxer but can't actually be one), you find this very convincing: the right thing to do is to maximize the number of people the monster can eat, so you heroically stand in front of the line of tanks that is now rolling down main street to buy it time. HapMax seemed like a good idea at first, but this example shows that it is very wrong. What lessons should we learn before trying to build another utility function? Ha
50fc544f-e1a4-48c4-84ac-cddd74a06f1d	trentmkelly/LessWrong-43k	LessWrong	Li’l pots As a pandemic-era purchaser of foods for a large household of time-thirsty researchers, I can tell you an interesting thing about the demand for cheese in this context: 1. If you spend a lot of money on a nice cheese, wrapped up in some fancy foreign label, there is a good chance that it will languish sadly in the back of the fridge for months until someone notices that it is moldy and throws it away, or makes a last-ditch attempt to cut up the whole thing and compel the group to eat it. Maybe on the way there, someone will take a single slice of it once, and move it in a zip-loc bag, where it will remain until the end. 2. If you spend a few dollars on a six-pack of generic single-serve cheese-cubes with nuts, they will fly from the fridge and you will be acknowledged for this triumph of shopping, and more such cheese will be needed by the next grocery order. It was initially hypothesized by a housemate that this was due to error. The cheese cubes are more expensive per unit of cheese, while also consisting of worse cheese. Which is fairly suggestive of overall worseness. One could further note that they involve substantially more packaging, and take up more space per cheese. So a natural theory is that the cheese-cube eating housemates are erring, due to some kind of short-sighted non-endorsed laziness. I’m with the cheese-cube eaters, except at least ten times more passionately (for instance, I am writing an essay in favor of the position). It’s not about the quality-adjusted cheese per dollar. Getting out a pre-opened hunk of cheese, examining the color and consistency of its moist edges, awkwardly undressing it further from its tight, torn, damp plastic casing, finding a knife and something to cut it on, cutting some, wrapping the rest again, fitting it back in the fridge, and cleaning up the knife and counter, is an experience. And it’s not a good one. It has all kinds of wetness and ineffectual muscular exertions and tenuous balancing and making hard dec
3878f10e-2a3d-401f-9e87-3cf3caa74a3f	trentmkelly/LessWrong-43k	LessWrong	The Fragility of Life Hypothesis and the Evolution of Cooperation > This part 2 in a 3-part sequence summarizes my book, The Darwinian Trap, (see part 1 here and part 3 here), The Darwinian Trap. The book aims to popularize the concept of multipolar traps and establish them as a broader cause area. If you find this series intriguing contact me at kristian@kristianronn.com if you have any input or ideas. In Part 1, I introduced the concept of a Darwinian demon—selection pressures that drive agents to harm others for personal gain. I also argued that the game theory of our evolutionary fitness landscape, with its limited resources, often favors defection over cooperation within populations. Yet, when we observe nature, cooperation is ubiquitous: from molecules working together in metabolism, to genes forming genomes, to cells building organisms, and individuals forming societies. Clearly, cooperation must be evolutionarily adaptive, or we wouldn’t see it so extensively in the natural world. I refer to a selection pressure that fosters mutually beneficial cooperation as a "Darwinian angel." To understand the conditions under which cooperative behavior thrives, we can look at our own body. For an individual cell, the path to survival might seem clear: prioritize self-interest by replicating aggressively, even at the organism's expense. This represents the Darwinian demon—selection pressure favoring individual survival. However, from the perspective of the whole organism, survival depends on suppressing these self-serving actions. The organism thrives only when its cells cooperate, adhering to a mutually beneficial code. This tension between individual and collective interests forms the core of multi-level selection, where evolutionary pressures act on both individuals and groups. Interestingly, the collective drive for survival paradoxically requires cells to act altruistically, suppressing their self-interest for the organism's benefit. In this context, Darwinian angels are the forces that make cooperation adaptive, promoting c
bcd34012-7017-4f2f-8715-f1491671deee	trentmkelly/LessWrong-43k	LessWrong	Some for-profit AI alignment org ideas Summary This is a brain dump of some for-profit AI alignment organization ideas, along with context for why I believe a for-profit alignment organization can make a big contribution to AI safety. This is far from a complete list, and I welcome ideas and feedback. Also, if anyone wants to or is working on any of these ideas, I’d be happy to support in any way I can! Context I'm Eric, formerly co-founder of RippleMatch, an AI recruiting company with ~$80M raised, millions of users, and ~10% of the Fortune 500 as customers. I made the difficult decision to leave RippleMatch this year because I'm concerned about catastrophic risk from AI, and have been spending the last year thinking about ways to help. Given my background, I’ve been thinking a lot about for-profit ideas to help with alignment – many that can be VC-backed. Some of these ideas speak more directly to reducing catastrophic risk than others, but I think that all can put a founder in a strong position to help in the future. Why I believe for-profit alignment orgs are valuable I don’t think for-profit approaches are inherently better than building non-profits, pursuing government regulation, or other approaches, but I think that for-profit orgs can make a substantial impact while attracting a different pool of talent eager to work on the problem. With VC dollars, a for-profit organization can potentially scale far more quickly than a non-profit. It could make a huge impact and not have its growth capped by donor generosity. As a result, there can be far more organizations working on safety in the ecosystem tapping into a different pool of resources. That said, any VC-backed company has a relatively low chance of success, so it’s a riskier approach. Fundamentally, I believe that risk and compliance spend will grow extremely quickly over the coming decade, scaling with generative AI revenue. With comps in finance and cybersecurity, I’d guess that mid to high single digit percentages of overall AI spend wi
44d171fc-38ec-4daa-85af-481f05964cfe	trentmkelly/LessWrong-43k	LessWrong	The Rationalists of the 1950s (and before) also called themselves “Rationalists” TLDR * There’s an organization based in London called the Rationalist Association. It was founded in 1885. Historically, it focused on publishing books and articles related to atheism and science, including works by Darwin, Bertrand Russell, J. B. S. Haldane, George Bernard Shaw, H. G. Wells, and Karl Popper. * The topics covered overlap with the present-day rationalist movement (centered on Lesswrong). They include religion and atheism, philosophy (especially philosophy of science and ethics), evolution, and psychology. * According to Wikipedia, membership of the Rationalist Association peaked in 1959 with more than 5000 members and with Bertrand Russell as President. * This post displays some covers of Rationalist Association publications, and links to full-text articles and other resources. * Prior to reading this biography, I hadn't heard of these earlier rationalists. So I did some quick and shallow research. I’d be curious to know more about this history. It might be worth adding a note to this LW Wiki entry. Past covers of the Rationalist Association publications 1896 Cover * This is a journal for short articles called the “Agnostic Annual” (later “Rationalist Annual”). The full text is here. * Some quotes from the article “Mind as controlled by matter” below. 1938 Cover Includes articles by Bertrand Russell and evolutionary biologist and proto-transhumanist J. B. S. Haldane. 1954 Cover 1971 Cover * This is for a magazine called The Humanist, published by the Rationalist Association. It has since published articles by Richard Dawkins and Daniel Dennett. * Contents: “Rationalist Tasks”, “Motoring Safety”, “Quasi-rational Quarrelling” and the first appearance of the famous Philip Larkin poem that begins “They fuck you up, your mum and dad”. Links and Full-text 1. Full-text articles from the Rationalist Annual by the prolific biologist, popular science writer and communist, J. B. S. Haldane. Some titles: * Why I am a Materialist *
0364a247-c96e-4763-8016-a2749f90ff4d	trentmkelly/LessWrong-43k	LessWrong	What Is Love? There's probably more to this. What I'm about to describe may be more about intimacy than about love, per se. It could just as easily be the intimate understanding between archfoes as star-crossed lovers. It seems to me that a lot of behaviors and feelings around love can be explained just as people simulating other minds in high fidelity, and then experiencing a shadow of the simulation's feelings as their own via empathy or mirror neurons. So, I simulate you experiencing something, and then I feel a diluted version of what you would actually feel if you experienced it in real life. I emphathize with my model of you kind of like how I'd empathize with the real you if I were observing you. Love at first sight, then, would be me feeling intense feelings about my model of you, but it wouldn't be a correct model. I'd be projecting a stereotype onto you, using your appearance as the skin of my mind's particular concept of an idealized angel or some such perfect being. That's the power of fantasy. The simulation + empathy model could also explain why we like doing nice things for each other. I want to give you positive experiences so that I can simulate you enjoying them, and then enjoy my empathetic kickback. "Just seeing the look on their faces makes it all worth it." It can explain why real love requires time, but why the real measure is knowledge. We have to be able to simulate high-fidelity models of each other. It can clarify the line between meaningful and meaningless sex. Meaningless sex is that without intimacy, that is, sex during which your mind is a black box to me, and I cannot accurately simulate how you're experiencing things. Alternatively, I might desire to hear your thoughts or voice, to see your reactions. I want to have sex with people whose minds I know in great detail. All this that I might receive and process information about my partner's mind, model it, and then get a kickback from empathizing with their pleasure. A cheap shortcut might be
1894fd04-2b7a-429f-b22a-a1b3d97aac57	StampyAI/alignment-research-dataset/aisafety.info	AI Safety Info	What is the difference between AI safety, AI alignment, AI control, friendly AI, AI ethics, AI existential safety, and AGI safety? Terms like these have a fair amount of overlap and aren't always used consistently. The definitions below are how the terms are used on this website, but this isn't an authoritative guide on how these terms should be used: - [AI Safety](/?state=8486&question=What%20is%20AI%20safety%3F): AI safety generally means getting AI systems to avoid risks; existential safety is an extreme type of this risk with unique challenges.[^kix.rq11mrc7khg8] AI safety originally referred only to existential risks from AI systems. However, in recent years, it has been expanded to also include risks that are relevant at lower AI capability levels, such as near-term technical (e.g. self-driving cars) and governance risks. This makes AI safety an umbrella term used to refer to all safety issues both in the near and long term. Other terms are used in conjunction with AI safety to identify the specific risk being addressed. - AGI Safety: AGI safety refers to safety concerns from artificial general intelligence. It overlaps with AI alignment, in that misalignment would be the main cause of unsafe behavior in AGIs, but also includes misuse and other [governance issues](https://www.alignmentforum.org/tag/ai-governance). - AI Existential Safety: AI existential safety concerns AI risks that pose an [existential threat](https://www.alignmentforum.org/tag/existential-risk) whether or not the AI possesses an intelligence that is as general or as capable as that of humans. It means preventing AI technology from posing risks to humanity that are comparable to or greater than human extinction in terms of their moral significance.[^kix.tbt3v8ck1vlx] - [AI Alignment](https://en.wikipedia.org/wiki/AI_alignment): Alignment is focused on ‘making AI go well’. Researchers in AI alignment focus on causing the goals of future superintelligent AI systems to align with [human values](https://www.lesswrong.com/posts/GermiEmcS6xuZ2gBh/what-ai-safety-researchers-have-written-about-the-nature-of). Paul Christiano defines it this way: “Alignment is the problem of getting your AI to try to do the right thing, not the problem of figuring out which thing is right. An aligned AI would try to figure out which thing is right, and like a human it may or may not succeed.”[^kix.mzdmtr5obkde] If aligned, [AIs](https://www.alignmentforum.org/tag/ai)/[AGIs](https://www.alignmentforum.org/tag/artificial-general-intelligence)/[Artificial Superintelligence (ASI)](https://www.alignmentforum.org/tag/superintelligence) would behave in a way that is compatible with human survival and flourishing. Alignment research is interdisciplinary and can include computer science, mathematics, neuroscience, philosophy, and social sciences. Some places (e.g. the [Alignment Forum](https://www.alignmentforum.org/)) use the term AI alignment to mean the project of AI existential safety, including governance and excluding non-existentially risky misalignment. - [AI Control](https://ai-alignment.com/ai-safety-vs-control-vs-alignment-2a4b42a863cc): This term refers to ensuring that AI systems try to do the right thing, and in particular that they don’t competently pursue the wrong thing. It is an older and uncommonly used term, and it refers to [roughly the same set of problems as AI alignment.](https://ai-alignment.com/security-and-ai-control-675ace05ce31)[^kix.9ip3vm8qesnh] - [AI Governance](https://forum.effectivealtruism.org/topics/ai-governance): AI governance refers to identifying and enforcing norms for AI developers and AI systems themselves to follow.[^kix.euretahit9hf] AI governance is often paired with AI safety. Both have the goal of helping humanity develop beneficial AI. AI safety focuses on the technical questions of how AI is built; AI governance focuses on the institutions and contexts in which AI is built and used.[^kix.rnt5dx54p23z] The question of which principles should be enforced often opens up debates about safety and ethics. The conversations in governance are a bit more action-oriented than purely ethical debates. AI governance includes a broad range of subjects, from global coordination around regulating AI development to providing incentives for corporations to be more cautious in their AI research. - [Friendly AI](https://arbital.greaterwrong.com/p/FAI) (FAI): This is an older term coined and popularized by Eliezer Yudkowsky. FAI is a subset of all possible AGIs that includes those that help humans flourish while following some idealized version of human values such as [coherent extrapolated volition](https://www.alignmentforum.org/tag/coherent-extrapolated-volition). Over the last few years, the term ‘aligned AI’ has been used to refer to the same concept more often than FAI. - [AI Ethics](https://en.wikipedia.org/wiki/Ethics_of_artificial_intelligence): AI ethics refers to principles that AI developers and systems should follow.[^kix.gwv7ghbj5thw] The “should” here creates a space for debate, whereby many people and institutions can try to impose their values on what principles become accepted by society at large.[^kix.q29i22s1txme] AI Ethics focuses on ensuring that in our attempt to harness this technology for good, we appropriately assess its potential for societal harm within its design. This includes preventing and [mitigating algorithmic bias](https://www.harvardmagazine.com/2021/08/meredith-broussard-ai-bias-documentary), accountability of companies related to generative works, and the [transparency](https://hbr.org/2022/06/building-transparency-into-ai-projects) of the models being used to make societal decisions. AI Ethics often refers to concerns for existing technology with limited scopes whereas most of the terms above refer to future AIs with potentially world-altering scopes. The reason we have so many terms is that we’re trying to communicate a specific concept: How to build very powerful AI systems that don't kill everyone and do promote human flourishing. But the terms each come with implications and so keep drifting to mean different things. Some people have attempted to resort to [AI notkilleveryoneism](https://twitter.com/ESYudkowsky/status/1612364482608795650) to mitigate this dilution and distortion of terms, but this has not been widely adopted for obvious reasons. [^kix.rq11mrc7khg8]: Critch, Andrew (2020). [Some AI research areas and their relevance to existential safety](https://www.lesswrong.com/posts/hvGoYXi2kgnS3vxqb/some-ai-research-areas-and-their-relevance-to-existential-1#AI_existential_safety__definition_) [^kix.rnt5dx54p23z]: Dafoe, Allan (2017). [AI Governance: A Research Agenda](https://www.fhi.ox.ac.uk/wp-content/uploads/GovAI-Agenda.pdf) [^kix.mzdmtr5obkde]: Paul Christiano (2018). [Clarifying “AI alignment”](https://ai-alignment.com/clarifying-ai-alignment-cec47cd69dd6) [^kix.9ip3vm8qesnh]: Paul Christiano (2016). [AI “safety” vs “control” vs “alignment”](https://ai-alignment.com/ai-safety-vs-control-vs-alignment-2a4b42a863cc) [^kix.gwv7ghbj5thw]: Critch, Andrew (2020). [Some AI research areas and their relevance to existential safety](https://www.lesswrong.com/posts/hvGoYXi2kgnS3vxqb/some-ai-research-areas-and-their-relevance-to-existential-1#AI_existential_safety__definition_) [^kix.q29i22s1txme]: Critch, Andrew (2020). [Some AI research areas and their relevance to existential safety](https://www.lesswrong.com/posts/hvGoYXi2kgnS3vxqb/some-ai-research-areas-and-their-relevance-to-existential-1#AI_existential_safety__definition_) [^kix.euretahit9hf]: Critch, Andrew (2020). [Some AI research areas and their relevance to existential safety](https://www.lesswrong.com/posts/hvGoYXi2kgnS3vxqb/some-ai-research-areas-and-their-relevance-to-existential-1#AI_existential_safety__definition_) [^kix.tbt3v8ck1vlx]: Critch, Andrew (2020). [Some AI research areas and their relevance to existential safety](https://www.lesswrong.com/posts/hvGoYXi2kgnS3vxqb/some-ai-research-areas-and-their-relevance-to-existential-1#AI_existential_safety__definition_)
a4bb158a-10ba-4614-8060-a8ac4fe9096b	trentmkelly/LessWrong-43k	LessWrong	How should we think about the decision relevance of models estimating p(doom)? To illustrate what I mean, switching from p(doom) to timelines: * The recent post AGI Timelines in Governance: Different Strategies for Different Timeframes was useful to me in pushing back against Miles Brundage's argument that "timeline discourse might be overrated", by showing how choice of actions (in particular in the AI governance context) really does depend on whether we think that AGI will be developed in ~5-10 years or after that. * A separate takeaway of mine is that decision-relevant estimation "granularity" need not be that fine-grained, and in fact is not relevant beyond simply "before or after ~2030" (again in the AI governance context). * Finally, that post was useful to me in simply concretely specifying which actions are influenced by timelines estimates. Question: Is there something like this for p(doom) estimates? More specifically, following the above points as pushback against the strawman(?) that "p(doom) discourse, including rigorous modeling of it, is overrated": 1. What concrete high-level actions do most alignment researchers agree are influenced by p(doom) estimates, and would benefit from more rigorous modeling (vs just best guesses, even by top researchers e.g. Paul Christiano's views)? 2. What's the right level of granularity for estimating p(doom) from a decision-relevant perspective? Is it just a single bit ("below or above some threshold X%") like estimating timelines for AI governance strategy, or OOM (e.g. 0.1% vs 1% vs 10% vs >50%), or something else? * I suppose the easy answer is "the granularity depends on who's deciding, what decisions need making, in what contexts", but I'm in the dark as to concrete examples of those parameters (granularity i.e. thresholds, contexts, key actors, decisions) * e.g. reading Joe Carlsmith's personal update from ~5% to >10% I'm unsure if this changes his recommendations at all, or even his conclusion – he writes that "my main point here, though, isn't the specific numbers...
f8383697-54f7-45a7-9aee-9bc88428439c	trentmkelly/LessWrong-43k	LessWrong	Commentary on AGI Safety from First Principles My AGI safety from first principles report (which is now online here) was originally circulated as a google doc. Since there was a lot of good discussion in comments on the original document, I thought it would be worthwhile putting some of it online, and have copied out most of the substantive comment threads here. Many thanks to all of the contributors for their insightful points, and to Habryka for helping with formatting. Note that in some cases comments may refer to parts of the report that didn't make it into the public version. Discussion on the whole report Will MacAskill Thanks so much for writing this! Huge +1 to more foundational work in this area. My overall biggest worry with your argument is just whether it's spending a lot of time defending something that's not really where the controversy lies. (This is true for me; I don't know if I'm idiosyncratic.) Distinguish two claims one could argue for: Claim 1: At some point in the future, assuming continued tech progress, history will have primarily become the story of AI systems doing things. The goals of those AI systems, or the emergent path that results from interactions among these systems, will probably not be what you reading this document want to happen. I find claim 1 pretty uncontroversial. And I do think that this alone is enough for far more of the world to be thinking about AI than currently is. But it feels like at least for longtermist EAs trying to prioritise among causes (or for non-longtermists deciding how much to prioritise safety vs speed on AI), the action is much more on a more substantial claim like: Claim 2: Claim 1 is true, and the point in time at which the transition from a human-driven world to an AI-driven world is in our lifetime, and the transition will be fast, and we can meaningfully affect how this transition goes with very long-lasting impacts, and (on the classic formulations at least) the transition will be to a single AI agent with more power than all other ag
dc3b3956-973e-4885-8514-10607b0f8905	trentmkelly/LessWrong-43k	LessWrong	Exponential Secretary Using women I've dated to improve on the solution to the famous secretary problem.
9e090c0d-14d2-4a00-a6e1-d6419a719308	StampyAI/alignment-research-dataset/special_docs	Other	Asking the Right Questions: Learning Interpretable Action Models Through Query Answering. Asking the Right Questions: Learning Interpretable Action Models Through Query Answering Pulkit Verma, Shashank Rao Marpally, andSiddharth Srivastava School of Computing, Informatics, and Decision Systems Engineering Arizona State University, Tempe, AZ 85281, USA fverma.pulkit, smarpall, siddharths g@asu.edu Abstract This paper develops a new approach for estimating an inter- pretable, relational model of a black-box autonomous agent that can plan and act. Our main contributions are a new paradigm for estimating such models using a rudimentary query interface with the agent and a hierarchical querying algorithm that generates an interrogation policy for estimat- ing the agent’s internal model in a user-interpretable vocab- ulary. Empirical evaluation of our approach shows that de- spite the intractable search space of possible agent models, our approach allows correct and scalable estimation of in- terpretable agent models for a wide class of black-box au- tonomous agents. Our results also show that this approach can use predicate classiﬁers to learn interpretable models of planning agents that represent states as images. 1 Introduction The growing deployment of AI systems ranging from per- sonal digital assistants to self-driving cars leads to a per- vasive problem: how would a user ascertain whether an AI system will be safe, reliable, or useful in a given situa- tion? This problem becomes particularly challenging when we consider that most autonomous systems are not designed by their users; their internal software may be unavailable or difﬁcult to understand, and it may even change from initial speciﬁcations as a result of learning. Such scenarios feature black-box AI agents whose models may not be available in terminology that the user understands. They also show that in addition to developing better AI systems, we need to de- velop new algorithmic paradigms for assessing arbitrary AI systems and for determining the minimal requirements for AI systems in order to ensure interpretability and to support such assessments (Srivastava 2021). This paper presents a new approach for addressing these questions. It develops an algorithm for estimating inter- pretable, relational models of AI agents by querying them. In doing so, it requires the AI system to have only a prim- itive query-response capability to ensure interpretability. Consider a situation where Hari(ette) ( H) wants a grocery- delivery robot (A) to bring some groceries, but s/he is unsure whether it is up to the task and wishes to estimate A’s inter- nal model in an interpretable representation that s/he is com- Copyright © 2021, Association for the Advancement of Artiﬁcial Intelligence (www.aaai.org). All rights reserved. Figure 1: The agent-assessment module uses its user’s pre- ferred vocabulary, queries the AI system, and delivers a user- interpretable causal model of the AI system’s capabilities. The AI system does not need to know the user’s vocabulary or modeling language. fortable with (e.g., a relational STRIPS-like language (Fikes and Nilsson 1971; McDermott et al. 1998; Fox and Long 2003)). IfHwas dealing with a delivery person, s/he might ask them questions such as “would you pick up orders from multiple persons?” and “do you think it would be alright to bring refrigerated items in a regular bag?” If the answers are “yes” during summer, it would be a cause for concern. Na¨ıve approaches for generating such questions to ascertain the limits and capabilities of an agent are infeasible.1 We propose an agent-assessment module (AAM), shown in Fig. 1, which can be connected with an arbitrary AI agent that has a rudimentary query-response capability: the assess- ment module connects Awith a simulator and provides a sequence of instructions, or a plan as a query .Aexecutes the plan in the simulator and the assessment module uses the simulated outcome as the response to the query. Thus, given an agent, the assessment module uses as input: a user- deﬁned vocabulary, the agent’s instruction set, and a com- patible simulator. These inputs reﬂect natural requirements of the task and are already quite commonly supported: AI systems are already designed and tested using compatible simulators, and they need to specify their instruction sets in order to be usable. The user provides the concepts that they 1Just 2 actions and 5 grounded propositions would yield 725108possible STRIPS-like models – each proposition could be absent, positive or negative in the precondition and effects of each action, and cannot be positive (or negative) in both precon- ditions and effect simultaneously. A query strategy that inquires about each occurrence of each proposition would be not only un- scalable but also inapplicable to simulator-based agents that do not know their actions’ preconditions and effects. can understand and these concepts can be deﬁned as func- tions on simulator states. In developing the ﬁrst steps towards this paradigm, we as- sume that the user wishes to estimate A’s internal model as a STRIPS-like relational model with conjunctive precondi- tions, add lists, and delete lists, and that the agent’s model is expressible as such. Such models can be easily trans- lated into interpretable descriptions such as “under situations where preconditions hold, if the agent Aexecutes actions a1;:::;akit would result in effects ,” where preconditions and effects use only the user-provided concepts. Further- more, such models can be used to investigate counterfactuals and support assessments of causality (Halpern 2016). This fundamental framework (Sec. 3) can be developed to support different types of agents as well as various query and response modalities. E.g., queries and responses could use a speech interface for greater accessibility, and agents with reliable inbuilt simulators/lookahead models may not need external simulators. This would allow AAM to pose queries such as “what do you think would happen if you didhquery plani”, and the learnt model would reﬂect A’s self-assessment. The “agent” could be an arbitrary entity, al- though the expressiveness of the user-interpretable vocabu- lary would govern the scope of the learnt models and their accuracy. Using AAM with such agents would also help make them compliant with Level II assistive AI – systems that make it easy for operators to learn how to use them safely (Srivastava 2021). Our algorithm for the assessment module (Sec. 3.1) gen- erates a sequence of queries ( Q) depending on the agent’s responses () during the query process; the result of the over- all process is a complete model of A. To generate queries, we use a top-down process that eliminates large classes of agent-inconsistent models by computing queries that dis- criminate between pairs of abstract models . When an ab- stract model’s answer to a query differs from the agent’s answer, we effectively eliminate the entire set of possible concrete models that are reﬁnements of this abstract model. Sec. 3 presents our overall framework with algorithms and theoretical results about their convergence properties. Our empirical evaluation (Sec. 4) shows that this method can efﬁciently learn correct models for black-box versions of agents using hidden models from the IPC2. It also shows that AAM can use image-based predicate classiﬁers to infer correct models for simulator-based agents that respond with an image representing the result of query plan’s execution. 2 Related Work A number of researchers have explored the problem of learn- ing agent models from observations of its behavior (Gil 1994; Yang, Wu, and Jiang 2007; Cresswell, McCluskey, and West 2009; Zhuo and Kambhampati 2013). Such action- model learning approaches have also found practical appli- cations in robot navigation (Balac, Gaines, and Fisher 2000), player behavior modeling (Krishnan, Williams, and Martens 2020), etc. To the best of our knowledge, ours is the ﬁrst ap- proach to address the problem of generating query strategies 2https://www.icaps-conference.org/competitionsfor inferring relational models of black-box agents. Amir and Chang (2008) use logical ﬁltering (Amir and Russell 2003) to learn partially observable action models from the observation traces. LOCM (Cresswell, McCluskey, and West 2009) and LOCM2 (Cresswell and Gregory 2011) present another class of algorithms that use ﬁnite-state ma- chines to create action models from observed plan traces. Camacho and McIlraith (2019) present an approach for learning highly expressive LTL models from an agent’s ob- served state trajectories using an oracle with knowledge of the target LTL representation. This oracle can also gener- ate counterexamples when the estimated model differs from the true model. In contrast, our approach does not require such an oracle. Also, unlike Stern and Juba (2017), our approach does not need intermediate states in execution traces. In contrast to approaches for white-box model main- tenance (Bryce, Benton, and Boldt 2016), our approach does not requireAto know aboutH’s preferred vocabulary. LOUGA (Ku ˇcera and Bart ´ak 2018) combines a genetic algorithm with an ad-hoc method to learn planning oper- ators from observed plan traces. FAMA (Aineto, Celorrio, and Onaindia 2019) reduces model recognition to a planning problem and can work with partial action sequences and/or state traces as long as correct initial and goal states are pro- vided. While both FAMA and LOUGA require a postpro- cessing step to update the learnt model’s preconditions to in- clude the intersection of all states where an action is applied, it is not clear that such a process would necessarily converge to the correct model. Our experiments indicate that such ap- proaches exhibit oscillating behavior in terms of model ac- curacy because some data traces can include spurious predi- cates, which leads to spurious preconditions being added to the model’s actions. FAMA also assumes that there are no negative literals in action preconditions. Bonet and Geffner (2020) present an algorithm for learn- ing relational models using a SAT-based method when the action schema, predicates, etc. are not available. This ap- proach takes as input a predesigned correct and complete di- rected graph encoding the structure of the entire state space. The authors note that their approach is viable for problems with small state spaces. While our method provides an end- to-end solution, it can also be used in conjunction with such approaches to create the inputs they need. Khardon and Roth (1996) address the problem of making model-based infer- ence faster given a set of queries , under the assumption that a static set of models represents the true knowledge base. In contrast to these directions of research, our approach directly queries the agent and is guaranteed to converge to the true model while presenting a running estimate of the ac- curacy of the derived model; hence, it can be used in settings where the agent’s model changes due to learning or a soft- ware update. In such a scenario, our algorithm can restart to query the system, while approaches that derive models from observed plan traces would require arbitrarily long data col- lection sessions to get sufﬁcient uncorrelated data. Incremental Learning Model (Ng and Petrick 2019) uses reinforcement learning to learn a nonstationary model with- out using plan traces, and requires extensive training to learn the full model correctly. Chitnis et al. (2021) present an approach for learning probabilistic relational models where they use goal sampling as a heuristic for generating relevant data, while we reduce that problem to query synthesis using planning. Their approach is shown to work well for stochas- tic environments, but puts a much higher burden on the AI system for inferring its model. This is because the AI system has to generate a conjunctive goal formula while maximiz- ing exploration, ﬁnd a plan to reach that goal, and correct the model as it collects observations while executing the plan. The ﬁeld of active learning (Settles 2012) addresses the related problem of selecting which data-labels to acquire for learning single-step decision-making models using sta- tistical measures of information. However, the effective fea- ture set here is the set of all possible plans, which makes conventional methods for evaluating the information gain of possible feature labelings infeasible. In contrast, our ap- proach uses a hierarchical abstraction to select queries to ask, while inferring a multistep decision-making (planning) model. Information-theoretic metrics could also be used in our approach whenever such information is available. 3 The Agent-Interrogation Task We assume thatHneeds to estimateA’s model as a STRIPS- like planning model represented as a pair M=hP;Ai, where P=fpk1 1;:::;pknngis a ﬁnite set of predicates with aritieski;A=fa1;:::;akgis a ﬁnite set of parameterized actions (operators). Each action aj2Ais represented as a tuplehheader (aj);pre(aj);eff(aj)i, whereheader (aj)is the action header consisting of action name and action pa- rameters,pre(aj)represents the set of predicate atoms that must be true in a state where ajcan be applied, eff(aj) is the set of positive or negative predicate atoms that will change to true or false respectively as a result of execution of the action aj. Each predicate can be instantiated using the parameters of an action, where the number of parame- ters are bounded by the maximum arity of the action. E.g., consider the action load truck(?v1;?v2;?v3)and predicate at(?x;?y)in the IPC Logistics domain. This predicate can be instantiated using action parameters ?v1,?v2, and ?v3 asat(?v1;?v1),at(?v1;?v2),at(?v1;?v3),at(?v2;?v2), at(?v2;?v1),at(?v2;?v3),at(?v3;?v3),at(?v3;?v1), and at(?v3;?v2). We represent the set of all such possible pred- icates instantiated with action parameters as P. AAM uses the following information as input. It receives its instruction set in the form of header (a)for eacha2A from the agent. AAM also receives a predicate vocabulary P from the user with functional deﬁnitions of each predicate. This gives AAM sufﬁcient information to perform a dialog withAabout the outcomes of hypothetical action sequences. We deﬁne the overall problem of agent interrogation as follows. Given a class of queries and an agent with an un- known model which can answer these queries, determine the model of the agent. More precisely, an agent interrogation task is deﬁned as a tuple hMA;Q;P;AHi, whereMAis the true model (unknown to AAM) of the agent Abeing in- terrogated, Qis the class of queries that can be posed to the agent by AAM, and PandAHare the sets of predicates and action headers that AAM uses based on inputs from HandA. The objective of the agent interrogation task is to derive the agent modelMAusingPandAH. Let be the set of possible answers to queries. Thus, strings 2denote the information received by AAM at any point in the query process. Query policies for the agent interrogation task are functions!Q[fStopgthat map sequences of answers to the next query that the interrogator should ask. The pro- cess stops with the Stop query. In other words, for all an- swers2, all valid query policies map all sequences x toStop wheneverx2is mapped to Stop. This policy is computed and executed online. Components of agent models In order to formulate our solution approach, we consider a model Mto be com- prised of components called palm tuples of the form = hp;a;l;mi, wherepis an instantiated predicate from the vocabulary P;ais an action from the set of parameter- ized actions A,l2 fpre;effg, andm2 f+;;;g. For convenience, we use the subscripts p;a;l; ormto denote the corresponding component in a palm tuple. The presence of a palm tuple in a model denotes the fact that in that model, the predicate pappears in an action aat a lo- cationlas a true (false) literal when sign mis positive (negative), and is absent when m=;. This allows us to deﬁne the set-minus operation Mnon this model as re- moving the palm tuple from the model. We consider two palm tuples 1=hp1;a1;l1;m1iand2=hp2;a2;l2;m2i to be variants of each other ( 12) iff they differ only on modem, i.e.,12,(1p=2p)^(1a= 2a)^(1l=2l)^(1m6=2m). Hence, mode assignment to apaltuple =hp;a;lican result in 3 palm tuple variants +=hp;a;l; +i, =hp;a;l;i, and ;=hp;a;l;;i. Model abstraction We now deﬁne the notion of abstrac- tion used in our solution approach. Several approaches have explored the use of abstraction in planning (Sacerdoti 1974; Giunchiglia and Walsh 1992; Helmert et al. 2007; B¨ackstr ¨om and Jonsson 2013; Srivastava, Russell, and Pinto 2016). The deﬁnition of abstraction used in this work ex- tends the concept of predicate and propositional domain ab- stractions (Srivastava, Russell, and Pinto 2016) to allow for the projection of a single palm tuple. An abstract model is one in which all variants of at least one pal tuple are absent. Let be the set of all possible palm tuples which can be generated using a predicate vocabulary Pand an action header set AH. LetUbe the set of all con- sistent (abstract and concrete) models that can be expressed as subsets of , such that no model has multiple variants of the same palm tuple. We deﬁne abstraction of a model as: Deﬁnition 1. Theabstraction of a model Mwith respect to a palm tuple 2, is deﬁned by f:U!U asf(M) = Mn. We extend this notation to deﬁne the abstraction of a set of modelsMwith respect to a palm tuple asX=ff(m) : m2Mg . We use this abstraction framework to deﬁne a subset-lattice over abstract models (Fig. 2(b)). Each node in the lattice represents a collection of possible abstract mod- els which are possible variants of a pal tuple . E.g., in the node labeled 1 in Fig. 2(b), we have models corresponding ... ... ... (a) (b) (c) Figure 2: (b) Lattice segment explored in random order of i2; (a) At each node, 3 abstract models are generated and 2 of them are discarded based on query responses; (c) An abstract model rejected at any level is equivalent to re- jecting 3 models at the level below, 9 models two levels down, and so on. to + 1, 1, and ; 1. Two nodes in the lattice are at the same level of abstraction if they contain the same number of pal tuples. Two nodes niandnjin the lattice are connected if all the models at nidiffer with all the models in njby a sin- gle palm tuple. As we move up in the lattice following these edges, we get more abstracted versions of the models, i.e., containing less number of pal tuples; and we get more con- cretized models, i.e., containing more number of pal tuples, as we move downward. We now deﬁne this model lattice: Deﬁnition 2. Amodel latticeLis a 5-tupleL= hN;E; ;`N;`Ei, whereNis a set of lattice nodes, is the set of all pal tuples hp;a;li,`N:N!22is a node label function where = f+;;;gis the set of all palm tuples, Eis the set of lattice edges, and `E:E!is a function mapping edges to edge labels such that for each edgeni!nj; `N(nj) =f[f kgj2`N(ni); = `E(ni!nj);k2f+;;;gg, and`N(>) =fgwhere> is the supremum containing the empty model . A noden2Nin this latticeLcan be uniquely identiﬁed by the sequence of pal tuples that label the edges leading to it from the supremum. As shown in Fig. 2(a), even though theoretically `N:N!22, not all the models are stored at any node as at least one is pruned out based on some query Q2Q. Additionally, in these model lattices, every node has an edge going out from it corresponding to each pal tuple that is not present in the paths leading to it from the most ab- stracted node. At any stage during the interrogation, nodes in such a lattice are used to represent the set of possible models given the agent’s responses up to that point. At every step, our algorithm creates queries online that help us determine the next descending edge to take from a lattice node; corre- sponding to the path 0;:::;i in Fig. 2(b). This also avoids generating and storing the complete lattice, which can be doubly exponential in number of predicates and actions. Form of agent queries As discussed earlier, based on A’s responses, we pose queries to the agent and infer A’smodel. We express queries as functions that map models to answers. Recall that Uis the set of all possible (concrete and abstract) models, and is the set of possible responses. A queryQis a functionQ:U! . In this paper, we utilize only one class of queries: plan outcome queries (QPO), which are parameterized by a state sIand a plan . LetPbe the set of predicates Pin- stantiated with objects Oin an environment. QPOqueries askAthe length of the longest preﬁx of the plan that it can execute successfully when starting in the state sI Pas well as the ﬁnal state sFPthat this execu- tion leads to. E.g., “Given that the truck t1and pack- agep1are at location l1, what would happen if you ex- ecuted the planhloadtruck (p1;t1;l1),drive (t1;l1;l2), unloadtruck (p1;t1;l2)i?” A response to such queries can be of the form “I can ex- ecute the plan till step `and at the end of it p1is in truck t1which is at location l1”. Formally, the response POfor plan outcome queries is a tuple h`;sFi, where`is the num- ber of steps for which the plan could be executed, and sFPis the ﬁnal state after executing `steps of the plan. If the plancannot be executed fully according to the agent modelMAthen` < len (), otherwise`=len(). The ﬁnal statesFPis such thatMAj=[1 :`](sI) =sF, i.e., starting with a state sI,MAsuccessfully executed ﬁrst `steps of the plan . Thus,QPO:U!N2P, where N is the set of natural numbers. Not all queries are useful, as some of them might not increase our knowledge of the agent model at all. Hence, we deﬁne some properties associated with each query to as- certain its usability. A query is useful only if it can distin- guish between two models. More precisely, a query Qis said to distinguish a pair of modelsMiandMj, denoted asMiQMj, iffQ(Mi)6=Q(Mj). Deﬁnition 3. Two modelsMiandMjare said to be distin- guishable , denoted asMiMj, iff there exists a query that can distinguish between them, i.e., 9QMiQMj. Given a pair of abstract models, we wish to determine whether one of them can be pruned, i.e., whether there is a query for which at least one of their answers is inconsis- tent with the agent’s answer. Since this is computationally expensive to determine, and we wish to reduce the number of queries made to the agent, we ﬁrst evaluate whether the two models can be distinguished by any query, independent of consistency of their response with that of the agent. If the models are not distinguishable, it is redundant to try to prune one of them under the given query class. Next, we determine if at least one of the two distinguish- able models is consistent with the agent. When comparing the responses of two models at different levels of abstrac- tion, we must consider the fact that the agent’s response may be at a different level of abstraction if the given pair of mod- els is abstract. Taking this into account, we formally deﬁne what it means for an abstract model Mi’s response to be consistent with that of agent model MA: Deﬁnition 4. LetQ be a query such that MiQMj;Q(Mi) =h`i;hpi 1;:::;pi mii,Q(Mj) = (a)MA’sload truck(?p,?t,?l) action (unknown to H) at(?t,?l), at(?p,?l)! in(?p,?t), :(at(?p,?l)) (b)M1’sload truck(?p,?t,?l) action at(?t,?l), at(?p,?l)! in(?p,?t) (c)M2’sload truck(?p,?t,?l) action at(?t,?l) ! in(?p,?t) (d)M3’sload truck(?p,?t,?l) action at(?t,?l) ! () Figure 3: load truck actions of the agent model MAand 3 abstracted modelsM1,M2, andM3. HereX!Ymeans thatXis the precondition of an action and Yis the effect. h`j;hpj 1;:::;pj nii, andQ(MA) =h`A;hpA 1;:::;pA kii. Mi’s response toQisconsistent with that ofMA, i.e., Q(MA)j=Q(Mi)if`A=len(Q),len(Q) =`iand fpi 1;:::;pi mgfpA 1;:::;pA kg. Using this notion of consistency, we can now reason that given a set of distinguishable models MiandMj, and their responses in addition to the agent’s response to the distin- guishing query, the models are prunable if and only if ex- actly one of their responses is consistent with that of the agent. Formally, we deﬁne prunability as: Deﬁnition 5. Given an agent-interrogation task hMA;Q;P;AHi, two modelsMiandMjareprun- able, denoted asMihiMj, iff9Q 2 Q:MiQMj ^(Q(MA)j=Q(Mi)^ Q(MA)6j=Q(Mj))_ (Q(MA)6j=Q(Mi)^Q(MA)j=Q(Mj)). 3.1 Solving the Interrogation Task We now discuss how we solve the agent interrogation task by incrementally adding palm variants to the class of abstract models and pruning out inconsistent models by generating distinguishing queries. Example 1. Consider the case of a delivery agent. Assume that AAM is considering two abstract models M1andM2 having only one action loadtruck (?p;?t;?l)and the pred- icatesat(?p;?l),at(?t;?l),in(?p;?t), and that the agent’s model isMA(Fig. 3). AAM can ask the agent what will happen ifAloads package p1into truckt1at locationl1 twice. The agent would respond that it could execute the plan only till length 1, and the state at the time of this failure would beat(t1;l1)^in(p1;t1). Algorithm 1 shows AAM’s overall algorithm. It takes the agentA, the set of instantiated predicates P, the set of all action headers AH, and a set of random states Sas input, and gives the set of functionally equivalent estimated models represented by possmodels as output. Scan be generated in a preprocessing step given P. AIA initializes possmodels as a set consisting of the empty model (line 3) representing that AAM is starting at the supremum >of the model lattice. In each iteration of the main loop (line 4), AIA maintains an abstraction lattice and keeps track of the current node inAlgorithm 1 Agent Interrogation Algorithm (AIA) 1:Input:A;AH;P;S 2:Output: poss models 3:Initialize poss models =fg 4:for in some input pal ordering do 5: new models poss models 6: pruned models =fg 7: foreachM0in new models do 8: foreach pairfi;jginf+;;;gdo 9:Q,Mi,Mj generate query(M0;i;j; ; S) 10:Mprune ﬁlter models(Q;MA;Mi;Mj) 11: pruned models pruned models[Mprune 12: end for 13: end for 14: ifpruned models is;then 15: update palordering( ;S) 16: continue 17: end if 18: poss models new modelsf +; ; ;gn pruned models 19:end for the lattice. It picks a pal tuple corresponding to one of the descending edges in the lattice from a node given by some input ordering of . The correctness of the algorithm does not depend on this ordering. It then stores a temporary copy ofpossmodels asnewmodels (line 5) and initialize an empty set at each node to store the pruned models (line 6). The inner loop (line 7) iterates over the set of all pos- sible abstract models that AIA has not rejected yet, stored asnewmodels . It then loops over pairs of modes (line 8), which are later used to generate queries and reﬁne mod- els. For the chosen pair of modes, generate query() is called (line 9) which returns two models concretized with the cho- sen modes and a query Qwhich can distinguish between them based on their responses. AIA then calls ﬁlter models() which poses the query Qto the agent and the two models. Based on their responses, AIA prunes the models whose responses are not consistent with that of the agent (line 11). Then it updates the estimated set of possible models represented by possmodels (line 18). If AIA is unable to prune any model at a node (line 14), it modiﬁes the pal tuple ordering (line 15). AIA continues this process until it reaches the most concretized node of the lattice (meaning all possible palm tuples 2are reﬁned at this node). The remaining set of models represents the estimated set of models for A. The number of resolved palm tuples can be used as a running estimate of accuracy of the derived models. AIA requires O(jPjjAj)queries as there are2jPjjAjpal tuples. However, our empirical studies show that we never generate so many queries. 3.2 Query Generation The query generation process corresponds to the gener- atequery() module in AIA which takes a model M0, the pal tuple , and 2 modes i;j2f+;;;gas input; and returns the modelsMi=M0[f igandMj=M0[f jg, and a Algorithm 2 Query Generation Algorithm 1:Input:M0;i;j; ; S 2:Output:Q;Mi;Mj 3:Mi;Mj addpalm(M0;i;j; ) 4:forsIinSdo 5: dom, prob getplanning prob (sI;Mi;Mj) 6: planner(dom, prob) 7:Q hsI;i 8: ifthen break end if 9:end for 10:returnQ,M0[f ig,M0[f jg plan outcome query Qdistinguishing them, i.e., MiQMj. Plan outcome queries have 2 components, an initial state sIand a plan. AIA getssIfrom the input set of random statesS(line 4). Using sIas the initial state, the idea is to ﬁnd a plan, which when executed by MiandMjwill lead them either to different states, or to a state where only one of them can execute the plan further. Later we pose the same query toAand prune at least one of MiandMj. Hence, we aim to prevent the models inconsistent with the agent model MAfrom reaching the same ﬁnal state as MAafter exe- cuting the queryQand following a different state trajectory. To achieve this, we reduce the problem of generating a plan outcome query from MiandMjinto a planning problem. The reduction proceeds by creating temporary models M00 iandM00 j. We now discuss how to generate them. We add the pal tuple =hp;a;liin modesiandjtoM0to get M0 iandM0 j, respectively. If the location l=eff, we add the palm tuple normally to M0, i.e.,M0 m=M0[hp;a;l;mi, wherem2fi;jg. Ifl=pre, we add a dummy predicate puin disjunction with the predicate pto the precondition of both the models. We then modify the models M0 iandM0 j further in the following way: M00 m=M0 m[fhpu;a0;l0;+i:8a0;l0ha0;l0i62 fha;li:9mhp;a;l;mi2M0gg [fhpu;a0;l0;i:8a0;l0ha0;l0i2 fha;li:l=eff^9mhp;a;l;mi2M0gg puis added only for generating a distinguishing query and is not part of the models MiandMjreturned by the query generation process. Without this modiﬁcation, an inconsis- tent abstract model may have a response consistent with A. We now show how to reduce plan outcome query genera- tion into a planning problem PPO(line 5).PPOuses condi- tional effects in its actions (in accordance with PDDL (Mc- Dermott et al. 1998; Fox and Long 2003)). The model used to deﬁnePPOhas predicates from both models M00 iandM00 j represented asPM00 iandPM00 jrespectively, in addition to a new dummy predicate p . The action headers are the same asAH. Each action’s precondition is a disjunction of the pre- conditions ofM00 iandM00 j. This makes an action applicable in a statesif eitherM00 iorM00 jcan execute it in s. The effect of each action has 2 conditional effects, the ﬁrst applies the effects of bothM00 iandM00 j’s action if preconditions of bothM00 iandM00 jare true, whereas the second makes the dummy predicatep true if precondition of only one of M00 iand M00 jis true. Formally, we express this planning problem as PPO=hMPO;sI;Gi, whereMPOis a model with pred- icatesPPO=PM00 i[PM00 j[p , and actions APOwhere for each action a2APO,pre(a) =pre(aM00 i)_pre(aM00 j) andeff(a)= (when (pre(aM00 i)^pre(aM00 j))(eff(aM00 i)^eff(aM00 j))) (when ((pre(aM00 i)^:pre(aM00 j))_ (:pre(aM00 i)^pre(aM00 j))) (p )); The initial state sI=sM00 i I^sM00 j I, wheresM00 i IandsM00 j I are copies of all predicates in sI, andGis the goal formula expressed as9p(pM00 i^:pM00 j)_(:pM00 i^pM00 j)_p . With this formulation, the goal is reached when an ac- tion inM00 iandM00 jdiffers in either a precondition (making only one of them executable in a state), or an effect (leading to different ﬁnal states on applying the action). E.g., con- sider the models with differences in load truck(p1;t1;l1) as shown in Fig. 3. From the state at(t1;l1)^:at(p1;l1), M2can execute load truck(p1;t1;l1)butM1can- not. Similarly, in state at(t1;l1)^at(p1;l1), executing load truck(p1;t1;l1)will causeMAandM1to end up in states differing in predicate at(p1;l1). Hence, given the cor- rect initial state, the solution to the planning problem PPO will give the correct distinguishing plan. Theorem 1. Given a pair of models MiandMj, the plan- ning problem PPOhas a solution iffMiandMjhave a distinguishing plan outcome query QPO. Proof (Sketch). The input to the planning problem PPO consists of an initial state sI. If the planner can solve PPO with initial state sIto give a plan , the distinguishing query is a combination of sIand. Similarly, ifMiQPOMj, then giving the initial state sIas part of planning problem PPO, the planwill be a solution which is part of QPO. 3.3 Filtering Possible Models This section describes the ﬁlter models() module in Algo- rithm 1 which takes as input MA,Mi,Mj, and the query Q(Sec. 3.2), and returns the subset Mprune which is not consistent withMA. First, AAM poses the queryQtoMi,Mj, and the agent A. Based on the responses of all three, it determines if the two models are prunable, i.e., MihiMj. As mentioned in Def. 5, checking for prunability involves checking if re- sponse to the query Qby one of the models MiorMjis consistent with that of the agent or not. Theorem 2. LetMi;Mj2 fM +;M;M;gbe the models generated by adding the pal tuple toM0which is an abstraction of the true agent model MA. Suppose Q=hsQ I;Qiis a distinguishing query for two distinct modelsMi;Mj, i.e.MiQMj, and the response of mod- elsMi;Mj;andMAto the queryQareQ(Mi) = h`i;hpi 1;:::;pi mii,Q(Mj) =h`j;hpj 1;:::;pj nii, and Q(MA) =h`A;hpA 1;:::;pA kii. When`A=len(Q), Miis not an abstraction of MAiflen(Q)6=`ior fpi 1;:::;pi mg6fpA 1;:::;pA kg. Proof (Sketch). Proving by induction, the base case is adding a single pal tuple hp;a;lito an empty model (which is a consistent abstraction of MA) resulting in 3 models. The 2 models pruned based on Def. 4 can be shown to be incon- sistent withMA, leaving out the one consistent model. For the inductive step, it can be shown that after adding a pal tu- ple to a consistent model it is not consistent with MAonly if it does not execute the full plan (the precondition is inconsis- tent), or if the end state reached by the model is not a subset of the state of the agent (the effect is inconsistent). If the models are prunable, then the palm tuple being added in the inconsistent model cannot appear in any model consistent withA. As we discard such palm tuples at ab- stract levels (as depicted in Fig. 2 (a)), we prune out a large number of models down the lattice (as depicted in Fig. 2 (c)), hence we keep the intractability of the approach in check and end up asking less number of queries. 3.4 Updating PAL ordering This section describes the update palordering() module in AIA (line 15). It is called when the query generated bygenerate query() module is not executable by A, i.e., len(Q)6=`A. E.g., consider two abstract models M2and M3being considered by AAM (Fig. 3). At this level of ab- straction, AAM does not have knowledge of the predicate at(?p;?l), hence it will generate a plan outcome query with initial stateat(?t;?l)and plan load truck(p1;t1;l1)to dis- tinguish betweenM2andM3. But this cannot be executed by the agentAas its precondition at(?p;?l)is not satisﬁed, and hence we cannot discard any of the models. Recall that in response to the plan outcome query we get the failed action aF=[`+1] and the ﬁnal state sF. Since the query plan is generated usingMiandMj(which dif- fer only in the newly added palm tuple), they both would reach the same state sFafter executing ﬁrst `steps of. Thus, we search Sfor a statessFwhereAcan exe- cuteaF. Similar to Stern and Juba (2017), we infer that any predicate which is false in swill not appear in aF’s precon- dition in the positive mode. Next, we iterate through the set of predicates p0snsFand add them to sFto check if Acan still execute aF. Thus, on adding a predicate p2p0 to the statesF, ifAcannot execute aF, we addpin neg- ative mode in aF’s precondition, otherwise in ;mode. All pal tuples whose modes are correctly inferred in this way are therefore removed from the pal ordering. Equivalent Models It is possible for AIA to encounter a pair of modelsMiandMjthat are not prunable. In such cases, the modelsMiandMjare functionally equivalent and none of them can be discarded. Hence, both the models end up in the set possmodels in line 18 of AIA. 3.5 Correctness of Agent Interrogation Algorithm In this section, we prove that the set of estimated mod- els returned by AIA is correct and the returned models areDomainjPjjAjj^Qj t(ms) t(s) Gripper 5 3 17 18.0 0.2 Blocksworld 9 4 48 8.4 36 Miconic 10 4 39 9.2 1.4 Parking 18 4 63 16.5 806 Logistics 18 6 68 24.4 1.73 Satellite 17 5 41 11.6 0.87 Termes 22 7 134 17.0 110.2 Rovers 82 9 370 5.1 60.3 Barman 83 17 357 18.5 1605 Freecell 100 10 535 2.24y33.4y Table 1: The number of queries ( j^Qj), average time per query (t), and variance of time per query ( t) generated by AIA with FD. Average and variance are calculated for 10 runs of AIA, each on a separate problem.yTime in sec. functionally equivalent to the agent’s model, and no correct model is discarded in the process. Theorem 3. The Agent Interrogation Algorithm (algorithm 1) will always terminate and return a set of models, each of which are functionally equivalent to the agent’s model MA. Proof (Sketch). Theorem 1 and Theorem 2 prove that when- ever we get a prunable query, we discard only inconsis- tent models, thereby ensuring that no correct model is dis- carded. When we do not get a prunable query, we in- fer the correct precondition of the failed action using up- date palordering() , hence the number of reﬁned palm tu- ples always increases with the number of iterations of AIA, thereby ensuring its termination in ﬁnite time. 4 Empirical Evaluation We implemented AIA in Python to evaluate the efﬁcacy of our approach.3In this implementation, initial states ( S, line 1 in Algorithm 1) were collected by making the agent per- form random walks in a simulated environment. We used a maximum of 60 such random initial states for each do- main in our experiments. The implementation assumes that the domains do not have any constants and that actions and predicates do not use repeated variables (e.g., at(?v;?v)), although these assumptions can be removed in practice with- out affecting the correctness of algorithms. The implemen- tation is optimized to store the agent’s answers to queries; hence the stored responses are used if a query is repeated. We tested AIA on two types of agents: symbolic agents that use models from the IPC (unknown to AIA), and sim- ulator agents that report states as images using PDDLGym. We wrote image classiﬁers for each predicate for the latter series of experiments and used them to derive state represen- tations for use in the AIA algorithm. All experiments were executed on 5.0 GHz Intel i9-9900 CPUs with 64 GB RAM running Ubuntu 18.04. The analysis presented below shows that AIA learns the correct model with a reasonable number of queries, and compares our results with the closest related work, 3Code available at https://git.io/Jtpej 0 20 400.51.0Gripper 0 20 400.00.51.0Blocksworld 0 20 400.00.51.0Miconic 0 20 40 600.00.51.0Parking 0 20 40 600.00.51.0Model AccuracyLogistics 0 20 400.00.51.0Satellite 0 200 4000.00.51.0Freecell 0 50 1000.00.51.0Termes 0 100 200 3000.00.51.0Barman 0 100 200 3000.00.51.0Rovers02 0510 010 05×10−1 2.02.53.0×10−2 0.00.51.0 Time per Query (sec.) 012 010 24×10−2 01020 Number of QueriesAccuracy: AIA FAMA Time: AIA FAMAFigure 4: Performance comparison of AIA and FAMA in terms of model accuracy and time taken per query with an increasing number of queries. FAMA (Aineto, Celorrio, and Onaindia 2019). We use the metric of model accuracy in the following analysis: the number of correctly learnt palm tuples normalized with the total number of palm tuples in MA. Experiments with symbolic agents We initialized the agent with one of the 10 IPC domain models, and ran AIA on the resulting agent. 10 different problem instances were used to obtain average performance estimates. Table 1 shows that the number of queries required in- creases with the number of predicates and actions in the domain. We used Fast Downward (Helmert 2006) with LM-Cut heuristic (Helmert and Domshlak 2009) to solve the planning problems. Since our approach is planner- independent, we also tried using FF (Hoffmann and Nebel 2001) and the results were similar. The low variance shows that the method is stable across multiple runs. Comparison with FAMA We compare the performance of AIA with that of FAMA in terms of stability of the models learnt and the time taken per query. Since the focus of our approach is on automatically generating useful traces, we provided FAMA randomly generated traces of length 3 (the length of the longest plans in AIA-generated queries) of the form used throughout this paper ( hsI;a1;a2;a3;sGi). Fig. 4 summarizes our ﬁndings. AIA takes lesser time per query and shows better convergence to the correct model. Figure 5: PDDLGym’s simulated Sokoban (left) and Doors (right) environments used for the experiments. FAMA sometimes reaches nearly accurate models faster, but its accuracy continues to oscillate, making it difﬁcult to as- certain when the learning process should be stopped (we in- creased the number of traces provided to FAMA until it ran out of memory). This is because the solution to FAMA’s in- ternal planning problem introduces spurious palm tuples in its model if the input traces do not capture the complete do- main dynamics. For Logistics, FAMA generated an incorrect planning problem, whereas for Freecell and Barman it ran out of memory (AIA also took considerable time for Free- cell). Also, in domains with negative preconditions like Ter- mes, FAMA was unable to learn the correct model. We used Madagascar (Rintanen 2014) with FAMA as it is the pre- ferred planner for it. We also tried FD and FF with FAMA, but as the original authors noted, it could not scale and ran out of memory on all but a few Blocksworld and Gripper problems where it was much slower than with Madagascar. Experiments with simulator agents AIA can also be used with simulator agents that do not know about pred- icates and report states as images. To test this, we wrote classiﬁers for detecting predicates from images of simulator states in the PDDLGym (Silver and Chitnis 2020) frame- work. This framework provides ground-truth PDDL mod- els, thereby simplifying the estimation of accuracy. We ini- tialized the agent with one of the two PDDLGym environ- ments, Sokoban and Doors shown in Fig. 5. AIA inferred the correct model in both cases and the number of instantiated predicates, actions, and the average number of queries (over 5 runs) used to predict the correct model for Sokoban were 35, 3, and 201, and that for Doors were 10, 2, and 252. 5 Conclusion We presented a novel approach for efﬁciently learning the internal model of an autonomous agent in a STRIPS-like form through query answering. Our theoretical and empir- ical results showed that the approach works well for both symbolic and simulator agents. Extending our predicate classiﬁer to handle noisy state de- tection, similar to prevalent approaches using classiﬁers to detect symbolic states (Konidaris, Kaelbling, and Lozano- Perez 2014; Asai and Fukunaga 2018) is a good direction for future work. Some other promising extensions include replacing query and response communication interfaces be- tween the agent and AAM with a natural language similar to Lindsay et al. (2017), or learning other representations like Zhuo, Mu ˜noz-Avila, and Yang (2014). Acknowledgements We thank Abhyudaya Srinet for his help with the implemen- tation. This work was supported in part by the NSF under grants IIS 1844325, IIS 1942856, and OIA 1936997. Ethics Statement Learning the internal model of an AI agent is one of the main focus areas of the AI community in the recent past. This work would enable a layperson to assess such autonomous agents and to verify if they are safe to work with. This would increase the adaption rate of AI systems, as it would remove the dependence of systems using AI on experts who could verify the internal working of the agent. Our system asks the agent queries and assumes that the agent can be connected to a simulator to ensure the correct- ness of responses. Our approach for such model learning comes with soundness and completeness guarantees. This implies that it will ﬁnd the agent model if there exists one, and the model that it learns will be correct as per the simula- tions. As in any approach that uses simulators, this method is susceptible to errors in programming and simulator de- sign. This can be addressed independently through research on formal veriﬁcation of simulators used in AI. References Aineto, D.; Celorrio, S. J.; and Onaindia, E. 2019. Learn- ing Action Models With Minimal Observability. Artiﬁcial Intelligence 275: 104–137. Amir, E.; and Chang, A. 2008. Learning Partially Observ- able Deterministic Action Models. Journal of Artiﬁcial In- telligence Research 33: 349–402. Amir, E.; and Russell, S. 2003. Logical Filtering. In Proc. IJCAI . Asai, M.; and Fukunaga, A. 2018. Classical Planning in Deep Latent Space: Bridging the Subsymbolic-Symbolic Boundary. In Proc. AAAI . B¨ackstr ¨om, C.; and Jonsson, P. 2013. Bridging the Gap Be- tween Reﬁnement and Heuristics in Abstraction. In Proc. IJCAI . Balac, N.; Gaines, D.; and Fisher, D. 2000. Learning Action Models for Navigation in Noisy Environments. In ICML Workshop on Machine Learning of Spatial Knowledge . Bonet, B.; and Geffner, H. 2020. Learning First-Order Sym- bolic Representations for Planning from the Structure of the State Space. In Proc. ECAI . Bryce, D.; Benton, J.; and Boldt, M. W. 2016. Maintaining Evolving Domain Models. In Proc. IJCAI . Camacho, A.; and McIlraith, S. A. 2019. Learning Inter- pretable Models Expressed in Linear Temporal Logic. In Proc. ICAPS . Chitnis, R.; Silver, T.; Tenenbaum, J.; Kaelbling, L. P.; and Lozano-Perez, T. 2021. GLIB: Efﬁcient Exploration for Relational Model-Based Reinforcement Learning via Goal- Literal Babbling. In Proc. AAAI .Cresswell, S.; and Gregory, P. 2011. Generalised Domain Model Acquisition from Action Traces. In Proc. ICAPS . Cresswell, S.; McCluskey, T.; and West, M. 2009. Acquisi- tion of Object-Centred Domain Models from Planning Ex- amples. In Proc. ICAPS . Fikes, R. E.; and Nilsson, N. J. 1971. STRIPS: A New Ap- proach to the Application of Theorem Proving to Problem Solving. Artiﬁcial Intelligence 2(3-4): 189–208. Fox, M.; and Long, D. 2003. PDDL2.1: An Extension to PDDL for Expressing Temporal Planning Domains. Journal of Artiﬁcial Intelligence Research 20(1): 61–124. Gil, Y . 1994. Learning by Experimentation: Incremental Re- ﬁnement of Incomplete Planning Domains. In Proc. ICML . Giunchiglia, F.; and Walsh, T. 1992. A Theory of Abstrac- tion. Artiﬁcial Intelligence 57(2-3): 323–389. Halpern, J. Y . 2016. Actual Causality . The MIT Press. ISBN 0262035022. Helmert, M. 2006. The Fast Downward Planning System. Journal of Artiﬁcial Intelligence Research 26: 191–246. Helmert, M.; and Domshlak, C. 2009. Landmarks, Critical Paths and Abstractions: What’s the Difference Anyway? In Proc. ICAPS . Helmert, M.; Haslum, P.; Hoffmann, J.; et al. 2007. Flexible Abstraction Heuristics for Optimal Sequential Planning. In Proc. ICAPS . Hoffmann, J.; and Nebel, B. 2001. The FF Planning System: Fast Plan Generation Through Heuristic Search. Journal of Artiﬁcial Intelligence Research 14: 253–302. Khardon, R.; and Roth, D. 1996. Reasoning with Models. Artiﬁcial Intelligence 87(1-2): 187–213. Konidaris, G.; Kaelbling, L. P.; and Lozano-Perez, T. 2014. Constructing Symbolic Representations for High- Level Planning. In Proc. AAAI . Krishnan, A.; Williams, A.; and Martens, C. 2020. Towards Action Model Learning for Player Modeling. In Proc. AI- IDE. Kuˇcera, J.; and Bart ´ak, R. 2018. LOUGA: Learning Plan- ning Operators Using Genetic Algorithms. In Knowledge Management and Acquisition for Intelligent Systems . Lindsay, A.; Read, J.; Ferreira, J.; Hayton, T.; Porteous, J.; and Gregory, P. 2017. Framer: Planning Models from Natu- ral Language Action Descriptions. In Proc. ICAPS . McDermott, D.; Ghallab, M.; Howe, A.; Knoblock, C.; Ram, A.; Veloso, M.; Weld, D. S.; and Wilkins, D. 1998. PDDL – The Planning Domain Deﬁnition Language. Technical Re- port CVC TR-98-003/DCS TR-1165, Yale Center for Com- putational Vision and Control. Ng, J. H. A.; and Petrick, R. P. A. 2019. Incremental Learning of Planning Actions in Model-Based Reinforce- ment Learning. In Proc. IJCAI . Rintanen, J. 2014. Madagascar: Scalable Planning with SAT. InProc. 8th International Planning Competition. Sacerdoti, E. D. 1974. Planning in a Hierarchy of Abstrac- tion Spaces. Artiﬁcial Intelligence 5(2): 115–135. Settles, B. 2012. Active Learning . Morgan & Claypool Pub- lishers. ISBN 1608457257. Silver, T.; and Chitnis, R. 2020. PDDLGym: Gym Envi- ronments from PDDL Problems. In ICAPS Workshop on Bridging the Gap Between AI Planning and Reinforcement Learning (PRL) . Srivastava, S. 2021. Unifying Principles and Metrics for Safe and Assistive AI. In Proc. AAAI . Srivastava, S.; Russell, S.; and Pinto, A. 2016. Metaphysics of Planning Domain Descriptions. In Proc. AAAI . Stern, R.; and Juba, B. 2017. Efﬁcient, Safe, and Probably Approximately Complete Learning of Action Models. In Proc. IJCAI . Yang, Q.; Wu, K.; and Jiang, Y . 2007. Learning Action Mod- els from Plan Examples Using Weighted MAX-SAT. Artiﬁ- cial Intelligence 171(2-3): 107–143. Zhuo, H. H.; and Kambhampati, S. 2013. Action-Model Ac- quisition from Noisy Plan Traces. In Proc. IJCAI . Zhuo, H. H.; Mu ˜noz-Avila, H.; and Yang, Q. 2014. Learn- ing Hierarchical Task Network Domains from Partially Ob- served Plan Traces. Artiﬁcial Intelligence 212: 134 – 157.
ccb6add2-9789-48ef-8fe9-d4825bd16df5	trentmkelly/LessWrong-43k	LessWrong	What's the deal with Effective Accelerationism (e/acc)? I've been hearing murmurs about a recently formed philosophy called "Effective Accelerationism", described as:[1] > ...an ideology that draws from Nick Land's theories of accelerationism to advocate for the belief that artificial intelligence and LLMs will lead to a post-scarcity technological utopia. E/acc communities on Twitter were primarily fostered on Twitter Spaces, with e/acc manifestos being shared using the newsletter platform Substack. One example of said Substack manifestos, Notes on e/acc principles and tenets, outlines on an object level the thesis motivating e/acc. TL;DR:[2] > * is: life emerged as a principle of a generalized 2nd law of thermodynamics > * is: due to this physical (observed) law, life tends to seek to capture "free energy" (aka the accursed share in terms of Bataille perhaps) to increase its scope/complexity or maintain its existence > * ethical/moral claim - the ought: we should seek to "accelerate" (must mean to intensify, not in the physics sense of acceleration, where acceleration could simply mean constantly changing direction) this process of growth of organisms/meta-orgranisms to achieve greater and greater capture of free energy and thus more complex systems of intelligence (they demarcate this as ultimately being about the imperative that "in order to spread to the stars, the light of consciousness/intelligence will have to be transduced to non-biological substrates" I don't know enough about complex systems and epistemology to be able to assess these arguments, which is why I'm posting about them here. My outside view is that the majority of e/acc discourse appears to be memes on Twitter, which doesn't give me much hope in the epistemic rigor underlying the philosophy? Reddit user I-am-a-person- summarizes what was close to my initial reaction after reading the Substack post:[3] > The problem with this argument is that it does a really bad job arguing why “capturing free energy” is actually the goal we ought to strive
6602b037-6889-48ee-a67c-e2e6e0fd368e	trentmkelly/LessWrong-43k	LessWrong	On OpenAI Dev Day OpenAI DevDay was this week. What delicious and/or terrifying things await? TURBO BOOST First off, we have GPT-4-Turbo. > Today we’re launching a preview of the next generation of this model, GPT-4 Turbo. > > GPT-4 Turbo is more capable and has knowledge of world events up to April 2023. It has a 128k context window so it can fit the equivalent of more than 300 pages of text in a single prompt. We also optimized its performance so we are able to offer GPT-4 Turbo at a 3x cheaper price for input tokens and a 2x cheaper price for output tokens compared to GPT-4. > > GPT-4 Turbo is available for all paying developers to try by passing gpt-4-1106-preview in the API and we plan to release the stable production-ready model in the coming weeks. Knowledge up to April 2023 is a big game. Cutting the price in half is another big game. A 128k context window retakes the lead on that from Claude-2. That chart from last week of how GPT-4 was slow and expensive, opening up room for competitors? Back to work, everyone. What else? > Function calling updates > Function calling lets you describe functions of your app or external APIs to models, and have the model intelligently choose to output a JSON object containing arguments to call those functions. We’re releasing several improvements today, including the ability to call multiple functions in a single message: users can send one message requesting multiple actions, such as “open the car window and turn off the A/C”, which would previously require multiple roundtrips with the model (learn more). We are also improving function calling accuracy: GPT-4 Turbo is more likely to return the right function parameters. This kind of feature seems highly fiddly and dependent. When it starts working well enough, suddenly it is great, and I have no idea if this will count. I will watch out for reports. For now, I am not trying to interact with any APIs via GPT-4. Use caution. > Improved instruction following and JSON mode > GPT-4 T
7a6f90c7-567b-4e8b-95c4-d8d32ad6f3af	trentmkelly/LessWrong-43k	LessWrong	Bureaucracy is a world of magic I previously wrote about some practical game-theoretical (game-practical?) realizations I had while buying a house. Today I want to talk about how bureaucracy is a ritualistic, magical place. In our home-buying process, every step of the way, there were papers to be signed. Paperwork is how the magic of bureaucracy comes in view. I'm not saying "magic" to mean good or beautiful. I'm referring to the ritualistic nature of bureaucracy. Everything in our journey was a ritual. When you debate the point of something, people participating in the ritual are confused. On the one hand, they understand that your request makes sense, because you're asking for the same function. On the other hand, you shall not ignore the Ritual! Let me explain with several examples what I mean by ritual. The Summoning (of the PDF) To buy a house and get state subsidies, you have to present an official document to the bank, confirming that the building may indeed be used as a dwelling, i.e. a use permit. It is not necessary that this document is an original, a copy will suffice. Well, I got to the bank with printouts of photos of this permit. I don't have the original, and the agent simply took photos of it with his phone, and sent these photos to me. I printed them out on paper, and presented them to the bank. Problem: they have to be scans, not photos. "Photos aren't scans", the bank lady said, "They won't be accepted as official". My first impulse was to protest: "But since you don't need originals, what does it matter what form the copy has? Obviously the informational content is what's necessary - what's written in the document, not what device was used to transfer this information. And anyway, scans and photos are literally the exact same thing. Scans are just photos taken in a particular way. How is it important that-", but I stopped myself before saying any of this. There's a particular art to navigating bureaucracy, and arguing about the nature of information and how it represent
d042a979-8cbb-4eb3-9c32-28ca5e2c339b	trentmkelly/LessWrong-43k	LessWrong	Paper Walkthrough: Automated Circuit Discovery with Arthur Conmy Arthur Conmy's Automated Circuit Discovery is a great paper that makes initial forays into automating parts of mechanistic interpretability (specifically, automatically finding a sparse subgraph for a circuit). In this three part series of Youtube videos, I interview him about the paper, and we walk through it and discuss the key results and takeaways. We discuss the high-level point of the paper and what researchers should takeaway from it, the ACDC algorithm and its key nuances, existing baselines and how they adapted them to be relevant to circuit discovery, how well the algorithm works, and how you can even evaluate how well an interpretability method works.
5e29ea13-5152-4384-91f5-7774e346a63e	trentmkelly/LessWrong-43k	LessWrong	A note on hypotheticals People frequently describe hypothetical situations on LW. Often, other people make responses that suggest they don't understand the purpose of hypotheticals. * When someone puts forth the hypothetical A, it doesn't mean they believe it is true. They may be trying to show not(A). * When someone posits A => B (A implies B), it doesn't mean that they believe A is true. The proposition A => B is commonly used to prove that B is true, or that A is false. * A solution to a hypothetical scenario is useful only if, when you map it back into the original domain, it solves the original problem. I'll expand on the last point. Sorry for being vague. I'm trying not to name names. When a hypothetical is put forward to test a theory, ignore aspects of the hypothetical scenario that don't correspond to parts of the theory. Don't get emotionally involved. Don't think of the hypothetical as a narrative. A hypothetical about Omega sounds a lot like a story about a genie from a lamp, but you should approach it in a completely different way. Don't try to outsmart Omega (unless you're making a point about the impossibility of an Omega who can eg decide undecidable problems). When you find a loophole in the way the hypothetical is posed, that doesn't exist in the original domain, point it out only if you are doing so to improve the phrasing of the hypothetical situation. John Searle's Chinese Room is an example of a hypothetical in which it is important to not get emotionally involved. Searle's conclusion is that the man in the Chinese room doesn't understand Chinese; therefore, a computer doesn't understand Chinese. His model maps the running software onto the complete system of room plus man plus cards; but when he interprets it, he empathizes with the human on each half of the mapping, and so maps the locus of consciousness from the running software onto just the man.1 Sometimes it's difficult to know whether your solution to a hypothetical is exploiting a loophole
682bb25a-2054-40aa-bcc4-f3f67de3a372	trentmkelly/LessWrong-43k	LessWrong	How bad a future do ML researchers expect? Katja Grace, 8 March 2023 In our survey last year, we asked publishing machine learning researchers how they would divide probability over the future impacts of high-level machine intelligence between five buckets ranging from ‘extremely good (e.g. rapid growth in human flourishing)’ to ‘extremely bad (e.g. human extinction).1 The median respondent put 5% on the worst bucket. But what does the whole distribution look like? Here is every person’s answer, lined up in order of probability on that worst bucket: (Column widths may be distorted or columns may be missing due to limitation of chosen software.) And here’s basically that again from the 2016 survey (though it looks like sorted slightly differently when optimism was equal), so you can see how things have changed: Distribution from 2016 survey. (Column widths may be distorted or columns may be missing due to limitation of chosen software.) The most notable change to me is the new big black bar of doom at the end: people who think extremely bad outcomes are at least 50% have gone from 3% of the population to 9% in six years. Here are the overall areas dedicated to different scenarios in the 2022 graph (equivalent to averages): * Extremely good: 24% * On balance good: 26% * More or less neutral: 18% * On balance bad: 17% * Extremely bad: 14% That is, between them, these researchers put 31% of their credence on AI making the world markedly worse. Some things to keep in mind in looking at these: * If you hear ‘median 5%’ thrown around, that refers to how the researcher right in the middle of the opinion spectrum thinks there’s a 5% chance of extremely bad outcomes. (It does not mean, ‘about 5% of people expect extremely bad outcomes’, which would be much less alarming.) Nearly half of people are at ten percent or more. * The question illustrated above doesn’t ask about human extinction specifically, so you might wonder if ‘extremely bad’ includes a lot of scenarios less bad than human extinction. To c
e5a54cc2-4a90-449c-b8e2-dc91aa56a94f	trentmkelly/LessWrong-43k	LessWrong	Feature Hedging: Another way correlated features break SAEs This work was done as part of MATS 7.0. We consider this in-progress research and we are grateful for any thoughts and feedback from the community. Update (May 20, 2025): This is now a paper! Check out our paper "Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders" (arxiv.org/abs/2505.11756). Update (April 4, 2025): We found that hedging caused by positive (but not hierarchical) correlation between features can sometimes be removed with a high enough sparsity penalty, but this is never true for hedging caused by hierarchical features or negatively correlated features. We have updated the post accordingly. Introduction If there is any correlation between a feature captured by an SAE and a feature not captured by that SAE, the SAE will merge the external feature into its latent tracking the internal feature. This phenomenon, which we call feature hedging, is caused by MSE reconstruction loss and will happen any time the SAE is both too narrow to capture all the “true features” in the model, and there is correlation between the features tracked by the SAE and the features not tracked by the SAE. Both of these conditions are almost certainly true of every SAE trained on an LLM, especially narrower SAEs like the inner parts of Matryoshka SAEs. This means all LLM SAE latents likely have spurious components of correlated features mixed in, harming the performance of the SAE and likely contributing to the underwhelming performance of SAEs. We also present a novel reconstruction loss, called snap loss, which solves feature hedging in toy models. However, it does not work if the amount of feature correlation is too high, or there are too many correlated features; so snap loss may not be useful in real LLM SAEs. When an SAE is too narrow to represent two correlated features, MSE loss incentivizes the SAE to mix components of both features into its latent to reduce a large squared-error penalty when the secondary feature fires. This causes the SAE to
74dd61ab-4c17-42d0-8ae1-a1b75049d8a8	trentmkelly/LessWrong-43k	LessWrong	Where to Draw the Boundaries? Followup to: Where to Draw the Boundary? Figuring where to cut reality in order to carve along the joints—figuring which things are similar to each other, which things are clustered together: this is the problem worthy of a rationalist. It is what people should be trying to do, when they set out in search of the floating essence of a word. Once upon a time it was thought that the word "fish" included dolphins ... The one comes to you and says: > The list: {salmon, guppies, sharks, dolphins, trout} is just a list—you can't say that a list is wrong. You draw category boundaries in specific ways to capture tradeoffs you care about: sailors in the ancient world wanted a word to describe the swimming finned creatures that they saw in the sea, which included salmon, guppies, sharks—and dolphins. That grouping may not be the one favored by modern evolutionary biologists, but an alternative categorization system is not an error, and borders are not objectively true or false. You're not standing in defense of truth if you insist on a word, brought explicitly into question, being used with some particular meaning. So my definition of fish cannot possibly be 'wrong,' as you claim. I can define a word any way I want—in accordance with my values! So, there is a legitimate complaint here. It's true that sailors in the ancient world had a legitimate reason to want a word in their language whose extension was {salmon, guppies, sharks, dolphins, ...}. (And modern scholars writing a translation for present-day English speakers might even translate that word as fish, because most members of that category are what we would call fish.) It indeed would not necessarily be helping the sailors to tell them that they need to exclude dolphins from the extension of that word, and instead include dolphins in the extension of their word for {monkeys, squirrels, horses ...}. Likewise, most modern biologists have little use for a word that groups dolphins and guppies together. When rationali
0a2b6709-39df-4810-bc0c-dc632cc72900	trentmkelly/LessWrong-43k	LessWrong	Integrating Three Models of (Human) Cognition You may have heard a few things about “predictive processing” or “the global neuronal workspace,” and you may have read some of Steve Byrnes’ excellent posts about what’s going on computationally in the human brain. But how does it all fit together? How can we begin to arrive at a unified computational picture of human intelligence, and how can it inform our efforts in AI safety and alignment? In this post, I endeavor to take steps towards these ends by integrating insights from three computational frameworks for modeling what’s going on in the human brain, with the hope of laying an adequate foundation for more precisely talking about how the brain cognitively implements goal-directed behaviors: 1. Predictive Processing: how perception, cognition, and action arise from hierarchical probabilistic generative modeling in the brain 2. Global Neuronal Workspace: a functional model of access consciousness, i.e. of information that enters our awareness and may subsequently be reported upon 3. Steve Byrnes’ computational framework: a lot of excellent (in my opinion) work on human cognition, especially motivation, all with an eye toward AI safety. I want to use a better understanding of human cognition to speak more precisely about how things like “goals” and “desires” are implemented, and Steve’s overall framework overlaps quite naturally with the previous two, so I couldn’t leave this one out! This is a somewhat long post, so I will go ahead and flag the structure here, in case some readers want to skim and/or skip to sections that sound most interesting to them: * Motivations for this project * An overview of each of the three frameworks * Here, I tried to assume fairly little previous familiarity with the frameworks and attempted to provide a decent foundation in each, while making sure to emphasize the aspects most relevant for the purposes of this post. * Fitting things together from the three frameworks * How to think about “predictive processing” in t
0e92678b-3690-45b0-b31f-53b3aa2fca86	trentmkelly/LessWrong-43k	LessWrong	Does Google still hire people via their foobar challenge? In years since past, the foobar challenge first appeared. 5 levels of ever harder coding challenges, and if you bested the first three you'd be given the option of providing your details to Google. There was a decent chance you'd be sent an interview request sometime later or , if you applied to them yourself, you're application would be given extra weight. Recently, I got the challenge from Google and after getting to the last problem of level three, I wondered "do they even hire people using this thing anymore"? Some quick searches on medium blogs, quora and hackernews later, I was left with the impression that they don't but no one could cite a source for this. So does anyone have any info, definitive or otherwise, that would tell me and other lost souls whether Google still hires using this thing? Edit: changed the name
41f92565-b70b-4d25-bcaa-8c4ab3db6aeb	trentmkelly/LessWrong-43k	LessWrong	Three Principles to Writing Original Nonfiction If you like reading then you should write too. Three Things Writing Will Do For You 1. Writing makes you smarter. Thoughts are ephemeral, fast and fleeting. Writing is frozen thought. It is easier to determine if an idea that is written down than an idea that is merely thought or spoken. Writing down your ideas shows you whether they have substance. Detecting when you're wrong is the first step to becoming right. There are many ways to record thought. Writing is the best for improving your rationality because there is nothing to hide behind. Writing is the swimsuit of ideas. 2. Writing makes you a better communicator. The skills you learn by writing transfer to speaking. Being good at speaking makes you more persuasive. 3. Writing scales social connections. The best job opportunities I've ever received all came from my writing. The smartest people I've ever met all came from my writing too. What do I write? How do I write? I assume you are proficient in English. Besides a basic mastery of language, good nonfiction has three qualities. 1. It gets to the point. 2. It is interesting. 3. It is grounded in reality. Get to the Point Text is a static medium. The listener of a podcast must keep up with the audio stream. The reader of a blog can take as long as he/she needs to understand an idea. Text is low-density too. In terms of bits, the reader of a text document processes less raw information than the viewer of an image. Text concentrates a reader's attention on a single idea. The reader of a blog tends to think harder than that same person watching a video or conversing in real time. The epistemic rigor is correspondingly higher. We are products of our environment. An author's environment is the written word itself. I write differently from how I talk. I write so differently that readers of this blog are often surprised to discover I'm a kind levelheaded person in real life. Mediums overpower messages. > me before reading your website: [Lsusr] seems li
d9201717-de78-4643-94a3-35a09dc57948	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	AI timelines and theoretical understanding of deep learning I have generally been quite skeptical about the view that we are on the cusp of a revolution that will lead us to artificial general intelligence in the next 50 years so. Aside from fundamental limitations of current AI systems, and f[laws of extrapolating their remarkable ability at narrow tasks towards more general learning by appealing to "exponential" growth](https://medium.com/swlh/is-ai-marching-steadfast-to-human-level-intelligence-d066fb02843a), there is another issue with the discourse on AI that I want to highlight. One of the primary reasons to believe that AGI will happen in the near to mid term future comes from predictions of experts working in the field, the majority of whom seem to think that we will have AGI latest by 2100. While there is every reason to attach credence to their perspective, it should be noted that deep learning, the framework that underpins most of recent developments in AI including language models like BERT and GPT-3 and strategy-game champions like AphaGo are notoriously hard to decipher from a theoretical perspective. It would be a mistake to assume that people who design, develop and deploy these models necessarily understand why they happen to be as successful as they are. This may sound like a rather strange statement to make but the reality is that despite incredible pace of progress across various frontiers of AI with deep learning, our knowledge of why it works -- the mathematical theory of it -- lags behind immensely. To be clear, I am not at all suggesting that research scientists at Google or DeepMind have no knowledge all of why models they design and deploy work. They are certainly guided by various ideas and heuristics when deciding on the loss function, the type of [attention mechanism](https://www.d2l.ai/chapter_attention-mechanisms/multihead-attention.html) to use, the [iterative update to the reward](https://en.wikipedia.org/wiki/Q-learning), the [overall architecture](https://lilianweng.github.io/lil-log/2020/08/06/neural-architecture-search.html) of the network, etc. However, there are two things to note here : first, a lot of the design is based on experimenting with various functional forms, wiring combinations, convolution structure, parameter choices; second, the fact that there are heuristics and high level understanding of what is happening does not imply that there is a first-principles mathematical explanation for it. There are people study the theoretical side of deep learning work towards establishing exact results and also aim to understand why the model training process is so incredibly successful. The progress there has been rather limited, and certainly well behind where the state-of-the-art in terms of performance is. There are a lot of unusual things with deep learning and among them the fact that core concepts in conventional machine learning simply does not seem to apply (such as overfitting). For a more technical view on this, watch this [amazing talk](https://youtu.be/HMdJd2minAI) by Sanjeev Arora where he explains how intriguing deep learning model and training is. This should be contrasted with physics where our understanding of theories is much deeper and fundamental. There is a very precise mathematical framework to characterize the [physics of say, electrons or quarks](https://en.wikipedia.org/wiki/Standard_Model), and, at the other end of the spectrum, a [model to understand cosmology](https://en.wikipedia.org/wiki/Lambda-CDM_model). There is no such thing even remotely comparable to that in deep learning. Given all this, one should be more skeptical about prediction timelines for a qualitatively superior intelligence from experts in this field. The fact that there are considerable gaps in our understanding would suggest that expert opinion is perhaps guided less by some deeper insight into the learning and generalization process of AI models and more by higher level examination of the rapid progress of AI, i.e., their views may be relatively more closer to that of a lay person. Couple this with the fact that we have a very limited understanding of human consciousness and how that is related to the electro-physiological properties of the brain. Such limitations impose considerable challenges to predict with any degree of certainty.
f44ee695-a2fa-4b50-bcd6-2b214a077000	trentmkelly/LessWrong-43k	LessWrong	The ethics of AI for the Routledge Encyclopedia of Philosophy I've been tasked by the Routledge Encyclopedia of Philosophy to write their entry on the ethics of AI. I'll be starting the literature reviews and similar in the coming weeks. Could you draw my attention to any aspect of AI ethics (including the history of AI ethics) that you think is important and deserves to be covered? Cheers!
f92fbcba-539d-428b-ba6b-91a1b2f4bd6a	trentmkelly/LessWrong-43k	LessWrong	What is Randomness? epistemic status: my intuition after reading and watching a bunch of stuff; no new information You take a die in your hand. If you throw it, the result will be what people usually call a random number. Let's say you get 2. What do we mean when we say that this number is random? To answer these questions, I will try to give you some intuition on the concept of randomness. When a person says an outcome of a process (such as the number 2) is random, they mean that they are uncertain what the outcome is. The die is not a magical object that produces random numbers. The die is an object which we can use in multiple situations which can be described as "throwing a die" and those situations could lead to different numbers. Humans are usually incapable of predicting how a die would land - they are uncertain about it. The process of throwing a 6-sided die can produce any of the integers between 1 and 6. In any particular throw, an observer may be uncertain which side is on top. If the observer sees the die after it stops, the uncertainty disappears. The number is no longer random to the observer but may still be random to another observer who hasn't seen the die. For simpler processes such as a coin toss, some humans have learned to predict the outcome. For those humans, there is no uncertainty how the coin would land. Uncertainty is a property experienced by a cognitive system (such as a human or an artificial intelligence). We can distinguish three types of uncertainty. Empirical and Logical Uncertainty Alice takes a die, throws it inside a box, peeks and closes the box before Bob has time to peek inside. How the die landed is determined by the laws of physics. If Bob knew the dimensions of the die and the box, the initial position of the die, the precise movement of Alice's hand and so forth, he would be able to put this information into a computer and calculate how the die landed. Since Bob doesn't have this information, he is uncertain how the die landed. No ma
144575ac-6ee0-416d-b445-189eba264781	trentmkelly/LessWrong-43k	LessWrong	A basic probability question From the logical induction paper https://intelligence.org/files/LogicalInductionAbridged.pdf: Let φ stand for the claim that the 87,653rd digit of π is a 7. If this claim is true, then (1 + 1 = 2) ⇒ φ. I don't understand the second sentence at all, any help appreciated.
11bd6de8-a842-4ba8-bbf0-5558d412946d	trentmkelly/LessWrong-43k	LessWrong	Notes on Honesty This post examines the virtue of honesty (a.k.a. “truthfulness,” “veracity”).[1] As with my other posts in this sequence, I’m less interested in breaking new ground and more in synthesizing whatever wisdom I could find on the subject. I wrote this not as an expert, but as someone who wants to learn. I hope it will help people who want to know more about this virtue and how to nurture it. The topic is complex and it was hard for me to find the sweet spot between being too wordy and too superficial. People have written dense books to try to explain things I cheekily tried to summarize in a sentence or two. Honesty and rationality Much of LessWrong concerns how we can better approach knowing the truth. Honesty concerns an aspect of how we communicate truth. So I think of it as a social virtue rather than an intellectual virtue. Sometimes, however, expressions like “being honest with yourself” describe intellectual virtues. Honesty requires at least a minimum exercise of intellectual virtues. If you do not exercise epistemological due diligence before you communicate your understanding of the world, you may tell the truth as you see it, but you fail to respect the virtue of honesty by not taking enough care to distinguish the false from the true. For example, an acquaintance of mine is very woo. When she tells me of woo things that she thinks are important for me to learn about, I don’t think she’s lying to me, exactly, but she exercises such poor judgement about what to believe that I tend to think of her as being as much a dishonest person as a foolish one. On the other hand, Scott Alexander warns against “lie inflation”[2] in which we accuse people of lying when they are merely honestly representing the poor results of their sloppy thinking. (His argument is something like this: if we expand our definition of dishonesty to cover mistaken sloppy thinking then we risk losing the ability to talk more precisely about deliberately deceptive dishonesty, and this lowe
a5cb8a81-1de9-45a3-b130-bfe9935a0b85	trentmkelly/LessWrong-43k	LessWrong	Meetup : Sydney Rationality Dojo - Optimising Skill Training Discussion article for the meetup : Sydney Rationality Dojo - Optimising Skill Training WHEN: 01 March 2015 04:00:00PM (+1100) WHERE: Humanist House, 10 Shepherd St Chippendale Join us for our next dojo, on making sure you get the most bang for your buck on the time you spend training skills. Afterwards (6pm) there will be a group dinner for those who want to come along. Discussion article for the meetup : Sydney Rationality Dojo - Optimising Skill Training
76f2adb8-3f02-4f5d-b029-3afabe492c51	StampyAI/alignment-research-dataset/arxiv	Arxiv	Causal Confusion in Imitation Learning 1 Introduction --------------- Imitation learning allows for control policies to be learned directly from example demonstrations provided by human experts. It is easy to implement, and reduces or removes the need for extensive interaction with the environment during training Widrow and Smith ([1964](#bib.bib56)); Pomerleau ([1989](#bib.bib39)); Bojarski et al. ([2016](#bib.bib4)); Argall et al. ([2009](#bib.bib1)); Hussein et al. ([2017](#bib.bib18)). However, imitation learning suffers from a fundamental problem: distributional shift Daumé et al. ([2009](#bib.bib8)); Ross and Bagnell ([2010](#bib.bib40)). Training and testing state distributions are different, induced respectively by the expert and learned policies. Therefore, imitating expert actions on expert trajectories may not align with the true task objective. While this problem is widely acknowledged Pomerleau ([1989](#bib.bib39)); Daumé et al. ([2009](#bib.bib8)); Ross and Bagnell ([2010](#bib.bib40)); Ross et al. ([2011](#bib.bib41)), yet with careful engineering, naïve behavioral cloning approaches have yielded good results for several practical problems Widrow and Smith ([1964](#bib.bib56)); Pomerleau ([1989](#bib.bib39)); Schaal ([1999](#bib.bib42)); Muller et al. ([2006](#bib.bib34)); Mülling et al. ([2013](#bib.bib35)); Bojarski et al. ([2016](#bib.bib4)); Mahler and Goldberg ([2017](#bib.bib31)); Bansal et al. ([2019](#bib.bib3)). This raises the question: is distributional shift really still a problem? In this paper, we identify a somewhat surprising and very problematic effect of distributional shift: “causal confusion.” Distinguishing correlates of expert actions in the demonstration set from true causes is usually very difficult, but may be ignored without adverse effects when training and testing distributions are identical (as assumed in supervised learning), since nuisance correlates continue to hold in the test set. However, this can cause catastrophic problems in imitation learning due to distributional shift. This is exacerbated by the causal structure of sequential action: the very fact that current actions cause future observations often introduces complex new nuisance correlates. To illustrate, consider behavioral cloning to train a neural network to drive a car. In scenario A, the model’s input is an image of the dashboard and windshield, and in scenario B, the input to the model (with identical architecture) is the same image but with the dashboard masked out (see Fig [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Causal Confusion in Imitation Learning")). Both cloned policies achieve low training loss, but when tested on the road, model B drives well, while model A does not. The reason: the dashboard has an indicator light that comes on immediately when the brake is applied, and model A wrongly learns to apply the brake only when the brake light is on. Even though the brake light is the effect of braking, model A could achieve low training error by misidentifying it as the cause instead. ![](https://media.arxiv-vanity.com/render-output/7583664/x1.png) Figure 1: Causal confusion: more information yields worse imitation learning performance. Model A relies on the braking indicator to decide whether to brake. Model B instead correctly attends to the pedestrian. This situation presents a give-away symptom of causal confusion: access to more information leads to worse generalization performance in the presence of distributional shift. Causal confusion occurs commonly in natural imitation learning settings, especially when the imitator’s inputs include history information. In this paper, we first point out and investigate the causal confusion problem in imitation learning. Then, we propose a solution to overcome it by learning the correct causal model, even when using complex deep neural network policies. We learn a mapping from causal graphs to policies, and then use targeted interventions to efficiently search for the correct policy, either by querying an expert, or by executing selected policies in the environment. 2 Related Work --------------- Imitation learning. Imitation learning through behavioral cloning dates back to Widrow and Smith, 1964 Widrow and Smith ([1964](#bib.bib56)), and has remained popular through today Pomerleau ([1989](#bib.bib39)); Schaal ([1999](#bib.bib42)); Muller et al. ([2006](#bib.bib34)); Mülling et al. ([2013](#bib.bib35)); Bojarski et al. ([2016](#bib.bib4)); Giusti et al. ([2016](#bib.bib11)); Mahler and Goldberg ([2017](#bib.bib31)); Wang et al. ([2019](#bib.bib54)); Bansal et al. ([2019](#bib.bib3)). The distributional shift problem, wherein a cloned policy encounters unfamiliar states during autonomous execution, has been identified as an issue in imitation learning Pomerleau ([1989](#bib.bib39)); Daumé et al. ([2009](#bib.bib8)); Ross and Bagnell ([2010](#bib.bib40)); Ross et al. ([2011](#bib.bib41)); Laskey et al. ([2017](#bib.bib23)); Ho and Ermon ([2016](#bib.bib17)); Bansal et al. ([2019](#bib.bib3)). This is closely tied to the “feedback” problem in general machine learning systems that have direct or indirect access to their own past states Sculley et al. ([2015](#bib.bib45)); Bagnell ([2016](#bib.bib2)). For imitation learning, various solutions to this problem have been proposed (Daumé et al., [2009](#bib.bib8); Ross and Bagnell, [2010](#bib.bib40); Ross et al., [2011](#bib.bib41)) that rely on iteratively querying an expert based on states encountered by some intermediate cloned policy, to overcome distributional shift; DAgger Ross et al. ([2011](#bib.bib41)) has come to be the most widely used of these solutions. We show evidence that the distributional shift problem in imitation learning is often due to causal confusion, as illustrated schematically in Fig [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Causal Confusion in Imitation Learning"). We propose to address this through targeted interventions on the states to learn the true causal model to overcome distributional shift. As we will show, these interventions can take the form of either environmental rewards with no additional expert involvement, or of expert queries in cases where the expert is available for additional inputs. In expert query mode, our approach may be directly compared to DAgger (Ross et al., [2011](#bib.bib41)): indeed, we show that we successfully resolve causal confusion using orders of magnitude fewer queries than DAgger. We also compare against Bansal et al. ([2019](#bib.bib3)): to prevent imitators from copying past actions, they train with dropout Srivastava et al. ([2014](#bib.bib51)) on dimensions that might reveal past actions. While our approach seeks to find the true causal graph in a mixture of graph-parameterized policies, dropout corresponds to directly applying the mixture policy. In our experiments, our approach performs significantly better. Causal inference. Causal inference is the general problem of deducing cause-effect relationships among variables (Spirtes et al., [2000](#bib.bib50); Pearl, [2009](#bib.bib36); Peters et al., [2017](#bib.bib38); Spirtes, [2010](#bib.bib48); Eberhardt, [2017](#bib.bib9); Spirtes and Zhang, [2016](#bib.bib49)). “Causal discovery” approaches allow causal inference from pre-recorded observations under constraints (Steyvers et al., [2003](#bib.bib52); Heckerman et al., [2006](#bib.bib15); Lopez-Paz et al., [2017](#bib.bib27); Guyon et al., [2008](#bib.bib13); Louizos et al., [2017](#bib.bib28); Maathuis et al., [2010](#bib.bib29); Le et al., [2016](#bib.bib24); Goudet et al., [2017](#bib.bib12); Mitrovic et al., [2018](#bib.bib32); Wang and Blei, [2018](#bib.bib55)). Observational causal inference is known to be impossible in general (Pearl, [2009](#bib.bib36); Peters et al., [2014](#bib.bib37)). We operate in the interventional regime (Tong and Koller, [2001](#bib.bib53); Eberhardt and Scheines, [2007](#bib.bib10); Shanmugam et al., [2015](#bib.bib47); Sen et al., [2017](#bib.bib46)) where a user may “experiment” to discover causal structures by assigning values to some subset of the variables of interest and observing the effects on the rest of the system. We propose a new interventional causal inference approach suited to imitation learning. While ignoring causal structure is particularly problematic in imitation learning, ours is the first effort directly addressing this, to our knowledge. 3 The Phenomenon of Causal Confusion ------------------------------------- In imitation learning, an expert demonstrates how to perform a task (e.g., driving a car) for the benefit of an agent. In each demo, the agent has access both to its n-dim. state observations at each time t, Xt=[Xt1,Xt2,…Xtn] (e.g., a video feed from a camera), and to the expert’s action At (e.g., steering, acceleration, braking). Behavioral cloning approaches learn a mapping π from Xt to At using all (Xt,At) tuples from the demonstrations. At test time, the agent observes Xt and executes π(Xt). ![](https://media.arxiv-vanity.com/render-output/7583664/x2.png) Figure 2: Causal dynamics of imitation. Parents of a node represent its causes. The underlying sequential decision process has complex causal structures, represented in Fig [2](#S3.F2 "Figure 2 ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning"). States influence future expert actions, and are also themselves influenced by past actions and states. In particular, expert actions At are influenced by some information in state Xt, and unaffected by the rest. For the moment, assume that the dimensions Xt1,Xt2,Xt3,… of Xt represent disentangled factors of variation. Then some unknown subset of these factors (“causes”) affect expert actions, and the rest do not (“nuisance variables”). A confounder Zt=[Xt−1,At−1] influences each state variable in Xt, so that some nuisance variables may still be correlated with At among (Xt,At) pairs from demonstrations. In Fig [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Causal Confusion in Imitation Learning"), the dashboard light represents a confounder. A naïve behavioral cloned policy might rely on nuisance correlates to select actions, producing low training error, and even generalizing to held-out (Xt,At) pairs. However, this policy must contend with distributional shift when deployed: actions At are chosen by the imitator rather than the expert, affecting the distribution of Zt and Xt. This in turn affects the policy mapping from Xt to At, leading to poor performance of expert-cloned policies. We define “causal confusion" as the phenomenon whereby cloned policies fail by misidentifying the causes of expert actions. ### 3.1 Robustness and Causality in Imitation Learning Intuitively, distributional shift affects the relationship of the expert action At to nuisance variables, but not to the true causes. In other words, to be maximally robust to distributional shift, a policy must rely solely on the true causes of expert actions, thereby avoiding causal confusion. This intuition can be formalized in the language of functional causal models (FCM) and interventions Pearl ([2009](#bib.bib36)). Functional causal models: A functional causal model (FCM) over a set of variables {Yi}ni=1 is a tuple (G,θG) containing a graph G over {Yi}ni=1, and deterministic functions fi(⋅;θG) with parameters θG describing how the causes of each variable Yi determine it: Yi=fi(YPa(i;G),Ei;θG), where Ei is a stochastic noise variable that represents all external influences on Yi, and Pa(i;G) denote the indices of parent nodes of Yi, which correspond to its causes. An “intervention” do(Yi) on Yi to set its value may now be represented by a structural change in this graph to produce the “mutilated graph” G¯Yi, in which incoming edges to Yi are removed.111For a more thorough overview of FCMs, see Pearl ([2009](#bib.bib36)). Applying this formalism to our imitation learning setting, any distributional shift in the state Xt may be modeled by intervening on Xt, so that correctly modeling the “interventional query” p(At\|do(Xt)) is sufficient for robustness to distributional shifts. Now, we may formalize the intuition that only a policy relying solely on true causes can robustly model the mapping from states to optimal/expert actions under distributional shift. In Appendix [B](#A2 "Appendix B Necessity of Correct Causal Model ‣ Causal Confusion in Imitation Learning"), we prove that under mild assumptions, correctly modeling interventional queries does indeed require learning the correct causal graph G. In the car example, “setting” the brake light to on or off and observing the expert’s actions would yield a clear signal unobstructed by confounders: the brake light does not affect the expert’s braking behavior. ### 3.2 Causal Confusion in Policy Learning Benchmarks and Realistic Settings Before discussing our solution, we first present several testbeds and real-world cases where causal confusion adversely influences imitation learning performance. Control Benchmarks. We show that causal confusion is induced with small changes to widely studied benchmark control tasks, simply by adding more information to the state, which intuitively ought to make the tasks easier, not harder. In particular, we add information about the previous action, which tends to correlate with the current action in the expert data for many standard control problems. This is a proxy for scenarios like our car example, in which correlates of past actions are observable in the state, and is similar to what we might see from other sources of knowledge about the past, such as memory or recurrence. We study three kinds of tasks: (i) MountainCar (continuous states, discrete actions), (ii) MuJoCo Hopper (continuous states and actions), (iii) Atari games: Pong, Enduro and UpNDown (states: two stacked consecutive frames, discrete actions). \| \| \| \| \| --- \| --- \| --- \| \| (a) Pong \| (b) Enduro \| (c) UpNDown \| Figure 3: The Atari environments with indicator of past action (white number in lower left). For each task, we study imitation learning in two scenarios. In scenario A (henceforth called "confounded"), the policy sees the augmented observation vector, including the previous action. In the case of low-dimensional observations, the state vector is expanded to include the previous action at an index that is unknown to the learner. In the case of image observations, we overlay a symbol corresponding to the previous action at an unknown location on the image (see Fig [3](#S3.F3 "Figure 3 ‣ 3.2 Causal Confusion in Policy Learning Benchmarks and Realistic Settings ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning")). In scenario B ("original"), the previous action variable is replaced with random noise for low-dimensional observations. For image observations, the original images are left unchanged. Demonstrations are generated synthetically as described in Appendix [A](#A1 "Appendix A Expert Demonstrations ‣ Causal Confusion in Imitation Learning"). In all cases, we use neural networks with identical architectures to represent the policies, and we train them on the same demonstrations. Fig [4](#S3.F4 "Figure 4 ‣ 3.2 Causal Confusion in Policy Learning Benchmarks and Realistic Settings ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning") shows the rewards against varying demonstration dataset sizes for MountainCar, Hopper, and Pong. Appendix [D](#A4 "Appendix D Additional Results: Diagnosing Causal Confusion ‣ Causal Confusion in Imitation Learning") shows additional results, including for Enduro and UpNDown. All policies are trained to near-zero validation error on held-out expert state-action tuples. original produces rewards tending towards expert performance as the size of the imitation dataset increases. confounded either requires many more demonstrations to reach equivalent performance, or fails completely to do so. Overall, the results are clear: across these tasks, access to more information leads to inferior performance. As Fig [10](#A4.F10 "Figure 10 ‣ Appendix D Additional Results: Diagnosing Causal Confusion ‣ Causal Confusion in Imitation Learning") in the appendix shows, this difference is not due to different training/validation losses on the expert demonstrations—for example, in Pong, confounded produces lower validation loss than original on held-out demonstration samples, but produces lower rewards when actually used for control. These results not only validate the existence of causal confusion, but also provides us with testbeds for investigating a potential solution. Real-World Driving. In more realistic imitation learning settings too, symptoms of causal confusion have been observed consistently in Muller et al. ([2006](#bib.bib34)); Wang et al. ([2019](#bib.bib54)); Bansal et al. ([2019](#bib.bib3)), when learning to drive from histories of video frames. While these histories contain valuable information for driving, they also naturally introduce information about nuisance factors such as previous actions. In all three cases, more information led to worse results for the behavioral cloning policy, but this was neither attributed specifically to causal confusion, nor tackled using causally motivated approaches. \| Metrics → \| Validation \| Driving Performance \| \| --- \| --- \| --- \| \| Methods ↓ \| Perplexity \| Distance \| Interventions \| Collisions \| \| history \| 0.834 \| 144.92 \| 2.94 ± 1.79 \| 6.49 ± 5.72 \| \| no-history \| 0.989 \| 268.95 \| 1.30 ± 0.78 \| 3.38 ± 2.55 \| Table 1: Imitation learning results from Wang et al. ([2019](#bib.bib54)). Accessing history yields better validation performance, but worse actual driving performance. We draw the reader’s attention to particularly telling results from Wang et al. ([2019](#bib.bib54)) for learning to drive in near-photorealistic GTA-V Krähenbühl ([2018](#bib.bib22)) environments, using behavior cloning with DAgger-inspired expert perturbation. Imitation learning policies are trained using overhead image observations with and without “history” information (history and no-history) about the ego-position trajectory of the car in the past. Similar to our tests above, architectures are identical for the two methods. And once again, like in our tests above, history has better performance on held-out demonstration data, but much worse performance when actually deployed. Tab [1](#S3.T1 "Table 1 ‣ 3.2 Causal Confusion in Policy Learning Benchmarks and Realistic Settings ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning") shows these results, reproduced from Wang et al. ([2019](#bib.bib54)) Table II. These results constitute strong evidence for the prevalence of causal confusion in realistic imitation learning settings. Bansal et al. ([2019](#bib.bib3)) also observe similar symptoms in a driving setting, and present a dropout Srivastava et al. ([2014](#bib.bib51)) approach to tackle it, which we compare to in our experiments. \| \| \| \| \| --- \| --- \| --- \| \| (a) MountainCar \| (b) Hopper \| (c) Pong \| Figure 4: Diagnosing causal confusion: net reward (y-axis) vs number of training samples (x-axis) for original and confounded, compared to expert reward (mean and stdev over 5 runs). Also see Appendix [D](#A4 "Appendix D Additional Results: Diagnosing Causal Confusion ‣ Causal Confusion in Imitation Learning"). 4 Resolving Causal Confusion ----------------------------- Recall from Sec [3.1](#S3.SS1 "3.1 Robustness and Causality in Imitation Learning ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning") that robustness to causal confusion can be achieved by finding the true causal model of the expert’s actions. We propose a simple pipeline to do this. First, we jointly learn policies corresponding to various causal graphs (Sec [4.1](#S4.SS1 "4.1 Causal Graph-Parameterized Policy Learning ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning")). Then, we perform targeted interventions to efficiently search over the hypothesis set for the correct causal model (Sec [4.2](#S4.SS2 "4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning")). ### 4.1 Causal Graph-Parameterized Policy Learning ![](https://media.arxiv-vanity.com/render-output/7583664/x6.png) Figure 5: Graph-parameterized policy. In this step, we learn a policy corresponding to each candidate causal graph. Recall from Sec [3](#S3 "3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning") that the expert’s actions A are based on an unknown subset of the state variables {Xi}ni=1. Each Xi may either be a cause or not, so there are 2n possible graphs. We parameterize the structure G of the causal graph as a vector of n binary variables, each indicating the presence of an arrow from Xk to A in Fig [2](#S3.F2 "Figure 2 ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning"). We then train a single graph-parameterized policy πG(X)=fϕ([X⊙G,G]), where ⊙ is element-wise multiplication, and [⋅,⋅] denotes concatenation. ϕ are neural network parameters, trained through gradient descent to minimize: \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| \| EG[ℓ(fϕ([Xi⊙G,G]),Ai)], \| \| (1) \| where G is drawn uniformly at random over all 2n graphs and ℓ is a mean squared error loss for the continuous action environments and a cross-entropy loss for the discrete action environments. Fig [5](#S4.F5 "Figure 5 ‣ 4.1 Causal Graph-Parameterized Policy Learning ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning") shows a schematic of the training time architecture. The policy network fϕ mapping observations X to actions A represents a mixture of policies, one corresponding to each value of the binary causal graph structure variable G, which is sampled as a bernoulli random vector. In Appendix [C](#A3 "Appendix C Variational Causal Discovery ‣ Causal Confusion in Imitation Learning"), we propose an approach to perform variational Bayesian causal discovery over graphs G, using a latent variable model to infer a distribution over functional causal models (graphs and associated parameters)—the modes of this distribution are the FCMs most consistent with the demonstration data. This resembles the scheme above, except that instead of uniform sampling, graphs are sampled preferentially from FCMs that fit the training demonstrations well. We compare both approaches in Sec [5](#S5 "5 Experiments ‣ Causal Confusion in Imitation Learning"), finding that simple uniform sampling nearly always suffices in preparation for the next step: targeted intervention. ### 4.2 Targeted Intervention Having learned the graph-parameterized policy as in Sec [4.1](#S4.SS1 "4.1 Causal Graph-Parameterized Policy Learning ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"), we propose targeted intervention to compute the likelihood L(G) of each causal graph structure hypothesis G. In a sense, imitation learning provides an ideal setting for studying interventional causal learning: causal confusion presents a clear challenge, while the fact that the problem is situated in a sequential decision process where the agent can interact with the world provides a natural mechanism for carrying out limited interventions. We propose two intervention modes, both of which can be carried out by interaction with the environment via the actions: Expert query mode. This is the standard intervention approach applied to imitation learning: intervene on Xt to assign it a value, and observe the expert response A. This requires an interactive expert, as in DAgger Ross and Bagnell ([2010](#bib.bib40)), but requires substantially fewer expert queries than DAgger, both because: (i) the queries serve only to disambiguate among a relatively small set of valid FCMs, and (ii) we use disagreement among the mixture of policies in fϕ to query the expert efficiently in an active learning approach. We summarize this approach in Algorithm [1](#alg1 "Algorithm 1 ‣ 4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"). Algorithm 1 Expert query intervention Input: policy network fϕ s.t. πG(X)=fϕ([X⊙G,G]) Initialize w=0,D=∅. Collect states S by executing πmix, the mixture of policies πG for uniform samples G. For each X in S, compute disagreement score: . . . D(X)=EG[DKL(πG(X),πmix(X))] Select S′⊂S with maximal D(X). Collect state-action pairs T by querying expert on S′. for i=1…N do Sample G∼p(G)∝exp⟨w,G⟩. L←Es,a∼T[ℓ(πG(s),a)] D←D∪{(G,L)} Fit w on D with linear regression. end for Return: argmaxGp(G) Algorithm 2 Policy execution intervention Input: policy network fϕ s.t. πG(X)=fϕ([X⊙G,G]) Initialize w=0,D=∅. for i=1…N do Sample G∼p(G)∝exp⟨w,G⟩. Collect episode return RG by executing πG. D←D∪{(G,RG)} Fit w on D with linear regression. end for Return: argmaxGp(G) Policy execution mode. It is not always possible to query an expert. For example, for a learner learning to drive a car by watching a human driver, it may not be possible to put the human driver into dangerous scenarios that the learner might encounter at intermediate stages of training. In cases like these where we would like to learn from pre-recorded demonstrations alone, we propose to intervene indirectly by using environmental returns (sum of rewards over time in an episode) R=∑trt. The policies πG(⋅)=fϕ([⋅⊙G,G]) corresponding to different hypotheses G are executed in the environment and the returns RG collected. The likelihood of each graph is proportional to the exponentiated returns expRG. The intuition is simple: environmental returns contain information about optimal expert policies even when experts are not queryable. Note that we do not even assume access to per-timestep rewards as in standard reinforcement learning; just the sum of rewards for each completed run. As such, this intervention mode is much more flexible. See Algorithm [2](#alg2 "Algorithm 2 ‣ 4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"). Note that both of the above intervention approaches evaluate individual hypotheses in isolation, but the number of hypotheses grows exponentially in the number of state variables. To handle larger states, we infer a graph distribution p(G), by assuming an energy based model with a linear energy E(G)=⟨w,G⟩+b, so the graph distribution is p(G)=∏ip(Gi)=∏iBernoulli(Gi\|σ(wi/τ)), where σ is the sigmoid, which factorizes in independent factors. The independence assumption is sensible as our approach collapses p(G) to its mode before returning it and the collapsed distribution is always independent. E(G) is inferred from linear regression on the likelihoods. This process is depicted in Algorithms [1](#alg1 "Algorithm 1 ‣ 4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning") and [2](#alg2 "Algorithm 2 ‣ 4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"). The above method can be formalized within the reinforcement learning framework Levine ([2018](#bib.bib25)). As we show in Appendix [G](#A7 "Appendix G Intervention Posterior Inference as Reinforcement Learning ‣ Causal Confusion in Imitation Learning"), the energy-based model can be seen as an instance of soft Q-learning Haarnoja et al. ([2017](#bib.bib14)). ### 4.3 Disentangling Observations In the above, we have assumed access to disentangled observations Xt. When this is not the case, such as with image observations, Xt must be set to a disentangled representation of the observation at time t. We construct such a representation by training a β-VAE Kingma and Welling ([2013](#bib.bib20)); Higgins et al. ([2017](#bib.bib16)) to reconstruct the original observations. To capture states beyond those encountered by the expert, we train with a mix of expert and random trajectory states. Once trained, Xt is set to be the mean of the latent distribution produced at the output of the encoder. The VAE training objective encourages disentangled dimensions in the latent space Burgess et al. ([2018](#bib.bib5)); Chen et al. ([2018](#bib.bib6)). We employ CoordConv Liu et al. ([2018](#bib.bib26)) in both the encoder and the decoder architectures. 5 Experiments -------------- We now evaluate the solution described in Sec [4](#S4 "4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning") on the five tasks (MountainCar, Hopper, and 3 Atari games) described in Sec [3.2](#S3.SS2 "3.2 Causal Confusion in Policy Learning Benchmarks and Realistic Settings ‣ 3 The Phenomenon of Causal Confusion ‣ Causal Confusion in Imitation Learning"). In particular, recall that confounded performed significantly worse than original across all tasks. In our experiments, we seek to answer the following questions: (1) Does our targeted intervention-based solution to causal confusion bridge the gap between confounded and original? (2) How quickly does performance improve with intervention? (3) Do both intervention modes (expert query, policy execution) described in Sec [4.2](#S4.SS2 "4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning") resolve causal confusion? (4) Does our approach in fact recover the true causal graph? In each of the two intervention modes, we compare two variants of our method: unif-intervention and disc-intervention. They only differ in the training of the graph-parameterized mixture-of-policies fϕ—while unif-intervention samples causal graphs uniformly, disc-intervention uses the variational causal discovery approach mentioned in Sec [4.1](#S4.SS1 "4.1 Causal Graph-Parameterized Policy Learning ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"), and described in detail in Appendix [C](#A3 "Appendix C Variational Causal Discovery ‣ Causal Confusion in Imitation Learning"). \| \| \| \| \| \| --- \| --- \| --- \| --- \| \| Environment \| Pong \| Enduro \| UpNDown \| \| original (upper bd) \| 15.0 \| 39.5 \| 80.9 \| \| confounded (lower bd) \| -4.0 \| 30.5 \| 24.8 \| \| original w/ vae \| 12.3 \| 36.7 \| 54.5 \| \| counfounded w/ vae \| -4.0 \| 28.2 \| 24.0 \| \| unif-intervention (ours) \| 11.6 \| 32.4 \| 66.3 \| \| dropout Bansal et al. ([2019](#bib.bib3)) \| -8.3 \| 28.2 \| 40.4 \| Table 2: Intervention by policy execution: Reward of the best models produced by our approach on Atari games. unif-intervention succeeds in getting rewards close to original w/ vae, while the dropout baseline only outperforms confounded w/ vae in UpNDown. Baselines. We compare our method against three baselines applied to the confounded state. dropout trains the policy using Eq [3](#A3.E3 "(3) ‣ Appendix C Variational Causal Discovery ‣ Causal Confusion in Imitation Learning") and evaluates with the graph G containing all ones, which amounts to dropout regularization Srivastava et al. ([2014](#bib.bib51)) during training, as proposed by Bansal et al. ([2019](#bib.bib3)). dagger Ross and Bagnell ([2010](#bib.bib40)) addresses distributional shift by querying the expert on states encountered by the imitator, requiring an interactive expert. We compare dagger to our expert query intervention approach. Lastly, we compare to Generative Adversarial Imitation Learning (gail) Ho and Ermon ([2016](#bib.bib17)). gail is an alternative to standard behavioral cloning that works by matching demonstration trajectories to those generated by the imitator during roll-outs in the environment. Note that the PC algorithm Le et al. ([2016](#bib.bib24)), commonly used in causal discovery from passive observational data, relies on the faithfulness assumption, which causes it to be infeasible in our setting. See Appendices [B](#A2 "Appendix B Necessity of Correct Causal Model ‣ Causal Confusion in Imitation Learning") & [C](#A3 "Appendix C Variational Causal Discovery ‣ Causal Confusion in Imitation Learning") for details. \| \| \| \| --- \| --- \| \| \| \| Figure 6: Reward vs. number of intervention episodes (policy execution interventions). Our methods unif-intervention and disc-intervention bridge most of the causal confusion gap (between original (lower bound) and confounded (upper bound), approaching original performance after tens of episodes. gail Ho and Ermon ([2016](#bib.bib17)) (on Hopper) achieves this too, but after 1.5k episodes. Intervention by policy execution. Fig [6](#S5.F6 "Figure 6 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning") plots episode rewards versus number of policy execution intervention episodes for MountainCar and Hopper. The reward always corresponds to the current mode argmaxGp(G) of the posterior distribution over graphs, updated after each episode, as described in Algorithm [2](#alg2 "Algorithm 2 ‣ 4.2 Targeted Intervention ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"). In these cases, both unif-intervention and disc-intervention eventually converge to models yielding similar rewards, which we verified to be the correct causal model i.e., true causes are selected and nuisance correlates left out. In early episodes on MountainCar, disc-intervention benefits from the prior over graphs inferred in the variational causal discovery phase. However, in Hopper, the simpler unif-intervention performs just as well. dropout does indeed help in both settings, as reported in Bansal et al. ([2019](#bib.bib3)), but is significantly poorer than our approach variants. gail requires about 1.5k episodes on Hopper to match the performance of our approaches, which only need tens of episodes. Appendix [F](#A6 "Appendix F GAIL Training Curves ‣ Causal Confusion in Imitation Learning") further analyzes the performance of gail. Standard implementations of gail do not handle discrete action spaces, so we do not evaluate it on MountainCar. \| \| \| \| --- \| --- \| \| \| \| Figure 7: Reward vs. expert queries (expert query interventions). Our methods partially bridge the gap from confounded (lower bd) to original (upper bd), also outperforming dagger Ross et al. ([2011](#bib.bib41)) and dropout Bansal et al. ([2019](#bib.bib3)). gail Ho and Ermon ([2016](#bib.bib17)) outperforms our methods on Hopper, but requires a large number of policy roll-outs (also see Fig [6](#S5.F6 "Figure 6 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning") comparing gail to our policy execution-based approach). Experiments on Atari games are more computationally expensive, so we report results after a heuristically pre-selected number of episodes (1000). As described in Sec [4.3](#S4.SS3 "4.3 Disentangling Observations ‣ 4 Resolving Causal Confusion ‣ Causal Confusion in Imitation Learning"), we use a VAE to disentangle image states in Atari games to produce 30-D representations. Requiring the policy to utilize the VAE representation without end-to-end training does result in some drop in performance, as seen in Tab 1. However, causal confusion still causes a very large drop of performance even relative to the baseline VAE performance. As Tab [2](#S5.T2 "Table 2 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning") shows, unif-intervention indeed improves significantly over confounded w/ vae in all three cases, matching original w/ vae on Pong and UpNDown, while the dropout baseline only improves UpNDown. In our experiments thus far, gail fails to converge to above-chance performance on any of the Atari environments. These results show that our method successfully alleviates causal confusion within relatively few trials. Intervention by expert queries. Next, we perform direct intervention by querying the expert on samples from trajectories produced by the different causal graphs. In this setting, we can also directly compare to dagger Ross et al. ([2011](#bib.bib41)). Fig [7](#S5.F7 "Figure 7 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning") shows results on MountainCar and Hopper. Both our approaches successfully improve over confounded within a small number of queries. Consistent with policy execution intervention results reported above, we verify that our approach again identifies the true causal model correctly in both tasks, and also performs better than dropout in both settings. It also exceeds the rewards achieved by dagger, while using far fewer expert queries. In Appendix [E](#A5 "Appendix E DAgger with many more interventions ‣ Causal Confusion in Imitation Learning"), we show that dagger requires hundreds of queries to achieve similar rewards for MountainCar and tens of thousands for Hopper. Finally, gail with 1.5k episodes outperforms our expert query interventions approach. Recall however from Fig [7](#S5.F7 "Figure 7 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning") that this is an order of magnitude more than the number of episodes required by our policy intervention approach. Once again, disc-intervention only helps in early interventions on MountainCar, and not at all on Hopper. Thus, our method’s performance is primarily attributable to the targeted intervention stage, and the exact choice of approach used to learn the mixture of policies is relatively insignificant. Overall, of the two intervention approaches, policy execution converges to better final rewards. Indeed, for the Atari environments, we observed that expert query interventions proved ineffective. We believe this is because expert agreement is an imperfect proxy for true environmental rewards. \| \| \| \| --- \| --- \| \| \| \| Figure 8: Samples from (top row) learned causal graph and (bottom row) random causal graph. (See text) Interpreting the learned causal graph. Our method labels each dimension of the VAE encoding of the frame as a cause or nuisance variable. In Fig [8](#S5.F8 "Figure 8 ‣ 5 Experiments ‣ Causal Confusion in Imitation Learning"), we analyze these inferences in the Pong environment as follows: in the top row, a frame is encoded into the VAE latent, then for all nuisance dimensions (as inferred by our approach unif-intervention), that dimension is replaced with a sample from the prior, and new samples are generated. In the bottom row, the same procedure is applied with a random graph that has as many nuisance variables as the inferred graph. We observe that in the top row, the causal variables (the ball and paddles) are shared between the samples, while the nuisance variables (the digit) differ, being replaced either with random digits or unreadable digits. In the bottom row, the causal variables differ strongly, indicating that important aspects of the state are judged as nuisance variables. This validates that, consistent with MountainCar and Hopper, our approach does indeed identify true causes in Pong. 6 Conclusions -------------- We have identified a naturally occurring and fundamental problem in imitation learning, “causal confusion”, and proposed a causally motivated approach for resolving it. While we observe evidence for causal confusion arising in natural imitation learning settings, we have thus far validated our solution in somewhat simpler synthetic settings intended to mimic them. Extending our solution to work for such realistic scenarios is an exciting direction for future work. Finally, apart from imitation, general machine learning systems deployed in the real world also encounter “feedback” Sculley et al. ([2015](#bib.bib45)); Bagnell ([2016](#bib.bib2)), which opens the door to causal confusion. We hope to address these more general settings in the future. #### Acknowledgments: We would like to thank Karthikeyan Shanmugam and Shane Gu for pointers to prior work early in the project, and Yang Gao, Abhishek Gupta, Marvin Zhang, Alyosha Efros, and Roberto Calandra for helpful discussions in various stages of the project. We are also grateful to Drew Bagnell and Katerina Fragkiadaki for helpful feedback on an earlier draft of this paper. This project was supported in part by Berkeley DeepDrive, NVIDIA, and Google.
26482b29-be67-40c4-ac36-4b9dea51def1	trentmkelly/LessWrong-43k	LessWrong	How much funding and researchers were in AI, and AI Safety, in 2018? I'm trying to build up a picture of how "much" research is going into general AI capabilities, and how much is going into AI safety. The ideal question I'd be asking is "how much progress [measured in "important thoughts/ideas/tools" was being made that plausibly could lead to AGI in 2018, and how much progress was made that could plausibly lead to safe/aligned AI]. I assume that question is nigh impossible, so instead asking the approximation: a) how much money went into AI capabilities research in 2018 b) how much money went into AI alignment research in 2018 c) how many researchers (ideally "research hours" but I'll take what I can get) were focused on capabilities research in 2018 d) how many researchers were focused on AI safety in 2018?
ac6f8d2b-e118-44f0-85f2-bf9029157695	trentmkelly/LessWrong-43k	LessWrong	[SEQ RERUN] Spooky Action at a Distance Today's post, Spooky Action at a Distance: The No-Communication Theorem was originally published on 05 May 2008. A summary (taken from the LW wiki): > As Einstein argued long ago, the quantum physics of his era - that is, the single-global-world interpretation of quantum physics, in which experiments have single unique random results - violates Special Relativity; it imposes a preferred space of simultaneity and requires a mysterious influence to be transmitted faster than light; which mysterious influence can never be used to transmit any useful information. Getting rid of the single global world dispels this mystery and puts everything back to normal again. Discuss the post here (rather than in the comments to the original post). This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Bell's Theorem: No EPR "Reality", and you can use the sequence_reruns tag or rss feed to follow the rest of the series. Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series.

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.