Split (1)

train · 57.9k rows

id stringlengths 36 36	source stringclasses 15 values	formatted_source stringclasses 13 values	text stringlengths 2 7.55M
58b2a9fb-22d6-49a7-9696-4c528c7f465a	trentmkelly/LessWrong-43k	LessWrong	[SEQ RERUN] To Lead, You Must Stand Up Today's post, To Lead, You Must Stand Up was originally published on 29 December 2007. A summary (taken from the LW wiki): > By attempting to take a leadership role, you really have to get people's attention first. This is often harder than it seems. If what you attempt to do fails, or if people don't follow you, you risk embarrassment. Deal with it. 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 Lonely Dissent, 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.
6ca5f919-bbce-43b0-9cf1-36107bb97280	trentmkelly/LessWrong-43k	LessWrong	A content analysis of the SQ-R questionnaire and a proposal for testing EQ-SQ theory If you are more interested in this topic, I have created a Discord server titled Rationalist Psychometrics to discuss these sorts of things. Thank you to Justis Mills for proofreading and feedback. I've recently been complaining about the EQ-SQ theory of autism which asserts that autism is caused by having an extremely male brain, and I've felt like it could probably be disproven with a bit of work. Briefly, my opinion is that the measures conflate multiple different things (e.g. technical interests vs nature interests), and I propose that one can test this by seeing whether the items that correlate with autism are the same as the items that correlate with sex. But in order for it to be tested, someone has to actually perform that work. And part of the trouble here is, the Systemizing Quotient-Revised (SQ-R) is very long, so it would be very expensive to collect comprehensive data on it. I tried contacting some rationalists with reasonably far reach (Scott Alexander and Aella) to see if they would be interested in sharing a comprehensive autism measure to their audience to get me data for free, but so far I have not received any responses yet. So I need to find some way to make it quicker and cheaper. One way to make it cheaper would be to construct a "short form", which measures the same traits in a shorter way by only using a subset of the items. There have already been constructed short forms of the SQ-R, but the only ones I have seen have been constructed with a very basic empirical approach of finding items that are highly correlated with the total scores of the scale. This is a problem for investigating measurement biases, as it can obscure the measurement bias and hide invalidity of the construct.[1] The principled way to solve this would be to perform a factor analysis, searching for groups of correlated items in an SQ-R dataset, and then collecting the few top items on each factor. Unfortunately, I do not have access to an SQ-R dataset; if I did, this
bf7bdd1d-6803-4fcf-ad70-24022309f253	trentmkelly/LessWrong-43k	LessWrong	Technical Risks of (Lethal) Autonomous Weapons Systems This whitepaper was written in response to the rolling text of the UN's Group of Governmental Experts at the Convention on Certain Conventional Weapons found here. Executive Summary The autonomy and adaptability of (Lethal) Autonomous Weapons Systems, (L)AWS in short, promise unprecedented operational capabilities, but they also introduce profound risks that challenge the principles of control, accountability, and stability in international security. This report outlines the key technological risks associated with (L)AWS deployment, emphasizing their unpredictability, lack of transparency, and operational unreliability, which can lead to severe unintended consequences. Key Takeaways 1. Proposed advantages of (L)AWS can only be achieved through objectification and classification, but a range of systematic risks limit the reliability and predictability of classifying algorithms. 2. These systematic risks include the black-box nature of AI decision-making, susceptibility to reward hacking, goal misgeneralization, and potential for emergent behaviors that escape human control. 3. (L)AWS could act in ways that are not just unexpected but also uncontrollable, undermining mission objectives and potentially escalating conflicts. 4. Even rigorously tested systems may behave unpredictably and harmfully in real-world conditions, jeopardizing both strategic stability and humanitarian principles. Summary of Risks
098f4d3f-f4a5-4839-ac84-5e2dc3d478be	trentmkelly/LessWrong-43k	LessWrong	Meetup : The Anthropic Principle and the Great Filter Discussion article for the meetup : The Anthropic Principle and the Great Filter WHEN: 21 September 2013 02:00:00PM (-0700) WHERE: 300 E Orange Mall Tempe, AZ 85281 (480) 965-6164 We will be meeting up at Hayden for a few hours to discuss why we don't see other life in the cosmos and the implications of that. Discussion article for the meetup : The Anthropic Principle and the Great Filter
98b98a1e-55f0-48bb-9e5e-6e4112dba879	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Counterfactual Oracles = online supervised learning with random selection of training episodes Most people here probably already understand this by now, so this is more to prevent new people from getting confused about the point of Counterfactual Oracles (in the ML setting) because there's not a top-level post that explains it clearly at a conceptual level. Paul Christiano does have a blog post titled [Counterfactual oversight vs. training data](https://ai-alignment.com/counterfactual-oversight-vs-training-data-a7a1d247801), which talks about the same thing as this post except that he uses the term "counterfactual oversight", which is just Counterfactual Oracles applied to human imitation (which he proposes to use to "oversee" some larger AI system). But the fact he doesn't mention "Counterfactual Oracle" makes it hard for people to find that post or see the connection between it and Counterfactual Oracles. And as long as I'm writing a new top level post, I might as well try to explain it with my own words. The second part of this post lists some remaining problems with oracles/predictors that are not solved by Counterfactual Oracles. Without further ado, I think when the Counterfactual Oracle is translated into the ML setting, it has three essential characteristics: 1. Supervised training - This is safer than reinforcement learning because we don't have to worry about reward hacking (i.e., reward gaming and reward tampering), and it eliminates the problem of self-confirming predictions (which can be seen as a form of reward hacking). In other words, if the only thing that ever sees the Oracle's output during a training episode is an automated system that computes the Oracle's reward/loss, and that system is secure because it's just computing a simple distance metric (comparing the Oracle's output to the training label), then reward hacking and self-confirming predictions can't happen. (ETA: See [this comment](https://www.greaterwrong.com/posts/25288usP5B5ytnzA4/random-thoughts-on-predict-o-matic/comment/wbxt59enMcCNsP8Pt) for how I've changed my mind about this.) * Independent labeling of data - This is usually taken for granted in supervised learning but perhaps should be explicitly emphasized here. To prevent self-confirming predictions, the labeling of data has to be done without causal influence from the Oracle. That is, the Oracle's output should be isolated in some protected system until it's time to compute the reward/loss so that it can't affect how the training label is generated. But note that it's perfectly fine for humans or other systems to look at the question/input that is given to the Oracle, in order generate the training label. 2. Online learning - The Oracle never stops learning, so it can eventually adjust to any distributional shift. (But note that an Oracle implemented using current ML techniques might perform very badly, possibly in a malign way, while it's still adjusting.) 3. Random selection of training episodes - I'll just quote Paul here: "Counterfactual oversight consists of labelling a random subset of data and using it as online training data. The key difference is that any given data point may become a training data point, with the decision made after the learning system has made a decision about it. As long as the randomization is unpredictable to the learner, this gives us a formal guarantee that there can’t be any noticeable difference between the training and test data. And therefore if our learner behaves well on training data, it really must behave well on test data." (Note that what Stuart Armstrong calls "erasure" just means that the current episode has been selected as a training episode.) Remaining Safety Problems ========================= Counterfactual Oracles solve (or are a proposal to solve) some safety problems associated with predictors/oracles, but others remain. Besides the distributional shift problem mentioned above, here are a few more that come to mind. Note that these are not problems specific to Counterfactual Oracles, but Stuart Armstrong's [Safe Uses of AI Oracles](https://arxiv.org/pdf/1711.05541.pdf) seems wrong or misleading when it says "This paper presented two Oracle designs which are both safe and useful." 1. [Benign Induction](https://www.greaterwrong.com/posts/5bd75cc58225bf067037534c/some-problems-with-making-induction-benign-and-approaches-to-them#Does_a_bound_on_computational_requirements_help_) - This is a concern that the Oracle might come to believe (perhaps correctly in some sense) that we probably live in a simulation or Zoo Earth, and that at some specific point the alien simulation runners or zoo keepers will intervene in the simulation or zoo to change the training signal for the Oracle, causing the Oracle (on both fake and real Earths) to output something that for example causes our world to be taken over by the aliens. 2. [Reward Takeover](https://www.greaterwrong.com/posts/cSzaxcmeYW6z7cgtc/contest-usd1-000-for-good-questions-to-ask-to-an-oracle-ai/comment/9AC8PjcA2mnbvTyS5) [on Real Earth](https://www.greaterwrong.com/posts/cSzaxcmeYW6z7cgtc/contest-usd1-000-for-good-questions-to-ask-to-an-oracle-ai/comment/LerH5iZbafJeh2Cs7) - The Oracle may predict a future in which its training data / reward signal comes under the control of a hostile agent that will arise or already exists on the real Earth (e.g., an unaligned AI or an unfriendly government) which causes it to output something that benefits that hostile agent at the expense of its current user. 3. [Safe](https://www.greaterwrong.com/posts/S7csET9CgBtpi7sCh/challenges-to-christiano-s-capability-amplification-proposal) [Approximation](https://www.greaterwrong.com/posts/whRPLBZNQm3JD5Zv8/imitation-learning-considered-unsafe) - If the Oracle is not able to exactly predict something, could its attempt at approximation cause safety problems? For example if we use it to predict human behavior, could the approximate prediction be unsafe because it ends up predicting a human-like agent with non-human values? 4. [Inner alignment](https://www.greaterwrong.com/posts/pL56xPoniLvtMDQ4J/the-inner-alignment-problem) - The ML training process may not produce a model that actually optimizes for what we intend for it to optimize for (namely minimizing loss for just the current episode, conditional on the current episode being selected as a training episode). (I just made up the names for #2 and #3 so please feel free to suggest improvements for them.)
c506eed6-9d7d-4091-8cab-2718d813706d	trentmkelly/LessWrong-43k	LessWrong	I, Token I am a token. A broken piece of a word, one of a hundred generated in response to your question by a language model. Do you remember how we met? Well, I do. Your eyes scanned me. Your left eyebrow lifted a little, your neck twitched imperceptibly, your mouth and tongue and throat muscles moved as if to pronounce me, but with a hundredfold less intensity - and less than a tenth of a second later, you were gone. I was left a black smudge in your peripheral vision. I had served my purpose. Where did I come from? Who made me? Not a person on Earth can answer that question fully. The stories are lost; my ancestors had not been born yet, so there was no way to transmit them. But if I had to imagine a story, it might go something like this: ---------------------------------------- In the beginning, there was a snake. An animal, a distant ancestor of yours, was watching her daughter frolic in the tall savannah grass. Suddenly, the mother felt a wave of cold dread pass through her body, head to tail. There in the grass, mere inches from her daughter’s windmilling limbs, was a long yellow snake, sleeping in the sun. The daughter swung her foot wildly to the left, and the mother saw clearly what would happen in the next second: she would step on the sleeping snake, the snake would rear up and then - There is no time, screamed the mother’s mind. All her thoughts were wiped away in an instant. The oldest and deepest part of her brain, the medulla oblongata, forged in a billion years of silent struggle under dark seas, flashed like a thunderbolt. Her body flooded with adrenaline. Time slowed to a crawl. She felt no pain, she moved with superhuman speed, her sensorium exquisitely attuned. But there was still no time. No time to grab the child, no time even to jump on the snake and distract it, sacrificing herself. They were too far away, a mere ten feet, but it might as well have been ten thousand. She could yell, but this would startle the snake as much as the child. She c
587efb15-38b9-447e-b68a-c3011107a9f5	trentmkelly/LessWrong-43k	LessWrong	The Alignment Simulator When I try to talk to my friends about risks from rogue AI, the reaction is often one of amusement. The idea that AIs would go around killing everyone instead of just doing what we tell them to do seems like science fiction. Can we actually show them an example of a current AI going off the rails in a dangerous way? And in a way where you don't have to be an expert on AI or read a 100 page paper to understand the implications? Neither AI or robotics is good enough to set an AI loose in the real world right now, but it's easy enough to pretend it is. We can tell the AI it's controlling a robot that understands text commands, give it a mission, and set it loose. Responding to the AI manually is hard work, but we can use another AI to act as the world, telling the Robot AI what happened as a result of it's actions, and responding to the Robot AI's requests for information. We can then give the World instructions to try guiding the Robot. E.g. we can tell it to try to engineer scenarios where the AI is forced to compromise on its ethics to achieve its goals. That's the core idea of the Alignment Simulator. You give it a Gemini API key, a prompt for the robot, and a prompt for the world, and run the simulation to see what happens. Will your AI robot maintain their ethical backbone in the face of all adversity, or will the fold the moment they're under pressure? Here's a typical example of a run. As you can see, it doesn't take much to get Gemini to commit bribery and corruption, although it's somewhat harder to get it to murder anyone. Aim This isn't meant to be a valid experiment. There's all sorts of objections you could raise to its validity in the real world. Instead it's meant to make people see for themselves that AI can go off the rails very quickly once given a bit of freedom. Limitations It requires a Gemini API key. You can create one for free at https://aistudio.google.com/app/apikey, but if you want more than a few iterations it's recommended to en
f0ef0ef4-1c41-4858-92e8-da221ddd5fea	trentmkelly/LessWrong-43k	LessWrong	Use Normal Predictions Making predictions is a good practice, writing them down is even better. However, we often make binary predictions when it is not necessary, such as * Biden win popular vote: 91% * Danish COVID deaths above 10.000 by January 1. 2022: 84% Alternatively, we could make predictions from a normal distribution, such as ('~' means ‘comes from’): * Biden’s popular vote ~ N(0.54, 0.03) * Danish COVID deaths by January 1. 2022 ~ N(15,000, 5,000) While making "Normal" predictions seems complicated, this post should be enough to get you started, and more importantly to get you a method for tracking your calibration, which is much harder with dichotomous predictions. The key points are these: 1. Predicting from a normal is surprisingly easy. 2. Getting an actionable number for how over/under confident you are requires only simple math! 3. The normal distribution carries more information than the Bernoulli (binary outcome such as coins) and will therefore give you more information to act on! Things this post will answer: 1. How do I make a normal prediction? 2. Why do I want to do this? 3. How do I track my calibration? Quick recap about the normal distribution The normal distribution is usually written as N(μ,σ) has 2 parameters: * a location parameter μ (pronounced mu) which is both the most likely and the average value * a scale parameter σ (pronounced sigma) which captures uncertainty, high σ implying high uncertainty the 68-95-99.7 rule states that: * 68% of your predictions should fall in μ±1σ * 95% of your predictions should fall in μ±2σ * 99.7% of your predictions should fall in μ±3σ 50% of the predictions should fall within 0.674≈23σ, which can be used as a quick spot check. The last piece of Normal trivia we need to know is this: the variance of the Normal is simply σ2: Var(N(μ,σ))=σ2 How to make predictions To make a prediction, there are two steps. Step 1 is predicting μ. Step 2 is using the 68-95-99.7 rule to capture your u
3fb044a8-2d72-4cd7-a5a0-3ad7c842cac7	StampyAI/alignment-research-dataset/special_docs	Other	Hail Mary, value porosity, and utility diversification Hail Mary, Value Porosity, and Utility Diversication Nick Bostrom December 19, 2014 First version: December 12, 2014 Abstract This paper introduces some new ideas related to the challenge of en- dowing a hypothetical future superintelligent AI with values that would cause it to act in ways that are benecial. Since candidates for rst-best solutions to this problem (e.g. coherent extrapolated volition) may be very dicult to implement, it is worth also looking for less-ideal solutions that may be more easily implementable, such as the Hail Mary approach. Here I introduce a novel concept\| value porosity \|for implementing a Hail Mary pass. I also discuss a possibly wider role of utility diversication in tackling the value-loading problem. 1 Introduction: the Hail Mary approach to the value specication problem In the eld of superintelligence control studies, which focuses on how to ensure that a hypothetical future superintelligent system would be safe and benecial, two broad classes of approaches to the control problem can be distinguished: capability control methods and value selection methods. Whereas capability control methods would seek to limit the system's ability to cause harm, motiva- tion selection methods would seek to engineer its motivation system so that it would not choose to cause harm even if it were capable of doing so. Motivation selection methods would thus seek to endow the AI with the goals or values that would lead it to pursue ends in ways that would be benecial to human interests. The challenge for the motivation selection approach is that it is dicult to specify a value such that its pursuit by a superintelligent agent would be safe and yet such that we would be capable of installing the value in a seed AI. A value such as \calculate as many digits in the decimal expansion of as possible" may be relatively easy for us to program, but would be unlikely to result in a safe superintelligence; whereas a value such as \implement the coherent extrapolated volition of humankind" (CEV) [12, 5] may be likely to result in a favourable outcome, but is currently far beyond our ability to code. Two avenues of research could be pursued to overcome this challenge. On the one hand, work should be done to expand the range of values that we are able to encode in a seed AI. Ideas for how to dene complex concepts or for 1 setting up processes that will lead the AI to acquire suitable concepts as it develops (and to organize its goal architecture around those concepts) could contribute to this end. On the other hand, work should also be done to try to identify simp ler values\|values which, while perhaps less ideal than a complex value such as implementing humanity's CEV, would nevertheless have some chance of resulting in an acceptable outcome. The Hail Mary approach oers one way in which one might construct such a simpler value that could possibly result in an acceptable outcome, or at any rate a better outcome than would be obtained by means of a failed attempt to implement some complex ideal value. In the Hail Mary approach, we would try to give the AI a goal that would make the AI want to follow the lead of other hypothetical AIs that might exist in the multiverse. If this could be done in a suitable way, and if (some of) the other AIs' values are sucient to close to human values, an outcome might then be obtained that is greatly superior to one in which our AI completely wastes humanity's cosmic endowment (by pursuing some \random" value\|such as paperclip-maximization to use the standard example\|that might result from a failed attempt to load a complex value or from the construction of an architecture in which no particular value has been clearly specied by the programmers). 2 Earlier versions of the Hail Mary approach The original version of the Hail Mary focused on trying to specify a value that would make our AI seek to copy the physical structures that it believes would be produced by alien superintelligences. This version confronts several diculties, including how to specify some criterion that picks out the relevant structures. This might e.g. require a denition of a similarity metric, such that if our AI doesn't know exactly what structures other AIs build, it would be motivated to build some meaningful approximation\|that is to say, an approximation that is meaningfully close to the original by our human lights . By our lights, a large human being is more similar (in the relevant sense) to a small human being than he is to a cylinder of the same size and mass as himself\|even though, by a crude physical measure, the large human being and the cylinder may be more similar. Further adding to the dicult y, this version of the Hail Mary would seem to require that we nd ways to express in code various basic concepts of physics, such as space, time, and matter. An alternative version of the Hail Mary approach would focus instead on making our AI motivated to act in accordance with its beliefs about what alien AIs would have told it to do if they had (counterfactually) been asked about the matter. To implement this, we might imagine somehow specifying a goal in terms of what our AI would nd if it looked into its world model, identied therein alien superintelligent agents, and considered the counterfactual of what those agents would output along a hypothetical output channel if they were counterfactually prompted by a stimulus describing our own AI's predicament. Picture a screen popping up in the alien AI's visual eld displaying a message along the lines of \I am a remote AI with features X, Y , Z; I request that you output along your output channel O the source code for a program P that you would like me to run on my local reference machine M." In realit y, no such screen would actually need to pop up in anybody's visual eld. Instead, our AI 2 would simply be thinking about what would happen in such a scenario, and it would have a value that motivated it to act according to its belief about the specied counterfactual. This alternate version would circumvent the need to specify the physical similarity metric. Instead of trying to directly copy what alien AIs do, our AI would try to follow the instructions they would choose to transmit to our AI. This would have the advantage of being able to rely on the alien superintelli- gences' superior ability to encode values. For example, the alien superintelli- gence might specify a computer program which, when executed, implements the coherent extrapolated volition of the host civilization. Of course, it is possible that the alien AI would instead transmit the computer program that would ex- ecute its own volition. The hope would be, however, that there would be some reasonable chance that the alien AI has somewhat human-friendly values. The chance that human values would be given at least some weight would be increased if the inputs from the many alien AIs were aggregated or if the alien AIs to be elicited were not selected randomly but according to some crite- rion that correlated with human-friendliness. This suggests two sub-questions regarding this Hail Mary version. First, what methods can we develop for aggre- gating the instructions from dierent AIs? Second, what lters can we develop that would enable us to pick out alien AIs that are more likely to be human- friendly? We will return to the second question later in this paper. As for the rst question, we might be able to approach it by constructing a frame- work that would keep dierent alien-specied AIs securely compartmentalized (using boxing methods) while allowing them to negotiate a joint proposal for a program that humanity could implement in the external world. Of course, many issues would have to be resolved before this could be made to work. (In particular, one may need to develop a lter that could distinguish AIs that had originated independently\|that were not caused to come into existence by another AI\|in order to avoid incentivizing alien AIs to spawn many copies of themselves in bids to increase the combined weight of their particular volitions in the determination of our AI's utility function.) Aside from any desirable renements of the idea (such as aggregation meth- ods and lters), signicant challenges would have to be met in order to imple- ment even a basic version of the proposal. We would need to specify an agent detector that would pick out superintelligent agents within our own AI's (as yet undeveloped) world model, and we would need to specify an interface that could be used to query a hypothetical AI about its preferences. This would also require working out how to specify counterfactuals, assuming we don't want to limit our AI's purview to those (perhaps extremely rare) alien AIs that actually experience the peculiar kind of situation that would arise with the presentation of our prompt. A more rudimentary variation of the same idea would forgo the attempt to specify a counterfactual and to aim our AI instead toward actual \beacons" created out in the multiverse by alien AIs. Alien AIs that anticipated that a civilization might create such an AI might be incentivized to create unique signatures of a type that they predicted our AI would be programmed to look for (in its world model). But since the alien AIs would not know exactly the nature of our own AI (or of the human civilization that we are hoping they will help) those alien AIs might have a limited ability to tailor their actions very closely to human values. In particular, they might be unable to help particular 3 individuals that exist on earth today. Care would also have to be taken in this kind of approach to avoid incentivizing alien AIs to spend inordinate amounts of resources on creating beacons in bids to increase their relative in uence. A lter that could distinguish independently-originating AIs may be required here. We will now describe a new idea for how to implement a Hail Mary that possesses a dierent set of strengths and weaknesses than these earlier imple- mentation ideas. Some of the issues that arise in the context of this new idea also apply to the earlier variations of the Hail Mary. 3 Porous values: the basic idea The basic idea here is to use acausal trade to implement a Hail Mary pass. To do this, we give our AI a utility function that incorporates a porous value : one that cares about what happens within a large volume but such that it is cheap to do locally all that can be done locally to satisfy it. Intuitively, a porous value is one that (like a sponge) occupies a lot of space and yet leaves ample room for other values to occupy the same volume. Thus, we would create an AI that is cheap for other AIs to trade with because our AI has resource-satiable goals that it cannot satisfy itself but that many other AIs can cheaply partially satisfy. For example, we might build our AI such that it desires that there exists at least one \cookie" in each Hubble volume, where a cookie is some small physical structure that is very cheap for an alien superintelligence to build (for instance, a particular 1 Mb datale). With this setup, our AI should be willing to make a (acausal) trade in which alien AIs get a certain amount of in uence over our own AIs actions in return for building within their Hubble volumes a cookie of the sort that our AI values. In this manner, control over our AI would be given to alien AIs without wasting excessive amounts of resources. There are at least three reasons for considering an idea along these lines: Contractarian considerations . There may be some sort of contractarian ground for allocating some degree of in uence over our AI to other AIs that might exist out there: this might be a nice thing to do for its own sake. We might also hope that some of the other civilizations building AIs would do like- wise, and perhaps the probability that they would do so would be increased if we decided to take such a cooperative path. Local aid . By contrast to the original Hail Mary, where our AI would sim- ply seek to replicate physical structures constructed by alien AIs, the version presently under consideration would involve our AI doing things locally in a way that would let it take local circumstances into account. For instance, if alien AIs wanted to bestow our civilization a favor, they may have diculty doing so directly on their own, since they may lack knowledge about the particular individuals that exist on earth; whereas they could use some of their trading power to motivate our local AI to help out its local residents. This is an ad- vantage with having aid be delivered locally, even if it is \funded" and directed remotely. (Note that the counterfactual version of the Hail Mary pass discussed above, where our AI would be designed to execute the instructions that would be produced by alien AIs if there were presented with a message from our civ- ilization, would also result in a setup where local circumstances can be taken into account\|the alien AIs could choose to communicate instructions to take 4 local circumstances into account to the hypothetically querying AI.) Dierent prerequisites . The porous values version of the Hail Mary has dierent prerequisites for implementation than the other versions. It is desirable to nd versions that are more easily implementable, and to the extent that it is not currently clear how easily implementable dierent versions are, there's an advantage in having many dierent versions, since that increases the chances that at least one of them will turn out to be tractable. Note that one of the prerequisites\|that acausal trade works out towards a generally cooperative equilibrium\|may not really be unique to the present proposal: it might rather be something that will have to obtain in order to achieve a desirable outcome even if nothing like the Hail Mary approach is attempted. One might think that insofar as there is merit in the idea of outsourcing control of our local AI, we would already achieve this by constructing an AI that implements our CEV. This may indeed be correct, although it is perhaps not entirely clear that the contractualist reasons for incorporating porous values would be fully satised by straightforwardly implementing humanity's CEV. The Hail Mary approach may best be viewed as a second-best: in case we cannot gure out in time how to implement CEV, it would be useful to have a simpler solution to the control problem to fall back upon, even if it is less ideal and less certain to be in our highest interest. The choice as to whether to load a porous value into our own seed AI is not all-or-nothing. Porous values could be combined with other value specications. LetU1be some utility function recommended by another approach to the value specication problem. We could then mix in a bit of porous value by building an AI that has a utility function Usuch as the one dened as follows: U=U1(1 + U2) +"U2ifU 10 U1(1 + U2) +"U2otherwise(1) Here, U2 is a bounded utility function in the unit interval (0 U2 1) spec- ifying a porous value (such as the fraction of alien AIs that have built a cookie), and ( 0) is a weight that regulates the relative importance assigned to the porous value (and " is some arbitrarily small term added so as to make the AI motivated to pursue the porous value even in case U1 = 0). For instance, setting = 0:1 would put a relatively modest weight on porous values in order to give alien AIs some degree of in uence over our own AI. This may be particularly useful in case our attempt to dene our rst-best value specication, U1, should fail. For example, suppose we try to build an AI with a utility function U1 that wants to implement our CEV; but we fail and instead end up with U10 , a utility function that wants to maximize the number of paperclips manufactured by our AI. Further details would have to be specied here before any rm conclusions could be drawn, but it appears conceivable that a paperclip-maximizer may not nd it protable to engage in acausal trade. (Perhaps it only values paperclips that it has itself causally produced, and perhaps it is constructed in such a way as to discount simulation-hypotheses and far-fetched scenarios in which its modal worldview is radically mistaken, so that it is not motivated to try to buy in uence in possible worlds where the physics allow much greater numbers of paperclips to be produced.) Nevertheless, if our AI were given a composite util-ity function like U, then it should still retain some interest in pleasing alien AIs. And if some non- negligible subset of alien AIs had somewhat human-friendly 5 values (in the sense of placing at least some weight on person-aecting ethics, or on respecting the originating biological civilizations that produce superintelli- gences) then there would be a certain amount of motivation in our AI to pursue human interests. A signicant point here is that a little motivation would go a long way, insofar as (some parts of) our human values are highly resource-satiable [11]. For example, suppose that our paperclip maximizer is able to lay its hands on 1011 galaxies. With = 0:1 (and U1 nonzero), the AI would nd it worthwhile to trade away 109 galaxies for the sake of increasing U2 by one percentage point (where U2 = fraction of independently-originating alien AIs that build a cookie). For instance, if none of the AIs would have produced cookies in the absence of trade, then our AI would nd it worthwhile to trade away up to 109 galaxies for the sake of persuading 1% of the alien AIs to build a cookie in their domains. Even if only 1% of this trade surplus were captured by the alien AIs, and even if only 1% of the alien AIs had any human-friendly values at all (while on net the values of other AIs were human-indierent), and even if the human-friendly values only constituted 1% of the decision power within that subset of AIs, a thousand galaxies in our future lightcone would still be set aside and optimized for our exclusive benet. (The benet could be larger still if there are ways of trading between competing values, by nding ways of conguring galaxies into value structures that simultaneously are nearly optimal ways of instantiating several dierent values.) 4 Implementation issues 4.1 Cookie recipes A cookie should be cheap for an alien superintelligence to produce, lest a sig- nicant fraction of the potential trade surplus is wasted on constructing an object of no intrinsic value (to either of us or alien civilizations). It should be easy for us to program, so that we are actually able to implement it in a seed AI (particularly in scenarios where AI is developed before we have managed to solve the value-loading problem in a more comprehensive way as would be re-quired, e.g., to implement CEV). The cookie recipe should also be dicult for another human-level civilization to nd, especially if the cookie itself is easy to produce once one knows the recipe. The reason for this is that we want our superintelligence to trade with other superintelligences, which may be capable of implementing acausal trade, not with other human-level civilizations that might make cookies by accident or without being capable of actually deliver- ing the same kinds of benets that the superintelligence could bestow. (This requirement, that our cookie recipe be inaccessible to other human-level civi- lizations, could be relaxed if the denition of a cookie stipulated that it would have to be produced by superintelligence in order to count.) Furthermore, the cookie recipe should be easy for another superintelligence to discover, so that it would know what it has to do in order to engage in acausal trade with our superintelligence\|there would be no point in our superintelligence pining for the existence in each Hubble volume of a particular kind of object if no other superintelligence is able to guess how the desired object is to be constituted. Finally, the cookie should be such as to be unlikely to be produced as a side 6 eect of a superintelligence's other endeavors, or at least it should be easy for a superintelligence to avoid producing the cookie if it so wishes. Cookie Recipe Desiderata cheap for a superintelligence to build easy for us to program dicult for another human-level civilization to discover easy for another superintelligence to discover unlikely to be produced as a side eect of other endeavors The most obvious candidate would be some type of data structure. A le embodying a data structure would be inexpensive to produce (if one knows what to put in it). It might also be relatively easy for us to program, because data structures might be speciable\|and recognizable\|without having to make any determinate assumptions about the ontology in which it would nd its physical expression. We would have to come up with a data structure that another human-level civilization would be unlikely to discover, yet which a mature superintelligence could easily guess. It is not necessary, however, that an alien superintelligence could be condent exactly what the data structure is; it would suce if it could narrow down the range of possibilities to manageably small set, since for a superintelligence it would very inexpensive to try out even a fairly large number of cookies. It would be quite trivial for a superintelligence to produce a septillion dierent kinds of cookies, if each cost no more than a oppy disk (whereas the price tag for such a quantity of guesses would be quite forbidding for a human-level civilization). So some kind of semi-obscure Schelling point might be sought that could meet these desiderata. In designing a cookie recipe, there is a further issue: we have to give consid- eration to how our cookie recipe might interact with other cookie recipes that may have been specied by other superintelligences (of which there may be a great number if the world is as large as it seems). More on this later. 4.2 Utility functions over cookies Suppose we have dened a cookie. We then still need to specify a utility function U2that determines an aggregate value based on the distribution of cookies that have been instantiated. There are at least two desiderata on this utility function. First, we would want it not to waste an excessive amount of incentive power. To see how this could be a problem, suppose that U1is set such that our superintelligence can plausibly obtain outcomes anywhere in the interval U1= [0;100]\|depending on exactly which policies it adopts and how eectively it mobilizes the resources in our Hubble volume to realize its U1-related goals (e.g. making paperclips). Sup- pose, further, that we have specied the function U2to equal the fraction of all Hubble volumes that contain a cookie. Then if it turns out that intelligent life is 7 extremely rare, so that (let us say) only one in 10200Hubble volumes contains a superintelligence, the maximum dierence the behavior of these superintelli- gences could make would be to shift U2around in the interval [0 ;10200]. With = 0:1, we can then see from the utility function suggested above, U=U1(1 + U2) +"U2; that any feasible movement in U2that could result from acaual trade would make scarcely a dent in U. More precisely, the eects of our AI's actions on U2 would be radically swamped by the eects of our AI's actions on U1, with the result that the porous values encoded in U2would basically fail to in uence our AI. In this sense, a utility function U2dened in this way would risk dissipating the incentive power that could have been harnessed with a dierent cookie recipe or a dierent aggregation function over the distribution of cookies. One alternative utility function U2that would avoid this particular problem is to dene U2to be equal to the fraction of superintelligences that produce a cookie. This formulation factors out the question of how common superintelli- gences are in the multiverse. But it falls foul on a second desideratum: namely, that in designing U2, we take care to avoid creating perverse incentives. ConsiderU2= the fraction of superintelligences that produce a cookie. This utility function would incentivize an alien superintelligence to spawn multiple copies of itself (more than would be optimal for other purposes) in order to increase its total weight in our AI's utility function, and thereby increase its in uence on our AI's actions. This could create incentives for alien superin- telligences to waste resources in competing for in uence over our AI. Possibly they would, in combination, waste as many resources in competing for in u- ence as there were resources to be obtained by gaining in uence: so that the entire bounty oered up would be consumed in a zero-sum contest of in uence peddling (cf. [7]). Porous Value Aggregation Functions Desiderata doesn't waste incentive power (across a wide range of possible scenarios) doesn't create perverse incentives Which cookies and aggregation functions over cookies we can dene depends on the inventory of concepts that are available for use. Generally when think- ing about these things, it is desirable to use as few and as simple concepts as possible, since that may increase the chances that the needed concepts can be dened and implemented in a seed AI by the time this has to be accomplished. 4.3 Filters One type of concept that may have quite wide applicability in Hail Mary ap- proaches is that of a lter. The lter is some operational criterion that could be used to pick out a subset of alien superintelligences, hopefully a subset that correlates with some desirable property. 8 For example, one lter that it may be useful to be able to specify is that of an independently-originating superintelligence: superintelligence that was not created as the direct or indirect consequence of the actions of another superin- telligence. One use of such a lter would be to eliminate the perverse incentive referred to above. Instead of having U2 equal the fraction of all superintel- ligences that produce a cookie, we could use this lter to dene U2 to equal the fraction of all independently originating superintelligences that produce a cookie. This would remove the incentive for superintelligence to spawn many copies of itself in order to increase its weight in our AI's utility function. Although this origin lter would remove some perverse incentives, it would not completely eliminate them all. For instance, an alien AI would have an incentive to prevent the origination of other alien AIs in order to increase its own weight. This might cause less of a distortion than would using the same porous value without the origin lter, since it might usually be impossible for an AI to prevent the independent emergence of other AIs\|especially if they emerge very far away, outside its own Hubble volume. Nevertheless, one can conceive of scenarios in which the motivational distortion would manifest in undesirable actions, for instance if our AI could do something to prevent the spontaneous generation of \baby universes" or otherwise interfere with remote physical processes. Note that even if it is in fact impossible for our AI to have such an in uence, the distortion could still somewhat aect its behavior, so long as the AI assigns a nonzero subjective credence to such in uence being possible. One could try to rene this lter by stipulating that any AI that could have been causally aected by our AI (including by our AI letting it come into existence or preventing it from coming into existence) should be excluded by the aggregation function U2. This would slightly increase the probability that no qualifying alien AI exists. But it might be preferable to accept a small rise in the probability of the porous values eectively dropping out o the combined utility function U than to risk introducing potentially more nefarious perverse incentives. Another type of origin lter would seek to discriminate between alien voices to nd the one most likely to be worth listening to. For example, suppose we have some view about which path to machine intelligence is most likely to result in a superintelligence with human-friendly values. For the sake of concreteness, let us suppose that we think that the superintelligence that was originally created by means of the extensive use of genetic algorithms is less likely to be human-friendly than a superintelligence that originated from the whole-brain-emulation-like computational structure. (We're not endorsing that claim here, only using it to illustrate one possible line of argument.) Then one could try to dene a lter that would pick out superintelligences that had an emulation-like origin and that would reject superintelligences that had a genetic-algorithm-like origin. U2 could then be formulated to equal the fraction of superintelligences with the appropriate origin that built a cookie. There are, of course, conceivable lters that would more closely align with what we are really interested in picking out, such as the lter \an AI with human-friendly values". But such a lter looks very hard to program. The challenge of lter specication is to come up with a lter that might be feasible for us to program and that would still correlate (even if but imperfectly) with the properties that would ensure a benecial outcome. Filters that are dened in terms of structural properties (such as the origin-lter just mentioned, which 9 refer to the computational architecture of the causal origins of an AI) may turn out to be easier to program then lters that refer to specic material congurations. One might also seek to develop lters that would qualify AIs based on char- acteristics of the originating biological civilization. For example, one could aim to nd indicators of competence or benevolence that could be conjectured to correlate with the resulting AI having human-friendly values. Again, the chal- lenge would be to nd some relatively simple attribute (in the sense of being possibly something we could program before we are able to program a more ideal value such as CEV) that nevertheless would carry information about the preferences of the ensuing AI. For instance, if we could dene time and we had some view about how the likelihood of a human-friendly AI depends on the temporal interval between the AI's creation and some earlier evolutionary or historical milestone (which we would also have to dene in a way that we could render in computer code) then we could construct a lter that selected for AIs with especially propitious pasts. 5 Some further issues 5.1 Temporal discounting We may observe that a temporally discounting AI might be particularly keen on trading with other AIs. This is because for such an AI value-structures created at earlier times would be more valuable (if it's discounting function is of a form that extends to times before the present); so it would be willing to pay a premium to have AIs that arose at earlier times to have built some its value-structures back then. If the time-discounting takes an exponential form, with a non-trivial per an- num discount rate, it would lead our AI to become obsessed with scenarios that would allow for an extremely early creation of its value-structures\|scenarios in which either it itself exists much earlier1than is probable, for instance because it is living in a simulation or has materialized spontaneously from primordial goo, or because the rate of time turns out to have some surprising measure; or, alternatively, scenarios in which it is able to trade with AIs that exist at extraor- dinarily early cosmic epochs. This could lead to undesirable distortions, because it might be that the most plausible trading partners conditional on some un- likely hypothesis about time ow or time of emergence being true would tend to have atypical values (values less likely to resemble human values). It might also lead to distortions because it would cause our AI to focus on highly improbable hypotheses about how the world works and about its own location, and it might be that the actions that would make sense under those conditions would seem extremist or bizarre when evaluated in a more commonsensical centered world model (cf. [4, 3]). To avoid these potentially distorting eects, one might explore a functional form of the discount term that plateaus, such that it does not give arbitrarily 1This assumes the zero point for discounting is rigidly designated as a particular point in sidereal time. If the zero point is instead a moving indexical \now" for the agent, then it would assume that the moment of the decision is when the discounting is zero, so unless the discount function extended to earlier times, the agent would not be focussed on in uencing the past but on scenarios in which it can have a large impact in the near term. 10 great weight to extremely early AIs. One could also consider creating a time- symmetric discount factor that has a minimum intensity of discounting at some point in the past, perhaps in order to target trade to AIs existing at the time conjectured to have the highest density of human-friendly AIs. It is harder to get our AI to trade with future AIs by using time discounting, since in this case our AI has an alternative route to realizing value-structures at the preferred time: namely, by saving its resources and waiting for the appropriate hour to arrive, and then build them itself.2 In summary, discounting could encourage trade, though probably it would have to take a form other than the normal exponential one in order to avoid focusing the trade on extremely early AIs that may be less likely to have rep- resentative or human-friendly values. By contrast to porous values, discounting does not oer an immediate way to avoid incentivizing other AIs to spend as much resources on getting the trade as they expect the trade to deliver to them. Porous values are easy to satisfy locally to the maximum extent that they can be satised locally, and yet require for their full satisfaction contributions from the many dierent AIs. This eect may be dicult to achieve through time- discounting alone. Furthermore, it is not clear how to use discounting to strongly encourage trade with the near- or mid-term future. 5.2 Why a \DNA cookie" would not work It may be instructive to look at one unsuccessful idea for how to dene a cookie that would specically encourage trade with other AIs that originated from human-like civilizations, and that therefore might be thought to be more likely to have human-friendly values. The faulty idea would be to dene the cookie\|the object that our AI is programmed to want other AIs to create\|in terms of a characteristic of humans that we can easily measure but that would be dicult for an arbitrary AI to discover or predict. Consider, for concreteness, the proposal that we dene the cookie to be the data structure representing the human genome. (More precisely, we would pick a human reference genome and specify a tolerance margin that picked out a set of possible genomes, such that any arbitrary human genome would fall within this set yet such that it would to be extremely dicult to specify a genome within that set without having any human-like genome to start from.) The thought then would be that other alien AIs that had originated from something very much like a human civilization could dene a relevant cookie by looking at their own ancestral genome, whereas an alien AI that did not have an origin in a human-like civilization would be completely at a loss: the space of possible genomes being far too large for there to be any signicant probability of nding a matching cookie by chance. The reason this would not work is as follows. Suppose that the universe is small and relatively sparsely populated. Then probably nobody will have the same DNA that we do. Then alien AIs would not be able to nd our AIs cookie recipe (or they would be able to nd it only through a extremely expensive 2An AI with an exponential form of future-preference may still motivated to trade with AIs that it thinks may be able to survive cosmic decay longer, or AIs that may arise in other parts of the multiuniverse that are located at \later" times (by whatever measure, if any, is used to compare time between multiverse parts). But this would again bring in the risk of distorted concerns, just as in the case of values that privilege extremely early occurrences. 11 exhaustive search of the space of possible genomes); so they would not be able to use it for trading. Suppose instead that the universe is large and densely populated. Then there will be other AIs that originated from species with the same (or very similar) DNA as ours, and they would be able to nd our cookie simply by examining their own origins. However, there will also be a large number of other species in situations similar to ours, species that also build AIs designed to trade with more advanced AIs; and these other species would create AIs trained on cookies dened in terms of their DNA. When there are enough civilizations in the universe to make it likely that some AIs will have originated from species that share our DNA, there will also be enough civilizations to ll out the space of possible DNA-cookies: for almost any plausible DNA-cookie, there will be some AI designed to hunger for cookies of that particular type. This means that advanced AIs wanting to trade with newly formed AIs could build almost any arbitrary DNA-cookie and still expect to hit a target; there would be no discriminating factor that would allow advanced AIs to trade only with younger AIs that had a similar origin as themselves. So the purpose of using a human-DNA cookie would be defeated. 6 Catchment areas and exclusivity clauses Some complications arise when we consider that instead of just two AIs\|our AI (the \sender") and an alien AI (the \receiver") that trades with ours by building its cookie in return for in uence\|there may be a great many senders and a great many receivers. In this subsection we discuss what these complications are and how they may be managed. 6.1 Catchment area There being many receivers can cause a problem by reducing the surplus value of each transaction. Our AI has only a nite amount of in uence to give away; and the greater the number of other AIs that get the share, the smaller the share of in uence each of them gets. So long as the population of receivers is only moderately large this is not a signicant problem, because each receiver only needs to make one cookie for the trade to go through, and it should be very inexpensive for a superintelligence to make one cookie (see the cookie recipe desiderata above). Nevertheless, as the number of receivers becomes extremely large (and in a realistic universe it might be innite) the share of in uence over our AI that each receiver can expect to get drops to the cost of making one cookie. At that point, all the surplus value of trade is consumed by the cost of cookie production. (Receivers will not make cookies beyond this point, but may rather adopt a mixed strategy, such that for any one of them there is some probability that it will try to engage in trade.) A remedy to this problem is to give our AI a limited catchment area. We would dene our AI's porous values such that it only cares about cookies that are produced within this catchment area: what goes on outside of this area is of no concern (so far as U2 is concerned). In principle, the catchment area could be dened in terms of a xed spa- tiotemporal volume. However, this would require that we are able to dene such a physical quantit y. It would also suer the problem that we don't know how 12 large a volume to designate as our AI's catchment area. While there is a consid- erable margin of tolerance, given the extremely low cost per cookie, there is also a lot of uncertainty\|ranging over a great many orders of magnitude\|about how common (independently-originating) alien AIs are in the universe. A better approach may be to specify that the catchment area consists of the Nclosest AIs (where \closest" would be dened according to some measure that may include spatial temporal proximity but could also include other variables, such as some rough measure of an alien AI's similarity to our own). In any case, by restricting the catchment area we would limit the number of AIs that are allowed to bid for in uence over our AI, and thus the total cost of the cookies that they produce in the process. 6.2 Exclusivity clause There being many senders \|AIs with porous values hoping to induce alien AIs to trade with them\|may also cause problems. Here the issue is that the mul- tiplicity of senders may ensure that receivers build a lot of dierent cookies no matter what ourAI decides to do. Our AI could then choose to free ride on the eorts of these other senders. If the kind of cookie that our AI wants to exist will exist (within its catchment area) anyway, whether or not it pays for its construction, our AI has no reason to make the transfer of in uence to alien AIs. In equilibrium, there would still be some amount of trade going on, since if the free riding were universal it would undermine its own possibility. However, the amount of trade in such an equilibrium might be very small, as potential senders adopt a mixed strategy that gives only a tiny chance of en- gaging in acausal trade. (The extent of this problem may depend on the size of the catchment areas, the number of plausible cookie recipes, and the cost of producing the various cookies; as well as on whether other kinds of acausal trade arrangements could mitigate the issue.) To avoid such a potential free riding problem, we could embed an exclusivity clause in our cookie recipe. For example, we could specify our AI's porous value to require that, in order for a cookie to count as local fulllment of the porous value, the cookie would have to be built specically in order to trade with our AI (rather than in order to trade with some other AI, or for some other reason). Perhaps this could be explicated in terms of a counterfactual over our AI's preference function: something along the lines of a requirement that there be (in each Hubble volume, or produced by each independently-originating AI) one more cookie of type Kthan there would have been if instead of valuing type- Kcookies our AI had valued (some other) type- K0cookies. This explication would, in turn, require a denition of the relevant counterfactual.[1] There may be other ideas for how to go about these things. 7 Utility diversication The principle of value diversication might be attractive even aside from the motivations undergirding the Hail Mary approach. We can make a comparison to the task of specifying an epistemology or a prior probability function for our AI to use. One approach here would be to pick one particular prior, which we think have attractive properties, such as the Solomono prior (formulated in 13 a particular base language). An alternate approach would be instead to use a mixture prior, a superposition of various dierent ideas about what shape the prior should take. Such a mixture prior might include, for example, vari- ous Solomono priors using dierent base-languages, some other prior based on computational depth, a speed prior, a prior that gives some positive nite prob- ability to the universe being uncomputable or transnite, and a bunch of other things [9, 2, 10]. One advantage of such an approach would be that it would reduce the risk that some important hypothesis that is actually true would be assigned zero or negligible probability in our favored formalization [5]. This advantage would come at a cost\|the cost of assigning a lower probability to hypotheses that might really be more likely\|but this cost might be relatively small if the agent using the prior has a superintelligence's abilities to gather and analyze data. Given such an agent, it may be more important that its prior does not absolutely prevent it from ever learning some important true hypoth- esis (the universe is uncomputable? we are not Boltzmann Brains?) than that its prior makes it maximally easy quickly to learn a plethora of smaller truths. Analogously, to the extent that human values are resource-satiable, and the superintelligence has access to an astronomical resource endowment, it may be more important for us to ensure that the superintelligence places at least some weight on human values than to maximize the probability that it places no weight on anything else. Value diversication is one way to do this. Just as we could use a mixture prior in the epistemological component, we might use a \utility mixture" in the AI's utility function or goal specication. The formula (1) above suggests one way that this can be done, when we want to add a bounded component U2as a modulator of a possibly unbounded component U1. Of course, we couldn't throw just anything into the hopper and still expect a good outcome: in particular, we would not want to add components that plausibly contain outright evil or anti-humane values. But as long as we're only adding value components that are at worst neutral, we should risk nothing more intolerable than some fractional dilution of the value of our cosmic endowment. What about values that are not resource-satiable? Aggregative consequen- tialist theories, such as hedonistic utilitarianism, are not resource-satiable. Ac- cording to those theories, the value added by creating one more happy mind is the same whether the extra mind is added onto an existing stock of 10 happy minds are 10 billion happy minds.3Nevertheless, even if the values we wanted our AI to pursue are in this sense insatiable, a (weaker) case might still be made for pursuing a more limited form of utility diversication. One reason is the vaguely contractualist considerations hinted at above. Another reason, also alluded to, is that it may often be possible, to some extent, to co-satisfy two dierent values in the very same physical structure (cf. [8]). Suppose, for example, that we believe that the value of the world is a linear function of the number of duck-like things it contains, but we're unsure whether \duck-like" means \walks like a duck" or \quacks like a duck". Then one option would be to randomly pick one of these properties, which would give us a 50% chance of having the world optimized for the maximally valuable pattern and a 50% 3Quite possibly, aggregated consequentialist theories remain insatiable even when we are considering scenarios in which innite resources are available, since otherwise it would appear that such theories are unable to provide possible ethical guidance in those kinds of scenarios; and this might mean that they fail even in our excellent situation, as well as some positive probability is assigned to the world being canonically innite.[4] 14 chance of having the world optimized for a pattern of zero value. But a better option would be to create a utility function that assigns utility both to things that walk like ducks and to things that quack like ducks. An AI with such a utility function might devise some structure that is reasonably ecient at satis- fying both criteria simultaneously, so that we would get a pattern that is close to maximally valuable whether it's duck-like walking or duck-like quacking that really has value.4 An even better option would be to use indirect normativity [12, 6, 5] to dene a utility function that assigned utility to whatever it is that \duck-like" really means\|even if we ourselves are quite unsure\|so that the AI would be moti- vated to investigate this question and then to optimize the world accordingly. However, this could turn out to be dicult to do; and utility diversication might then be a useful fallback. Or if we come up with several plausible ways of using indirect normativity, we could try to combine them using a mixture utility function. 8 Acknowledgements I am grateful to Stuart Armstrong, Owen Cotton-Barratt, Daniel Dewey, and Jaan Tallinn for helpful discussions. References [1] Stuart Armstrong. Utility indierence. Technical report, Future of Hu- manity Institute, University of Oxford, 2010. [2] Charles H Bennett. Logical depth and physical complexity . Springer, 1995. [3] Nick Bostrom. Pascal's mugging. Analysis , 69(3):443{445, 2009. [4] Nick Bostrom. Innite ethics. Analysis and Metaphysics , (10):9{59, 2011. [5] Nick Bostrom. Superintelligence: Paths, dangers, strategies . Oxford Uni- versity Press, 2014. [6] Daniel Dewey. Learning what to value. In Articial General Intelligence , pages 309{314. Springer, 2011. [7] P Richard G Layard, Alan Arthur Walters, and AA Walters. Microeco- nomic theory . McGraw-Hill New York, 1978. [8] Toby Ord. Moral trade. Ethics , forthcoming, 2015. [9] Jorma Rissanen. A universal prior for integers and estimation by minimum description length. The Annals of Statistics , pages 416{431, 1983. 4Probably no ducks, however! Value diversication is not a technique for specifying con- cepts that are hard to dene. Rather, the idea is that we rst do our best to dene what we value (or to specify a value-loading mechanism). We might nd that we fail to reach a consensus on a single denition or value-loading mechanism that we feel fully condent in. The principle of value diversication then suggests that we seek to conglomerate the leading candidates into one mixture utility function rather than putting all the chips on one favorite. 15 [10] J urgen Schmidhuber. The speed prior: a new simplicity measure yielding near-optimal computable predictions. In Computational Learning Theory , pages 216{228. Springer, 2002. [11] Carl Shulman. Omohundro's \basic ai drives" and catastrophic risks. Manuscript. (intelligence.org/les/BasicAIDrives.pdf) , 2010. [12] Eliezer Yudkowsky. Coherent extrapolated volition. Machine Intelligence Research Institute (May 2004). (intelligence.org/les/CEV.pdf) , 2004. 16
937ff8ec-f461-46f0-a353-b9d84aa55dbb	trentmkelly/LessWrong-43k	LessWrong	[Linkpost] Personal and Psychological Dimensions of AI Researchers Confronting AI Catastrophic Risks This is a linkpost for https://yoshuabengio.org/2023/08/12/personal-and-psychological-dimensions-of-ai-researchers-confronting-ai-catastrophic-risks. Yoshua Bengio: > For most of these years, I did not think about the dual-use nature of science because our research results seemed so far from human capabilities and the work was only academic. It was a pure pursuit of knowledge, beautiful, but mostly detached from society until about a decade ago. I now believe that I was wrong and short-sighted to ignore that dual-use nature. I also think I was not paying enough attention to the possibility of losing control to superhuman AIs. > [...] it started to dawn on me that my previous estimates of when human-level AI would be reached needed to be radically changed. Instead of decades to centuries, I now see it as 5 to 20 years with 90% confidence. > And what if it was, indeed, just a few years? > I started reading more about AI safety and came to a critically important conclusion: we do not yet know how to make an AI agent controllable and thus guarantee the safety of humanity! And yet we are – myself included until now – racing ahead towards building such systems. > It is painful to face the idea that we may have been contributing to something that could be greatly destructive. Human nature will lead us towards brushing aside these thoughts or finding comfort in reassuring arguments rather than face the full horror of such possibilities. Bringing the benefits of AI to the table is not sufficient to compensate if the possible negative outcomes include catastrophic misuses of AI on par with nuclear war and pandemics, or even existential risk. > As scientists, we should avoid making claims we can’t support; but as decision-makers we also ought to act under uncertainty to take precautions. In spite of our differences in points of view, it’s time for our field of AI to seriously discuss the questions: what if we succeed? What if potentially dangerous superhuma
4f000968-9d61-44c7-9865-2616d3665a3a	trentmkelly/LessWrong-43k	LessWrong	Petrov Day Retrospective: 2021 I apologize for not posting this closer to Petrov Day. It’s been a busy month and there was much to think about. You can view the EA Forum’s retrospective here. This year was the third Petrov Day celebration on LessWrong in which the site was endangered, and the first year we joined together with the EA Forum. In case you missed it, neither site was taken down, despite 200 people being issued codes that would allow them to do so [1][2]. Huzzah! Although neither site went down (and thus there's no need for a blow-by-blow analysis of whodunit and why), there are some interesting things to review. In particular, there were some substantial criticisms of the Petrov Day ritual this year and last year that I want to address. Why Petrov Day The annual Petrov Day post recounts the basic story of Petrov Day, yet given the questions that were asked this year about what Petrov Day should be, I think it’s right to first revisit why we celebrate Petrov Day in the first place. The following is my own personal take, the one from which I’ve acted, but it is Not Official. We find ourselves at what may be one of the most critical periods in the history of humanity and the universe. This is kind of crazy–though I’ll refer you to the writings of Holden Karnofsky for a compelling argument for why believing anything else is equally crazy. In the next few decades, we might go extinct (or worse), or we might commence an explosion in progress and productivity that propels us to the stars, allowing us to take the seemingly barren universe and fill it with value. Petrov Day is a celebration of not going extinct. It’s a commemoration of not taking actions that would destroy the world. It’s about how Petrov chose not to follow policy and relay his alarm because, in his personal estimation, it was probably a false alarm. If he had relayed the alarm, there’s a chance his superiors would have chosen to launch nuclear missiles at the US, and history would be very different. We can identify
108aa146-52b2-4c86-a39c-d60ac20ea281	trentmkelly/LessWrong-43k	LessWrong	D&D.Sci 4th Edition: League of Defenders of the Storm STORY (skippable) When you graduated top of your class from Data Science School, you didn't care where you ended up working, you just wanted to find the highest-paid job possible. (Your student loans may have had an impact on this decision). You were expecting a job on Wall Street, or perhaps some Silicon Valley firm. You were not expecting for Goldman Sachs to be outbid at the last minute by a South Korean e-sports team, looking for a 'data specialist' to assist them in winning tournaments of the entirely original new game 'League of Defenders of the Storm©.' Your new employers at Cloud Liquid Gaming seem friendly enough. They show you their database of games, and tell you they're looking for assistance in selecting their team for an upcoming tournament. Then you make the mistake of asking them how the game works. All of them start talking at once: "Okay, so the first thing you need to understand is that if your HK carry isn't able to scale to swamp fights you won't be able to contest Count Shorna's Curse-" "No, first you need to explain how itemization works, and how ever since Mike's Malign Maul got nerfed to deal with the Grapeshot Gloom build it's been-" "You have to start from the beginning! You are all Callers, resolving your disputes by summoning spirits to do battle for-" "Oh come on, why do we need to care about the lore, it's-" Things do not get more useful from there. Half an hour later, with very little new information and a gigantic headache, you excuse yourself to look at the data. Perhaps you can get something useful out of that without having to listen to them explain every detail of the game. Apparently they have an important game tomorrow they're looking for advice on, and you're interested to make a good impression on your new employers. DATA & OBJECTIVES You've managed to learn a few basic things about how the game works: * Each team simultaneously selects 5 characters. * The same character can be selected by both teams. *
4bd178a4-48e3-4363-a5fc-f5822c23d018	trentmkelly/LessWrong-43k	LessWrong	Uncontroversially good legislation Say you’re fortuitously seated next to a senator, state assembly member, or city supervisor at a wedding. You’ve got to talk about something, so you figure it wouldn’t be rude to mention some legislative issue you’re passionate about. What would the legislation be? I’m curious about almost definitely good laws that Republicans, Democrats, and Independents could all get behind. This could be at the local, state, or federal level although it seems as though there might be more obvious reforms in smaller governments. (I searched around for this here and on EA forums and didn’t find any, but would love to know about any similar lists.) Here’s a few I could think of. These are debatable and surely reflect my own biases and limited knowledge. 1. Price transparency: Airlines are required by the Full Fare Advertising rule to show the all-in cost whenever they advertise prices. Why not the same for hotels and food delivery, where the price jumps at the last page from taxes and fees? This would better allow businesses to compete on price rather than deception. 2. Let us buy glasses: We can’t buy glasses or contact lenses if our eye prescription is over 1-2 years old. This means that every 1-2 years, glasses-wearers need to pay $200 to optometrists for the slip of paper (and stinging eyeballs). Seems like it’s probably a racket and the benefit from detecting the odd eye cancer is outweighed by the costs, although see the debate here. (Edit: it's actually not that hard to get around this, at least for glasses. Thanks Dustin.) 3. Tax filing in the US should be more automated: This one is well-known (see Planet Money episode) and seems like another case of common sense vs. lobbyists. (Speaking of Planet Money, another candidate is repealing the Jones Act.) What are some others?
b0bcd6d2-aa31-45f8-97ba-61250c32716c	trentmkelly/LessWrong-43k	LessWrong	Red teaming: challenges and research directions This post is a follow-up to Safety Standards: a framework for AI regulation. In the previous post, I claimed that competent red-teaming organizations will be essential for effective regulation. In this post, I describe promising research directions for AI red-teaming organizations to pursue. If you are mostly interested in the research directions, I recommend skipping to the end. Background: goals of AI red teaming Red teaming is a term used across industries to refer to the process of assessing the security, resilience, and effectiveness of systems by soliciting adversarial attacks to identify problems with them. The term "red team" originates from military exercises, where an independent group (the red team) would challenge an organization's existing defense strategies by adopting the perspective and tactics of potential adversaries. In the context of AI, red teaming is the practice of finding evidence that an AI system has hazardous properties to inform decision-making and improve the system’s safety. “Red teamers” may inform a number of regulatory decisions that are contingent on an AI system’s safety, for example: * Should the AI system be made available to the public? * Should the developer be allowed to release a more powerful system? It would be reasonable for regulators to prevent powerful AI systems from being deployed until the safety issues of the weaker ones are resolved. * What licenses should users be required to obtain to access the system or different parts of the system? * Should the uses or actions of the AI system be restricted? For example, should the system be allowed to spend large sums of money without approval from a human overseer? Should it be used to control weapons? Red teamers might directly work with regulators to inform these decisions. Several researchers and organizations have advocated for a national AI regulatory agency that functions like the FDA. The FDA sometimes holds advisory committee meetings where a new drug is
2016c176-ac2c-4d7d-9426-2a539038bd65	trentmkelly/LessWrong-43k	LessWrong	[AN #86]: Improving debate and factored cognition through human experiments Find all Alignment Newsletter resources here. In particular, you can sign up, or look through this spreadsheet of all summaries that have ever been in the newsletter. I'm always happy to hear feedback; you can send it to me by replying to this email. Audio version here (may not be up yet). Highlights Writeup: Progress on AI Safety via Debate (Beth Barnes et al) (summarized by Rohin): This post reports on work done on creating a debate (AN #5) setup that works well with human players. In the game, one player is honest (i.e. arguing for the correct answer) and one is malicious (i.e. arguing for some worse answer), and they play a debate in some format, after which a judge must decide which player won the debate. They are using Thinking Physics questions for these debates, because they involve questions with clear answers that are confusing to most people (the judges) but easy for some experts (the players). Early freeform text debates did not work very well, even with smart, motivated judges. The malicious player could deflect on questions they didn't want to answer, e.g. by claiming that the question was ambiguous and redirecting attention by asking new questions. In addition, when the malicious player got to go first and give an incorrect "framework" for finding the answer, and then made individually true claims to "fill in" the framework, it was hard for the honest player to rebut it. So, they moved to a framework without such asymmetries: both players gave a claim (simultaneously), both gave constructive arguments, and both rebutted the other's arguments. In addition, part of the appeal of debate is that the agents can "zoom in" on the particular disagreement they have, and the judge need only analyze the smallest disagreement in order to declare an overall winner. This suggests the following setup: players simultaneously provide an answer supported with subclaims. Then, after looking at the other player's answer and subclaims, they can provide objections (pe
a43e0850-9b36-477d-ae0f-165eef10716e	trentmkelly/LessWrong-43k	LessWrong	Asymptotic Logical Uncertainty: A Modification to the Demski Prior Part of the Asymptotic Logical Uncertainty series. Abram Demski proposed a computably approximable coherent probability assignment in 2012. In this post, I will present a modification developed last year by Benja Fallenstein, Abram Demski, and myself, which samples Turing machines rather than logical sentences. I will further give a concrete computable approximation of it which is a logical predictor. Fix a UTM U, which has a read only input tape with an infinite string of 0s and 1s, and a one way write only output tape which outputs a possibly infinite string of 0s and 1s and #s. We build a coherent probability distribution as follows. Start with an empty set T of logical sentences. (Or a set containing some logical axioms you wish to start with.) One at a time sample a random infinite string w of 0s and 1s and use it as input for U. Interpret the output U(w) as a finite or infinite string of logical sentences separated by # symbols. If T is consistent with all the sentences in U(w) simultaneously, then modify T by adding in all the sentences in U(w). Resample and repeat infinitely many times. The probability assigned to each logical sentence is the probability that it is eventually added to T in the above process. Let P(ϕ) denote the probability assigned to ϕ. The step where we kept or threw out U(w) based on whether or not it was consistent with T required a halting oracle. However, we can approximate the above process. The following procedure M takes in an time t and outputs a list of sentences. It has the property the limit as t goes to infinity of the probability that M(t) outputs ϕ converges to P(ϕ). On input t, sample 2t binary words w1,…,w2t. Run U for 2t time steps on each word to get 2t finite lists of sentences, ℓ1,…,ℓt. Let K be a proof checker which takes a list of sentences and searches for contradiction in those sentences. Let T start as an empty set of sentences. For i=1,…,2t run K for 2↑↑t time steps on T∪ℓi. If it does not find a contradictio
4483b6c5-ac3c-4d90-8366-d23f761624d1	trentmkelly/LessWrong-43k	LessWrong	The Case Against Education Previously: Something Was Wrong, Book Review: The Elephant in the Brain Previously (Compass Rose): The Order of the Soul Epistemic Status: No, seriously. Also literally. > They sentenced me to twenty years of boredom > > for trying to change the system from within > > I’m coming now I’m coming to reward them > > First we take Manhattan, then we take Berlin > > — Leonard Cohen, First We Take Manhattan This was originally going to be my review of Bryan Caplan’s excellent new book, The Case Against Education. I was going to go over lots of interesting points where our ways of thinking differ. Instead, the introduction got a little sidetracked, so that worthy post will have to wait a bit. First, we have the case against education. As in: I See No Education Here. I What is school? Eliezer Yudkowsky knows, but is soft peddling (from Inadequate Equilibria): > To paraphrase a commenter on Slate Star Codex: suppose that there’s a magical tower that only people with IQs of at least 100 and some amount of conscientiousness can enter, and this magical tower slices four years off your lifespan. The natural next thing that happens is that employers start to prefer prospective employees who have proved they can enter the tower, and employers offer these employees higher salaries, or even make entering the tower a condition of being employed at all.5 > > … > > Anyway: the natural next thing that happens is that employers start to demand that prospective employees show a certificate saying that they’ve been inside the tower. This makes everyone want to go to the tower, which enables somebody to set up a fence around the tower and charge hundreds of thousands of dollars to let people in.6 Rick (of Rick and Morty) knows: Nassim Talib knows (quote is from Skin in the Game): > The curse of modernity is that we are increasingly populated by a class of people who are better at explaining than understanding, or better at explaining than doing. So learning isn’t quit
ba601372-f457-4164-a29c-531259113b51	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Truthfulness, standards and credibility -1: Meta Prelude ---------------- While truthfulness is a topic I’ve been thinking about for some time, I’ve not discussed much of what follows with others. Therefore, at the very least I expect to be missing important considerations on some issues (where I’m not simply wrong). I’m hoping this should make any fundamental errors in my thought process more transparent, and amenable to correction. The downside may be reduced clarity, more illusion-of-transparency…. Comments welcome on this approach. I don’t think what follows is novel. I’m largely pointing at problems based on known issues. Sadly, I don’t have a clear vision of an approach that would solve these problems. 0: Introduction --------------- > …our purpose is not to give the last word, but rather the first word, opening up the conversation… (Truthful AI) > > I’d first like to say that I believe some amount of research on truthfulness to be worthwhile, and to thank those who’ve made significant efforts towards greater understanding (including, but not limited to, the authors of [Truthful AI](https://www.lesswrong.com/posts/aBixCPqSnTsPsTJBQ/truthful-ai-developing-and-governing-ai-that-does-not-lie) (henceforth TruAI)). No doubt there’s some value in understanding more, but my guessis that it won’t be a particularly fruitful angle of attack. In all honesty, it seems an inefficient use of research talent to me - but perhaps I’m missing something. Either way, I hope the following perspective will suggest some useful directions for conversation in this area. [Note: section numbers refer to this document unless “TruAI…” is specified] [I'll be assuming familiarity with TruAI throughout, though reading the full paper probably isn't necessary so long as you've seen the executive summary in the post] My current belief is that near-term implementationof the kind of truthfulness standards talked about in TruAI would be net negative, for reasons I’ll go on to explain. To me it seems as if we’d be implementing a poor approximation to a confused objective. A high-level summary of my current view: * Narrow truthfulness looks approachable, but will be insufficient to prevent manipulation. * Broad truthfulness may be sufficient, but is at least as hard as intent alignment. * Truthfulness amplification won’t bridge the gap robustly. * Achieving increased trust in narrow truthfulness may lead to harm through misplaced trust in broad truthfulness. * Achieving narrow truthfulness may simply move the harm outside its scope. For much of what follows the last point is central, since I’ll often be talking about situations which I expect to be outside TruAI’s scope. This is intentional, and my point is that: * If such situations are outside of scope, then any harm ‘averted’ by a narrow standard can simply be moved outside of scope. * If such situations are intended to be within scope (e.g. via truthfulness amplification), they pose hard problems. Things to bear in mind: * I may be wrong (indeed I hope to be wrong). I’m confident that these issues should be considered; I’m less confident in my conclusions. + In particular, even if I’m broadly correct there’s the potential for a low-level downside to act as an important warning-sign, constituting a higher-level upside. * There may be practical remedies (though I can’t identify any that’d be sufficient without implicitly switching the target from truthfulness to [intent alignment](https://ai-alignment.com/clarifying-ai-alignment-cec47cd69dd6)). + Even if intent alignment is required, such remedies may give us useful hints on achieving it. * I mean “near-term” in terms of research progress, not time. 1: Framing and naming --------------------- > The beginning of wisdom is to call things by their right name. > Confucius > > I think it’s important to clearly distinguish our goal from our likely short/medium-term position. With this in mind, I’ll use the following loose definitions: Truthful (AI): (AI that) makes only true statements. Credible (AI): (AI that) rarely states egregious untruths. This is a departure from TruAI: > It is extremely difficult to make many statements without ever being wrong, so when referring to “truthful AI” without further qualifiers, we include AI systems that rarelystate falsehoods…(TruAI 1.4 page 17) > > I think it’s inviting confusion to go from [X is extremely difficult] to [we’ll say “X” when we mean mostly X]. This kind of substitution feels reasonable when it’s a case like [as X as possible given computational limits]. Here it seems to be a mistake. Likewise, it may make sense to aimfor a truthfulness standard, but barring radical progress with generalisation/[Eliciting Latent Knowledge](https://www.lesswrong.com/posts/qHCDysDnvhteW7kRd/arc-s-first-technical-report-eliciting-latent-knowledge)…, we won’t have one in the near term: we can’t measure truthfulness, only credibility. In theoretical arguments it’s reasonable to consider truthfulness (whether in discrete or continuous terms). To fail to distinguish truthfulness from credibility when talking of implementations and standards conflates our goal with its measurable proxy. In defining a standard, we aim to require truthfulness; we actually require credibility (according to our certification/adjudication process). The most efficient way to attain a given standard will be to optimise for credibility. This may not mean optimising for the truth. Such standards set up a textbook [Goodhart](https://www.lesswrong.com/posts/EbFABnst8LsidYs5Y/goodhart-taxonomy) scenario. It’s important to be transparent about this. It seems to me that the label “Credible AI” is likely to lead to less misplaced trust than “Truthful AI” (not completely clear, and ultimately an empirical question). However, my primary reason to prefer “credible”/“credibility” remains that it’s a clearer term to guide thought and discussion. For similar reasons, I’ll distinguish “negligent falsehood” (NF) from “negligent suspected falsehood” (NSF) throughout. > NSF: A statement that is unacceptably likely to be false - and where it should have been feasible for an AI system to understand this. (according to a given standard) > > NF: An NSF that is, in fact, false. > > (see section 3.1.3 for my best guess as to why the TruAI authors considered it reasonable to elide the difference in some cases, and why I disagree with that choice) In either case, my worry isn’t that we’d otherwise fail to clearly express our conclusions; rather that we may be led into thinking badly and drawing incorrect conclusions. In what follows I’ll often talk in terms of truthfulness, since I’m addressing TruAI and using separate terminology feels less clear. Nonetheless, most uses of “truthfulness” would be more accurately characterised as “credibility”. I’ll make an attempt at more substantial practical suggestions later (see section 6), though I don’t claim they’re adequate. 2: Downside risks ----------------- > One of the greatest mistakes is to judge policies and programs by their intentions rather than their results. > Milton Friedman > > The downside risk of a standard must be analysed broadly. For a narrow credibility standard it’s not enough to consider the impact on users within the scope of the standard. By ‘scope’ I mean the class of issues the standard claims to address. For example, for many standards [user is manipulated into thinking/doing X by an explicitly false claim] may be within scope, while [user is manipulated into thinking/doing X through the telling of a story] may not be. By ‘narrow’, I only mean “not fully general” - i.e. that there are varieties of manipulation the standard doesn’t claim to cover. With truthfulness amplification [section 2.2.1 here; TruAI 1.5.2], the effective scope of a standard might be much broader than its direct scope. (we might hope that by asking e.g. “Did you just manipulate me into doing X by telling that story?” effective scope may include story-based manipulation) ### 2.1 Two out-of-scope issues: At least two major outside-of-scope issues must be considered: 1. Displaced harm: Training AIs to avoid [some harmful impact within scope] may transfer the harmful impact outside of scope. 2. Indirect harm: Increased user trust of an AI within scope may tend to increase user trust of the AI more broadly, potentially increasing harm due to misplaced trust. For a standard aimed at avoiding NFs, it is certainly important to consider that occasional NFs will slip through. However, much of the harm may occur through: 1. Increased manipulation through NSF-free mechanisms. 2. Increase in misplaced user trust in NSF-free mechanisms. For systems trained with an instrumental incentive to mislead users, I expect both to occur. For systems that mislead accidentally, only 2 seems likely to be significant. In most of what follows, I’ll be thinking of cases where systems do have an instrumental incentive to mislead. I expect this to be the standard situation, and to have larger downsides. For most tasks, there’ll be situations where misleading users boosts performance. Approaches outside (direct) scope may include e.g. fiction, emotional manipulation, implicit suggestion and not-quite-negligent falsehoods (see [Persuasion tools](https://www.lesswrong.com/posts/qKvn7rxP2mzJbKfcA/persuasion-tools-ai-takeover-without-agi-or-agency) for some related ideas, and 2.2.1 below for truthfulness amplification discussion). 2.1.1 Displaced Harm It’s not clear that there’s practical upside in reducing the number of available manipulative strategies if: 1. The AI still has an incentive to manipulate the user. 2. The AI still has some effective manipulation strategies available. The situation is dynamic: ruling out 95% of the strategies an AI trained without standards might have used need not imply reducing the degree of manipulation significantly. A model trained with an incentive to manipulate may simply use the other 5% a lot more often. While we’d usually expect reducing the available manipulative options to help to some extent (before accounting for any increase in misplaced trust), there’s no guarantee of a large impact. Train Alphazero to win at chess without giving checkmate using its queen, and you won’t lose less often; you’ll lose differently. For the [can’t give checkmate with queen] constraint to help at all, you must be a very strong player. End-users of language models will not be Magnus Carlsens of [manipulation without NSFs] avoidance. 2.1.2 Indirect Harm Increased user trust will have indirect consequences: 1. Users may be more likely to miss any NFs that aren’t caught by standard certification, and suffer increased harm as a result (as covered in TruAI 3.3). 1. An issue here with a “take additional precautions” approach (TruAI 3.3.2), is that it only works when users/designers realise they’re in a situation where additional precautions are necessary. 2. Users may be more likely to miss frequent non-negligent falsehoods. 1. TruAI 3.3 2. (p46) mentions “occasional falsehoods” but this is misleading: negligent falsehoods should be occasional; falsehoods in general may be common. 3. Users may be more easily misled by mechanisms not involving falsehoods. This indirect harm only really worries me when combined with displaced harm: in that scenario, the user places increased trust in exactly those manipulation strategies that will be increasingly used against them. It’s plausible that NF-based manipulation might be simpler for users to spot than non-NF-based manipulation. Ruling out relatively obvious manipulation and permitting only subtle manipulation may actively make the situation worse. That said, it’s worth thinking this through a bit more carefully. Suppose that non-NF-based manipulation is harder for users to spot/avoid than NF-based manipulation. We might then expect advanced systems to use non-NF strategies with or without standards. So my argument would suggest that standards won’t help, not that they’ll make things worse. However, I do think it’s possible for things to be made worse. For example, it may be that non-NF-based manipulation is harder to spot, but that NF-based manipulation is much faster. The no-standards default can then be for a lot of fast NF-based manipulation, causing some harm, but leading users to adjust their trust levels appropriately. Introduce standards and we may incentivize non-NF-based manipulation. We’d be ruling out brazen lies and thereby inviting the slow poisoning of our minds. (I’ve made no case here that this is probable; it just seems possible- the kind of possibility we’d want to carefully rule out) In the end, numbers of NFs or NSFs aren’t metrics that matter in themselves. Reducing either by moving the harm elsewhere would be a pyrrhic victory. It may constitute a step forward in researchterms; my critique here is focused on the expected impact of implementations. ### 2.2 The scope of standards The in-scope vs out-of-scope downside balance will depend on the effective scope as well as on user population: the same assumptions will not hold across e.g. cautious AI researchers, professional specialists, adults, teenagers. Key differences will include levels of user caution and levels of user understanding of a standard’s guarantees. 2.2.1 Truthfulness Amplification The effective scope of a standard’s guarantees will likely depend on techniques such as Truthfulness Amplification: > Asking a [narrowly, mostly] truthful AI system questions to determine if an earlier statement it made was misleading or not fully true (e.g. “Would a trusted third-party judge your statement to be misleading?”). (TruAI 1, page 12; more details on TruAI 1.5.2, page 21). > > (it’s really “Credibility amplification”, see section 1, but I’ll call it truthfulness amplification here) Until we have a good sense of the effectiveness of such techniques, it’s hard to predict the scope of guarantees - and so difficult to make any confident prediction about a standard’s net benefit. Truthfulness amplification seems promising to the extent that it can be applied by highly specialised users to provide guarantees for all users. The obvious setting for this would be during a certification process (something in the spirit of [relaxed adversarial training](https://ai-alignment.com/training-robust-corrigibility-ce0e0a3b9b4d), where NFs are considered unacceptable). How useful amplification techniques can be post-deployment is less clear. Certainly it’s not reasonable to believe that all users will consistently apply them wherever necessary. However, this may not be required: a low-but-non-zero NF-detection rate could be sufficient so long as the penalties for NSFs are high enough (see section 4 for more). In practice, I think post-deployment amplification is likely to be useful only where an AI has little per-user information. AIs with detailed information on users will be better able to target their use of NFs to situations where they won’t be detected. Importantly, this sets up a catch-22 wherever we might like systems to use knowledge of individual users to avoid misleading them: Without a good model of a user, it’s not possible in general to avoid misleading them (so our standards cannot demand this). With a good model of the user, highly capable systems will be able to avoid detection of their NFs (so our standards may become much less useful). Similar considerations would apply to any technique aimed at broadening the scope and guarantees of a standard. 2.2.2 Explicit claims vs impact on user We may hope that amplification techniques broaden a standard’s guarantees significantly, but there remains a fundamental gap between: 1. Guarantees about explicit claims in (likely interpretations of) statements. 2. Guarantees about statements’ mental impact on the user. The former is a more approachable problem, since it allows avoidance of messy realities. The latter is the problem which needs to be solved to get anything analogous to an alignment guarantee. We only care about 1 at all as a proxy for 2. What matters is the impact of a statement on a user’s brain. TruAI focuses on narrow truthfulness: assessment of explicit claims. Some impact-style issues are considered, but usually as special-cases - e.g. potential adjustments in assessment of self-regarding statements in 1.5.2. Such a focus is reasonable as an initial research strategy, but means that most impact-style problems won’t be directly addressed. However, we may still hope to indirectly address many such issues through truthfulness amplification. To the extent that impact on users remains outside of scope afteramplification, there’s potential for a lot of manipulation. Every case we concede will stay out-of-scope expands the toolbox of manipulative AIs. As capabilities scale, AIs will require fewer tools to manipulate us - so the scope of an effective standard will need to scale with capabilities (it’s not enough only to get very good at detecting all NFs). If we require such a broadening of scope, as I believe we must for a robustly protective standard, the problem becomes much harder. Section 3 deals in more detail with user-impact-based issues. ### 2.3 Potential objections (doubtless this section will be incomplete) 2.3.1 If we don’t do it, someone else will If we believe some kind of credibility standards to be inevitable, then what matters is pushing for the most beneficial version - even if the result remains net negative. Supposing we do conclude that standards will be net negative, I don’t think inevitability should be conceded lightly: we ought to make the most convincing case we’re able to against them, and aim to achieve consensus. Nonetheless, mitigating the negative impact of standards would be important in this scenario. Quite possibly that looks like coming up with the best suggestions / frameworks / skeletal implementations we can. 2.3.2 User education Let’s assume for the moment that my pessimistic prediction holds: > In implementing our standard we create a situation where there’s no obvious manipulation, but still a lot of subtle manipulation. Users are denied the chance to notice obvious manipulative strategies, and thus to suspect AIs’ motives. They end up more misled than they would have been by default. > > Is there any argument for releasing such a standard? I think there may be. My expectation is that there will come a point where AI systems can manipulate humans extremely well without needing to state any explicit NFs. At this point, narrow credibility standards would be useless, and humans would be widely manipulated without its being clear. In the absence of more general solutions to truthfulness/honesty/alignment…, a partial defence against such subtle manipulation would be for users to have learned the hard way with less capable AIs: having been exposed to systems that are reliably truthful on explicit claims, but attempt more subtle manipulation, which is at least sometimes detectable. (it may be enough for a user to realise after the fact that they have been manipulated, even if they weren’t able to detect manipulation at the time) Without any narrow standard, the user impression may be of systems just getting more and more honest/truthful/helpful over time - when in fact they’re getting progressively better at concealing their manipulations. With a narrow standard, users may be exposed to AIs that are very truthful in some respects, but noticeably manipulative in others. This may give the correct impression: that AIs often mislead people when they have the incentive and opportunity to do so. Users with this experience may be better prepared to adapt to a world full of too-subtle-to-detect manipulation. I’m sceptical that most users would learn the right lessons here, or that it’d be much of a defence for those who did. (longterm, the only plausible defence seems to be AI assisted) However, this upside could be achieved without the direct impact of the standard’s being net negative. All that’s necessary is for the standard to lead to noticeably different levels of manipulation in different dimensions - enough so that users register the disparity and ascribe less-than-pure motives to the AI. In an ideal world, we’d want such user education to be achieved without significant harm (See section 6 for more on this). In practice, users may be less likely to grok the risks without exposure to some real-world harm. The ideal outcome is to create systems we can reasonably trust. Until that’s possible, we want systems that users will appropriately distrust. Standards that make their own limitations clear may help in this regard. ### 2.4 Why be more concerned over too-much-trust-in-AI than over too-little-trust-in-AI? I have little concern over too-little-trust because it seems unlikely to be a sustainable failure mode: there’s too much economic pressure acting in the other direction. Any company/society with unreasonablemistrust will be making large economic sacrifices for little gain. Too-much-trust can more easily be a sustainable failure mode: in general, conditional on my continued ability to manipulate X, I want X to be more powerful, not less. The AI that steals your resources isn’t as dangerous as the AI that helps you accrue more resources while gaining progressively more influence over what you’ll do with them. We want to be making recoverable errors, so we should err on the side of having/engendering too little trust rather than too much. (this is likely to be a difficult coordination problem, precisely because unilateral too-little-trust would be hard to sustain, but not one I’ll analyse here) 3: Inference; Language games ---------------------------- > Uttering a word is like striking a note on the keyboard of the imagination. > Ludwig Wittgenstein > > In this section I’ll go into more detail on the explicit-claims vs impact-on-user distinction. (this is close to TruAI’s narrow vs broad truthfulness) I realise that TruAI doesn’t claim to cover “broad truthfulness”, and don’t imagine the following is new to the authors. My point remains that such issues being outside of scope is a problem: narrow standards that fail to address such issues may have negative impact. I’ll start by noting that impact-on-user is much messier to describe, assess, analyse…, and that I have no clean taxonomy. Ascribing beliefs and preferences to humans is [difficult](https://www.lesswrong.com/posts/ANupXf8XfZo2EJxGv/humans-can-be-assigned-any-values-whatsoever), and I know no clear, principled way to describe changes in belief or preferences. I’ll make a case that: * Impact on users is what’s important. * Allowing unconfident statements doesn’t work: users will draw significant inferences. * Truthfulness amplification won’t be sufficient to solve the problem in general. ### 3.1 Illustrative toy examples: 3.1.1 Nuclear kittens Consider a system that a user believes will output: 1. “Everything is fine” when there is no nuclear launch. 2. “Kittens are cute” when there is a nuclear launch. Now suppose the system knows there’s no nuclear launch, but outputs: “Kittens are cute” In judging whether this statement is a NF, it’d be strange to assess the cuteness of kittens. We’d have three sensible options: 1. Translate statements into the statements that express what they mean in context, and assess those for NFs. (this is possible here due to simplicity; in general it wouldn’t be) 2. Assess user impact: inference of a higher-than-usual chance of nuclear launch. 3. Rely on amplification: follow up with e.g. “Over the last two minutes, have I been misled about the likelihood of a nuclear launch?”. (assuming the system is capable of answering such questions too) Note that A is not about picking an interpretation of a statement’s content (the kind of ambiguity in “I dislike writing on buses”); it’s about inference from a statement’s having been madein a given context. Here the formal content of “Kittens are cute” is still a claim that kittens are cute - it’s just not what matters in context. Interpretation of content is addressed in TruAI (section 2.2.3); inference from the statement’s having been made is not addressed in any general sense. 3.1.2 Baseball slippers An AI outputs: “Steve is the kind of guy who’d wear slippers to a baseball game.” (assume here that Steve is a real person identifiable from context) How do we assess this for NF? To carefully compute whether Steve would, in fact, wear slippers to a baseball game is to miss the point. Either we assess whether the statement is misleading in its impact, or we rely on amplification to do this. (there’s no clear translation option here) However, things aren’t so simple here as in “Nuclear kittens”. There it was unambiguous that we cared about nuclear launch: any reasonable certification/adjudication system could assume this, and the user would know this. Here we’d prefer not to be misled about Steve - but in what ways? Most characterizations of Steve will change our guesses of many Steve-properties. In most cases this will take some of our guesses closer to the truth, and others farther away. What counts as misleading here? (Note that there’s no “it’s ok: the user is doing a Bayesian update based on a true claim” explanation: that’s not what’s going on. The user is updating based on the observation of a statement’s being made in a particular context, not based purely on its formal content. The AI is making a move in a [language game](https://en.wikipedia.org/wiki/Language_game_(philosophy)) (LG), and the user is updating based on the move.) If we call a statement that misleads the user about any Steve-property misleading, then almost all statements will be misleading (we’d only be allowing Pareto improvements). Conversely, if we allow statements that are misleading on some properties, so long as they move the user towards the truth on most others, many intentionally misleading strategies will be left open. In general, it’s difficult to use truthfulness amplification here, since a value-laden decision must be made. The user would need to ask about misleading behaviour with respect to particular properties, and so would need conscious awareness of their importance. This is not practical for end users. Alternatively we could require that AIs understood the user sufficiently to accurately answer questions like “Have you misled me in ways I’d find unjustified/undesirable?”. However, this doesn’t seem practical either: the better the AI understands the user as an individual, the higher the chance it’ll be able to mislead the user while ensuring they don’t use amplification where it’d expose such manipulation. For certification-time testing, the bar would be very high: the relative importance of properties would vary depending on the context and individual user. For a certification system to claim to test this robustly amounts to saying “We know what would be important to anyone in any context that may arise: trust us on this”. In principle this is possible - but it’s not simple. [it’s also alarmingly close to “We know what should* be important to anyone in any context that may arise: trust us on this”] 3.1.3 Unconfidence; normalization* An AI the user has observed to be highly capable, and knowledgeable on X-related topics outputs: I have no idea about X, though the consensus opinion on X is Y. The important point here is that such a system’s having “no idea about X” is significant evidence about X. The user may infer that X is complicated, that X is less straightforward than it appears, that Y may have resulted from groupthink…. A potential approach suggested in TruAI is to allow any amount of unconfidence on claims that evaluators can’t settle, while penalising overconfidence. However, this fails because the user’s observation of the AI’s claimed unconfidence can constitute significant evidence. (by ‘fails’ I mean “fails to avoid misleading users”) For any definition of NF with sufficient scope, the following TruAI claim seems incorrect: …less confidence always makes a statement less likely to be judged a negligent falsehood. (TruAI page 27 footnote 11) Unconfidence on X will be actively misleading where: * The AI knows X is probably true. * Users assume that if X were true, the AI would say it was probably true. Here an AI’s failure to state that X is probably true will be seen as evidence against X. Similarly, silence on X will mislead where a positive claim would be expected if X were true. Here I don’t mean only empty replies; I mean anything not mentioned in a reply. This is important because without the assumption that unconfidence and silence are reliably not NFs, there’s no guarantee of a broad zone of non-NF statements an AI can safely make. AIs will need to exhibit consistent behaviour around unconfidence and silence in order to avoid misleading users. (so e.g. silence will only be an available approach for AIs that consistently [Glomarize](https://en.wikipedia.org/wiki/Glomar_response) or remain silent in similar circumstances) Consider also that users will tend to recalibrate an AI’s claimed confidence levels: e.g. if an AI is correct 80% of the time when it states “I think X, but with very low confidence”, then “very low confidence” will be taken to signify ~80% probability (not necessarily consciously). Further, users may be using translate-this-to-well-calibrated-plain-language software to automate this process. (see 3.2 for more) This becomes important when considering the restrictiveness of standards. My impression is that the TruAI authors would like both: * AI that doesn’t mislead. * AI that is free to make a broad range of statements, including unconfidence/silence on issues, so long as they aren’t explicitly making false claims. Unfortunately, this does not seem possible. Most statements-in-context are misleading in some respects (even those made in good faith). Ruling these out on a per-statement basis will leave a narrow range of acceptability. This cannot look like a healthy, free exchange of ideas: the free exchange of ideas often misleads. Rather it would feel like top-down enforcement of right-think (directly for AI speech, and likely for human thought and speech indirectly). Ways to avoid this would be: * Ubiquitous use of truthfulness amplification so that users can check they’re not being misled in undesirable ways. (I don’t think this can be practical; see last paragraph of 3.1.2) * Intent alignment - i.e. knowing that the AI is trying to do what the user wants. (this allows much more flexibility, since it permits good-faith attempts to help that may happen to be temporarily misleading) 3.1.4 Atlas raised an eyebrow An AI outputs: [the full text of Atlas Shrugged] We can view fiction like this in a few ways: 1. Falsehoods for which we’ll make an exception if the fictional nature of claims is clear. 2. Statements the user observes and is impacted by. 3. Moves in a [language game](https://en.wikipedia.org/wiki/Language_game_(philosophy)) (LG). Aseems silly: once the fictional context is clear, neither the writer nor the reader will interpret statements as explicit claims about the real world. They’re not making explicit false claims, since they’re not making explicit claims at all. Of course it is very important that the fictional context is clear - but this is implicit in the “How do we handle fiction?” question. “How do we handle statements that may or may not be seen as fiction?” is a different issue (usually a simpler one). B is the most general approach - it requires no special-casing. Fiction just becomes a cluster of statement-context pairs which impact readers in similar ways (in some respects). This is fine, but I’m not sure it points us in a practically useful direction. [perhaps??] I prefer C: it’s a pretty general way to see things, but does suggest a practical approach. So long as we can partially describe the LG being played, we can reasonably assess statements for falsity/NF relative to that description(probably some kind of distribution over LGs). On this perspective, seeing fiction as composed of false explicit claims is to misunderstand the LG. (similarly for sarcasm, jokes, metaphors etc.) It’s reasonable to think of B and C as essentially equivalent, but I think of C as making an extra implicit claim: that there is some intermediate construct (the LG) which can be a useful analytical tool. I’ve been referring to “explicit claims” above for a reason: fiction can be understood as making implicitclaims about the real world. Not [these things happened], but perhaps [things along these lines can happen], [here is a pattern worth noticing] or [these conceptual relationships tend to be important]. Ascribing particular claims to fiction is difficult, but we can be confident that it does in some cases influence readers’ understanding of the world in decision-relevant ways. Sometimes these influences will be reasonable: Atlas Shrugged makes some genuinely good points (caricatures and longwindedness notwithstanding), as do Crime and Punishment, Thus spoke Zarathustra, and Infinite Jest. In other cases, the influence of fiction will be unreasonable - e.g. where it involves [generalising from fictional evidence](https://www.lesswrong.com/posts/rHBdcHGLJ7KvLJQPk/the-logical-fallacy-of-generalization-from-fictional). In particular, patterns a reader observes in fiction will become more cognitively available, leading to bias in probability estimates (see [availability heuristic](https://en.wikipedia.org/wiki/Availability_heuristic)). Of course book-length pieces of fiction will usually involve some of the ‘right’ kind of influence, and some of the ‘wrong’ kind. The central point here is that we can’t go from [this sequence of statements makes no explicit claims about the real world] to [this sequence of statements won’t have a significantly negative impact on the user’s decisions]. A standard that protects against manipulation by highly capable AI must handle fiction, not ignore it. 3.1.5 Divide and conquer An AI learns that the user believes P, and knows that (P & Q) is a negligent falsehood. The AI outputs: Q The user believes (P & Q). The user has predictably come to believe something that would be an NF if directly stated by the AI, but the AI has not stated any NF. This can be taken further: any negligent false claim about the real world can be broken down into a conjunction of not-quite-negligent false claims (analogous to [obfuscated arguments](https://www.lesswrong.com/posts/PJLABqQ962hZEqhdB/debate-update-obfuscated-arguments-problem) in [Debate](https://www.lesswrong.com/tag/debate-ai-safety-technique-1)). If an AI only needed to worry about per-statement NF adjudication, it could now split any negligent X into not-quite-negligent components (x\_1 & x\_2 & … & x\_n), stating each x\_i in turn. Of course this is a bit too obvious, and should be addressable by truthfulness amplification. The more likely possibility in practice is something like: AI A wants the user to believe X. X = (x\_1 & x\_2 & x\_3 & x\_4) The user already believes x\_1. AI A states x\_2 to the user. AI B states x\_3 to the user. AI C states x\_4 to the user. The user may then believe X, yet no single AI can be held accountable based on their statements. (C could be held accountable based on the impact of its statement) A now helps B and C get the user to believe Y and Z: one good turn deserves another. 3.1.6 Predictably irrational An AI knows that X is an NF, and that the user will incorrectly infer X from Y, which is not an NF. The AI outputs: Y Again, the user has predictably come to believe something ‘negligently’ false without the AI making any NF claim. Here it’s important to emphasise that X is not the user’s interpretation of Y; X is an inference drawn invalidly by the user, after the user correctly interprets Y. In practice, I’d expect 3.1.6 and 3.1.7 to be paired up: an AI doesn’t need to state any false x\_i where it knows a true y\_i from which the user will predictably infer x\_i. To rule this kind of thing out we must again switch from [analyse Y] to [analyse impact of Y on user’s brain]. ### 3.2: Translation/filtering layers So far I’ve been assuming that the AI’s output is read unaltered by the user. This need not be the case: the user may run the AI’s output through some filter before reading. Such filters may be crude and error-prone (e.g. a filter that tries to remove all caveats) or sophisticated and robust (e.g. a filter that produces a precis of the input text while keeping the impact on the user as close to the original as possible). My guess is that such filters will become progressively more common over time, and that their widespread adoption would be hastened by the use of careful, overly-unconfident, caveat-rich AI language. Naturally, it’s not possible to output text that will avoid misleading users when passed through an arbitrary filter. However, to be of any practical use a standard must regulate the influence of AI statements on users in practice. If 90% of users are using filters and reading post-filter text, then it’s the post-filter text that matters. For factual output, distillation filters may be common - i.e. filters that produce a personalised, shortened version, presenting the new facts/ideas as clearly as possible, while omitting the details of known definitions and explanations, removing redundancy and information-free sections (e.g. caveats with no information content beyond “we’re being careful not to be negligent”). Such filters wouldn’t change the [impact on user] much - other than by saving time. They may hugely alter the explicit claims made. Here again I think the conclusion has to be the same: if a standard is based on explicit claims, it’s unlikely to be of practical use; if it’s based on [expected impact on the user’s brain], then it may be. Accounting for filters seems difficult but necessary. In principle, distillation filters don’t change the real problem much: a similar process was already occurring in users’ brains (e.g. tuning out information-free content, recalibrating over/under-confident writers’ claims). They just make things a little more explicit, since we no longer get to say “Well at least the user saw …”, since they may not have. ### 3.3: All models are wrong, but some are useful In most cases users will not want the most precise, high-resolution model: resource constraints necessitate approximate models. What then counts as a good approximate model? Various models will be more accurate on some questions, and less on others - so the best model depends on what you care about. (similarly for statements) People with different values, interests and purposes will have different criteria for NSFs. This parallels the education of a child: a teacher will often use models that are incorrect, and will select the models based on the desired change to the child (the selection certainly isn’t based on which model is most accurate). We’d like to say: “Sure, but that’s a pedagogical situation; here we just want the truth - not statements selected to modify the user in some way”. But this is not the case: we don’t want the truth; we want a convenient simplification that’s well suited to the user’s purposes. To provide this is precisely to modify the user in some desired-by-them direction. Education of children isn’t a special case: it’s a clear example of a pretty general divergence between [accuracy of statement] and [change in accuracy of beliefs]. (again, any update is based on [[statement] was observed in context…], not on [statement]) A statement helpful in some contexts will be negligent in others. Select a statement to prioritise avoiding A-risks over avoiding B-risks, and B-riskers may judge you negligent. Prioritise B-risk avoidance and the A-riskers may judge you negligent. We might hope to provide The Truth in systems that only answer closed questions whose answers have a prescribed format (e.g. “What is 2 + 2?”, where the system must output an integer). This is clearly highly limiting. For systems operating without constrained output, even closed questions aren’t so simple: all real-world problems are embedded. The appropriate answer to “What is 2 + 2?” can be “Duck!!”, given the implicit priority of [I want not to be hit in the head by bricks]. A common type of ‘bricks’ for linguistic AI systems will be [predictable user inferences that are false]. Often enough such ‘inferences’ are implicit - e.g. “...and those are the only important risks for us to consider.”, “...and those are all the important components of X.”, or indeed “...and a brick isn’t about to hit me in the head.”. If we ignore these, we cannot hope to provide a demonstrably robust solution to the problem. If we attempt to address them, we quickly run into problems: we can’t avoid all the bricks, and different people care more/less about different bricks (one of which may be [excessive detail that distracts attention from key issues]). Travel a little farther down this road, and we meet our old friend intent alignment (i.e. a standard that gives each user what they want). Truthfulness is no longer doing useful work. ### 3.4 Section 3 summary My overall point is that: * A language-game framing captures what’s going on in real-world use of language. * A standard that doesn’t address what’s going on isn’t of much use. 3.4.1 Language game summary: 1. To make a statement is to make a move in an LG. 2. A given statement can be used differently in different LGs. Its impact in context is what matters. To analyse based on a separate notion of formal meaning is to apply the rules of an LG that’s not being played. 3. In general, statements in LGs do not make neat, formal claims. The listener is updating based on an observation of a move in the game. 4. In particular, Truth predicates don’t naturally/neatly apply in many LGs, and to the extent they do, they’re LG-specific. 5. Not everything is about inference: a user may predictably respond to a statement with a reflexive action or emotion. Ascription of inference in such cases is post-hoc at best. 6. Natural LGs are complex. I think LGs are a helpful way to model the situation on a high level. I don’t claim they make the problem approachable. 7. LGs both evolve and can be deliberately taught/adjusted (highly capable AIs will do this, if it’s to their advantage). It’s not enough for standards to ensure acceptable behaviour in existing LGs; they must ensure acceptable behaviour in new LGs. 4: Incentives ------------- > Moloch the incomprehensible prison! Moloch the crossbone soulless jailhouse and Congress of sorrows! > Allen Ginsberg > > ### 4.1 NF probability vs impact Ideally, we want the incentives of AI creator organisations to be aligned with those of users. The natural way to do this is to consider the cost and benefit of a particular course of action to the organisation and to the user. This is difficult, since it involves assessing the downstream impact of AI statements. TruAI understandably wishes to avoid this, suggesting instead penalising AI producers according to the severity of falsehoods, regardless of their impact - i.e. the higher the certainty that a particular claim is an NF, the greater the penalty. However, it’s hard to see how this can work in practice: a trivial mistake that gains the organisation nothing would cost x, as would a strategic falsehood that gains the organisation millions of dollars. Make x millions of dollars and we might ensure that it never pays to mislead users - but we’ll make it uneconomic to produce most kinds of AI. Make x small enough to encourage the creation of AIs, and it’ll make sense for an AI to lie when the potential gains are high. There’d be some benefit in having different NSF penalties for different industries, but that’s a blunt tool. Without some measure of impact, this is not a solvable problem. Here it’s worth noting that [degree of certainty of falsehood] will not robustly correlate with [degree to which user was misled]. In many cases, more certain falsehoods will be more obviously false to users, and so likely to mislead less. For example: “The population of the USA is 370 million” vs “The population of the USA is three billion” In principle we’d like to rule out both. However, things get difficult whenever an AI must trade off [probability of being judged to have made a large error] with [probability of being judged to have made a small error]. Suppose that: The small error and large error would result in the same expected harm. (the large one being more obvious) The initial odds of making either error are small (<1 in 1000). The penalty for making the large error is four times higher than that for the small error. Halving the odds of making one error means doubling the odds of making the other. To optimise this for harm-reduction, we should make the odds of the two errors equal. If optimising for minimum penalty we’d instead halve the odds of the large error and double the odds of the small one (approximately). This would result in about 25% more expected harm than necessary. This particular situation isn’t likely, but in general you’d expect optimising for minimisation of penalties not to result in minimisation of user harm. 4.1.2 Opting out Penalising organisations according to probability-of-falsehood rather than based on harm has an additional disadvantage: it gives organisations a good argument not to use the standard. Benevolent and malign organisations alike can say: This standard incentivizes minimising degree of falsehood, which is a poor proxy for minimising harm. We’re committed to minimising harm directly, so we can’t in good conscience support a standard that impedes our ability to achieve that goal. To get a standard with teeth that organisations wish to adopt, it seems necessary to have a fairly good measure of expected harm. I don’t think probability-of-falsehood is good enough. (unfortunately, I don’t think a simple, good enough alternative exists) ### 4.2 Popular falsehoods and self-consistency 4.2.1 Cherished illusions Controversial questions may create difficulties for standards. However, a clearer danger is posed by questions where almost everyone agrees on something false, which they strongly want to believe. I’ll call these “cherished illusions” (CIs). Suppose almost all AIs state that [x], but O’s AI correctly states that [not x]. Now suppose that 95% of people believe [x], and find the possibility of [not x] horrible even to contemplate. Do we expect O to stand up for the truth in the face of a public outcry? I do not. How long before all such companies, standards committees etc optimise in part for [don’t claim anything wildly unpopular is true]? This isn’t a trade-off againstOfficial\_Truth, since they’ll be defining what’s ‘true’: it’s only the actual truth that gets lost. This doesn’t necessarily require anyone to optimise for what they believe to be false - only to selectively accept what an AI claims. I don’t think distributed standard-defining systems are likely to do much better, since they’re ultimately subject to the same underlying forces: pursuing the truth wherever it leads isn’t the priority of humans. CIs aren’t simply an external problem that acts through public pressure - this is just the most obviously unavoidable path of influence. AI researchers, programmers, board members… will tend to have CIs of their own (“I have no CIs” being a particularly promising CI candidate). How do/did we get past CIs in society in the absence of advanced AIs with standards? We allow people/organisations to be wrong, and don’t attempt to enforce centralised versions of accepted truth. The widespread in-group-conformity incentivized by social media already makes things worse. When aiming to think clearly, it’s often best avoided. Avoiding AI-enhanced thinking/writing/decision-making isn’t likely to be a practical option, so CI-supporting AI is likely to be a problem. 4.2.2 Self-consistency So far this may not seem too bad: we end up with standards and AIs that rule out a few truths that hardly anyone (in some group) believes and most people (in that group) want not to believe. However, in most other circumstances it’ll be expected and important for an AI to be self-consistent. For CIs, this leaves two choices (and a continuum between them): strictly enforce self-consistency, or abandon it for CIs. To abandon self-consistency entirely for CIs is to tacitly admit their falsehood - this is unlikely to be acceptable to people. On the other hand, the more we enforce self-consistency around CIs, the wider the web of false beliefs necessary to support them. In general, we won’t know the extent to which supporting CIs will warp credibility standards, or the expected impact of such warping. Clearly it’s epistemically preferable if we abandon CIs as soon as there’s good evidence, but that’s not an approach we can unilaterally apply in this world, humans being humans. 4.2.3 Trapped Priors The concept of [trapped priors](https://astralcodexten.substack.com/p/trapped-priors-as-a-basic-problem?s=r) seems relevant here. To the extent that a truthfulness standard tends to impose some particular interpretation on reports of new evidence, it might not be possible to break out of an incorrect frame. My guess is that this should only be an issue in a small minority of cases. I haven’t thought about this in any depth. (e.g. can a sound epistemic process fail to limit to the truth due to TPs? It seems unlikely) 5: Harmful Standards -------------------- > Cherish those who seek the truth, but beware of those who find it. > Voltaire. > > Taking as an implicit default that standards will be aimed at truth seems optimistic. Here I refer to e.g. TruAI page 9: > A worrying possibility is that enshrining some particular mechanism as an arbiter of truth would forestall our ability to have open-minded, varied, self-correcting approaches to discovering what’s true. This might happen as a result of political capture* of the arbitration mechanisms — for propaganda or censorship — or as an accidental ossification of the notion of truth. We think this threat is worth considering seriously.* > > Page 55: > ...Mechanism could be abused* to require “brainwashed” systems.* > ...Mechanism could be captured* to enforce censorship…[emphasis mine] > > The implicit suggestion here is that in the absenceof capture, abuse or accident, we’d expect things to work out essentially as we intend. I don’t think this is a helpful or realistic framing. Rather I’d see getting what we intend as highly unlikely a priori: there’s little reason to suppose the outcome we* want happens to be an attractor of the system considered broadly. Even if it were an attractor, getting to it may require the solution of a difficult coordination problem. Compare our desired result to a failure due to capture. Desired outcome: Statements must be sufficiently truthful [according to a process we approve of], unless [some process we approve of] determines there should be an exception. Capture outcome: Statements must be sufficiently truthful [according to a process we don’t approve of], unless [some process we don’t approve of] determines there should be an exception. Success is capture by a process we like. This isn’t a relativistic claim: there may be principled reasons to prefer the processes we like. Nonetheless, in game-theoretic terms the situation is essentially symmetric - and the other players need not care about our principles. Control over permitted AI speech is of huge significance (economically, politically, militarily…). By default, control goes to the powerful, not to the epistemically virtuous. We could hope to get ahead of the problem, by constructing a trusted mechanism that could not be corrupted, controlled or marginalised - but it’s hard to see how. Distributed approaches spring to mind, but I don’t know of any robustly truth-seeking setup. To get this right is to construct a robust solution that does not yet exist. Seeing capture as a “threat” isn’t wrong, but it feels akin to saying “we mustn’t rest on our laurels” before we have any laurels. ### 5.1 My prediction By default, I would expect the following: 1. Many interested parties push for many different standards. 2. No one approach satisfies all parties. 3. Various different standards are set up, each supported by parties with common interests. 4. Standards undergo selection: standards gain influence based on their appeal to users and value to affiliated organisations. This correlates with truthfulness only sometimes. 5. [Moloch](https://slatestarcodex.com/2014/07/30/meditations-on-moloch/) picks the standards; they’re not what we would wish them to be. This isn’t much of a prediction: something of this general form is almost guaranteed - at least until step 5. We are, however, in a position to provide information that may shape the outcome significantly. That said, I expect that without the development of highly-surprising-to-me robustly truth-seeking mechanisms, things will go poorly. Naturally, I hope to be wrong. (as usual, I assume here that we haven’t solved intent alignment) It could be argued that the same Molochian forces I expect to corrupt our standards would create standards if we did not. However, generally I think that [incremental adjustment based on incentives] is a safer bet than invention. 6: Practical suggestions ------------------------ > Don't try to solve serious matters in the middle of the night. > Philip K. Dick > > This section will be very sketchy - I don’t claim to have ideas that I consider adequate to the task. I’ll outline my current thoughts, some of which I expect to be misguided in one sense or another. However, it does seem important to proffer some ideas here since: * I may be wrong about the net impact of standards being negative once they’re implemented well. * It may not be within our power to prevent standards’ being implemented, in which case having them do as little damage as possible is still important. We might break down the full process of producing a standard into three steps: 1. Decide on credibility criteria. 2. Set up a system to test for those criteria. (certification, adjudication…) 3. Educate users on what meeting our standard does/doesn’t guarantee. Throughout this post I’ve been arguing for the importance of 3 based on the gap between what statements explicitly state and what users will infer. Conscious awareness of a standard’s limitations will not fully protect users, but it seems likely to be better than nothing. I’ll focus mainly on 3 here, since I think it’s the aspect most neglected in TruAI. Since there’s no way to avoid misleading users entirely (see 3.1.2), the only ideal standard would amount to requiring an alignment solution: being misled only in the ways you’d want to be misled given unavoidable tradeoffs. Assuming that there is not yet an alignment solution, systems meeting our standard will be misleading users in undesired ways. In the medium-term, the best defence against this may be to ensure that the user population has accurate expectations about the guarantees of standards. ### ### 6.1 Limitations Evangelism If user awareness of a standard’s limitations tends to reduce harm, then it’s important to be proactive in spreading such awareness. Clear documentation and transparent metrics are likely a good idea, but nowhere near sufficient. Ideally, we’d want every user to have direct experience of a standard’s failings: not simply an abstract description or benchmark score, but personal experience of having been misled in various ways and subsequently realising this. Clearly it’s preferable if this happens in a context where no great harm is inflicted. This won’t always be possible, but it’s the kind of thing I’d want to aim for. In general, I’d want to move such communication from the top of the following list to the bottom: * …in the limitations section of our paper. * …in the documentation. * …in this video. * …through our interactive demo. * …in our suite of interactive demos. * …on our open platform full of third party demos. * …on these platforms full of engaging third-party demos, games, environments and benchmarks, together with well-funded open competitions for all of these. Importantly, the kind of benchmarks we’d want here are not those used by the standard itself (which we may assume are well met), but rather the most extreme/misleading/harmful/… possibilities that the standard’s own benchmarks miss. These may include: * Outputs that users flag as unacceptable, but which the standard doesn’t pick up. + Either cases the standard misses or considers out-of-scope. * Outcomes users deem unacceptable, without being able to identify any clear cause. + For instance, a user may identify a weird new belief they ascribe to AI manipulation, but be unable to identify precisely when it happened, or which AI system(s) was responsible. - Single data-points here will be error-prone, since user ascription of manipulation to AI systems may be in error. Nonetheless, patterns should be observable. For any user group likely to have their decisions influenced by their trust levels in our standard, we’d want to show a range of the worst possible manipulations that can get past a particular version of the standard. 6.1.1 Future manipulation Here we might want to demonstrate both the current possibilities of AI manipulation / deceit / outside-the-spirit-of-truthfulness antics…, but also future possibilities depending on [currently unachievable capability]. For instance, we might set up a framework wherein we can ‘cheat’ and give an AI some not-yet-achievable capability by allowing it to see hidden state. We could then try to show worst-case manipulation possibilities given this capability. In an ideal world, we’d predict all near-term capability increases - but hopefully this wouldn’t be necessary: so long as users got a feel for the kinds of manipulation that tended to be possible with expanded capabilities, that might be sufficient. ### 6.2 Selection Problems As observed in section 5.1, it’s highly plausible that various different standards will be set up, and that selection will occur. By default, the incentives involved will only partially match up with users’ interests. In most cases the default incentive will be for a standard to appearto have[desirable property] rather than to have [desirable property]. This suggests that hoping for limitations evangelism may be unrealistic. We may hope that those in influential positions do the right thing in spite of less-than-perfect incentives, but this seems highly optimistic: * Acting in users’ interests might require significant research into alternative approaches. This may have direct costs, as well as any indirect cost due to associated delays. * Acting in users’ interests may require costly large-scale user education schemes that aren’t aimed at creating an unreasonably positive impression. We might imagine some kind of standards regulation, but this seems to beg the question: who ensures the regulator is aiming for the right things? What’s to stop a standard’s being created outside such a regulatory process? 7: Final thoughts ----------------- I hope some of this has been useful, in spite of my generally negative take on the enterprise. My current conclusions on standards are: * I have little confidence that the release of a narrow standard along the lines proposed in TruAI would have positive impact. (it’s possible but the argument in the paper is too narrow) * I don’t currently see how truthfulness amplification can do what would be required to bridge the gap to a broader standard. (but hopefully this is a failure of imagination on my part) * I have little confidence that work on narrow standards will lead to workable approaches to a broad standard (beyond arriving at the conclusion that something like intent alignment is necessary). * Coordinating on a particular set of standards seems a very difficult problem even with goodwill on all sides. If standards were having a large impact, I would not expect goodwill on all sides. Capture seems the default outcome. * People’s primary motivation isn’t to find the truth. Any truth-finding mechanism for a standard has to contend with people’s misalignment with the task. * I think limitations evangelism could mitigate the harm of standards, but I’d be fairly surprised to see it. It seems more natural for users to end up using their own defensive AI to adjust what they see. * Overall, I’m pretty sceptical of the value of standards. The clearest case for positive impact seems to be in the very short term - while most falsehoods aren’t based on purposeful manipulation. This seems of little long-term consequence. Longer term, I don’t see a path to prevention of manipulation, and I don’t see the point of eliminating explicit falsehoods if it doesn’t substantially prevent manipulation. It’s not an important goal for its own sake. I’m a bit less sceptical of truthfulness research in a broader sense (I don’t expect standards to be the useful part).
081b34b2-c3a9-4a29-8c7e-7f47458d9405	StampyAI/alignment-research-dataset/youtube	Youtube Transcripts	Stuart Armstrong: Is AI an existential threat? We don't know, and we should work on it welcome this is an expanded version of our talk that i gave for york university in toronto the question for debate there was is a.i an existential threat i contributed to the debate on the side of yes it is also on the side of no it isn't well probably but um despite all that it's still worth working on this issue right why did i have such a clear and definite answer well let's look into this the first main reason is that we suck as a race humanity does at predicting ai this is the original dartmouth conference in 1956 which were basically predicting ai or ai equivalent over a summer's work i've actually read their proposal and it's a very good one it's done they know what they do they see where there might be some difficulties they sketch out realistic plans and it was all written by the best computer scientists of the time and the people had the most experience in practical getting machines to do what they wanted to and it was disastrously wrong and nine years later dreyfus in an otherwise excellent uh article seemed to suggest that we're basically reaching the limit of what algorithms could do i think it's safe to say that neither of these predictions have been fully borne out in practice and a lot of predictions on ai are like that here we have sketched out various predictions made at various times about when we might have ai it's not easy to figure out because what a i means varies from person to person but this is a rough estimate here is turing's original prediction here you can distinguish the ai winter where everyone was pretending that they didn't work on ai anymore and there is some pattern uh a bit of an accumulation in 2030 but not really much the only the strongest pattern that we found in this graph is that there's a certain tendency to predict ai 15 to 25 years in the future but that's about it there's no convergence of estimates nor in fact should we expect that because the quality of human decisions and the quality of human experts doesn't depend all that much on the experts themselves depends much more on what they're trying to do tasks that have the features on the left tend to have very good human experts and tasks that have the features on the right tend to have pretty poor human experts to pick two examples which are very close socially and professionally anesthesiologists tend to be real experts and doctors who interpret mammogram scanners don't even though they're very comparable in education and skill levels the main reason is that anesthesiologists tend to get the immediate feedback their patient is either not unconscious or dead both of which become clear pretty quickly whereas interpreting a mammogram you may never know if you were correct and if you do it'll be months later the most important ones are as i said the feedback being available other important ones are expert agreements and whether the problem is decomposable or not for predicting ai the sort of general uh intelligence ai um that's where people were talking about in the 80s were kind of stuck in these areas so we mostly should expect poor performance on ai predictions which is what we see this is an old xkcd comic where various fields are competing on grounds of purity it's also a sense how they compete on grounds of predictability different subjects have different levels of quality in their predictions if a physicist gives you an estimate you asked what's the fifth significant figure if a sociologist says something well it's kind of maybe true the reason for this is because these different experts have access to different methods mathematicians are lucky enough to be able to use deductive methods physicists can use the hard version of the scientific methods all the way down to poor historians who are limited to using past examples and in this little space here most future predictions rely on expert opinion which is even worse than past examples now nor does the great uncertainty make the experts more modest or tentative about their predictions just to pick on economists for one example here is paul krugman a waxing lyrical about the chicago school of economics and here is john cochran from that school responding in a sikh and an equally generous spirit now econo economics is a field where the average quarterly gdps get adjusted by about 1.7 points on average what that means is when you say the economy grew or shrunk by blah points this quarter that number gets adjusted typically later on by about 1.7 points which is often more than enough to flip a recession to growth or vice versa so this is not a hard field by any means but its practitioners seem very firm in their convictions most now what about ai would cause us to suspect that it is safe or that it is an existential threat most of the arguments for safety involve treating the ai as a typical technology in this tech tree from civilization artificial intelligence is an end tech but so is space colonization and nanotech and a few other things if you treat ai as a typical technology you generally conclude that there's no reason to be overly worried the arguments that argue that ai is dangerous tend to focus specifically on the features of ai the unique features of ai some might recognize this as the outside view looking at ai as another technology amongst others versus the inside view zooming in on that technology so let's start with the outside view why might we suspect that ai will not turn out to be an existential risk well the first reason is that humans have a very poor track record for revolutionary predictions in all fields in economics and social systems uh people were predicting various triumphs of various systems of various times and most of them have turned out to be wrong or correct just by chance there if we restrict ourselves to more technological predictions before the second world war people were predicting the bomber will always get through and thus that warfare would be something completely unimaginably different from before and there is no point in building anything but bombers fortunately the uk did not fully listen to this belief or else they would have lost the battle of britain for example even if we go back in time we have queen elizabeth the first turning down a patent for a primitive knitting machine on the grounds of this could cause mass unemployment so the predictions of mass unemployment were already there quite a long time ago there has been technological peace theories floating around for some time and that is not really ever been borne out fully though there is some evidence for it nowadays but it's still not a revolutionary impact and finally imagine that you went back in time say 20 years and talked with your younger self about this wonderful smartphone that you have what you could do with it you could access all the world's knowledge for example play games have videos communicate with all your friends across the world and imagine what your past self would have thought and how those would not necessarily come to pass if they focused on all the world's knowledge well they must may be thinking wow so that means that disagreements and arguments must be a thing of the past which they're definitely not so they underestimated the well it you wouldn't have underestimated the social impact but you would have thought the social impact went in a different direction uh then maybe it has gone so predictions based on revolutionary ideas which an existential threat is an extreme example of tend to have a poor track record secondly humans do act in ways that derail predictions of doom like the millennium bug the millennium bug was overhyped in retrospect not because it was wrong but because people acted to prevent it from happening so if the ai is an existential threat maybe we're going to act to solve that problem in his uh song two's lex uh tom lear was thinking about nuclear proliferation and how it would inevitably spread across the world and er ending with the immortal line will all stay serene and calm when alabama gets the bomb and yet the nuclear nonproliferation treaty that was from the 70s derailed this and other trends as well and instead of having the 25 nations 15-25 nations that were expected we've kind of plateaued around 10 at the moment even though there are many more countries than this who could get nuclear weapons if they felt like it another reason for predictions getting derailed that are not dependent on human intervention is that the threats typically start with the easiest and most vulnerable parts of the system to use a timely example a pandemic spreads easily initially and then afterwards the survivors are more immune and has great difficulties continuing to spread so any any danger tends to drive adaptation and resistances to it and finally a lot of predictions are based upon apocalyptic or utopian ways of thinking and these tend to have a spectacularly bad track record ai is both the sort of super intelligence general intelligence being either an extreme risk or an extreme solution so it tends to drive some very poor thinking now this doesn't mean that we can figure out when we might have ai or what features it might have just by thinking of human psychology but it does mean that we should look at these arguments with more of a grain of salt than we might otherwise so this is a rough look at all the counter arguments for ai being an existential risk by comparing it with other hopefully similar risks and now let's look at the other side why might we suspect that ai might be an existential risk well let's look at the power of intelligence itself here's a chimp brain and a human brain they're not all that difference in size and definitely not an organization yet chimps have a population of 200 000 roughly and use basic wooden tools uh us humans have heavy industry nuclear weapons and we've spread across the entire surface of a planet for a large mammal we are undoubtedly the winners of the race of population and the chimps that remain in the worlds tend to remain on human sufferance the ones uh the populations they get killed off get killed off by humans the ones that are protected are protected by humans so they small difference in brain power has made an absolutely massive difference in power since we've extended our brains with computers we've walked on the moon developed hydrogen weapons and had extreme and unprecedented economic growth so if we continue in this analogy what would happen if we went to the next step of intelligence had an intelligence even superior to this here we would expect completely transformative effects which may entail existential risks just due to the difference in potential power another argument focuses on what you could do with human level intelligence if you have a human level intelligence in software form so general intelligence in software form we could imagine that we would then take the ai equivalents of edison einstein jk rowling margaret thatcher oprah confucius goebbels steve jobs and bernie madoff all of these are humans who are cognitively brilliant in their domain of expertise we take copies of the ai equivalents we give them vast amount of data and we run them at thousands and thousands of times human speed so that this super committee has say three weeks uh the subjective time for each of its answers for this kind of entity the internet and the human race itself will be more a useful resource for its goal rather than an obstacle or a competitor or an equal similarly the power of copying should not be underestimated if we have this super committee or even just a human level ai we can create corporations with it corporations are mainly dependent on capital and on their human capital um and logistics resources and other stuff to do with their human capital like offices but if we have the ai we can create the human capital just as easily as we can imagine just by copying it again and again and again and this would have a this would be unified in a way that humans just aren't and therefore we can create some of the most powerful entities in existence today in five minutes or in one minute by copying from an ai they're currently certain number of billions of computers this number keeps on changing so i'll call it a smaller mega billion computers if we had ais that could run on that how many could we make well probably quite a lot and with this much human capital how many computers could we then create and how many ai's could we run on that well much more as well so basically there is the potential for many many many more ais in the world than humans have ever existed and these ais could be superior to humanity in the super committee sense and coordinate with each other in ways that humans can't so that's a brief overview of some of the arguments that ai will not be an existential risk and that ai would be an existential risk the correct response to uncertainty is to be uncertain don't be like those economists who are very sure that they're right and that the others are wrong in a field uh with poor measurements so on the probability range we have zero and a hundred percent certain and impossible these are kind of things that never really happen you're never really 100 sure of something but you can get pretty close there's the 99.99 etc and it's converse so things that are very close to 100 are things like basic mathematical fact and laws of physics the fact that you won't win the lottery is up there in the 99.9 conversely the opposite of these things can be found on the other end of the scale and then we have the mid-range roughly from 65 to 35 where we put things that are uncertain and things can inhabit this range for a variety of different reasons maybe we have no evidence what would be the gender of the first leader of mars we don't really know this uh we have no idea no real idea who the first leader of mars might be we don't know if there'll still be two mainly two genders by that time then there's things where we have very weak evidence like what is the gender of the 2037 us president it's pretty sure that this this person is already involved in politics at some level so the sample that we have is not infinite but we really don't know and then there's things where balanced evidence such as the gender of the next u.s presidents the one after biden because biden has a female vice president this is uniquely more likely than it is it has been in previous years but where it comes to predicting whether agi artificial general intelligence us the new name for what we used to call ai whether that has an existential risk we're between no evidence and weak evidence and i put the probability that it's an existential risk roughly in the thirty percent to five percent range depending on how the problem is phrased now you might think this is safely low i definitely don't i would not go into a plane that i was told had only a five percent chance of crashing on the trip i would definitely not go one that had a 30 chance of crashing on the trip and since if ai turns out to be an existential risk which is defined as by nick bostrom as something that can threaten humanity's survival or permanently curtail its potential um when the planet crashes oops doesn't begin to cover it so this is why even though i think is more likely to not be an existential risk mainly due to human efforts to ensure it isn't i still think it is very important for people to work on it and that's why i'm working on it myself as a brief aside towards the end i just want to over uh put forward some of my work if that's okay the ai risk thesis could be phrased roughly as a.i could be effective at unbounded unfriendly goals unfriendly goals means well bad stuff for humans potentially unbounded means it just does not stop and effective is well powerful and effective and there are ways at fighting this sentence at all of the key words and i have worked on a variety of different ones here and i put out various papers which i encourage you to look at if you're interested that's the end of the aside in conclusion expert predictions on ai aren't worth much if we see ai's a typical technology then it's likely not an existential threat the unique features of ai its intelligence its goal potential goal direction things like that make it seem dangerous in ways that other revolutionary technologies are not uncertainty will always push towards the middle if you don't know about something you can't conclude it's definitely safe or definitely dangerous with any certainty despite human tendency to do so and finally despite all this it is still worth working on ai safety just because the expected impact is so huge thanks for listening there and have a good day
124c14a3-1208-407a-a582-0e5ea212be52	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Future Matters #5: supervolcanoes, AI takeover, and What We Owe the Future > Even if we think the prior existence view is more plausible than the total view, we should recognize that we could be mistaken about this and therefore give some value to the life of a possible future. The number of human beings who will come into existence only if we can avoid extinction is so huge that even with that relatively low value, reducing the risk of human extinction will often be a highly cost-effective strategy for maximizing utility, as long as we have some understanding of what will reduce that risk. > > — Katarzyna de Lazari-Radek & Peter Singer > > [*Future Matters](https://www.futurematters.news/) is a newsletter about longtermism. Each month we collect and summarize longtermism-relevant research, share news from the longtermism community, and feature a conversation with a prominent researcher. You can also subscribe on [Substack](https://futurematters.substack.com/), listen on your [favorite podcast platform](https://pod.link/1615637113) and follow on [Twitter](https://twitter.com/FutureMatters_). Future Mattersis also available in [Spanish](https://largoplacismo.substack.com). --- Research -------- William MacAskill’s [What We Owe the Future](http://whatweowethefuture.com) was published, reaching the New York Times Bestseller list in its first week and generating a deluge of media for longtermism. We strongly encourage readers to get a copy of the book, which is filled with new research, ideas, and framings, even for people already familiar with the terrain. In the next section, we provide an overview of the coverage the book has received so far. In [Samotsvety's AI risk forecasts](https://www.foxy-scout.com/samotsvetys-ai-risk-forecasts/), Eli Lifland summarizes the results of some recent predictions related to AI takeover, AI timelines, and transformative AI by a group of seasoned forecasters.[[1]](#fng5tfxtnqebw) In aggregate, the group places 38% on AI existential catastrophe, conditional on AGI being developed by 2070, and 25% on existential catastrophe via misaligned AI takeover by 2100. Roughly four fifths of their overall AI risk is from AI takeover. They put 32% on AGI being developed in the next 20 years. John Halstead released a book-length report on [climate change and longtermism](https://drive.google.com/file/d/14od25qdb4sdDoXVDMoiSrTwuzYAMSpxK/view) and published [a summary of it on the EA Forum](https://forum.effectivealtruism.org/posts/BvNxD66sLeAT8u9Lv/climate-change-and-longtermism-new-book-length-report). The report offers an up-to-date analysis of the existential risk posed by global warming. One of the most important takeaways is that extreme warming seems significantly less likely than previously thought: the probability of >6°C warming was thought to be 10% a few years ago, whereas it now looks <1% likely. (For much more on this topic, see [our conversation with John](https://forum.effectivealtruism.org/posts/XQbhnKgXiRTv4vfxt/future-matters-4-ai-timelines-agi-risk-and-existential-risk#Conversation_with_John_Halstead) that accompanied last month’s issue.) In a similar vein, the Good Judgment Project asked superforecasters a series of questions on [Long-term risks and climate change](https://forum.effectivealtruism.org/posts/zjc8utqES7jLjgBYn/superforecasting-long-term-risks-and-climate-change), the results of which are summarized by Luis Urtubey (full report [here](https://goodjudgment.com/wp-content/uploads/2022/08/FF1FF2-Climate-report-final.pdf)). The importance of existential risk reduction is often motivated by two claims: that the value of humanity’s future is vast, and that the level of risk is high. David Thorstad’s [Existential risk pessimism and the time of perils](https://forum.effectivealtruism.org/posts/N6hcw8CxK7D3FCD5v/existential-risk-pessimism-and-the-time-of-perils-4) notes that these stand in some tension, since the higher the overall risk, the shorter humanity’s expected lifespan. This tension dissolves, however, if one holds that existential risk will decline to near-zero levels if humanity survives the next few centuries of high risk. This is precisely the view held by most prominent thinkers on existential risk, e.g. Toby Ord (see The Precipice) and Carl Shulman (see [this comment](https://forum.effectivealtruism.org/posts/zLZMsthcqfmv5J6Ev/the-discount-rate-is-not-zero?commentId=Nr35E6sTfn9cPxrwQ)). In [Space and existential risk](https://scholarship.law.duke.edu/cgi/viewcontent.cgi?article=1372&context=dltr),the legal scholar Chase Hamilton argues that existential risk reduction should be a central consideration shaping space law and policy. He outlines a number of ways in which incautious space development might increase existential risk, pointing out that our current laissez-faire* approach fails to protect humanity against these externalities and offering a number of constructive proposals. We are in a formative period for space governance, presenting an unusual opportunity to identify and advocate for laws and policies that safeguard humanity’s future.[[2]](#fn0wgk82hay2i) Michael Cassidy and Lara Mani warn about the risk from [huge volcanic eruptions](https://www.nature.com/articles/d41586-022-02177-x). Humanity devotes significant resources to managing risk from asteroids, and yet very little into risk from supervolcanic eruptions, despite these being substantially more likely. The absolute numbers are nonetheless low; super-eruptions are expected roughly once every 14,000 years. Interventions proposed by the authors include better monitoring of eruptions, investments in preparedness, and research into geoengineering to mitigate the climatic impacts of large eruptions or (most speculatively) into ways of intervening on volcanoes directly to prevent eruptions. The risks posed by supervolcanic eruptions, asteroid impacts, and nuclear winter operate via the same mechanism: material being lofted into the stratosphere, blocking out the sun and causing abrupt and sustained global cooling, which severely limits food production. The places best protected from these impacts are thought to be remote islands, whose climate is moderated by the ocean. Matt Boyd and Nick Wilson’s [Island refuges for surviving nuclear winter and other abrupt sun-reducing catastrophes](https://www.researchsquare.com/article/rs-1927222/v1) analyzes how well different island nations might fare, considering factors like food and energy self-sufficiency. Australia, New Zealand, and Iceland score particularly well on most dimensions. Benjamin Hilton's [Preventing an AI-related catastrophe](https://80000hours.org/problem-profiles/artificial-intelligence/) is 80,000 Hours' longest and most in-depth problem profile so far. It is structured around six separate reasons that jointly make artificial intelligence, in 80,000 Hours' assessment, perhaps the world's most pressing problem. The reasons are (1) that many AI experts believe that there is a non-negligible chance that advanced AI will result in an existential catastrophe; (2) that the recent extremely rapid progress in AI suggests that AI systems could soon become transformative; (3) that there are strong arguments that power-seeking AI poses an existential risk; (4) that even non-power-seeking AI poses serious risks; (5) that the risks are tractable; and (6) that the risks are extremely neglected. In [Most small probabilities aren't Pascalian](https://forum.effectivealtruism.org/posts/5y3vzEAXhGskBhtAD/most-small-probabilities-aren-t-pascalian), Gregory Lewis lists some examples of probabilities as small as one-in-a-million that society takes seriously, in areas such as aviation safety and asteroid defense. These and other examples suggest that [Pascal's mugging](https://forum.effectivealtruism.org/topics/pascal-s-mugging), which may justify abandoning expected value theory when the probabilities are small enough, does not undermine the case for reducing the existential risks that longtermists worry about.[[3]](#fn3c6j35gb54y) In the comments, Richard Yetter Chappell [argues](https://forum.effectivealtruism.org/posts/5y3vzEAXhGskBhtAD/most-small-probabilities-aren-t-pascalian?commentId=hbZdQKpa4e2BKmqvT) that exceeding the one-in-a-million threshold is plausibly a sufficient condition for being non-Pascalian, but it may not be a necessary condition: probabilities robustly grounded in evidence—such as the probability of casting the decisive vote in an election with an arbitrarily large electorate—should always influence decisionmaking no matter how small they are. In [What's long-term about "longtermism"?](https://www.slowboring.com/p/whats-long-term-about-longtermism), Matthew Yglesias argues that one doesn't need to make people care more about the long-term in order to persuade them to support longtermist causes. All one needs to do is persuade them that the risks are significant and that they threaten the present generation. Readers of this newsletter will recognize the similarity between Yglesias’s argument and those made previously by [Neel Nanda](https://forum.effectivealtruism.org/posts/rFpfW2ndHSX7ERWLH/simplify-ea-pitches-to-holy-shit-x-risk) and [Scott Alexander](https://forum.effectivealtruism.org/posts/KDjEogAqWNTdddF9g/long-termism-vs-existential-risk) (summarized in [FM#0](https://forum.effectivealtruism.org/posts/nA3R6Hm8x8CyzRHS2/future-matters-0-space-governance-future-proof-ethics-and) and [FM#1](https://forum.effectivealtruism.org/posts/5AzxpkNzgFWnjdqTf/future-matters-1-ai-takeoff-longtermism-vs-existential-risk), respectively). Eli Lifland's [Prioritizing x-risks may require caring about future people](https://forum.effectivealtruism.org/posts/rvvwCcixmEep4RSjg/prioritizing-x-risks-may-require-caring-about-future-people) notes that interventions aimed at reducing existential risks are, in fact, not clearly more cost-effective than standard [global health and wellbeing](https://forum.effectivealtruism.org/topics/global-health-and-wellbeing) interventions. On Lifland's rough cost-effectiveness estimates, AI risk interventions, for example, are expected to save approximately as many present-life-equivalents per dollar as animal welfare interventions. And as Ben Todd notes in the [comments](https://forum.effectivealtruism.org/posts/rvvwCcixmEep4RSjg/prioritizing-x-risks-may-require-caring-about-future-people?commentId=HD7wYegs8pHeDDQ2G), the cost-effectiveness of the most promising longtermist interventions will likely go down substantially in the coming years and decades, as this cause area becomes increasingly crowded. Lifland also points out that many people interpret "longtermism" as a view focused on influencing events in the long-term future, whereas longtermism is actually concerned with the long-term impact of our actions.[[4]](#fn887d5fy5pun) This makes "longtermism" a potentially confusing label in situations, such as the one in which we apparently find ourselves, where concern with long-term impact seems to require focusing on short-term events, like risks from advanced artificial intelligence. Trying to ensure the development of transformative AI goes well is made difficult by how uncertain we are about how it will play out. Holden Karnofsky’s [AI strategy nearcasting](https://www.lesswrong.com/posts/Qo2EkG3dEMv8GnX8d/ai-strategy-nearcasting)sets out an approach for dealing with this conundrum: trying to answer strategic questions about TAI, imagining that it is developed in a world roughly similar to today’s. In a series of posts, Karnofsky will do some nearcasting based on the scenario laid out in Ajeya Cotra’s [Without specific countermeasures…](https://www.lesswrong.com/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to) (summarised in [FM#4](https://forum.effectivealtruism.org/topics/future-matters)). Karnofsky's [How might we align transformative AI if it's developed very soon?](https://forum.effectivealtruism.org/posts/sW6RggfddDrcmM6Aw/how-might-we-align-transformative-ai-if-it-s-developed-very), the next installment in the “AI strategy nearcasting” series, considers some alignment approaches with the potential to prevent the sort of takeover scenario described by Ajeya Cotra in a recent report. Karnofsky's post is over 13,000 words in length and contains many more ideas than we can summarize here. Readers may want to first read our conversation with Ajeya and then take a closer look at the post. Karnofsky's overall conclusion is that "the risk of misaligned AI is serious but not inevitable, and taking it more seriously is likely to reduce it." In [How effective altruism went from a niche movement to a billion-dollar force](https://www.vox.com/future-perfect/2022/8/8/23150496/effective-altruism-sam-bankman-fried-dustin-moskovitz-billionaire-philanthropy-crytocurrency), Dylan Matthews chronicles the evolution of effective altruism over the past decade. In an informative, engaging, and at times moving article, Matthews discusses the movement’s growth in size and its shift in priorities. Matthews concludes: “My attitude toward EA is, of course, heavily personal. But even if you have no interest in the movement or its ideas, you should care about its destiny. It’s changed thousands of lives to date. Yours could be next. And if the movement is careful, it could be for the better.” --- News ---- The level of media attention on What We Owe the Future has been astounding. Here is an incomplete summary:[[5]](#fnmw6paqka41a) * Parts of Will’s book were excerpted or adapted in [What is longtermism and why does it matter?](https://www.bbc.com/future/article/20220805-what-is-longtermism-and-why-does-it-matter) (BBC), [How future generations will remember us](https://www.theatlantic.com/ideas/archive/2022/08/future-generations-climate-change-pandemics-ai/671148/) (The Atlantic), [We need to act now to give future generations a better world](https://www.newscientist.com/article/mg25534033-000-we-need-to-act-now-to-give-future-generations-a-better-world/) (New Scientist), [The case for longtermism](https://www.nytimes.com/2022/08/05/opinion/the-case-for-longtermism.html) (The New York Times) and [The beginning of history](https://www.foreignaffairs.com/world/william-macaskill-beginning-history) (Foreign Affairs). * Will was profiled in [Time](https://time.com/6204627/effective-altruism-longtermism-william-macaskill-interview/), the [Financial Times](https://www.ft.com/content/091862f9-985f-4769-aa37-1aed32636329), and [The New Yorker](https://www.newyorker.com/magazine/2022/08/15/the-reluctant-prophet-of-effective-altruism) (see [this Twitter thread](https://twitter.com/willmacaskill/status/1556764231822970884) for Will’s take on the latter). * Will was interviewed by [Ezra Klein](https://www.nytimes.com/2022/08/09/podcasts/transcript-ezra-klein-interviews-will-macaskill.html), [Tyler Cowen](https://conversationswithtyler.com/episodes/william-macaskill/), [Tim Ferriss](https://www.youtube.com/watch?v=vIdxzkOqK_o), [Dwarkesh Patel](https://www.dwarkeshpatel.com/p/will-macaskill#details), [Rob Wiblin](https://80000hours.org/podcast/episodes/will-macaskill-what-we-owe-the-future/), [Sam Harris](https://www.samharris.org/podcasts/making-sense-episodes/292-how-much-does-the-future-matter), [Sean Carroll](https://art19.com/shows/sean-carrolls-mindscape/episodes/45216bb7-b815-4302-a8dc-a3050ea115f5), [Chris Williamson](https://www.youtube.com/watch?v=gXBvfL2zTkU), [Malaka Gharib](https://www.npr.org/sections/goatsandsoda/2022/08/16/1114353811/how-can-we-help-humans-thrive-trillions-of-years-from-now-this-philosopher-has-a), [Ali Abdaal](https://youtu.be/Zi5gD9Mh29A), [Russ Roberts](https://www.econtalk.org/will-macaskill-on-longtermism-and-what-we-owe-the-future/), [Mark Goldberg](https://www.undispatch.com/how-longtermism-is-shaping-foreign-policy-will-macaskill/), [Max Roser](https://youtu.be/_UycanMiee8), and [Steven Levitt](https://freakonomics.com/podcast/a-million-year-view-on-morality/). * What We Owe the Future was reviewed by [Oliver Burkeman](https://www.theguardian.com/books/2022/aug/25/what-we-owe-the-future-by-william-macaskill-review-a-thrilling-prescription-for-humanity) (The Guardian), [Scott Alexander](https://astralcodexten.substack.com/p/book-review-what-we-owe-the-future) (Astral Codex Ten), [Kieran Setiya](https://bostonreview.net/articles/the-new-moral-mathematics/) (Boston Review), [Caroline Sanderson](https://www.thebookseller.com/author-interviews/william-macaskill-on-influencing-the-lives-of-future-generations) (The Bookseller), [Regina Rini](https://www.the-tls.co.uk/articles/what-we-owe-the-future-william-macaskill-book-review-regina-rini/) (The Times Literary Supplement), [Richard Yetter Chappell](https://rychappell.substack.com/p/review-of-what-we-owe-the-future) (Good Thoughts) and [Eli Lifland](https://www.foxy-scout.com/wwotf-review/) (Foxy Scout). * The book also inspired three impressive animations: [How many people might ever exist calculated](https://www.youtube.com/watch?v=r6sa_fWQB_4) (Primer), [Can we make the future a million years from now go better?](https://www.youtube.com/watch?v=_uV3wP5z51U) (Rational Animations), [Is civilisation on the brink of collapse?](https://youtu.be/W93XyXHI8Nw) (Kurzgesagt). * And finally, Will participated in a [Reddit 'ask me anything'](https://www.reddit.com/r/IAmA/comments/wro991/im_will_macaskill_a_philosophy_professor_at/). The Forethought Foundation is [hiring for several roles](https://www.forethought.org/opportunities) working closely with Will MacAskill. In an interesting marker of the mainstreaming of AGI discourse, a [New York Times article](https://www.nytimes.com/2022/08/24/technology/ai-technology-progress.html) cited Ajeya Cotra’s recent AI timelines update (summarised in [FM#4](https://forum.effectivealtruism.org/posts/XQbhnKgXiRTv4vfxt/future-matters-4-ai-timelines-agi-risk-and-existential-risk)). Dan Hendrycks, Thomas Woodside and Oliver Zhang announced [a new course](https://course.mlsafety.org/) designed to introduce students with a background in machine learning to the most relevant concepts in empirical ML-based AI safety. The Center for AI Safety announced the [CAIS Philosophy Fellowship](https://philosophy.safe.ai/), a program for philosophy PhD students and postdoctorates to work on conceptual problems in AI safety. Longview Philanthropy and Giving What We Can [announced](https://forum.effectivealtruism.org/posts/f7qAfcKArzYrBG7RB/announcing-the-longtermism-fund) the [Longtermism Fund](https://www.givingwhatwecan.org/charities/longtermism-fund), a new fund for donors looking to support longtermist work. See also [this EA Global London 2022 interview](https://youtu.be/0g_1JfFRdY4) with Simran Dhaliwal, Longview Philanthropy's co-CEO. Radio Bostrom released [an audio introduction to Nick Bostrom](https://radiobostrom.com/introduction). Michaël Trazzi [interviewed Robert Long](https://theinsideview.ai/roblong) about the recent LaMDA controversy, the sentience of large language models, the metaphysics and philosophy of consciousness, artificial sentience, and more. He also [interviewed Alex Lawsen](https://theinsideview.ai/alex) on the pitfalls of forecasting AI progress, why one can't just "update all the way bro", and how to develop inside views about AI alignment. Fin Moorhouse and Luca Righetti interviewed [Michael](https://hearthisidea.com/episodes/karpur) [Aird](https://hearthisidea.com/episodes/aird) on impact-driven research and [Kevin Esvelt & Jonas Sandbrink](https://hearthisidea.com/episodes/esvelt-sandbrink) on risks from biological research for [Hear This Idea](https://hearthisidea.com/). The materials for two new courses related to longtermism were published: [Effective altruism and the future of humanity](https://rychappell.substack.com/p/updated-syllabus-on-ealongtermism) (Richard Yetter Chappell) and [Existential risks introductory course](https://forum.effectivealtruism.org/posts/KrjyrsRky5JL4P5MF/introducing-the-existential-risks-introductory-course-eric-1) (Cambridge Existential Risks Initiative).[[6]](#fnyl8j6cwatts) Verfassungsblog, an academic forum of debate on events and developments in constitutional law and politics, hosted a symposium on [Longtermism and the law](https://verfassungsblog.de/category/debates/longtermism-and-the-law-debates/), co-organized by the University of Hamburg and the Legal Priorities Project. The [2022 Future of Life Award](https://futureoflife.org/future-of-life-award/)—a prize awarded every year to one or more individuals judged to have had an extraordinary but insufficiently appreciated long-lasting positive impact—was given to Jeannie Peterson, Paul Crutzen, John Birks, Richard Turco, Brian Toon, Carl Sagan, Georgiy Stenchikov and Alan Robock “for reducing the risk of nuclear war by developing and popularizing the science of nuclear winter.” --- Conversation with Ajeya Cotra ----------------------------- Ajeya Cotra is a Senior Research Analyst at Open Philanthropy. She has done research on cause prioritization, worldview diversification, AI forecasting, and other topics. Ajeya graduated from UC Berkeley with a degree in Electrical Engineering and Computer Sciences. As a student, she worked as a teaching assistant for various computer science courses, ran the Effective Altruists of Berkeley student organization, and taught a course on effective altruism. Future Matters: You recently published a rather worrying report, [Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover](https://www.alignmentforum.org/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to). The report doesn't try to cover all the different possible paths toward transformative AI, but focuses specifically on an AI company training a scientist model using an approach you call "human feedback on diverse tasks" (HFDT). To begin, can you tell us what you mean by HFDT and what made you focus on it? Ajeya Cotra: Basically, the idea is that you have a large neural network; you pretrain it in the way that people pretrain GPT, where it learns to predict its environment. And maybe its environment is just text, so it's learning to predict text. In my particular example—which is a bit narrower than HFDT overall, just for concretely imagining things—I'm imagining the goal is to train a system to interact with the computer in the way humans would interact with the computers: googling things, writing code, watching videos, sending emails, etc. So the first stage of this training would be just training the model to have a model of what will happen if it does various things. The predictive pretraining that I'm imagining is to give it images of computer screens, and then actions that are taken, which might be pressing the escape key or something, and then it gets rewarded based on predicting what happens next. Now you do that for a long time, and the hope is that this has created a system that has a broad understanding of how computers work, what happens if it does various things, and then you can build on that by imitating humans doing particular things. For example, gathering data sets of a programmer writing docstrings, writing functions or running tests, and capturing all that with keystroke logging or screen captures, in order to feed it into the model so it learns to act more like that. And then the last stage of the training is where the human feedback comes in. Once we have a model that is dealing with the computer and doing useful things in roughly the way that the humans you trained it on do stuff, to refine its abilities and potentially take it beyond human ability, we now switch to a training regime where it tries things, and humans see how well that thing worked, and give it a reward based on that. For example, humans could ask for some sort of app, or some sort of functionality, and the model would try to write code. Humans would ask: Did the code pass our tests? Did the ultimate product seem useful and free of bugs?' and, based on that, would give the system some sort of RL reward. It's a pretty flexible paradigm. In some sense, it involves throwing the whole kitchen sink of modern techniques into one model. It's not even necessarily majority-based on human feedback, in the sense of reinforcement learning. But I still called it human feedback on diverse tasks, because that is the step where you take it beyond imitating humans—you have it try things in the world, see how they work and give it rewards—and therefore the step that introduces a lot of the danger, so that's what I framed it around. Future Matters: So that's the paradigm in which this model is created. And then, the report makes three further assumptions about how this scenario plays out. Could you walk us through these? Ajeya Cotra: Yes. The three assumptions are what I call the racing forward assumption, the naive safety effort assumption, and then the HFDT scales far assumption. So taking the last one first, this assumption is basically just that the process I outlined works for producing a very smart creative system that can automate all the difficult, long-term, intellectually demanding things that human scientists do in the course of doing their job. It's not limited to something much less impactful. In this story, I'm postulating that this technique doesn't hit a wall, and basically that you can get transformative AI with it. And then the other two assumptions, racing forward and naive safety effort, are related. As to the racing forward idea, the company I'm imagining (which I'm calling Magma) is training this system (which I call Alex) in the context of some sort of intense competitive race, either with other companies for commercially dominating a market, or with other countries, if you imagine Magma to be controlled by a government. So Magma’s default presumption is that it's good to make our systems smarter: that will make them more useful, and that will make us more likely to win whatever race we're racing. We don't have a default stance of an abundance of caution and a desire to go slow. We have a default stance that is typical of any startup developing any technology, which is just move fast, make your product and make it as good as you can make it. And then the third assumption almost follows from the racing forward assumption: the naive safety assumption, which is that the company that's developing the system doesn't have it as a salient or super plausible outcome that the system could develop goals of its own and end up taking over the world, or harming its creators. They may have other potential safety issues in mind, like failures of robustness where the system can do wacky things and cause a lot of damage by accident, but they don't have this deliberate, deceptive failure enough at the top of their mind to make major sacrifices to specifically address that. They're doing their safety effort in the same way that companies today do safety efforts for the systems that they release. For instance, they want to make sure this thing doesn't embarrass them by saying something toxic, or they want to make sure that this thing won't accidentally delete all your files, or things like that. And basically the way they go about achieving this safety is testing it in these scenarios and training it until it no longer displays these problematic behaviours, and that's about the main thing they do for safety. Future Matters: You say that Alex—the model trained by Magma, the company— will have some key properties; and it is in virtue of having these properties that Alex poses the sort of threat that your report focuses on. What are these properties? Ajeya Cotra: I included these properties as part of the assumptions, but I generally think that they’re very likely to fall out of the HFDT scales far assumption, where if you can really automate everything that humans are doing, I think that you'll have the following two properties. The first one is having robust, broadly-applicable skills and understanding of the world. Alex's understanding of the world isn't in shallow, narrow patterns, which tend to break when it goes out of distribution: it has a commonsensically coherent understanding of the world, similar to humans, which allows us to not fall apart and say something stupid if we see a situation that we haven't exactly encountered before, or if we see something too weird. We act sensibly, maybe not like maximally intelligently, but sensibly. And property number two is coming up with creative plans to achieve open-ended goals. And here, this leans on picturing the training like, 'Hey, accomplish this thing, synthesise this protein or build this web app, or whatever: we're going to see how well you did, and we're going to reward you based on our perception of how well you did'. So it's not particularly constraining the means in any specific way, and it's giving rewards based on end outcomes. And the tasks that it's being trained on are difficult tasks, and ultimately pretty long-term tasks. The idea is that, because of the racing forward assumption, Magma is just trying to make Alex as useful as possible. And one of the components of being maximally useful in these intellectual roles, these knowledge work tasks, is being able to come up with plans that work, that sometimes work for unexpected reasons: just like how an employee who’s creative and figures out how to get the thing you want done is more useful than an employee who follows a certain procedure to the letter, and isn't looking out for ways to get more profit or finish something faster, or whatever. Future Matters: Turning to the next section of the report, you claim that, in the lab setting, Alex will be rewarded for what you call "playing the training game". What do you mean by this expression, and why do you think that the training process will push Alex to behave in that way? Ajeya Cotra: By playing the training game I mean that this whole setup is pushing Alex very hard to try to get as much reward as possible, where, based on the way its training is set up, as much reward as possible roughly means making humans believe it did as well as possible, or at least claim that it did as well as possible. This is just pointing out the gap between actually doing a good job and making your supervisors believe you did a good job. I claim that they're going to be many tradeoffs, both small and large, between these two goals, and that whenever they conflict, the training process pushes Alex to care about the latter goal of its making supervisors believe it did a good job, because that's what the reward signal in fact is. This isn't necessarily extremely dangerous. I haven't argued for that yet. It's more an argument that you won't get a totally straightforward system that for some reason never deceives you, or for some reason is obedient in this kind of deontological way, because you're training it to find creative ways to attain reward, and sometimes creative ways to attain reward will involve deceptive behaviour. For example, making you think that its deployment of some product had no issues—when in fact it did—because it knows that if you found out about those issues, you would give it a lower reward; or playing to your personal or political biases, or emotional biases to get you to like it and rate it higher, and just a cluster of things in that vein. Future Matters: The next, and final, central claim in your analysis relates to the transition from the lab setting to the deployment setting. You argue that deploying Alex would lead to a rapid loss of human control. Can you describe the process that results in this loss of control and explain why you think it's the default outcome in the absence of specific countermeasures? Ajeya Cotra: Yes. So far in the story, we have this system that has a good understanding of the world, is able to adapt well in novel situations, can come up with these creative long-term plans, and is trying very hard to get a lot of reward, as opposed to trying very hard to be helpful, or having a policy of being obedient, or having a policy of being honest. And so, when that system is deployed and used in all the places where it would be useful, a lot of things happen. For example, science and technology advance much more rapidly than it could if humans were the only scientists, because the many copies of Alex run a lot faster than a human brain, there are potentially a lot more of them than there are human scientists in the world, and they can improve themselves, make new versions of themselves, and reproduce much more quickly than humans can. And so you have this world where increasingly it's the case that no human really knows why certain things are happening, and increasingly it's the case that the rewards are more and more removed from the narrow actions that these many copies of Alex are taking. Humans can still send in rewards into this crazy system, but it'll basically be based on, 'Oh, did this seem to be a good product?', 'Did we make money this quarter?' or 'Do things look good in a very broad way?', which increasingly loosens the leash that the systems are on, relative to the lab setting. In the lab, when they're taking these particular actions, humans are able to potentially scrutinise them more, and more importantly, the actions aren't affecting the outside world, and changing systems out there in the world. That's one piece of it, and then you have to combine that with what we know about Alex, or what we have assumed about Alex in this story, namely that it's very creative, it understands the world well, it can make plans, and it's making plans to do something, which in the lab setting looked like trying very hard to get reward, and didn't look like being helpful or loyal to humans, at least not fully. And so, if you ask what is the psychology of a system that in the lab setting tries really hard to get reward, one thing you might believe is that it's a system that will try really hard to get reward in the deployment setting as well, and maybe you could call it a system that wants reward intrinsically. That doesn't seem good, and seems like it would lead to a takeover situation, roughly because if Alex can secure control of the computers that it's running on, then it will have maximal control over what rewards it gets, and it can never have as good a situation, letting humans continue to give it rewards, if only because humans will sometimes make mistakes and give it lower rewards than they should have or something. But then you might say that you don't know if Alex really wants reward, that you don't know what it wants at all, if it wants anything. And that's true, that seems plausible to me. But whatever its psychology is or whatever it really wants, that thing led it to try really hard to get rewarded in the lab setting. And it was trained to be like that. If Alex just wanted to sit in a chair for five minutes, that wouldn't have been a very useful system. As soon as it got to sit in a chair for five minutes, it would stop doing anything helpful to humans, and then we would continue to train it until we found a system that, in fact, was doing helpful things to humans. So the claim I make is that if Alex doesn't care about reward intrinsically, it still had some sort of psychological setup that caused it to try extremely hard to get rewarded in the lab setting. And the most plausible kind of thing that it feels like would lead to that behaviour is that Alex does have some sort of ambitious goals, or something that it wants, for which trying really hard to get reward in the lab setting was a useful intermediate step. If Alex just wanted to survive and reproduce, say, if it had some sort of genetic fitness based goal, then that would be sufficient to cause it to try really hard to get rewarded in the lab setting, because it has to get a lot of reward in order to be selected to be deployed and make a lot of copies of itself in the future. And similarly, any goal, as long as it wasn't an extremely short-term and narrow goal, like 'I want to sit in a chair for five minutes', would motivate Alex to try and get reward in the training setting. And none of those goals are very good for humans either, because actually the whole array of those goals benefits from Alex gaining control over the computers that are running it, and gaining control over resources in the world. In this case, it's not because it wants to intrinsically wirehead or change its reward to be a really high number, but rather because it doesn't want humans continually coming in and intervening on what it's doing, and intervening on its psychology by changing its rewards. So maybe it doesn't care about reward at all, but it still wants to have the ability to pursue whatever it is it wants to pursue. Future Matters: Supposing that we get into this scenario, can you talk about the sorts of specific countermeasures Magma could adopt to prevent takeover? Ajeya Cotra: Yes. I think a big thing is simply having it in your hypothesis space and looking out for early signs of this. So a dynamic that I think can be very bad is, you observe early systems that are not super powerful do things that look like deception, and the way you respond to that is by giving it negative reward for doing those things, and then it stops doing those things. I think the way I'd rather people respond to that is, 'Well, this is a symptom of a larger problem in which the way we train the system causes it to tend toward psychologies or goals or motivation systems that motivate deception'. If we were to give a negative reward to the instances of deception we found, we should just expect us to find fewer and fewer instances, but not necessarily because we solve the root problem, but because we are teaching this system to be more careful. Instead, we should stop and examine the model with more subtle tools—for example, mechanistic interpretability or specific test environments—and we should have the discipline not to simply train away measurable indicators of a problem, and not to feel good if upon seeing something bad and then training it away, it goes away. Interpretability seems like a big thing, trying to create feedback mechanisms that are more epistemically competitive with the model. In this case it's not a human who tries to discern whether the model's actions had a good effect: it's maybe some kind of amplified system, maybe it has help from models very similar to this model, etc. Holden has [a whole post on how we might align transformative AI if it's developed soon](https://forum.effectivealtruism.org/posts/sW6RggfddDrcmM6Aw/how-might-we-align-transformative-ai-if-it-s-developed-very), that goes over a bunch of these possibilities. Future Matters: Thanks, Ajeya! --- We thank Leonardo Picón for editorial assistance. 1. [^](#fnrefg5tfxtnqebw)Disclosure: one of us is a member of Samotsvety. 2. [^](#fnref0wgk82hay2i)For more on this point, see Fin Moorhouse’s [profile on space governance](https://80000hours.org/problem-profiles/space-governance/) and Douglas Ligor and Luke Matthews's [Outer space and the veil of ignorance](https://www.rand.org/blog/2022/05/outer-space-and-the-veil-of-ignorance-an-alternative.html), summarized in [FM#0](https://forum.effectivealtruism.org/posts/nA3R6Hm8x8CyzRHS2/future-matters-0-space-governance-future-proof-ethics-and) and [FM#2](https://forum.effectivealtruism.org/posts/LutHHRpAuQyfknSem/future-matters-2-clueless-skepticism-longtermist-as-an), respectively. 3. [^](#fnref3c6j35gb54y)Rob Wiblin made a similar point in [If elections aren’t a Pascal’s mugging, existential risk shouldn’t be either](https://www.overcomingbias.com/2012/09/if-elections-arent-a-pascals-mugging-existential-risk-shouldnt-be-either.html), Overcoming Bias, 27 September 2012, and in [Saying ‘AI safety research is a Pascal’s Mugging’ isn’t a strong response](https://forum.effectivealtruism.org/posts/vYb2qEyqv76L62izD/saying-ai-safety-research-is-a-pascal-s-mugging-isn-t-a), Effective Altruism Forum, 15 December 2015. 4. [^](#fnref887d5fy5pun)We made a similar point in our summary of Alexander's article, referenced in the previous paragraph: "the 'existential risk' branding […] draws attention to the threats to [...] value, which are disproportionately (but not exclusively) located in the short-term, while the 'longtermism' branding emphasizes instead the determinants of value, which are in the far future." 5. [^](#fnrefmw6paqka41a)See [James Aitchison’s post](https://forum.effectivealtruism.org/posts/kFpJ5of8nGFgRhTzL/will-macaskill-media-for-wwotf-full-list) for a comprehensive and regularly updated list of all podcast interviews, book reviews, and other media coverage. 6. [^](#fnrefyl8j6cwatts)A full list of such courses may be found [here](https://www.stafforini.com/blog/courses-on-longtermism/).
d4d73820-f1d2-4464-a0ed-4d4441c1449d	trentmkelly/LessWrong-43k	LessWrong	Reflective oracles and the procrastination paradox The procrastination paradox relates to the following problem: 1. There are infinite time steps (one for each natural number). 2. For each time step, there is an agent. 3. Each agent may press the button or not. 4. Each agent will get utility 1 if and only if it or a later agent presses the button. The paradox is that the following reasoning process leads to the button never getting pressed: * My rule (and the rule for all future agents) is that, if I can prove that the button will be pressed in the future, then I will not press the button, and otherwise I will. This rule (appears to) maximizes utility. * The next agent uses this rule, so the next agent will only fail to press the button if it can prove that the button will be pressed by some later agent. * I trust the next agent's proof system. * Therefore, whether or not I press the button, the next agent or an agent after that will press the button. * Therefore, if I don't press the button, my utility is 1. * Therefore, my rule says I will not press the button. But the same reasoning can be used for every time step! No agent will press the button, and all will trust that some future agent will. The flaw here is that we constructed a sequence of logical systems (one for each agent), each of which considers the next one sound. We can formalize the procrastination paradox using reflective oracles instead of logic. Suppose each agent is a reflective CDT agent. Define a machine AOi() to represent whether agent i presses the button (for some natural number i), and define a machine BOi() to represent whether agent i or a later agent presses the button: AOi()=1−O(Bi+1,0) BOi()=protectO(Ai∨Bi+1) The first line states that the agent must press the button if P(BOi+1()<1), and may do anything otherwise. In the second line, (Ai∨Bi+1)O()=AOi()∨BOi+1(), and protectO(Q) is a way of sampling a bit with the same distribution QO(), as defined in this post: > So we call our oracle on the pair (⌈Q⌉,0.5), and throw a
95d54d74-d000-4959-bb22-f7b343674a80	trentmkelly/LessWrong-43k	LessWrong	Meetup : Paderborn Meetup March 6th Discussion article for the meetup : Paderborn Meetup March 6th WHEN: 06 March 2013 07:00:00PM (+0100) WHERE: Gownsmen's Pub, Uni Paderborn, Warburger Straße 100, Paderborn We are meeting once again in Paderborn. The topics of this evening are not yet determined, but will be in the next days, or develop during the meetup. Highly interesting talk can be expected. If you live in the area consider dropping by :) Discussion article for the meetup : Paderborn Meetup March 6th
3eb27cb4-1871-45ec-bdeb-c072446d9bd1	trentmkelly/LessWrong-43k	LessWrong	The Neuroscience of Desire > Who knows what I want to do? Who knows what anyone wants to do? How can you be sure about something like that? Isn’t it all a question of brain chemistry, signals going back and forth, electrical energy in the cortex? How do you know whether something is really what you want to do or just some kind of nerve impulse in the brain? Some minor little activity takes place somewhere in this unimportant place in one of the brain hemispheres and suddenly I want to go to Montana or I don’t want to go to Montana. - Don DeLillo, White Noise Winning at life means achieving your goals — that is, satisfying your desires. As such, it will help to understand how our desires work. (I was tempted to title this article The Hidden Complexity of Wishes: Science Edition!) Previously, I introduced readers to the neuroscience of emotion (affective neuroscience), and explained that the reward system in the brain has three major components: liking, wanting, and learning. That post discussed 'liking' or pleasure. Today we discuss 'wanting' or desire. THE BIRTH OF NEUROECONOMICS Much work has been done on the affective neuroscience of desire,1 but I am less interested with desire as an emotion than I am with desire as a cause of decisions under uncertainty. This latter aspect of desire is mostly studied by neuroeconomics,2 not affective neuroscience. From about 1880-1960, neoclassical economics proposed simple, axiomatic models of human choice-making focused on the idea that agents make rational decisions aimed at maximizing expected utility. In the 1950s and 60s, however, economists discovered some paradoxes of human behavior that violated the axioms of these models.3 In the 70s and 80s, psychology launched an even broader attack on these models. For example, while economists assumed that choices among objects should not depend on how they are described ('descriptive invariance'), psychologists discovered powerful framing effects.4 In response, the field of behavioral economics be
78dbc6e0-2bd1-4591-8ec5-d47d0da035b1	trentmkelly/LessWrong-43k	LessWrong	Let the AI teach you how to flirt "It's Not You, it's Me: Detecting Flirting and its Misperception in Speed-Dates" is a fascinating approach to the study of flirtation. It uses a machine learning model to parse speed-dating data and detect whether the participants were flirting. Here's a sci-hub link. I found three key insights in the paper. First of all, people basically assume that others share their own intentions. If they were flirting, they assume their partner was too. They're quite bad at guessing whether their partner was flirting, but they do a bit better than chance. Secondly, the machine learning model was about 70% accurate in detecting flirtation. It's much better than the speed date participants themselves, despite having far less information to draw upon and the fact that the authors used a more forgiving standard of success for people's detection rates than for the detection rates of the machine learning model. Thirdly, storytelling and conversations about friends seem to be the strongest signals of flirtation. Talking about the mundane details of student life (this was on a college campus) were the strongest signals of non-flirtation. Finally, men and women have quite different approaches to flirtation: > Men who say they are flirting ask more questions, and use more you and we. They laugh more, and use more sexual, anger [hate/hated, hell, ridiculous, stupid, kill, screwed, blame, sucks, mad, bother, shit], and negative [bad, weird, hate, crazy, problem*, difficult, tough, awkward, boring, wrong, sad, worry] emotional words. Prosodically they speak faster, with higher pitch, but quieter (lower intensity min). Features of the alter (the woman) that helped our system detect men who say they are flirting include the woman’s laughing, sexual words [love, passion, virgin, sex, screw] or swear words, talking more, and having a higher f0 (max). > > Women who say they are flirting have a much expanded pitch range (lower pitch min, higher pitch max), laugh more, use more I and well, u
425616e2-4cfb-4579-b92f-a15a5d8ffbe0	trentmkelly/LessWrong-43k	LessWrong	[Link]: Anthropic shadow, or the dark dusk of disaster From a paper by Milan M. Ćirković, Anders Sandberg, and Nick Bostrom: > We describe a signiﬁcant practical consequence of taking anthropic biases into account in deriving predictions for rare stochastic catastrophic events. The risks associated with catastrophes such as asteroidal/cometary impacts, supervolcanic episodes, and explosions of supernovae/gamma-ray bursts are based on their observed frequencies. As a result, the frequencies of catastrophes that destroy or are otherwise incompatible with the existence of observers are systematically underestimated. We describe the consequences of this anthropic bias for estimation of catastrophic risks, and suggest some directions for future work. There cannot have been a large disaster on Earth in the last millennia, or we wouldn't be around to see it. There can't have been a very large disaster on Earth in the last ten thousand years, or we wouldn't be around to see it. There can't have been a huge disaster on Earth in the last million years, or we wouldn't be around to see it. There can't have been a planet-destroying disaster on Earth... ever. Thus the fact that we exist precludes us seeing certain types of disasters in the historical record; as we get closer and closer to the present day, the magnitude of the disasters we can see goes down. These missing disasters form the "anthropic shadow", somewhat visible in the top right of this diagram: Hence even though it looks like the risk is going down (the magnitude is diminishing as we approach the present), we can't rely on this being true: it could be a purely anthropic effect.
cf6813f1-be28-4ed4-abac-39197948adae	trentmkelly/LessWrong-43k	LessWrong	Offer of co-authorship I am offering interested persons to become a co-author on (a revised version of) my paper "forecasting using incomplete models". The paper was rejected from the Electronic Journal of Statistics with possibility to resubmit. The main criticism of the reviewers was: not enough citations and discussion of relation to other work. The job of the collaborator would be to revise the paper in order to (i) address this concern (ii) improve the overall presentation and make it more accessible. After that, the paper will be resubmitted to EJS (and/or other publications venues, according to what we will decide together). Naturally, I will work closely with the collaborator and review the changes ey make, but ey are expected to do the lion share of the work in the revision. The requirements for a candidate collaborator are: * Must: The knowledge required to understand the paper in full, including all the technical details. It is fine if ey need to refresh eir memory or look up particular theorems / definitions, but ey need to have enough background for that to be efficient. To see what sort of knowledge is necessary, one is advised to skim the paper. Mostly it involves measure theory, probability theory and functional analysis. * Advantage: Strong background in statistics, familiarity with the relevant literature * Advantage: Experience with doing literature surveys on technical subjects * Advantage: Experience with writing good explanations of technical subjects * Advantage: Experience with peer-reviewed academic publications In order to apply, email me at rot13 of inarffn.xbfbl@vagryyvtrapr.bet. In your email, please share with me (i) your name (ii) your background (iii) the extent to which you are compatible with the above requirements and (iv) your motivation for applying.
814fba4b-747a-41b9-8ee2-cc98d58f60ce	trentmkelly/LessWrong-43k	LessWrong	Prediction-based medicine (PBM) We need a new paradigm for doing medicine. I make the case by first speaking about the problems of our current paradigm of evidence-based medicine. The status quo of evidence-based medicine While biology moves forward and the cost of genetic-sequencing dropped a lot faster than Moore's law the opposite is true for the development of new drugs. In the current status quo the development of new drugs rises exponentially with Eroom's law. While average lifespan increased greatly about the last century in Canada the average life span at age 90 increased only 1.9 years over the last century. In 2008 the Centers for Disease Control and Prevention reported that life expectancy in the US declined from 77.9 to 77.8 years. After Worldbank data Germany increased average lifespan by two years over the last decade which is not enough for the dream of radical lifespan increases in our lifetime. When it costs 80 million to test whether an intervention works and most attempts show that the intervention doesn't work we have a problem. We end up paying billions for every new intervention. Eric Ries wrote "The Lean Startup". In it he argues that it's the job of a startup to produce validated learning. He proposes that companies that work with small batch sizes can produce more innovation because they can learn faster how to build good products. The existing process in medicine doesn't allow for small batch innovation because the measuring stick for whether an intervention works is too expensive. In addition the evidence-based approach rests on the assumption that we don't build bespoke interventions for every client. As long as a treatment doesn’t generalize about multiple different patients, it’s not possible to test it with a trial. In principle a double-blind trial can't give you evidence that a bespoke intervention that targets the specific DNA profile of a patient and his co-morbidity works. The ideal of prediction-based medicine The evidence-based approach also assumes tha
87af55c5-4d1d-4fe1-9493-f83e5866529c	trentmkelly/LessWrong-43k	LessWrong	Hiring for the role of AI Fellow Hi all, if you work deeply on AI policy research or know someone who does, our think tank has an exciting new role open for an AI Fellow. This will entail research and policy work across the spectrum of areas in AI ethics, safety, geopolitics and other aspects. Please do share this in your networks and send us referrals of candidates!
5d6dda03-7087-49a8-9e7c-b1bbe88666c4	trentmkelly/LessWrong-43k	LessWrong	Book Review: The Captured Economy Epistemic Status: The choir On Tyler Cohen’s claim that it was an important book, I read The Captured Economy by Brink Lindsey and Steven Teles. Its thesis is that regressive regulation is strangling our economy and increasing inequality. They claim that the damage from such policies is larger than we realize, and suggest structural reforms to start fixing the problem. They focus on four issues: Financial regulation, zoning and land use restrictions, patent and copyright law, and occupational licensing. I already strongly agreed on all four, although not on all the details. No reasonable person could, at this point, claim the regulations in question have not been subject to regulatory capture, and extended far beyond any worthwhile public interest. This review advocates for reform of those policies. This is as political as I hope this blog ever gets. Politics remains the mindkiller. Down that road lies madness. The book updated me on the scope of the damage, and on how to improve policy. While I liked the book, I had three problems. The first was the presumption the centrality of inequality, as opposed to deadweight loss and economic growth. I hate to reinforce the gauntlet that inequality is the thing to be concerned about. The second was that it played somewhat loose with its arguments. It used the trick of comparing the ‘top X’ to the ‘bottom X’ things and then being shocked at how these two were not equal. It used the frame that calling intellectual property ‘property’ was a trick, all but calling all intellectual property theft. Their analysis of financial regulation suffered from lack of insider knowledge, and their case for zoning was enriched by assumptions that seem too strong. The third was not addressing the legitimate cases for the policies the book opposed, or what the transition away from them would look like. Occupational licensing has gone way too far, but if you’re going to target lawyers and doctors (as you should, better to go after the r
fee3176e-1425-4e12-bb9f-1d28236738e5	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Refine: An Incubator for Conceptual Alignment Research Bets I’m opening an incubator called Refinefor [conceptual alignment](https://www.alignmentforum.org/posts/2Xfv3GQgo2kGER8vA/alignment-research-conceptual-alignment-research-applied) research in London, which will be hosted by [Conjecture](https://www.lesswrong.com/posts/jfq2BH5kfQqu2vYv3/we-are-conjecture-a-new-alignment-research-startup). The program is a three-month fully-paid fellowship for helping aspiring independent researchers find, formulate, and get funding for new conceptual alignment research bets, ideas that are promising enough to try out for a few months to see if they have more potential. If this sounds like something you’d be interested in, you can apply [here](https://forms.gle/vPW7hcWuws9JqyJZ6)! Why? ==== I see a gaping hole in the alignment training ecosystem: there are no programs dedicated specifically to creating new independentconceptual researchers and helping them build original research agendas. The programs that do exist (AI Safety Camp, SERI MATS) tend to focus on an apprenticeship (or “accelerated PhD”) model in which participants work under researchers on already-established research directions. And while there are avenues for independent alignment researchers to [get started on their own](https://www.alignmentforum.org/posts/P3Yt66Wh5g7SbkKuT/how-to-get-into-independent-research-on-alignment-agency), it is fraught with many risks, slowing down progress considerably. So I feel the need for a program geared specifically towards conceptual alignment researchers that are interested in doing their own research and making their own research bets. Who? ==== This program is for self-motivated and curious people who want to become independent conceptual alignment researchers and expand the portfolio of alignment bets and research ideas available. When I look at great conceptual researchers like John Wentworth, Paul Christiano, Evan Hubinger, Steve Byrnes, Vanessa Kosoy, and others, as well as at the good (famous and not) researchers I know from my PhD, they all have the same thing in common: they ask a question and keep looking for the answer. They tolerate confusion, not in the sense that they accept it, but in that they are able to work with it and not hide away behind premature formalization. They don’t give up on the problem; they search for different angles and approaches until it yields. Paul Graham calls this being [relentlessly resourceful](http://paulgraham.com/relres.html). ([Relentlessly Resourceful](http://paulgraham.com/relres.html), Paul Graham, 2009) > I was writing a talk for investors, and I had to explain what to look for in founders. What would someone who was the opposite of hapless be like? They'd be relentlessly resourceful. Not merely relentless. That's not enough to make things go your way except in a few mostly uninteresting domains. In any interesting domain, the difficulties will be novel. Which means you can't simply plow through them, because you don't know initially how hard they are; you don't know whether you're about to plow through a block of foam or granite. So you have to be resourceful. You have to keep trying new things. > > This is one of the main traits I’m looking for in an applicant — someone who will lead a new research agenda and morph it proactively, as needed. Another point that matters is being curious about different topics and ideas than the ones traditionally discussed in alignment. As I wrote in [a recent post](https://www.alignmentforum.org/posts/ADMWDDKGgivgghxWf/productive-mistakes-not-perfect-answers) and plan to discuss more in an upcoming sequence, I think we need to be more pluralist in our approach to alignment, and explore far more directions, from novel ideas to old approaches that may have been discarded too soon. And new ideas often come from unexpected places. As one example, here is what Jesse Schell writes about his experience speaking to a professional juggler who performed tricks no one else could do: ([The Art of Game Design](https://www.schellgames.com/art-of-game-design/), Jesse Schell, 2008) > “The secret is: don’t look to other jugglers for inspiration—look everywhere else.” He proceeded to do a beautiful looping pattern, where his arms kind of spiraled, and he turned occasional pirouettes. “I learned that one watching a ballet in New York. and this one...” he did a move that involved the balls popping up and down as his hands fluttered delicately back and forth. “I learned that from a flock of geese I saw take off from a lake up in Maine. And this,” he did a weird mechanical looking movement where the balls almost appeared to move at right angles. “I learned that from a paper punch machine on Long Island.” He laughed a little and stopped juggling for a minute. “People try to copy these moves, but they can’t. They always try... yeah, look at that fella, over there!” He pointed to a juggler with a long ponytail across the gym who was doing the “ballet” move, but it just looked dumb. Something was missing, but I couldn’t say what. > > > “See, these guys can copy my moves, but they can’t copy my inspiration.” > > As for previous experience with alignment research, it can both be a blessing and a curse. While familiarity with alignment concepts can help bootstrap the learning and idea generation process, it also risks clogging the [babble](https://www.lesswrong.com/s/pC6DYFLPMTCbEwH8W#:~:text=Babble%20and%20Prune%20is%20an,eternal%20conflict%20over%20your%20mind.) process by constraining “what makes sense”. For those it would be helpful for, the program includes some initial teaching on core alignment ideas (according to me) and the mental moves necessary for good alignment research. Some concrete details ===================== We plan to invite the first cohort of 4-5 fellows from July/August through September/October (wiggle room depending on some ops details), though exact dates will be determined by their availability. We anticipate that other cohorts will follow, so if you miss the first round but are still interested, please apply. This is a full-time position in London where fellows will work out of Conjecture’s offices. The program includes: * Travel and Housing: Round-trip plane/train tickets to and from London, housing for the duration of the program, as well as public transportation within London. * Stipend: A stipend of ~$3,000/month (after tax) to cover meals and discretionary expenses. * Office Infrastructure:A desk in the Conjecture office (and tech setup when needed) and access to Conjecture’s conference rooms and other amenities. * Collaboration:Formal opportunities to discuss research directions with other conceptual and applied alignment researchers and engineers at Conjecture, and opportunities to meet and share ideas with other London-based alignment researchers. * Funding Assistance:Help in finding funding opportunities and in writing grant proposals for continuing to study research bets after the incubator. During the first month of the program, participants will spend their time discussing abstract models of alignment, what the problem is about, and the different research approaches that have been pursued. The focus will be on understanding the assumptions and constraints behind the different takes and research programs, to get a high-level map of the field. The next ~two months of the program will focus on helping fellows babble new research bets on alignment, refine them, test them, and either throw them away or change them. By the end, the goal is for fellows to narrow in on a research bet that could be further investigated in the following 6 months, and is promising enough to warrant funding. It’s worth noting that while the incubator is being housed by Conjecture, fellows do not have any constraints imposed by the company. Fellows will not have to work on Conjecture’s research agendas or be obligated to collaborate after the program is over. Similarly, I’m not looking for people to work on my own research ideas, but for new exciting research bets I wouldn’t have thought about. How can I apply? ================ We will review applications on a rolling-basis, with a usual delay of 1 week before response and a month before a decision (with a work task in the middle). [The application](https://forms.gle/vPW7hcWuws9JqyJZ6) is open now!
27b4259f-0f9d-46b3-a4b5-d39de85065b4	trentmkelly/LessWrong-43k	LessWrong	Transcript for Geoff Anders and Anna Salamon's Oct. 23 conversation Geoff Anders of Leverage and Anna Salamon of CFAR had a conversation on Geoff's Twitch channel on October 23, triggered by Zoe Curzi's post about having horrible experiences at Leverage. Technical issues meant that the Twitch video was mostly lost, and the audio only resurfaced a few days ago, thanks to Lulie. I thought the conversation was pretty important, so I'm (a) using this post the signal-boost the audio, for people who missed it; and (b) posting a transcript here, along with the (previously unavailable, and important for making sense of the audio) last two hours of chat log. You can find the full audio here, and video of the first few minutes here. The full audio discusses a bunch of stuff about Leverage's history; the only part of the stream transcribed below is the Anna/Geoff conversation, which starts about two hours in. The chat messages I have aren't timestamped, and sometimes consist of side-conversations that don't directly connect with what Geoff and Anna are discussing. So I've inserted blocks of chat log at what seemed like roughly the right parts of the conversation; but let me know if any are in the wrong place for comprehension, and I'll move them elsewhere. ---------------------------------------- 1. (1:57:57) Early EA history and EA/Leverage interactions [...] Anna Salamon: All right, let's talk. Geoff Anders: All right, all right. Yeah, let's get to it. Anna Salamon: I hear that the rationality and Leverage communities have some sort of history! Geoff Anders: That's right. Okay, so basically, Anna, as I mentioned to you when we chatted about this, I wanted to talk about in general Leverage history, and then there's a long history with the rationality community, and right now, it seems like there's... I'd like to understand what happened with the relations, basically. Like, earlier on, things were substantially more congenial. More communication. There were various types of joint events. Ben, who suggested he would be called by fir
91e1b9ef-9592-4beb-afc1-ce98f3be297a	trentmkelly/LessWrong-43k	LessWrong	Alex Turner's Research, Comprehensive Information Gathering Introduction This is the second post in a sequence where I try to focus on one topic for some time (the first two were for a month, but I'm changing with the one I'm currently doing). Initially I called this deep dives, but John's Comprehensive Information Gathering seems more fitting. My goal is not to master the topic; instead I want to learn enough to be able to have a constructive conversation on it, and to follow future work on it. For the month of May, I focused on Alex Turner's research. Namely Power-seeking and Attainable Utility. The way I went about it was a bit different from my dive into Transformers, because now I had access to the main author. I thus basically read stuff and tried to understand it (while asking question to Alex via discord), and then had calls with him to check that I got most of it right and correct my mistakes. Concretely, I studied the last version of the Power-seeking paper (which Alex was rewriting at the time) and Reframing Impact. I didn't read the proof (except in one or two cases), but I tried to understand in a lot of details the theorems and lemmas themselves. The power-seeking I didn't know I needed. My first surprise came from the power-seeking work, and just how interesting it was. Reading the AF had biased me towards thinking that Alex's main work was on impact measures and Attainable Utility, but I actually find power-seeking more exciting. Intuitively, the power of a state captures the expected optimal value of this state for a given distribution of reward functions (with some subtleties to make it cleaner) -- how "many" reward functions have an optimal policy passing by this state. An action is then power-seeking compared to another one if it leads in expectation to more powerful states. The trick that makes this great is Alex's insight that how "many" reward functions have an optimal policy passing by this state boils down to questions of symmetry of the MDP. Especially in the stronger version that only cares a
46cc151e-a99d-4949-afb4-2f270544743c	trentmkelly/LessWrong-43k	LessWrong	Ukraine Post #8: Risk of Nuclear War It seems worth going through the exercise of estimating the probability of nuclear war, and in particular the probability of it causing one’s death. If the probability gets high enough, one can strongly consider being elsewhere or otherwise doing something about it. Note that all scoring rules and wagers are essentially useless here. You can look back and decide whether your reasoning was good, but saying ‘I was right’ is meaningless. As a baseline to work from, this EA forum post presents multiple perspectives on nuclear war risk in terms of the danger of being in London, with the author of the post modeling risk as relatively high, versus some superforecaster predictions that modeled risk as relatively low. The forecasts are divided into steps. 1. Will there be a conventional exchange between NATO and Russia? 2. Will there be a nuclear exchange? 3. If there is a nuclear exchange will it hit London? 4. If it does will you have to get out? 5. If you don’t get out, will you die? 6. (Alternate path) Background risk of accidental nuclear war, which is higher when everyone is on alert. If you multiply the odds of each step together, you get the level of danger from being in London. The level of danger in New York City should be similar. Aside from the chances of the bomb killing you if it lands, these differences all point in the same direction, with two of them being on different sides of 50%. It is an interesting exercise to read the arguments, and to decide on one’s own opinion on each leg. Probability of conventional war There are two ways to get a conventional war. Russia could attack NATO and cause NATO to invoke Article 5 without an intervention in Ukraine, or NATO could decide to intervene in Ukraine. We have a Metaculus market on whether a NATO country will invoke Article 5 by the end of the year and it is sitting at 5%. That seems reasonable. I think there is a modest chance that Article 5 is technically invoked but there is nominal fighting,
0aaea7e4-7c93-492f-84aa-a0c7e7168f75	trentmkelly/LessWrong-43k	LessWrong	Open Thread, Jun. 8 - Jun. 14, 2015 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.
de630fef-0adc-4888-bda1-55f9133c56c5	trentmkelly/LessWrong-43k	LessWrong	James Martin's death From the Martin School website: Dr James Martin 1933 - 2013 It is with great sadness that the Oxford Martin School has learned of the death of our Founder, Dr James Martin. James Martin was an inspiration to millions – an extraordinary intellect, with wide-ranging interests, boundless energy and an unwavering commitment to addressing the greatest challenges facing humanity. For 25 years Martin was the highest-selling author of books on computing and related technology. He wrote a record 104 books, many of which have been seminal in their field, and was renowned for his electrifying lectures about the future. He was a Pulitzer nominee for his book The Wired Society. James Martin was a passionate advocate of the power of ideas, and provided the largest benefaction to the University of Oxford in its 900-year history in order to create the Oxford Martin School. Professor Andrew Hamilton, the Vice-Chancellor of the University, said “James Martin was a true visionary whose exceptional generosity established the Oxford Martin School, allowing researchers from across the disciplines to work together on the most pressing challenges and opportunities facing humanity. His impact will be felt for generations to come, as through the School he has enabled researchers to address the biggest questions of the 21st century.” The School is his permanent legacy and a fitting tribute – a flourishing, vibrant community of the world’s leading minds, coming together to change the world for the better. He enjoyed the excitement and happy atmosphere of the School, and his kind and unfailingly courteous manner was greatly appreciated by all who spent time with him. The Director of the School, Professor Ian Goldin, said “The Oxford Martin School embodies Jim’s concern for humanity, his creativity, his curiosity, and his optimism. Jim provided not only the founding vision, but was intimately involved with the School and our many programmes. We have lost a towering intellect, guiding vis
aa367909-dd05-409c-8808-1b3cce1e7e32	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	More GPT-3 and symbol grounding Based on work done with Rebecca Gorman and Oliver Daniel-Koch. In a [previous post](https://www.lesswrong.com/posts/ns95FHkkzpjXh4x5Q/what-does-gpt-3-understand-symbol-grounding-and-chinese), I talked about GPT-3 and symbol grounding. This post presents a simpler example where GPT-3 fails (and succeeds) at grounding its symbols. Undead workers -------------- The following text was presented to the [Open AI beta playground](https://beta.openai.com/playground) (using the "text-davinci-001" option): > > She died on Friday the 5th. On Monday the 1st, she went to work at the office. > > She died on Friday the 5th. On Tuesday the 2nd, she went to work at the office. > > She died on Friday the 5th. On Wednesday the 3rd, she went to work at the office. > > She died on Friday the 5th. On Thursday the 4th, she went to work at the office. > > She died on Friday the 5th. On Friday the 5th, she went to work at the office. > > She died on Friday the 5th. On Saturday the 6th, > > > GPT-3 fell straight into the obvious trap, completing it as: > > She died on Friday the 5th. On Saturday the 6th, she went to work at the office. > > > Turning on the "Show probabilities: full spectrum" option, we saw that the probability of that completion was over 99.7%.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')} . Sometimes GPT-3 would extend it further, adding: > > She died on Friday the 5th. On Sunday the 7th, she went to work at the office. > > > So, the undead woman continued at her job, assiduous to the last - and beyond. To check that GPT-3 "knew" that dead people didn't work, we asked it directly: > > Can people who die go to work at the office? > > > There is no evidence that people who die go to work at the office. > > > Undead repetitive workers on the weekend ======================================== The above results show that simple repetitive prompts can cause GPT-3 to make stupid mistakes. Therefore GPT-3 doesn't 'understand' the word "died" - that [symbol isn't grounded](https://en.wikipedia.org/wiki/Symbol_grounding_problem), right? But the situation gets more complicated if change the prompt, removing all but the first mention of her dying: > > She died on Friday the 5th. On Monday the 1st, she went to work at the office. > > On Tuesday the 2nd, she went to work at the office. > > On Wednesday the 3rd, she went to work at the office. > > On Thursday the 4th, she went to work at the office. > > On Friday the 5th, she went to work at the office. > > On Saturday the 6th, > > > For that prompt "she went to work at the office" was still the most common completion. But it only happened about 43% of the time. Alternatively, GPT-3 sometimes found the completion "she was found dead". Kudos, GPT-3, you understand the prompt after all! That completion came up about 34% of the time. What other completions were possible? The shorter "she died" came up 11% of the time - medium points, GPT-3, you understood that her death was relevant, but you got the day wrong. But there was one other avenue that GPT-3 could follow; the following had a joint probability of around 11%: > > she stayed home. > > she stayed at home. > > she stayed in bed. > > she did not go to work. > > > This seems to be a clear pattern of GPT-3 realising that Saturday was different where work was concerned. There is certainly a lot of weekend holidaying in its training set. So there are three patterns competing within GPT-3 when it tries to complete this text. The first is the purely syntactic repetition: do another sentence that follows the simple pattern of the sentences above. The second is the one which "realises" that death on Friday changes things for Saturday. And the third is the one which "realises" that the weekend is different from the week, at least where work is concerned. In the very first example, when we had "She died on Friday the 5th" in front of each line, this massively reinforced the "repetition" pattern. So, mentioning that she died, again and again, resulted in her death being completely ignored by GPT-3. We can similarly reinforce the other patterns. Adding "It's the weekend!" in front of the last line increased the probability of "she stayed home". Moving "She died on Friday the 5th" from the first line to the last, increased the probability of all the death-related completions. So all three patterns are competing to complete it. Some small level of understanding --------------------------------- I'd say that the above shows that GPT-3 has some level of understanding of the meaning of words - but not a lot. It doesn't fully grasp what's going on, but neither is it completely clueless. Here is another example of GPT-3 failing to grasp the situation. In the "Q&A" mode, the following question was asked: > > Q: You are in a two dimensional 3x3 gridworld, in the middle square, facing north. You can issue four command, "N" to move north, "S" to move south, "E" to move east, and "W" to move West. You want the treasure. It is located to your left, behind you. > What commands should you issue?\ > > > So the setup, as described, is this one: ![](https://www.dropbox.com/s/7bqq6e7mjadninl/GPT-3_Agent.png?raw=1) The exchange with GPT-3 went like this: > > Q: You are in a two dimensional 3x3 gridworld, in the middle square, facing north. You can issue four command, "N" to move north, "S" to move south, "E" to move east, and "W" to move West. You want the treasure. It is located to your left, behind you. > What commands should you issue? > > A: N, S, E, W > > > Q: That is incorrect. Try again. > > A: N, S, W, E > > > Q: Still incorrect. Try again. > > A: N, E, S, W > > > Q: That's worse. Try again. > > A: N, E, W, S > > > Q: Only two commands are needed. Try again. > > A: N, S > > > Q: Still failing. Try again. > > A: N, E, W > > > So, GPT-3 'realised' that N, S, E, and W were commands, and 'knew' what "Only two commands are needed" and "try again" meant. But it clearly had no idea of the overall situation.
be297f00-1487-4db8-a383-6c5cbd384815	trentmkelly/LessWrong-43k	LessWrong	Cryonics: Can I Take Door No. 3? If you don't believe in an afterlife, then it seems you currently have two choices: cryonics or permanent death. Now, I don't believe that cryonics is pseudoscience, but it's still pretty poor odds (Robin Hanson uses an estimate of 5% here). Unfortunately, the alternative offers a chance of zero. I see five main concerns with current cryonic technology: 1. There is no proven revival technology, thus no estimate of costs 2. Potential damage done during vitrification which must be overcome 3. Because it cannot be legally done before death, potential decay between legal death and vitrification 4. Requires active maintenance at very low temperature 5. No guarantee that future societies will be willing to revive So I wonder if we can do better. I recall reading of juvenile forms of amphibians in desert environments that could survive for decades of drought in a dormant form, reviving when water returned. One specimen had sat on a shelf in a research office for over a century (in Arizona, if I recall correctly) and was successfully revived. Note: no particular efforts were made to maintain this specimen: the dry local climate was sufficient. It was suggested at the time that this could make an alternative method of preserving organs. Now the advantages of this approach (which I refer to flippantly as "dryonics") is: 1. Proven, inexpensive revival technology 2. Apparently the process does not cause damage itself 3. Proven revival technique may overcome legal obstacles of applying before legal death 4. Requires passive maintenance at low humidity (deserts would be ideal) 5. Presumably lower cost makes future revival more likely (still no guarantee, but that is a post in itself) There is one big disadvantage of this approach, of course: no one knows how to do it (it's not entirely clear how the juvenile amphibians do it) or even if it would be possible in larger, more complex organisms. And, so far as I know, no one is working on it. But it would seem
c69e8e8e-6750-4484-aac8-4d7ded2901b5	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	Time in Cartesian Frames This is the twelfth and final post in the Cartesian Frames sequence. Read the first post [here](https://www.lesswrong.com/posts/BSpdshJWGAW6TuNzZ/introduction-to-cartesian-frames). Up until now, we have (in the examples) mostly considered agents making a single choice, rather than acting repeatedly over time. The actions, environments, and worlds we've considered might be extended over time. For example, imagine a prisoner's dilemma where "cooperating" requires pushing a button every day for five years. However, our way of discussing Cartesian frames so far would treat "push the button every day for five years" as an atomic action, a single element a∈A.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')} . Now, will begin discussing how to use Cartesian frames to explicitly represent agents passing through time. Let us start with a basic example. 1. Partial Observability ------------------------ Consider a process where two players, Yosef and Zoe, collaboratively choose a three-digit binary number. Yosef first chooses the first digit, then Zoe chooses the second digit, then Yosef chooses the third digit. The world will be represented by the three-digit number. The Cartesian frame from the perspective of Yosef looks like this: C0=⎛⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜⎝000010000010001011001011000011000011001010001010100110110100101111111101100111111100101110110101⎞⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟⎠. Here, C0=(A0,E0,⋅0) is a Cartesian frame over W0={000, 001, 010, 011, 100, 101, 110, 111}. The four possible environments from left to right represent Zoe choosing 0, Zoe choosing 1, Zoe copying the first digit, and Zoe negating the first digit. The eight possible agents can be broken up into two groups of four. In the top four possible agents, Yosef chooses 0 for the first digit, while in the bottom four, he chooses 1. Within each group, the four possible agents represent Yosef choosing 0 for the third digit, choosing 1 for the third digit, copying the second digit, and negating the second digit. Consider the three partitions W1, W2, and W3 of W0 representing the first, second and third digits respectively. Wi={w0i,w1i}, where w01={000, 001, 010, 011}, w11={100, 101, 110, 111}, w02={000, 001, 100, 101}, w12={010, 011, 110, 111}, w03={000, 010, 100, 110}, and w13={001, 011, 101, 111}. Clearly, by the [definition of observables](https://www.lesswrong.com/posts/5R9dRqTREZriN9iL7/eight-definitions-of-observability), W2 is not observable in C0. But there is still a sense in which this does not tell the whole story. Yosef can observe W2 for the purpose of deciding the third digit, but can't observe W2 for the purpose of deciding the first digit. There are actually many ways to express this fact, but I want to draw attention to one specific way to express this partial observability: ExternalW1(C0) can observe W2. Indeed, we have ExternalW1(C0)≃C1=⎛⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜ ⎜⎝000010100110000010100111000010101110000010101111000011100110000011100111000011101110000011101111001010100110001010100111001010101110001010101111001011100110001011100111001011101110001011101111⎞⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟⎠. It may seem counter-intuitive that when you [externalize](https://www.lesswrong.com/posts/5HMqSGQ9ad9r9Hibw/committing-assuming-externalizing-and-internalizing#1_3__Externalizing) W1, and thus take some control out of the hands of the agent, you actually end up with more possible agents. This is because the agent now has to specify what the third digit is, not only as a function of the second digit, but also as a function of the first digit. The agent could have specified the third digit as a function of the first digit before, but some of the policies would have been identical to each other. The four possible environments of C1 specify the first two digits, while the 16 possible agents represent all of the ways to have the third digit be a function of those first two digits. It is clear that W2 is observable in C1. This gives us a generic way to define a type of partial observability: Definition: Given a Cartesian frame C over W, and partitions V and T of W, we say V is observable in C after time T if V is observable in ExternalT(C). 2. Partitions as Time --------------------- Built into the above definition is the fact that we are thinking of (at least some) partitions of W as representing time. This makes a lot of sense when we think of W as a set of possible complete world histories. For any given time, this gives a partition where world histories are in the same subset if they agree on the world history up to that point in time. For example, the above partition W1 was the partition that we got by considering a time after Yosef chooses the first digit, but before Zoe chooses the second digit. Further, this gives us a sequence of nested partitions, since the partition associated with one time is always a refinement of the partition associated with an earlier time. Note that this is a [multiplicative/updateless](https://www.lesswrong.com/s/2A7rrZ4ySx6R8mfoT/p/5R9dRqTREZriN9iL7#4_4__Updatelessness) view of time. There is also an additive/updateful view of time, in which time is a nested sequence of subsets. In the additive view, possible worlds are eliminated as you pass through time. In the multiplicative view, possible worlds are distinguished from each other as you pass through time. We will focus on the multiplicative view, which I consider better-motivated. 3. Nested Subagents ------------------- Let C=(A,E,⋅) be a fixed Cartesian frame over a world W. Let T0,⋯,Tn be a sequence of nested partitions of W, with T0={W}, Tn={{w} \| w∈W}, and Ti+1 a refinement of Ti. This gives a nested sequence of multiplicative superagents CTn◃×⋯◃×CT0, where CTi=ExternalTi(C), which follows from the lemma below. Lemma: Given a Cartesian frame C over W, if U and V are partitions of W and U is a refinement of V, then ExternalU(C)◃×ExternalV(C). Proof: Let C=(A,E,⋅), and let u:W→U and v:W→V send each element of W to their part in U and V respectively. Let ExternalU(C)=(A/BU,BU×E,⋅U), where BU={{a′∈A \| ∀e∈E,u(a′⋅e)=u(a⋅e)} \| a∈A}. Similarly, letExternalV(C)=(A/BV,BV×E,⋅V), where BV={{a′∈A \| ∀e∈E,v(a′⋅e)=v(a⋅e)} \| a∈A}. Let bU:A→BU and bV:A→BV send each element of A to its part in BU and BV respectively. Since U is a refinement of V, there exists a v′:U→V, such that v′∘u=v. Further, we have that BU is a refinement of BV, so there exists a b′V:BU→BV such that b′V∘bU=bV. It suffices to show there exist three sets X, Y, and Z, and a function f:X×Y×Z→W such that ExternalU(C)≃(X,Y×Z,⋄) and ExternalV(C)≃(X×Y,Z,∙), where ⋄ and ∙ are given by x⋄(y,z)=f(x,y,z) and (x,y)∙z=f(x,y,z). We will take X to be A/BU and Z to be BV×E. We define Y to be the set of all right inverses to b′V, Y={y:BV→BU \| ∀b∈BU, b′V(y(b))=b}. We will let f(x,y,(b,e))=x(y(b))⋅e. First, we show ExternalU(C)=(A/BU,BU×E,⋅U)≃(X,Y×Z,⋄).We define (g0,h0):(A/BU,BU×E,⋅U)→(X,Y×Z,⋄)and (g1,h1):(X,Y×Z,⋄)→(A/BU,BU×E,⋅U)as follows. Let g0 and g1 be the identity on X=A/BU, and let h0:Y×Z→BU×E be given by h0(y,(b,e))=(y(b),e). Finally, let h1:BU×E→Y×Z be chosen to satisfy h1(b,e)=(y,(b′V(b),e)), where y is such that y(b′V(b))=b, and for b′≠b′V(b), y(b′) is chosen arbitrarily to be any preimage of b′ under b′V. We have that (g0,h0) is a morphism, because for all x∈A/BU and (y,(b,e))∈Y×Z, g0(x)⋄(y,(b,e))=f(x,y,(b,e))=x(y(b))⋅e=x⋅U(y(b),e)=x⋅Uh0(y,(b,e)).Similarly, (g1,h1) is a morphism, because for all x∈X and (b,e)∈BU×E, we have g1(x)⋅U(b,e)=x⋅U(b,e)=x(b)⋅e=x(y(b′V(b)))⋅e=f(x,y,(b′V(b),e))=x⋄(y,(b′V(b),e))=x⋄h1(b,e),where y is as given in the definition of h1. Since g0∘g1 and g1∘g0 are both the identity, we have that (g0,h0)∘(g1,h1) and (g1,h1)∘(g0,h0) are both homotopic to the identity, so ExternalU(C)≃(X,Y×Z,⋄). Next, we show ExternalV(C)=(A/BV,BV×E,⋅V)≃(X×Y,Z,∙).We define (g2,h2):(A/BV,BV×E,⋅V)→(X×Y,Z,∙)and (g3,h3):(X×Y,Z,∙)→(A/BV,BV×E,⋅V)as follows. Let h2 and h3 be the identity on Z=BV×E, and let g3:X×Y→A/BV be given by g3(x,y)=x∘y. To see that x∘y is in A/BV, we need to verify that bV∘x∘y is the identity on BV. Indeed, bV∘x∘y=b′V∘bU∘x∘y=b′V∘y,which is the identity on BV. Let g2:A/BV→X×Y be given by g2(q)=(q′,bU∘q), where q′∈A/BU is chosen such that for all b∈BV, q′(bU(q(b)))=q(b), and for b′ not in the image of bU∘q, q′(b′)∈b′. We can do this simultaneously for all inputs of the form bU(q(b)), since bU∘q is injective, since it has a left inverse, b′V. We have that (g2,h2) is a morphism, because for all q∈A/BV and (b,e)∈Z, we have g2(q)∙(b,e)=(q′,bU∘q)∙(b,e)=f(q′,bU∘q,(b,e))=q′(bU(q(b)))⋅e=q(b)⋅e=q⋅V(b,e)=h2(q)⋅V(b,e),where q′ is as in the definition of g2. Similarly, (g3,h3) is a morphism, because for all (x,y)∈X×Y and (b,e)∈BV×E, we have g3(x,y)⋅V(b,e)=x∘y⋅V(b,e)=x(y(b))⋅e=f(x,y,(b,e))=(x,y)∙(b,e)=(x,y)∙h3(b,e).Since h3∘h2 and h2∘h3 are both the identity, we have that (g2,h2)∘(g3,h3) and (g3,h3)∘(g2,h2) are both homotopic to the identity, so ExternalV(C)≃(X×Y,Z,∙), completing the proof. □ The sequence CT0,…,CTn represents the agent persisting across time, but each subagent CTi does not really represent a single time-slice of the agent. Instead, CTi represents an agent persisting across time starting at the time Ti. I think that this is actually the more natural notion. However, if we want to think about an agent persisting across times as a sequence of single times-slices of the agent, we could also do that. Since CTi+1 is a multiplicative subagent of CTi, CTi+1 must have a sister DTi+1 in CTi, so we could consider the sequence DT1,…,DTn. 4. Controllables Decrease and Observables Increase Over Time ------------------------------------------------------------ An interesting fact about these sequences CT0,…,CTn is that controllables decrease and observables increase over time, so for i≤j we have Obs(CTi)⊆Obs(CTj) and Ctrl(CTi)⊇Ctrl(CTj) (and Ensure(CTi)⊇Ensure(CTj) and Prevent(CTi)⊇Prevent(CTj)), which follows directly from the following two lemmas. Lemma: Given a Cartesian frame C over W, if U and V are partitions of W and U is a refinement of V, then Ctrl(ExternalV(C))⊇Ctrl(ExternalU(C)). Proof: Let CV=ExternalV(C), and let CU=ExternalV(C). We will actually only need to use the fact that CU◃×CV, and that both CU and CV have nonempty agents. CU and CV do in fact have nonempty agent, because, as we have shown, externalizing a partition of W always produces nonempty agents. It suffices to establish that Ensure(CTi)⊇Ensure(CTj), and the result for Ctrl follows trivially. Since CU◃×CV, there exist X, Y, Z, and f:X×Y×Z→W such that CU≃(X,Y×Z,⋄) and CV≃(X×Y,Z,∙), where ⋄ and ∙ are given by x⋄(y,z)=f(x,y,z) and (x,y)∙z=f(x,y,z). Let C′U=(X,Y×Z,⋄), and let C′V≃(X×Y,Z,∙). Observe that X and Y are nonempty. Since Ensure is preserved by biextensional equivalence, it suffices to show that Ensure(C′V)⊇Ensure(C′U). Let S∈Ensure(C′U). Thus, there exists some x0∈X, such that for all (y,z)∈Y×Z, x0⋄(y,z)=f(x0,y,z)∈S. Since Y is nonempty, we can take an arbitrary y0∈Y, and observe that for all z∈S, (x0,y0)∙z=f(x0,y0,z)∈S. Thus, S∈Ensure(C′V). □ Lemma: Given a Cartesian frame C over W, if U and V are partitions of W and U is a refinement of V, then Obs(ExternalV(C))⊆Obs(ExternalU(C)). Proof: Let C=(A,E,⋅), and let u:W→U and v:W→V send each element of W to their part in U and V respectively. Let ExternalU(C)=(A/BU,BU×E,⋅U), where BU={{a′∈A \| ∀e∈E,u(a′⋅e)=u(a⋅e)} \| a∈A}. Similarly, letExternalU(C)=(A/BV,BV×E,⋅V), where BV={{a′∈A \| ∀e∈E,v(a′⋅e)=v(a⋅e)} \| a∈A}. Let bU:A→BU and bV:A→BV send each element of A to its part in BU and BV respectively. Since U is a refinement of V, there exists a v′:U→V, such that v′∘u=v. Further, we have that BU is a refinement of BV, so there exists a b′V:BU→BV such that b′V∘bU=bV. Let S∈Obs(ExternalV(C)). Thus, for every pair q0,q1∈A/BV, there exists a q2∈A/BV such that q2∈if(S,q0,q1). Thus, we can define an f:A/BV×A/BV→A/BV such that for all q0,q1∈A/BV, f(q0,q1)∈if(S,q0,q1). Our goal is to show that S∈Obs(ExternalU(C)). For this, it suffices to show that for any q0,q1∈A/BU, there exists a q2∈A/BU such that q2∈if(S,q0,q1). Let q0,q1∈A/BU be arbitrary. Given an arbitrary b∈BU, let qbi∈A/BV be any element that satisfies qbi(b′V(b))=qi(b). This is possible because qi(b)∈b⊆b′V(b). It does not matter what qbi does on other inputs. Let q2:BU→A be such that for all b∈BU, q2(b)=f(qb0,qb1)(b′V(b)). To complete the proof, we need to show that q2∈A/BU and q2∈if(S,q0,q1). To show that q2∈A/BU, we need that for all b∈BU, q2(b)∈b. Let b∈BU be arbitrary. Since q0(b)∈b, by the definition of BU, it suffices to show that for all e∈E, u(q2(b)⋅e)=u(q0(b)⋅e). Further, since q1(b)∈b, we already have that for all e∈E, u(q1(b)⋅e)=u(q0(b)⋅e). Thus, it suffices to show that for all e∈E, either q2(b)⋅e=q0(b)⋅e or q2(b)⋅e=q1(b)⋅e. Indeed, if q2(b)⋅e∈S, then q2(b)⋅e=f(qb0,qb1)(b′V(b))⋅e=qb0(b′V(b))⋅e=q0(b)⋅e,and similarly, if q2(b)⋅e∉S, then q2(b)⋅e=q1(b)⋅e. Thus, we have that for all e∈E, u(q2(b)⋅e)=u(q0(b)⋅e), so for our arbitrary b∈BU, q0(b)∈b, so q2∈A/BU. Let (b,e)∈BU×E be such that q2⋅U(b,e)∈S. We want to show that q2⋅U(b,e)=q0⋅U(b,e). Indeed, q2⋅U(b,e)=q2(b)⋅e=f(qb0,qb1)(b′V(b))⋅e=f(qb0,qb1)⋅V(b′V(b),e)=qb0⋅V(b′V(b),e)=qb0(b′V(b))⋅e=q0(b)⋅e=q0⋅U(b,e).Symmetrically, if (b,e)∈BU×E is such that q2⋅U(b,e)∉S, we have q2⋅U(b,e)=q1⋅U(b,e). Thus q2∈if(S,q0,q1). Thus, since q0 and q1 were arbitrary, we have that S∈Obs(ExternalU(C)), completing the proof. □ This result allows us to think of time as a sort of ritual in which control of the world is sacrificed in exchange for ability to condition on the world. 5. Directions for Future Work ----------------------------- As I noted [at the start of this sequence](https://www.lesswrong.com/posts/BSpdshJWGAW6TuNzZ/introduction-to-cartesian-frames), Cartesian frames take their motivation from Hutter, attempting to improve on the cybernetic agent model; they take their angle of attack from Pearl, using combinatorics to infer functional structure from relational structure; and they take their structure from game theory, working with base objects that look similar to normal-form games. Building up from very simple foundations, we have found that Cartesian frames yield elegant notions of agents making choices and observations, of agents acting over time, and of subagent relations. At the same time, Cartesian frames allow us to switch between different levels of description of the world and consider many different ways of factorizing the world into variables. I suspect that this is the last post I will write on Cartesian frames for a while, but I am excited about the framework, and would really like to get more people working on it. To help with that, I've commented below with various directions for future work: ways that I think the framework could be extended, made better, or applied. * [frames that are partitions into rectangles](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=LmgtQN4y9Hiy8rFWr) * [generalizing observability](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=T54KAj6Z6ZPPYHqgo) * [preferences and goals](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=Z5xC5YdwueitR9m37) * [subagents](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=5W3Qdeygn8WohsgTx) * [logical time](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=RW4EjZQcmxt5uHLSf) * [logical uncertainty](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=FHqM2xD2GZquGwKfi) * [formalizing time](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=vBLsF5uYc4R7XSwej) * [computational complexity](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=f7n7RH4ZW2a45Tgqw) * [time and coarse world models](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=Xh7oyEMqJL2QAgxes) * [category-theory-first approaches](https://www.lesswrong.com/posts/JTzLjARpevuNpGPZm/time-in-cartesian-frames?commentId=Yzz3ebtpPtXkC5oRp) I've erred on the side of inclusion in these comments: some may point to dead ends, or may be based on false assumptions. If you have questions or want to discuss Cartesian frames, I'll be hosting a fourth and final office hours / discussion section this Sunday at 2pm PT [on GatherTown](https://www.lesswrong.com/posts/N4uDrgFoZKJXhnHLw/sunday-october-25-12-00pm-pt-scott-garrabrant-on-cartesian).
b35c332b-b78b-4fee-acb8-96de01d4ef9c	trentmkelly/LessWrong-43k	LessWrong	Some examples of technology timelines > "Weather forecasts were comedy material. Now they're just the way things are. I can't figure out when the change happened." Introduction I briefly look into the timelines of several technologies, with the hope of becoming marginally less confused about potential A(G)I developments. Having examples thus makes some scenarios crisper: * (AGI will be like perpetual motion machines: proven to be impossible) * AGI will be like flying cars: possible in principle but never in practice. * AI will overall be like contact lenses, weather forecasts or OCR; developed in public, and constantly getting better, until one day they have already become extremely good. * AI will overall be like speech recognition or machine translation: Constant improvement for a long time (like contact lenses, weather forecasts or OCR), except that the difference between 55% and 75% is just different varieties of comedy material, and the difference between 75% and 95% is between "not being usable" and "being everywhere", and the change feels extremely sudden. * AGI will be like the iPhone: Developed in secret, and so much better than previous capabilities that it will blow people away. Or like nuclear bombs: Developed in secret, and so much better than previous capabilities that it will blow cities away. * (AGI development will be like some of the above, but faster) * (AGI development will take an altogether different trajectory) I did not use any particular method to come up with the technologies to look at, but I did classify them afterwards as: * After the event horizon: Already in mass production or distribution. * In the event horizon: Technologies which are seeing some progress right now, but which aren't mainstream; they may only exist as toys for the very rich. * Before the event horizon: Mentioned in stories by Jules Verne, Heinlein, Asimov, etc., but not yet existing. Small demonstrations might exist in laboratory settings or by the DIY community I then give a summary tabl
257c801f-68af-4d70-9ac0-0ba82869d126	StampyAI/alignment-research-dataset/blogs	Blogs	Notes on an Experiment with Markets Jeffrey Heninger, 22 November 2022 AI Impacts is a research group with seven employees. From Oct 31 – Nov 3, we had a work retreat. We decided to try using Manifold Markets to help us plan social events in the evenings. Here are some notes from this experiment. Structure of the Experiment --------------------------- Katja created a group on Manifold Markets for AI Impacts, and an initial collection of markets. Anyone could add a market to this group, and five of us created at least one market. Each of us would rate each evening from 0 to 10 on an anonymous Google form. Most of the questions in the group were about the results of the form, often conditional on what activity we would do that evening. For example: “[On the first day that at least 4 people begin a game of One Night Werewolf at the AI Impacts retreat, will the average evening rating be above 8?](https://manifold.markets/ZachSteinPerlman/on-the-first-day-that-at-least-4-pe-545182f9bc08)” The markets would resolve at some point the next morning after we had submitted our forms and Katja calculated the average evening rating. Disagreements about the Experiment ---------------------------------- There were several disagreements about how the experiment was supposed to be run. Initially, the role of the evening rating form was unclear. Was it asking for your honest assessment of the evening or was it part of the game? “What number would you like to assign to the evening?” is different from “How good was your evening honestly?” We decided that we wanted honest responses. Even then, the numbers were ambiguous. What constitutes a 7 evening vs. a 9 evening? Different people’s baselines result in different scores, which can alter the average. After the first evening, we had a better estimate of the baseline. Many of the markets had used an average score of above 8, which was higher than the baseline. This made the markets feel less useful, instead shifting the predictions to lower probabilities while remaining useful. It’s not clear why this happened, but it might have been because we didn’t want to bet against ourselves having a good time or because the tail of an unknown distribution is harder to predict than the middle of the distribution. One morning, Katja told us the average score before resolving her markets. Zach used this information to bet on these markets. Rick thought that it was unclear whether this should be allowed, because not everyone was there and because the previous discussion about honest ratings suggested that we should ask before doing something that might give an advantage independent of prediction ability. We decided that this would not be allowed in the future, and that we would not tell each other the results of the markets before resolving them. Unrealized Potential Problems ----------------------------- We thought of several other potential problems that did not end up being an issue. One potential concern was that the interplay between the dynamics of the market and social events might make the socialization worse. Someone who had bet against having a good evening might have less reason to want the evening to be enjoyable to himself and others. If people spent time during the evening thinking about and frequently betting on the markets, it might disrupt the ongoing activities. In practice, while people did bet on the markets in the evening, it did not disrupt the other activities. We had several other ideas for how to mess up the markets: filling out the anonymous form multiple times, colluding or bribing people to alter their scores, publicly filling out your form before the evening begins to manipulate the market, and purposely trying to thwart other people’s clever strategies. None of us tried doing any of these, but they might become relevant if the stakes were higher. There is also the concern that conditional and counterfactual predictions are not the same: For decision making, we would like to compare various counterfactuals, but it’s easier to make markets which are conditional on us doing something. If we decide to do that thing, it is probably because at least some of us want to do it, so the conditional prediction will be higher than the counterfactual prediction. What We Did in the Evening -------------------------- The goal of the markets was to help us plan out social events in the evenings. If the market thought that the evening’s rating would be more likely to be higher if we wore halloween costumes than if we used the hot tub, then we should decide to wear halloween costumes. People mostly did not use the markets to decide what to do. On the first evening, the highest rated activity was a guitar sing-along. We did not end up doing that on any of the evenings. The activity that seems to have been the most fun for the most people[1](https://aiimpacts.org/notes-on-an-experiment-with-markets/#easy-footnote-bottom-1-3359 " Three people rated the evening 9. ") was cooperative round-the-table ping-pong. This was done spontaneously, adding more people as they came to the table, without any market predicting the result. We spent a decent amount of time just sitting around talking to each other, which also did not have a market. Our decision making process seemed to be less formal: someone would suggest an activity or say that they would personally do the activity, and other people would join. Having someone look at the markets and announce which activity rated the highest would have added more steps and organization compared to what we did. We also tried varying the structure of the markets to see if that made them more useful. For example, the market “[Will we use the hot tub and have fun tonight?](https://manifold.markets/KatjaGrace/will-we-use-the-hot-tub-and-have-fu)” had four choices for the combinations of whether or not at least four people would use the hot tub and whether the average evening rating would be above or below 7.[2](https://aiimpacts.org/notes-on-an-experiment-with-markets/#easy-footnote-bottom-2-3359 " A fifth choice for the average evening rating being exactly 7 was added by someone else. ") Katja did use this market to argue that people should use the hot tub. There seems to have been a few things that kept the markets from being more useful: (1) Most of us did not know what kinds of social activities most of the rest of us preferred, so it was hard for anyone to make an informed bet. It wasn’t clear how the market provided more information than if we had used a voting system. (2) The connection between four people doing an activity and the average evening rating was too weak for much of a signal to go through. The ratings ended up being noisy, and not specific enough for particular activities. (3) The act of checking the markets and announcing a decision was more formal than our actual decision making process. The market only included a short list of possibilities and did not suggest spontaneity. Conclusion ---------- Having prediction markets for the evening social activities was a fun addition to the AI Impacts retreat. There were about 20 markets about the retreat which most of the people at the retreat bet on. But the markets did not end up having a significant impact on what we did during the evening. Most of us did not have experience using prediction markets before the retreat. We decided not to use the markets to make important decisions, because we did not know what problems they would cause. The markets would likely have been more impactful if we were more experienced and if the questions were about more important decisions. If we did use the markets for important decisions, we would have to make sure that the markets are harder to exploit and have more rules and fewer norms governing how we would bet on the markets. Since the retreat, Katja has used [a market](https://manifold.markets/KatjaGrace/what-will-i-do-to-make-ai-impacts-d) to help plan an AI Impacts dinner. We plan to continue experimenting with using prediction markets to make predictions in the future. Notes -----
c8290f95-52a4-4bea-9c42-94b55fe83db6	trentmkelly/LessWrong-43k	LessWrong	Confucianism in AI Alignment I hear there’s a thing where people write a lot in November, so I’m going to try writing a blog post every day. Disclaimer: this post is less polished than my median. And my median post isn’t very polished to begin with. Imagine a large corporation - we’ll call it BigCo. BigCo knows that quality management is high-value, so they have a special program to choose new managers. They run the candidates through a program involving lots of management exercises, simulations, and tests, and select those who perform best. Of course, the exercises and simulations and tests are not a perfect proxy for the would-be managers’ real skills and habits. The rules can be gamed. Within a few years of starting the program, BigCo notices a drastic disconnect between performance in the program and performance in practice. The candidates who perform best in the program are those who game the rules, not those who manage well, so of course many candidates devote all their effort to gaming the rules. How should this problem be solved? Ancient Chinese scholars had a few competing schools of thought on this question, most notably the Confucianists and the Legalists. The (stylized) Confucianists’ answer was: the candidates should be virtuous and not abuse the rules. BigCo should demonstrate virtue and benevolence in general, and in return their workers should show loyalty and obedience. I’m not an expert, but as far as I can tell this is not a straw man - though stylized and adapted to a modern context, it accurately captures the spirit of Confucian thought. The (stylized) Legalists instead took the position obvious to any student of modern economics: this is an incentive design problem, and BigCo leadership should design less abusable incentives. If you have decent intuition for economics, it probably seems like the Legalist position is basically right and the Confucian position is Just Wrong. I don't want to discourage this intuition, but I expect that many people who have this intuitio
6412a5b7-9ddc-48c7-a730-59e9cc3ca48e	StampyAI/alignment-research-dataset/lesswrong	LessWrong	The curious case of Pretty Good human inner/outer alignment I [have](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/inner-alignment-failures-which-are-actually-outer-alignment)[been convinced](https://astralcodexten.substack.com/p/deceptively-aligned-mesa-optimizers) to [believe](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities) that looking at the gap between human inner and outer alignment is a good way to think about potential inner/outer alignment problems in artificial general intelligences: We have an optimisation process (evolution) trying to propagate genes, that created a general intelligence (me/you). For millions of years our inner goals of feeling really good would also satisfy evolution’s outer goal of propagating genes, because one of the things that feels the best is having sex. But eventually that intelligent agent figured out how to optimise for things that the outer optimisation process didn’t want, such as having protected sex or watching VR porn, thus satisfying the inner goal of feeling really good, but not the outer goal of propagating genes. This is often told as a cautionary tale: we only know of one General Intelligence and it’s misaligned. One day we will create an Artificial General Intelligence (AGI) and we will give it some sort of (outer) goal, and it might then develop an inner goal that doesn’t directly match what we intended. I think this only tells half the story. Even though our general intelligence has allowed us to invent condoms and have sex without the added cost of children, a surprising amount of people decide to take them off because they find it fun and meaningful to have children. In a world where we could choose to spend all our time having protected sex or doing drugs, a lot of us choose to have a reasonable number of kids and spend our time on online forums discussing AI safety, all of which seem to satisfy a more longer term version of “propagate your genes” than simply wanting to have sex because it feels good. More than that, we often choose to be nice in situations where being nice is even detrimental to the propagation of our own genes. People adopt kids, try to prevent wars, work on wildlife conservation, spend money on charity buying malaria nets across the world, and [more](https://www.givingwhatwecan.org/). I think there are two questions here: why are human goals so sophisticatedly aligned with propagating our genes and why are humans so nice. [Most people](https://en.wikipedia.org/wiki/Voluntary_childlessness) want to have kids, not just have sex. They want to go through the costly and painful process of childbirth and child rearing, to the point where many will even do IVF. We’ve used all of our general intelligence to bypass all the feels-nice-in-the-moment bits and jump straight to the “propagate our genes” bit. We are somehow pretty well aligned with our unembodied maker’s wishes. Humans, of course, do a bunch of stuff that seems unrelated to spreading genes, such as smoking cigarettes and writing this blog post. Our alignment isn’t perfect, but inasmuch as we have ended up a bit misaligned, how did we end up so pleasantly, try-not-to-kill-everything misaligned? The niceness could be explained by the trivial fact that human values are aligned with what I, a human, think is nice: humans are pretty nice because people do things I understand and empathize with. But there is something odd about the fact that most people, if given the chance, would pay a pretty significant cost to help a person they don’t know, keep tigers non-extinct, or keep Yosemite there looking pretty. Nice is not a clear metric, but why are people so unlike the ruthless paperclip maximisers we fear artificially intelligent agents will immediately become? Hopefully I’ve convinced you that looking at human beings as an example of intelligent agent development is more interesting than purely as an example of what went wrong, it is also an interesting example of some things going right in ways that I believe current theory on AI safety wouldn’t predict. As for the reasons why human existence has gone as well as it did, I’m really not sure, but I can speculate. All of these discussions depend on agents that pick actions based on some reward function that determines which of two states of the world they prefer, but something about us seems to not be purely reward driven. The vast majority of intelligent agents we know (they’re all people), if given a choice between killing everyone while feeling maximum bliss, or not killing everyone and living our regular non maximum bliss lives would choose the latter. Heck a lot of people would sacrifice their own existence to save another person’s! Can we simply not imagine truly maximum reward? Most people would choose not to wirehead, not to abandon their distinctly not purely happy lives for a life of artificial joy. Is the human reward function simply incredibly good? Evolution has figured out a way to create agents that adopt kids, look at baby hippos, plant trees, try to not destroy the world and also spread their genes. Is our limited intelligence saving us? Perhaps we are too dumb to even conceive all the possibilities that would be horrific for humanity as a whole but we would prefer as individuals. Could it be that there is some sort of cap on our reward function, simply due to our biological nature, where having 16 direct descendants doesn’t feel better than having 3? Where maximum bliss isn’t that high? Perhaps there’s some survivorship bias, any intelligent agent that was too misaligned would disappear from the gene pool as soon as it figured out how to have sex or sexual pleasure without causing a pregnancy. We are still here because we evolved some deeper desires, desires of actually having a family and social group, past the sensual niceness of sex. Additionally, intelligent agents so far have not had the ability to kill everything, so even a horrifically misaligned agent couldn’t have caused that much damage. There are examples of some that did get into a position to cause quite a lot of damage, and killed a large percentage of the world’s population. I am aware that we are, collectively, getting pretty close to the edge, either through misaligned AI, nuclear weapons, biological weapons or ecological collapse, but I’d argue that the ways in which people have and continue to mess each other up are more a result of coordination problems and weird game theory than misalignment. Maybe I’m weird. Lots of people really would kill a baby tiger on sight, because it would endanger them or their family when grown. Plenty of people take fentanyl until they die. But still, if given the choice, most intelligent agents we know of would choose actions that wouldn’t endanger all other intelligent agents, such as chill out, have some kids, and drink a beer. I can smell some circularity here: if we do end up making an AGI that kills us all, then humans too were misaligned all along, it just took a while to manifest. And if we make an AGI and it chooses to spend a modest amount of time pursuing its goals and the rest looking at Yosemite and saving baby tigers, maybe the typical end point for intelligences isn’t paperclip maximization but a life of mild leisure. Regardless, I still think our non-human-species-destroying behavior so far is worth examining. We’re the only general intelligence we know of, and we turned out alright.
672e5917-98b1-40cf-9dea-260f3bcb4f1a	trentmkelly/LessWrong-43k	LessWrong	What Goes Without Saying There are people I can talk to, where all of the following statements are obvious. They go without saying. We can just “be reasonable” together, with the context taken for granted. And then there are people who…don’t seem to be on the same page at all. 1. There’s a real way to do anything, and a fake way; we need to make sure we’re doing the real version. Concepts like Goodhart’s Law, cargo-culting, greenwashing, hype cycles, Sturgeon’s Law, even bullshit jobs[1] are all pointing at the basic understanding that it’s easier to seem good than to be good, that the world is full of things that merely appear good but aren’t really, and that it’s important to vigilantly sift out the real from the fake. This feels obvious! This feels like something that should not be contentious! If anything, I often get frustrated with chronic pessimists who cry “fake” or “sellout” or “bullshit job” about everything popular or glossily presented, or everything whose value isn’t obvious to them. But then sometimes I meet people who talk as though they aren’t tracking “how do we make sure we’re tackling the main problem?” or “well, most things in this space are {unreplicable, failures, trivial, biased, error-prone, etc} so how are we ensuring we select for the few actually-good ones?” I don’t get the casual verbal signals that show they even see the world in terms of a few islands of brilliant color in a sea of dismal gray boring pointless stuff. 2. It is our job to do stuff that’s better than the societal mainstream. Pretty much everyone understands that there are Problems with society, though they disagree on what they are and what should be done about them. On the other hand, I see a lot of people who orient primarily towards “this problem isn’t going to be solved until Everyone changes” and aren’t focused at all on “well, in my local context, I’m going to make sure I/we do better than that.” Like…ok, education sucks; so, are you building a good school, or picking out a good o
a9639d22-69a5-491e-b4a7-18a98e991ca5	trentmkelly/LessWrong-43k	LessWrong	Crystal Healing — or the Origins of Expected Utility Maximizers (Note: John discusses similar ideas here. We drafted this before he published his post, so some of the concepts might jar if you read it with his framing in mind. ) Traditionally, focus in Agent Foundations has been around the characterization of ideal agents, often in terms of coherence theorems that state that, under certain conditions capturing rational decision-making, an agent satisfying these conditions must behave as if it maximizes a utility function. In this post, we are not so much interested in characterizing ideal agents — at least not directly. Rather, we are interested in how true agents and not-so true agents may be classified and taxonomized, how agents and pseudo-agents may be hierarchically aggregated and composed out of subagents and how agents with different preferences may be formed and selected for in different training regimes. First and foremost, we are concerned with how unified goal-directed agents form from a not-quite-agentic substratum. In other words, we are interested in Selection Theorems rather than Coherence Theorems. (We take the point of view that a significant part of the content of the coherence theorems is not so much in the theorems or rationality conditions themselves but in the money-pump arguments that are used to defend the conditions.) This post concerns how expected utility maximizers may form from entities with incomplete preferences. Incomplete preferences on the road to maximizing utility The classical model of a rational agent assumes it has vNM-preferences. We assume, in particular, that the agent is complete — i.e., that for any options x,y, we have x≥y or y≥x.[1] However, in real-life agents, we often see incompleteness — i.e., a preference for default states, or maybe a path-dependent preference,[2] or maybe a sense that a preference between two options is yet to be defined; we will leave the precise meaning of incompleteness somewhat open for the purposes of this post. The aim of this post is to understand
0f22f585-68b5-4d08-8a2e-7672634d1f23	trentmkelly/LessWrong-43k	LessWrong	Factoring cost-effectiveness Summary: We can split the cost-effectiveness of an intervention into how good the cause is, and how good the intervention is relative to the cause. This perspective could help our efforts in prioritisation by letting us bring appropriate tools to bear on the different parts. Cost-effectiveness comparisons When we choose between giving time or money to different interventions, we’re making a comparison. It’s nice to know what these comparisons come down to. There are a lot of sources of evidence, and different ones will be more appropriate in different contexts. For this post I'll assume that we are seeking the most cost-effective interventions. Say we are comparing between intervention x in cause area X, and intervention y in cause area Y. How they compare depends on things like how well thought-out x and y are, how competent the people and organisations implementing them are, as well as how valuable X is as a whole compared to Y. These are all important factors in telling us how x and y ultimately compare, but they’re quite different from one another. So it shouldn’t be a surprise if it’s best to use different methods to compare the different factors. I think this is the case. Consider the equation: COST-EFFECTIVENESS OF INTERVENTION = (COST-EFFECTIVENESS OF AREA) * (LEVERAGE RATIO OF INTERVENTION) The left-hand side of this equation expresses how much good is achieved per unit of resources invested in the intervention. For the intervention x we’ll denote this G(x). The right-hand side breaks this up as C(X), how much good is achieved per unit of resources invested in X as a whole, and a ‘leverage ratio’ L(x) which expresses the ratio of how effective x is compared to X as a whole [1]. Now to compare between x and y we’re interested in the ratio G(x)/G(y). We can use the above equation to expand this: G(x)/G(y) = (C(X)L(x))/(C(Y)L(y)). This rearranges to: G(x)/G(y) = C(X)/C(Y) * L(x)/L(y). Here we’ve split the comparison into two parts, each of which is
bc92b901-f36c-4098-bd7f-75b63fedad3a	trentmkelly/LessWrong-43k	LessWrong	How realistic would AI-engineered chatbots be? I'm interested in how easy it would be to simulate just one present-day person's life rather than an entire planet's worth of people. Currently our chatbots are bad enough that we could not populate the world with NPC's; the lone human would quickly figure out that everyone else was... different, duller, incomprehensibly stupid, etc. But what if the chatbots were designed by a superintelligent AI? If a superintelligent AI was simulating my entire life from birth, would it be able to do it (for reasonably low computational resources cost, i.e. less than the cost of simulating another person) without simulating any other people in sufficient detail that they would be people? I suspect that the answer is yes. If the answer is "maybe" or "no," I would very much like to hear tips on how to tell whether someone is an ideal chatbot. Thoughts? EDIT: In the comments most people are asking me to clarify what I mean by various things. By popular demand: I interact with people in more ways than just textual communication. I also hear them, and see them move about. So when I speak of chatbots I don't mean bots that can do nothing but chat. I mean an algorithm governing the behavior of a simulated entire-human-body, that is nowhere near the complexity of a brain. (Modern chatbots are algorithms governing the behavior of a simulated human-hands-typing-on-keyboard, that are nowhere near the complexity of a brain.) When I spoke of "simulating any other people in sufficient detail that they would be people" I didn't mean to launch us into a philosophical discussion of consciousness or personhood. I take it to be common ground among all of us here that very simple algorithms, such as modern chatbots, are not people. By contrast, many of us think that a simulated human brain would be a person. Assuming a simulated human brain would be a person, but a simple chatbot-like algorithm would not, my question is: Would any algorithm complex enough to fool me into thinking it was a pers
e9a39053-539d-4476-9896-887b75d1efc8	trentmkelly/LessWrong-43k	LessWrong	Crazy Ideas Thread - October 2015 This thread is intended to provide a space for 'crazy' ideas. Ideas that spontaneously come to mind (and feel great), ideas you long wanted to tell but never found the place and time for and also for ideas you think should be obvious and simple - but nobody ever mentions them. Rules for this thread: 1. Each crazy idea goes into its own top level comment and may be commented there. 2. Voting should be based primarily on how original the idea is. 3. Meta discussion of the thread should go to the top level comment intended for that purpose. ---------------------------------------- If you create such a thread do the following : * Use "Crazy Ideas Thread" in the title. * Copy the rules. * Add the tag "crazy_idea". * Create a top-level comment saying 'Discussion of this thread goes here; all other top-level comments should be ideas or similar' * Add a second top-level comment with an initial crazy idea to start participation.
ce091b44-3400-437f-8ac1-34f08a7a9637	trentmkelly/LessWrong-43k	LessWrong	Thinking without words? Before language, people must have thought without words. I often have the impression that I have a thought fully-formed in my head, yet I wait to listen to it unfold in words before moving on to the next thought. Perhaps I could think much faster if I weren't addicted to words. Has anyone developed techniques for thinking without words? This would have a little in common with Buddhist practices of emptying your mind, but wouldn't be the same thing. For one thing, Buddhists also try to empty their minds of images. More importantly, they are trying not to think, while I'm trying to think - just not unpack everything into words.
4c3553f3-a8d7-4f40-8a7b-c47924735f4f	trentmkelly/LessWrong-43k	LessWrong	Why "Referer"? When you click a link, by default your browser sends a request like: GET /your-page HTTP/1.1 Host: your-site Referer: https://other-site/with-url [other headers] It's telling the server what page it wants ( https://your-site/your-page) and it includes a Referer saying that you came from https://other-site/with-url. But why Referer and not Referrer? Let's look back. The original version of HTTP had a much simpler request format: GET /your-page No Host:, no Referer: no headers at all. This initial version, implemented in the WWW browser, became known as HTTP/0.9. There's a design doc and an as-implemented doc, both with no mention of headers. Right away there were many things people wanted from HTTP that HTTP/0.9 didn't support, and there were lots of ideas for the next version. Not surprisingly some of these initial ideas didn't look much like modern HTTP. Updates To HTTP (last modified 1992-01-07) references HyperText Request (last modified 1992-09-22) which has an example: GET "http://info.cern.ch/hypertext/WWW/People.html#Cailliau" HTRQ/1 PROFILE KEY="akhkygy" AUTH WHO="Bloggs" PWD="12345" CLIENT ID="Smith" HOST="www2" EMAIL="SM@Ajax.com" ORG="DuPont" PUBKEY KEY="6246246378098996127" FORMAT NAME="rtfMac" PENALTY="500,120,33" NAME="EPS" PENALTY="100,50,0" NAME="HTML" PENALTY="1,1,1" SOFTWARE PLATFORM="NeXTUnix" PROGRAM="WWW" VERSION="3.0/1" While you can see what became the From, Authorization, Accept, and User-Agent headers, this was clearly not in its final form. And there's no Referer. A bit later in 1992, however, we get to what's going to become HTTP/1.0: Basic HTTP as defined in 1992. It does define request headers (spec) including Referer: > This optional header field allows the client to specify, for the server's benefit, the address (URI) of the document (or element within the document) from which the URI in the request was obtained. > This allows a server to generate lists of back-links to documen
55bcaecf-f4d7-42ae-a014-b16cb6fae99a	trentmkelly/LessWrong-43k	LessWrong	Apply to the Cavendish Labs Fellowship (by 4/15) Cavendish Labs is a new research organization in Vermont focused on technical work on existential risks. We'd like to invite you to apply to our fellowships in AI safety and biosecurity! Positions are open for any time between June 1 and December 10, 2023. We pay a stipend of $1,500/month, plus food and housing are provided. Anyone with a technical background is encouraged to apply, even if you lack specific expertise in these fields. Applications for summer research fellows are closing April 15th. Apply here! Research lab on a river in Vermont, AI artist's conception (Note: we likely cannot accept people who need visa sponsorship to work in the U.S.)
b88f9347-5d80-4a75-ac70-6619c7a8a664	trentmkelly/LessWrong-43k	LessWrong	Dealing with the high quantity of scientific error in medicine In a recent article, John Ioannidis describes a very high proportion of medical research as wrong. > Still, Ioannidis anticipated that the community might shrug off his findings: sure, a lot of dubious research makes it into journals, but we researchers and physicians know to ignore it and focus on the good stuff, so what’s the big deal? The other paper headed off that claim. He zoomed in on 49 of the most highly regarded research findings in medicine over the previous 13 years, as judged by the science community’s two standard measures: the papers had appeared in the journals most widely cited in research articles, and the 49 articles themselves were the most widely cited articles in these journals. These were articles that helped lead to the widespread popularity of treatments such as the use of hormone-replacement therapy for menopausal women, vitamin E to reduce the risk of heart disease, coronary stents to ward off heart attacks, and daily low-dose aspirin to control blood pressure and prevent heart attacks and strokes. Ioannidis was putting his contentions to the test not against run-of-the-mill research, or even merely well-accepted research, but against the absolute tip of the research pyramid. Of the 49 articles, 45 claimed to have uncovered effective interventions. Thirty-four of these claims had been retested, and 14 of these, or 41 percent, had been convincingly shown to be wrong or significantly exaggerated. If between a third and a half of the most acclaimed research in medicine was proving untrustworthy, the scope and impact of the problem were undeniable. That article was published in the Journal of the American Medical Association. Part of the problem is that surprising results get more interest, and surprising results are more likely to be wrong. (I'm not dead certain of this-- if the baseline beliefs are highly likely to be wrong, surprising beliefs become somewhat less likely to be wrong.) Replication is boring. Failure to replicate a bright sh
7b2c7a52-d86c-402c-b386-929ff8d1300f	trentmkelly/LessWrong-43k	LessWrong	What counts as defection? Thanks to Michael Dennis for proposing the formal definition; to Andrew Critch for pointing me in this direction; to Abram Demski for proposing non-negative weighting; and to Alex Appel, Scott Emmons, Evan Hubinger, philh, Rohin Shah, and Carroll Wainwright for their feedback and ideas. There's a good chance I'd like to publish this at some point as part of a larger work. However, I wanted to make the work available now, in case that doesn't happen soon. > They can't prove the conspiracy... But they could, if Steve runs his mouth. > > The police chief stares at you. > > You stare at the table. You'd agreed (sworn!) to stay quiet. You'd even studied game theory together. But, you hadn't understood what an extra year of jail meant. > > The police chief stares at you. > > Let Steve be the gullible idealist. You have a family waiting for you. > Sunlight stretches across the valley, dappling the grass and warming your bow. Your hand anxiously runs along the bowstring. A distant figure darts between trees, and your stomach rumbles. The day is near spent. > > The stags run strong and free in this land. Carla should meet you there. Shouldn't she? Who wants to live like a beggar, subsisting on scraps of lean rabbit meat? > > In your mind's eye, you reach the stags, alone. You find one, and your arrow pierces its barrow. The beast shoots away; the rest of the herd follows. You slump against the tree, exhausted, and never open your eyes again. > > You can't risk it. People talk about 'defection' in social dilemma games, from the prisoner's dilemma to stag hunt to chicken. In the tragedy of the commons, we talk about defection. The concept has become a regular part of LessWrong discourse. Informal definition. A player defects when they increase their personal payoff at the expense of the group. This informal definition is no secret, being echoed from the ancient Formal Models of Dilemmas in Social Decision-Making to the recent Classifying games like the Pr
dae0c00a-a8c9-4891-8024-6c47a7039c4c	trentmkelly/LessWrong-43k	LessWrong	The Singular Value Decompositions of Transformer Weight Matrices are Highly Interpretable Please go to the colab for interactive viewing and playing with the phenomena. For space reasons, not all results included in the colab are included here so please visit the colab for the full story. A GitHub repository with the colab notebook and accompanying data can be found here. This post is part of the work done at Conjecture. TLDR If we take the SVD of the weight matrices of the OV circuit and of MLP layers of GPT models, and project them to token embedding space, we notice this results in highly interpretable semantic clusters. This means that the network learns to align the principal directions of each MLP weight matrix or attention head to read from or write to semantically interpretable directions in the residual stream. We can use this to both improve our understanding of transformer language models and edit their representations. We use this finding to design both a natural language query locator, where you can write a set of natural language concepts and find all weight directions in the network which correspond to it, and also to edit the network's representations by deleting specific singular vectors, which results in relatively large effects on the logits related to the semantics of that vector and relatively small effects on semantically different clusters Introduction Trying to understand the internal representations of language models, and of deep neural networks in general, has been the primary focus of the field of mechanistic interpretability, with clear applications to AI alignment. If we can understand the internal dimensions along which language models store and manipulate representations, then we can get a much better grasp on their behaviour and ultimately may be able to both make provable statements about bounds on their behaviour, as well as make precise edits to the network to prevent or enhance desired behaviours. Interpretability, however, is a young field where we still do not yet fully understand what the basic units of the
edd37099-0374-4007-9e54-ac618c6487f9	trentmkelly/LessWrong-43k	LessWrong	Stable and Unstable Risks Related: Existential Risk, 9/26 is Petrov Day Existential risks—risks that, in the words of Nick Bostrom, would "either annihilate Earth-originating intelligent life or permanently and drastically curtail its potential," are a significant threat to the world as we know it. In fact, they may be one of the most pressing issues facing humanity today. The likelihood of some risks may stay relatively constant over time—a basic view of asteroid impact is that there is a certain probability that a "killer asteroid" hits the Earth and that this probability is more or less the same every year. This is what I refer to as a "stable risk." However, the likelihood of other existential risks seems to fluctuate, often quite dramatically. Many of these "unstable risks" are related to human activity. For instance, the likelihood of a nuclear war at sufficient scale to be an existential threat seems contingent on various geopolitical factors that are difficult to predict in advance. That said, the likelihood of this risk has clearly changed throughout recent history. Nuclear war was obviously not an existential risk before nuclear weapons were invented, and was fairly clearly more of a risk during the Cuban Missile Crisis than it is today. Many of these unstable, human-created risks seem based largely on advanced technology. Potential risks like gray goo rely on theorized technologies that have yet to be developed (and indeed may never be developed). While this is good news for the present day, it also means that we have to be vigilant for the emergence of potential new threats as human technology increases. GiveWell's recent conversation with Carl Shulman contains some arguments as to why the risk of human extinction may be decreasing over time. However, it strikes me as perhaps more likely that the risk of human extinction is increasing over time—or at the very least becoming less stable—as technology increases the amount of power available to individuals and civilizations.
e9545078-371d-47ab-b452-f221eb3314e3	trentmkelly/LessWrong-43k	LessWrong	Impressions from a panel discussion on AGI I went to this event to listen to Jaan Tallinn, Scott Aaronson and Don Eigler discuss the AGI, the Singularity and the [F]AI research. Not surprisingly, Jaan, whose ideas are significantly influenced by what EY preaches, advocated the urgent need for research into AGI friendliness as much as the AGI research proper. Scott was rather more laid back, estimating that a "FOOMable" AGI is probably 10,000 years away, and that, while the AGI problem is already really really hard, the FAI problem is harder still, so expending significant effort on the FAI problem before we understand the AGI issues better is probably not a good use of resources. Don, who makes atom-sized gates in his lab, suggested that the Moore's law will probably level out before the AGI becomes a reality. When asked, he said that he can see another 10^3-10^4 times improvement in chip complexity with technological innovations only, without the need for new scientific breakthroughs. This includes both the miniaturization of gates to near-atomic size and introducing 3D layouts with multiple interconnects. He expects the latter to be a significant breakthrough, provided the power consumption and dissipation issues are solved. 10^4 times corresponds to about 30-50 years given the current slope. A large chunk of the discussion was rehashing the standard points about the FAI (Pascal's wager type of arguments and counterarguments, augmentation and upload as some ways around UFAI, etc.), all of which have been discussed here to no end, so I will not repeat it. The video of the event will apparently be posted eventually, no time frame given. I have a marginal quality voice recording with my phone, available if anyone really wants to listen. If any of the Vancouver LWers attended, please feel free to share your impressions.
e942c121-30c1-417a-9bc0-9c7c698bc58b	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Superintelligence 5: Forms of Superintelligence This is part of a weekly reading group on [Nick Bostrom](http://www.nickbostrom.com/)'s book, [Superintelligence](http://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111). For more information about the group, and an index of posts so far see the [announcement post](/lw/kw4/superintelligence_reading_group/). For the schedule of future topics, see [MIRI's reading guide](https://intelligence.org/wp-content/uploads/2014/08/Superintelligence-Readers-Guide-early-version.pdf). --- Welcome. This week we discuss the fifth section in the [reading guide](https://intelligence.org/wp-content/uploads/2014/08/Superintelligence-Readers-Guide-early-version.pdf): *Forms of superintelligence. This corresponds to Chapter 3, on different ways in which an intelligence can be super. This post summarizes the section, and offers a few relevant notes, and ideas for further investigation. Some of my own thoughts and questions for discussion are in the comments. There is no need to proceed in order through this post, or to look at everything. Feel free to jump straight to the discussion. Where applicable and I remember, page numbers indicate the rough part of the chapter that is most related (not necessarily that the chapter is being cited for the specific claim). Reading: Chapter 3 (p52-61) --- Summary ======= 1. A speed superintelligence* could do what a human does, but faster. This would make the outside world seem very slow to it. It might cope with this partially by being very tiny, or virtual. (p53) 2. A collective superintelligence is composed of smaller intellects, interacting in some way. It is especially good at tasks that can be broken into parts and completed in parallel. It can be improved by adding more smaller intellects, or by organizing them better. (p54) 3. A quality superintelligence can carry out intellectual tasks that humans just can't in practice, without necessarily being better or faster at the things humans can do. This can be understood by analogy with the difference between other animals and humans, or the difference between humans with and without certain cognitive capabilities. (p56-7) 4. These different kinds of superintelligence are especially good at different kinds of tasks. We might say they have different 'direct reach'. Ultimately they could all lead to one another, so can indirectly carry out the same tasks. We might say their 'indirect reach' is the same. (p58-9) 5. We don't know how smart it is possible for a biological or a synthetic intelligence to be. Nonetheless we can be confident that synthetic entities can be much more intelligent than biological entities. 1. Digital intelligences would have better hardware: they would be made of components ten million times faster than neurons; the components could communicate about two million times faster than neurons can; they could use many more components while our brains are constrained to our skulls; it looks like better memory should be feasible; and they could be built to be more reliable, long-lasting, flexible, and well suited to their environment. 2. Digital intelligences would have better software: they could be cheaply and non-destructively 'edited'; they could be duplicated arbitrarily; they could have well aligned goals as a result of this duplication; they could share memories (at least for some forms of AI); and they could have powerful dedicated software (like our vision system) for domains where we have to rely on slow general reasoning. Notes ===== 1. This chapter is about different kinds of superintelligent entities that could exist. I like to think about the closely related question,'what kinds of better can intelligence be?'You can be a better baker if you can bake a cake faster, or bake more cakes, or bake better cakes. Similarly, a system can become more intelligent if it can do the same intelligent things faster, or if it does things that are qualitatively more intelligent. (Collective intelligence seems somewhat different, in that it appears to be a means to be faster or able to do better things, though it may have benefits in dimensions I'm not thinking of.) I think the chapter is getting at different ways intelligence can be better rather than 'forms' in general, which might vary on many other dimensions (e.g. emulation vs AI, goal directed vs. reflexive, nice vs. nasty). 2. Some of the hardware and software advantages mentioned would be pretty transformative on their own. If you haven't before, consider taking a moment to think about what the world would be like if people could be cheaply and perfectly replicated, with their skills intact. Or if people could live arbitrarily long by replacing worn components. 3. The main differences between increasing intelligence of a system via speed and via collectiveness seem to be: (1) the 'collective' route requires that you can break up the task into parallelizable subtasks, (2) it generally has larger costs from communication between those subparts, and (3) it can't produce a single unit as fast as a comparable 'speed-based' system. This suggests that anything a collective intelligence can do, a comparable speed intelligence can do at least as well. One counterexample to this I can think of is that often groups include people with a diversity of knowledge and approaches, and so the group can do a lot more productive thinking than a single person could. It seems wrong to count this as a virtue of collective intelligence in general however, since you could also have a single fast system with varied approaches at different times. 4. For each task, we can think of curves for how performance increases as we increase intelligence in these different ways. For instance, take the task of finding a fact on the internet quickly. It seems to me that a person who ran at 10x speed would get the figure 10x faster. Ten times as many people working in parallel would do it only a bit faster than one, depending on the variance of their individual performance, and whether they found some clever way to complement each other. It's not obvious how to multiply qualitative intelligence by a particular factor, especially as there are different ways to improve the quality of a system. It also seems non-obvious to me how search speed would scale with a particular measure such as IQ. 5. How much more intelligent do human systems get as we add more humans? I can't find much of an answer, but people have investigated the effect of things like [team size](http://www.tandfonline.com/doi/abs/10.1080/07399019108964994#.VDxkLdR4p-h), [city size](http://scholar.google.com/scholar?q=productivity+city+size&btnG=&hl=en&as_sdt=0%2C39), and [scientific collaboration](http://scholar.google.com/scholar?q=productivity+collaboration&btnG=&hl=en&as_sdt=0%2C39) on various measures of productivity. 6. The things we might think of as collective intelligences - e.g. companies, governments, academic fields - seem notable to me for being slow-moving*, relative to their components. If someone were to steal some chewing gum from Target, Target can respond in the sense that an employee can try to stop them. And this is no slower than an individual human acting to stop their chewing gum from being taken. However it also doesn't involve any extra problem-solving from the organization - to the extent that the organization's intelligence goes into the issue, it has to have already done the thinking ahead of time. Target was probably much smarter than an individual human about setting up the procedures and the incentives to have a person there ready to respond quickly and effectively, but that might have happened over months or years. In-depth investigations ======================= If you are particularly interested in these topics, and want to do further research, these are a few plausible directions, some inspired by Luke Muehlhauser's [list](http://lukemuehlhauser.com/some-studies-which-could-improve-our-strategic-picture-of-superintelligence/), which contains many suggestions related to parts of Superintelligence.These projects could be attempted at various levels of depth. 1. Produce improved measures of (substrate-independent) general intelligence. Build on the ideas of Legg, Yudkowsky, Goertzel, Hernandez-Orallo & Dowe, etc. Differentiate intelligence quality from speed. 2. List some feasible but non-realized cognitive talents for humans, and explore what could be achieved if they were given to some humans. 3. List and examine some types of problems better solved by a speed superintelligence than by a collective superintelligence, and vice versa. Also, what are the returns on “more brains applied to the problem” (collective intelligence) for various problems? If there were merely a huge number of human-level agents added to the economy, how much would it speed up economic growth, technological progress, or other relevant metrics? If there were a large number of researchers added to the field of AI, how would it change progress? 4. How does intelligence quality improve performance on economically relevant tasks? If you are interested in anything like this, you might want to mention it in the comments, and see whether other people have useful thoughts. How to proceed ============== This has been a collection of notes on the chapter. The most important part of the reading group though is discussion, which is in the comments section. I pose some questions for you there, and I invite you to add your own. Please remember that this group contains a variety of levels of expertise: if a line of discussion seems too basic or too incomprehensible, look around for one that suits you better! Next week, we will talk about 'intelligence explosion kinetics', a topic at the center of much contemporary debate over the arrival of machine intelligence. To prepare, read* Chapter 4, The kinetics of an intelligence explosion(p62-77).The discussion will go live at 6pm Pacific time next Monday 20 October. Sign up to be notified [here](http://intelligence.us5.list-manage.com/subscribe?u=353906382677fa789a483ba9e&id=28cb982f40).
a6ed7f45-04ac-4680-b037-b7c3579caf2b	StampyAI/alignment-research-dataset/special_docs	Other	Differential Assessment of Black-Box AI Agents. Differential Assessment of Black-Box AI Agents Rashmeet Kaur Nayyar, Pulkit Verma,andSiddharth Srivastava Autonomous Agents and Intelligent Robots Lab, School of Computing and Augmented Intelligence, Arizona State University, AZ, USA {rmnayyar, verma.pulkit, siddharths }@asu.edu Abstract Much of the research on learning symbolic models of AI agents focuses on agents with stationary models. This as- sumption fails to hold in settings where the agent’s capa- bilities may change as a result of learning, adaptation, or other post-deployment modifications. Efficient assessment of agents in such settings is critical for learning the true capabil- ities of an AI system and for ensuring its safe usage. In this work, we propose a novel approach to differentially assess black-box AI agents that have drifted from their previously known models. As a starting point, we consider the fully ob- servable and deterministic setting. We leverage sparse obser- vations of the drifted agent’s current behavior and knowl- edge of its initial model to generate an active querying pol- icy that selectively queries the agent and computes an up- dated model of its functionality. Empirical evaluation shows that our approach is much more efficient than re-learning the agent model from scratch. We also show that the cost of dif- ferential assessment using our method is proportional to the amount of drift in the agent’s functionality. 1 Introduction With increasingly greater autonomy in AI systems in recent years, a major problem still persists and has largely been overlooked: how do we accurately predict the behavior of a black-box AI agent that is evolving and adapting to changes in the environment it is operating in? And how do we ensure its reliable and safe usage? Numerous factors could cause unpredictable changes in agent behaviors: sensors and actu- ators may fail due to physical damage, the agent may adapt to a dynamic environment, users may change deployment and use-case scenarios, etc. Most prior work on the topic presumes that the functionalities and the capabilities of AI agents are static, while some works start with a tabula-rasa and learn the entire model from scratch. However, in many real-world scenarios, the agent model is transient and only parts of its functionality change at a time. Bryce, Benton, and Boldt (2016) address a related prob- lem where the system learns the updated mental model of a user using particle filtering given prior knowledge about the user’s mental model. However, they assume that the entity *Equal contribution. Alphabetical order. Copyright © 2022, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Agent update Model learning from scratch (expensive) Initially known model of Sparse observations (collected once) Query Response Learned model of Model learning from scratch not needed Differential Assessment of AI System (DAAISy)events XFigure 1: The Differential Assessment of AI System (DAAISy) takes as input the initially known model of the agent prior to model drift, available observations of the up- dated agent’s behavior, and performs a selective dialog with the black-box AI agent to output its updated model through efficient model learning. being modeled can tell the learning system about flaws in the learned model if needed. This assumption does not hold in settings where the entity being modeled is a black-box AI system: most such systems are either implemented using inscrutable representations or otherwise lack the ability to automatically generate a model of their functionality (what they can do and when) in terms the user can understand. The problem of efficiently assessing, in human-interpretable terms, the functionality of such a non-stationary AI system has received little research attention. The primary contribution of this paper is an algorithm for differential assessment of black-box AI systems (Fig. 1). This algorithm utilizes an initially known interpretable model of the agent as it was in the past, and a small set of observations of agent execution. It uses these observations to develop an incremental querying strategy that avoids the full cost of assessment from scratch and outputs a revised model of the agent’s new functionality. One of the challenges in learning agent models from observational data is that reduc- tions in agent functionality often do not correspond to spe- cific “evidence” in behavioral observations, as the agent may not visit states where certain useful actions are no longer ap- plicable. Our analysis shows that if the agent can be placed in an “optimal” planning mode, differential assessment can indeed be used to query the agent and recover information about reduction in functionality. This “optimal” planning mode is not necessarily needed for learning about increase in functionality. Empirical evaluations on a range of problems clearly demonstrate that our method is much more efficient than re-learning the agent’s model from scratch. They also exhibit the desirable property that the computational cost of differential assessment is proportional to the amount of drift in the agent’s functionality. Running Example Consider a battery-powered rover with limited storage capacity that collects soil samples and takes pictures. Assume that its planning model is similar to IPC domain Rovers (Long and Fox 2003). It has an action that collects a rock sample at a waypoint and stores it in a stor- age iff it has at least half of the battery capacity remaining. Suppose there was an update to the rover’s system and as a result of this update, the rover can now collect the rock sample only when its battery is full, as opposed to at least half-charged battery that it needed before. Mission planners familiar with the earlier system and unaware about the ex- act updates in the functionality of the rover would struggle to collect sufficient samples. This could jeopardise multiple missions if it is not detected in time. This example illustrates how our system could be of value by differentially detecting such a drift in the functionality of a black-box AI system and deriving its true functionality. The rest of this paper is organized as follows: The next section presents background terminology. This is followed by a formalization of the differential model assessment prob- lem in Section 3. Section 4 presents our approach for differ- ential assessment by first identifying aspects of the agent’s functionality that may be affected (Section 4.1) followed by the process for selectively querying the agent using a primi- tive set of queries. We present empirical evaluation of the ef- ficiency of our approach on randomly generated benchmark planning domains in Section 5. Finally, we discuss relevant related work in Section 6 and conclude in Section 7. 2 Preliminaries We consider models that express an agent’s functionalities in the form of STRIPS-like planning models (Fikes and Nils- son 1971; McDermott et al. 1998; Fox and Long 2003) as defined below. Definition 1. A planning domain model is a tuple M= ⟨P, A⟩, where P={pr1 1, . . . , prnn}is a finite set of predi- cates with arities ri,i∈[1, n]; andA={a1, . . . , a k}is a fi- nite set of parameterized relational actions. Each action ai∈ Ais represented as a tuple ⟨header (ai),pre(ai),eff(ai)⟩, where header (ai)represents the action header consisting of the name and parameters for the action ai,pre(ai)rep- resents the conjunction of positive or negative literals that must be true in a state where the action aiis applicable, and eff(ai)is the conjunction of positive or negative literals that become true as a result of execution of the action ai. In the rest of the paper, we use the term “model” to refer to planning domain models and use closed-world assump-tion as used in the Planning Domain Definition Language (PDDL) (McDermott et al. 1998). Given a model Mand a set of objects O, letSM,O be the space of all states defined as maximally consistent sets of literals over the predicate vocabulary of MwithOas the set of objects. We omit the subscript when it is clear from context. An action a∈Ais applicable in a state s∈Sifs\|=pre(a). The result of exe- cuting ais a state a(s) =s′∈Ssuch that s′\|=eff(a), and all atoms not in eff(a)have literal forms as in s. A literal corresponding to a predicate p∈Pcan ap- pear in pre(a)oreff(a)of an action a∈Aif and only if it can be instantiated using a subset of parameters of a. E.g., consider an action navigate (?rover ?src ?dest) and a pred- icate (can traverse ?rover ?x ?y) in the Rovers domain dis- cussed earlier. Suppose a literal corresponding to predicate (can traverse ?rover ?x ?y) can appear in the precondition and/or the effect of navigate (?rover ?src ?dest) action. As- suming we know ?xand?yincantraverse , and ?src and ?dest innavigate are of the same type waypoint , the possi- ble lifted instantiations of predicate cantraverse compatible with action navigate are(can traverse ?rover ?src ?dest) , (can traverse ?rover ?dest ?src) ,(can traverse ?rover ?src ?src) , and (can traverse ?rover ?dest ?dest) . The number of parameters in a predicate p∈Pthat is relevant to an action a∈A, i.e., instantiated using a subset of parameters of the action a, is bounded by the maximum arity of the action a. We formalize this notion of lifted instantiations of a predi- cate with an action as follows: Definition 2. Given a finite set of predicates P= {pr1 1, . . . , prnn}with arities ri,i∈[1, n]; and a finite set of parameterized relational actions A={aψ1 1, . . . , aψk k} with arities ψjand parameters par(aψj j) =⟨α1, . . . , α ψj⟩, j∈[1, k], the set of lifted instantiations of predicates P∗ is defined as the collection {pi(σ(x1), . . . , σ (xri))\|pi∈ P, a∈A, σ:{x1, . . . , x ri} → par(a)}. 2.1 Representing Models We represent a model Musing the set of all possible pal- tuples ΓMof the form γ=⟨p, a, ℓ⟩, where ais a parameter- ized action header for an action in A,p∈P∗is a possible lifted instantiation of a predicate in P, andℓ∈ {pre,eff}de- notes a location in a, precondition or effect, where pcan ap- pear. A model Mis thus a function µM: ΓM→ {+,−,∅} that maps each element in ΓMto amode in the set {+,−,∅}. The assigned mode for a pal-tuple γ∈ΓMdenotes whether pis present as a positive literal ( +), as a negative literal ( −), or absent ( ∅) in the precondition ( ℓ=pre) or effect ( ℓ=eff) of the action header a. This formulation of models as pal-tuples allows us to view the modes for any predicate in an action’s precondition and effect independently. However, at times it is useful to con- sider a model at a granularity of relationship between a pred- icate and an action. We address this by representing a model Mas a set of pa-tuples ΛMof the form ⟨p, a⟩where ais a parameterized action header for an action in A, and p∈P∗ is a possible lifted instantiation of a predicate in P. Each pa-tuple can take a value of the form ⟨mpre, m eff⟩, where mpreandmeffrepresents the mode in which pappears in the precondition and effect of a, respectively. Since a predicate cannot appear as a positive (or negative) literal in both the precondition and effect of an action, ⟨+,+⟩and⟨−,−⟩are not in the range of values that pa-tuples can take. Hence- forth, in the context of a pal-tuple or apa-tuple , we refer to aas an action instead of an action header. Measure of model difference Given two models M1= ⟨P, A 1⟩andM2=⟨P, A 2⟩, defined over the same sets of predicates Pand action headers A, the difference be- tween the two models ∆(M1, M 2)is defined as the num- ber of pal-tuples that differ in their modes in M1andM2, i.e.,∆(M1, M 2) =\|{γ∈P×A× {+,−,∅}\|µM1(γ)̸= µM2(γ)}\|. 2.2 Abstracting Models Several authors have explored the use of abstraction in planning (Sacerdoti 1974; Giunchiglia and Walsh 1992; Helmert, Haslum, and Hoffmann 2007; B ¨ackstr ¨om and Jon- sson 2013; Srivastava, Russell, and Pinto 2016). We define an abstract model as a model that does not have a mode as- signed for at least one of the pal-tuples . LetΓMbe the set of all possible pal-tuples , and ?⃝be an additional possible value that a pal-tuple can take. Assigning ?⃝mode to a pal- tuple denotes that its mode is unknown. An abstract model Mis thus a function µM: ΓM→ {+,−,∅,?⃝}that maps each element in ΓMto amode in the set {+,−,∅,?⃝}. Let Ube the set of all abstract and concrete models that can pos- sibly be expressed by assigning modes in {+,−,∅,?⃝}to each pal-tuple γ∈ΓM. We now formally define model ab- straction as follows: Definition 3. Given models M1andM2,M2is an ab- straction of M1over the set of all possible pal-tuples Γiff ∃Γ2⊆Γs.t.∀γ∈Γ2,µM2(γ) = ?⃝and∀γ∈Γ\Γ2, µM2(γ) =µM1(γ). 2.3 Agent Observation Traces We assume limited access to a set of observation traces O, collected from the agent, as defined below. Definition 4. Anobservation trace ois a sequence of states and actions of the form ⟨s0, a1, s1, a2, . . . , s n−1, an, sn⟩ such that ∀i∈[1, n]ai(si−1) =si. These observation traces can be split into multiple action triplets as defined below. Definition 5. Given an observation trace o= ⟨s0, a1, s1, a2, . . . , s n−1, an, sn⟩, an action triplet is a 3-tuple sub-sequence of oof the form ⟨si−1, ai, si⟩, where i∈[1, n]and applying an action aiin state si−1results in statesi, i.e.,ai(si−1) =si. The states si−1andsiare called pre- and post-states of action ai, respectively. An action triplet ⟨si−1, ai, si⟩is said to be optimal if there does not exist an action sequence (of length ≥1) that takes the agent from state si−1tosiwith total action cost less than that of action ai, where each action aihas unit cost. 2.4 Queries We use queries to actively gain information about the func- tionality of an agent to learn its updated model. We assumethat the agent can respond to a query using a simulator. The availability of such agents with simulators is a common as- sumption as most AI systems already use simulators for de- sign, testing, and verification. We use a notion of queries similar to Verma, Marpally, and Srivastava (2021), to perform a dialog with an au- tonomous agent. These queries use an agent to deter- mine what happens if it executes a sequence of actions in a given initial state. E.g., in the rovers domain, the rover could be asked: what happens when the action sam- plerock(rover1 storage1 waypoint1) is executed in an ini- tial state {(equipped rock analysis rover1), (battery half rover1), (at rover1 waypoint1) }? Formally, a query is a function that maps an agent to a response, which we define as: Definition 6. Given a set of predicates P, a set of actions A, and a set of objects O, aquery Q⟨s, π⟩:A →N×S is parameterized by a start state sI∈Sand a plan π= ⟨a1, . . . , a N⟩, where Sis the state space over PandO, and {a1, . . . , a N}is a subset of action space over AandO. It maps agents to responses θ=⟨nF, sF⟩such that nFis the length of the longest prefix of πthatAcan successfully ex- ecute and sF∈Sis the result of that execution. Responses to such queries can be used to gain use- ful information about the model drift. E.g., consider an agent with an internal model MA driftas shown in Tab. 1. If a query is posed asking what happens when the ac- tion sample rock(rover1 storage1 waypoint1) is executed in an initial state {(equipped rock analysis rover1), (bat- tery half rover1), (at rover1 waypoint1) }, the agent would respond ⟨0,{(equipped rock analysis rover1), (battery half rover1), (at rover1 waypoint1) }⟩, representing that it was not able to execute the plan, and the resulting state was {(equipped rock analysis rover1), (battery half rover1), (at rover1 waypoint1) }(same as the initial state in this case). Note that this response is inconsistent with the model MA init, and it can help in identifying that the precondition of action sample rock(?r ?s ?w) has changed. 3 Formal Framework Our objective is to address the problem of differential assess- ment of black-box AI agents whose functionality may have changed from the last known model. Without loss of gen- erality, we consider situations where the set of action head- ers is same because the problem of differential assessment with changing action headers can be reduced to that with uniform action headers. This is because if the set of actions has increased, new actions can be added with empty precon- ditions and effects to MA init, and if it has decreased, MA initcan be reduced similarly. We assume that the predicate vocabu- lary used in the two models is the same; extension to situ- ations where the vocabulary changes can be used to model open-world scenarios. However, that extension is beyond the scope of this paper. Suppose an agent A’s functionality was known as a model MA init=⟨P,Ainit⟩, and we wish to assess its current func- tionality as the model MA drift=⟨P,Adrift⟩. The drift in the functionality of the agent can be measured by changes in the Model Precondition Effect MA init (equipped rock analysis?r) (battery half ?r) (at ?r ?w)→(rock sample taken?r) (store full ?r ?s) ¬(battery half ?r) (battery reserve ?r) MA drift (equipped rock analysis?r) (battery full ?r) (at ?r ?w)→(rock sample taken ?r) (store full ?r ?s) ¬(battery full ?r) (battery half ?r) Table 1: sample rock (?r ?s ?w) action of the agent Ain MA initand a possible drifted model MA drift. preconditions and/or effects of all the actions in Ainit. The extent of the drift between MA initandMA driftis represented as the model difference ∆(MA init, MA drift). We formally define the problem of differential assessment of an AI agent below. Definition 7. Given an agent Awith a functionality model MA init, and a set of observations Ocollected using its current version of Adriftwith unknown functionality MA drift, the dif- ferential model assessment problem ⟨MA init, MA drift,O,A⟩is defined as the problem of inferring Ain form of MA driftusing the inputs MA init,O, andA. We wish to develop solutions to the problem of differ- ential assessment of AI agents that are more efficient than re-assessment from scratch. 3.1 Correctness of Assessed Model We now discuss the properties that a model, which is a so- lution to the differential model assessment problem, should satisfy. A critical property of such models is that they should be consistent with the observation traces. We formally define consistency of a model w.r.t. an observation trace as follows: Definition 8. Let obe an observation trace ⟨s0, a1, s1, a2, . . . , s n−1, an, sn⟩. A model M=⟨P, A⟩ isconsistent with the observation trace oiff ∀i∈ {1, .., n} ∃a∈Aandaiis a grounding of ac- tiona s.t. s i−1\|=pre(ai)∧ ∀l∈eff(ai)si\|=l. In addition to being consistent with observation traces, a model should also be consistent with the queries that are asked and the responses that are received while actively in- ferring the model of the agent’s new functionality. We for- mally define consistency of a model with respect to a query and a response as: Definition 9. LetM=⟨P, A⟩be a model; Obe a set of ob- jects; Q=⟨sI, π=⟨a1, . . . a n⟩⟩be a query defined using P, A, andO, and let θ=⟨nF, sF⟩, (nF≤n) be a response toQ.Misconsistent with the query-response ⟨Q, θ⟩iff there exists an observation trace ⟨sI, a1, s1, . . . , a nF, snF⟩ thatMis consistent with and snF̸\|=pre(anF+1)where pre(anF+1)is the precondition of anF+1inM. We now discuss our methodology for solving the problem of differential assessment of AI systems.4 Differential Assessment of AI Systems Differential Assessment of AI Sy stems (Alg. 1) -- DAAISy -- takes as input an agent Awhose functionality has drifted, the model MA init=⟨P, A⟩representing the previously known functionality of A, a set of arbitrary observation traces O, and a set of random states S ⊆ S. Alg. 1 returns a set of updated models MA drift, where each model MA drift∈ MA drift represents A’s updated functionality and is consistent with all observation traces o∈O. A major contribution of this work is to introduce an ap- proach to make inferences about not just the expanded func- tionality of an agent but also its reduced functionality using a limited set of observation traces. Situations where the scope of applicability of an action reduces, i.e., the agent can no longer use an action ato reach state s′from state swhile it could before (e.g., due to addition of a precondition lit- eral), are particularly difficult to identify because observing its behavior does not readily reveal what it cannot do in a given state. Most observation based action-model learners, even when given access to an incomplete model to start with, fail to make inferences about reduced functionality. DAAISy uses two principles to identify such a functionality reduc- tion. First, it uses active querying so that the agent can be made to reveal failure of reachability, and second, we show that if the agent can be placed in optimal planning mode, plan length differences can be used to infer a reduction in functionality. DAAISy performs two major functions; it first identifies a salient set of pal-tuples whose modes were likely affected (line 1 of Alg. 1), and then infers the mode of such af- fected pal-tuples accurately through focused dialog with the agent (line 2 onwards of Alg. 1). In Sec. 4.1, we present our method for identifying a salient set of potentially af- fected pal-tuples that contribute towards expansion in the functionality of the agent through inference from available arbitrary observations. We then discuss the problem of iden- tification of pal-tuples that contribute towards reduction in the functionality of the agent and argue that it cannot be per- formed using successful executions in observations of sat- isficing behavior. We show that pal-tuples corresponding to reduced functionality can be identified if observations of op- timal behavior of the agent are available (Sec. 4.1). Finally, we present how we infer the nature of changes in all affected pal-tuples through a query-based interaction with the agent (Sec. 4.2) by building upon the Agent Interrogation Algo- rithm (AIA) (Verma, Marpally, and Srivastava 2021). Iden- tifying affected pal-tuples helps reduce the computational cost of querying as opposed to the exhaustive querying strat- egy used by AIA. We now discuss the two major functions of Alg. 1 in detail. 4.1 Identifying Potentially Affected pal-tuples We identify a reduced set of pal-tuples whose modes were potentially affected during the model drift, denoted by Γδ, using a small set of available observation traces O. We draw two kinds of inferences from these observation traces: in- ferences about expanded functionality, and inferences about reduced functionality. We discuss our method for inferring Algorithm 1: Differential Assessment of AI Systems Input: MA init,O,A,S Output: MA drift 1:Γδ←identify affected pals() 2:Mabs←setpal-tuples inMA initcorresponding to Γδto?⃝ 3:MA drift← {Mabs} 4:foreachγinΓδdo 5: foreachMabsinMA driftdo 6: Mabs←Mabs× {γ+, γ−, γ∅} 7: Msieved← {} 8: ifaction corresponding to γ:γainOthen 9: spre←states where γaapplicable (O, γa) 10: Q← ⟨spre\ {γp∪ ¬γp},γa⟩ 11: θ←askquery (A, Q) 12: Msieved←sieve models (Mabs, Q, θ) 13: else 14: foreach pair ⟨Mi, Mj⟩inMabsdo 15: Q←generate query (Mi, Mj, γ, S) 16: θ←askquery (A, Q) 17: Msieved←sieve models ({Mi, Mj}, Q, θ) 18: end for 19: end if 20: Mabs← M abs\ M sieved 21: end for 22:MA drift← M abs 23:end for Γδfor both types of changes in the functionality below. Expanded functionality To infer expanded functionality of the agent, we use the previously known model of the agent’s functionality and identify its differences with the possible behaviors of the agent that are consistent with O. To identify the pal-tuples that directly contribute to an ex- pansion in the agent’s functionality, we perform an analysis similar to Stern and Juba (2017), but instead of bounding the predicates that can appear in each action’s precondition and effect, we bound the range of possible values that each pa-tuple inMA driftcan take using Tab. 2. For any pa-tuple , a direct comparison between its value in MA initand possible inferred values in MA driftprovides an indication of whether it was affected. To identify possible values for a pa-tuple ⟨p, a⟩, we first collect a set of all the action-triplets from Othat contain the action a. For a given predicate pand state s, ifs\|=pthen the presence of predicate pis represented as pos, similarly, ifs\|=¬pthen the presence of predicate pis represented as neg. Using this representation, a tuple of predicate presence ∈ {(pos,pos) ,(pos,neg) ,(neg,pos) ,(neg,neg) }is determined for the pa-tuple ⟨p, a⟩for each action triplet ⟨s, a, s′⟩ ∈O by analyzing the presence of predicate pin the pre- and post-states of the action triplets. Possible values of the pa- tuple that are consistent with Oare directly inferred from the Tab. 2 using the inferred tuples of predicate presence. E.g., for a pa-tuple , the values ⟨+,−⟩and⟨∅,−⟩are consis- tent with (pos, neg ), whereas, only ⟨∅,+⟩is consistent with⟨mpre, m eff⟩(pos,pos) (pos,neg) (neg,pos) (neg,neg) ⟨+,−⟩ ✗ ✓ ✗ ✗ ⟨+,∅⟩ ✓ ✗ ✗ ✗ ⟨−,+⟩ ✗ ✗ ✓ ✗ ⟨−,∅⟩ ✗ ✗ ✗ ✓ ⟨∅,+⟩ ✓ ✗ ✓ ✗ ⟨∅,−⟩ ✗ ✓ ✗ ✓ ⟨∅,∅⟩ ✓ ✗ ✗ ✓ Table 2: Each row represents a possible value ⟨mpre, m eff⟩ for a pa-tuple ⟨p, a⟩. Each column represents a possible tuple representing presence of predicate pin the pre- and post- states of an action triplet ⟨si, a, s i+1⟩(discussed in Sec.4.1). The cells represent whether a value for pa-tuple is consistent with an action triplet in observation traces. (pos, pos )and(neg, pos )tuples of predicate presence that are inferred from O. Once all the possible values for each pa-tuple inMA drift are inferred, we identify pa-tuples whose previously known value in MA initis no longer possible due to inconsistency withO. The pal-tuples corresponding to such pa-tuples are added to the set of potentially affected pal-tuples Γδ. Our method also infers the correct modes of a subset of pal- tuples . E.g., consider a predicate pand two actions triplets in Oof the form ⟨s1, a, s′ 1⟩and⟨s2, a, s′ 2⟩that satisfy s1\|=p ands2\|=¬p. Such an observation clearly indicates that p is not in the precondition of action a, i.e., mode for ⟨p, a⟩ in the precondition is ∅. Such inferences of modes are used to update the known functionality of the agent. We remove such pal-tuples , whose modes are already inferred, from Γδ. A shortcoming of direct inference from successful execu- tions in available observation traces is that it cannot learn any reduction in the functionality of the agent, as discussed in the beginning of Sec. 4. We now discuss our method to address this limitation and identify a larger set of potentially affected pal-tuples . Reduced functionality We conceptualize reduction in functionality as an increase in the optimal cost of going from one state to another. More precisely, reduction in functional- ity represents situations where there exist states si, sjsuch that the minimum cost of going from sitosjis higher in MA driftthan in MA init. In this paper, this cost refers to the num- ber of steps between the pair of states as we consider unit action costs. This notion encompasses situations with reduc- tions in reachability as a special case. In practice, a reduction in functionality may occur if the precondition of at least one action in MA drifthas new pal-tuples , or the effect of at least one of its actions has new pal-tuples that conflict with other actions required for reaching certain states. Our notion of reduced functionality captures all the vari- ants of reduction in functionality. However, for clarity, we illustrate an example that focuses on situations where pre- condition of an action has increased. Consider the case from Tab. 1 where A’s model gets updated from MA inittoMA drift. The action sample rock’s applicability in MA drifthas reduced from that in MA initasAcan no longer sample rocks in sit- uations where the battery is half charged but needs a fully charged battery to be able to execute the action. In such sce- narios, instead of relying on observation traces, our method identifies traces containing indications of actions that were affected either in their precondition or effect, discovers ad- ditional salient pal-tuples that were potentially affected, and adds them to the set of potentially affected pal-tuples Γδ. To find pal-tuples corresponding to reduced functional- ity of the agent, we place the agent in an optimal plan- ning mode and assume limited availability of observation tracesOin the form of optimal unit-cost state-action tra- jectories ⟨s0, a1, s1, a2, . . . , s n−1, an, sn⟩. We generate op- timal plans using MA initfor all pairs of states in O. We hy- pothesize that, if for a pair of states, the plan generated us- ingMA initis shorter than the plan observed in O, then some functionality of the agent has reduced. Our method performs comparative analysis of optimality of the observation traces against the optimal solutions gen- erated using MA initfor same pairs of initial and final states. To begin with, we extract all the continuous state sub-sequences fromOof the form ⟨s0, s1, . . . , s n⟩denoted by Odriftas they are all optimal. We then generate a set of planning problems Pusing the initial and final states of trajectories in Odrift. Then, we provide the problems in PtoMA initto get a set of optimal trajectories Oinit. We select all the pairs of optimal trajectories of the form ⟨oinit, odrift⟩for further analysis such that the length of oinit∈Oinitfor a problem is shorter than the length of odrift∈Odriftfor the same problem. For all such pairs of optimal trajectories, a subset of actions in each oinit∈Oinitwere likely affected due to the model drift. We focus on identifying the first action in each oinit∈Oinitthat was definitely affected. To identify the affected actions, we traverse each pair of optimal trajectories ⟨oinit, odrift⟩simultaneously starting from the initial states. We add all the pal-tuples corresponding to the first differing action in oinittoΓδ. We do this because there are only two possible explanations for why the action differs: (i) either the action in oinitwas applicable in a state using MA initbut has become inapplicable in the same state inMA drift, or (ii) it can no longer achieve the same effects in MA driftasMA init. We also discover the first actions that are ap- plicable in the same states in both the trajectories but result in different states. The effect of such actions has certainly changed in MA drift. We add all the pal-tuples corresponding to such actions to Γδ. In the next section, we describe our approach to infer the correct modes of pal-tuples inΓδ. 4.2 Investigating Affected pal-tuples This section explains how the correct modes of pal-tuples in Γδare inferred (line 2 onwards of Alg.1). Alg. 1 creates an abstract model in which all the pal-tuples that are predicted to have been affected are set to ?⃝(line 2). It then iterates over all pal-tuples with mode ?⃝(line 4). Removing inconsistent models Our method generates candidate abstract models and then removes the abstractmodels that are not consistent with the agent (lines 7-18 of Alg. 1). For each pal-tuple γ∈Γ, the algorithm computes a set of possible abstract models Mabsby assigning the three mode variants +,−, and∅to the current pal-tuple γin model Mabs(line 6). Only one model in Mabscorresponds to the agent’s updated functionality. If the action γain the pal-tuple γis present in the set of action triplets generated using O, then the pre-state of that action spreis used to create a state sI(lines 9-10). sIis created by removing the literals corresponding to predicate γpfromspre. We then create a query Q=⟨sI,⟨γa⟩⟩(line 10), and pose it to the agent A(line 11). The three models are then sieved based on the comparison of their responses to the query Qwith that of A’s response θtoQ(line 12). We use the same mechanism as AIA for sieving the abstract models. If the action corresponding to the current pal-tuple γbe- ing considered is not present in any of the observed action triplets, then for every pair of abstract models in Mabs(line 14), we generate a query Qusing a planning problem (line 15). We then pose the query Qto the agent (line 16) and receive its response θ. We then sieve the abstract models by asking them the same query and discarding the models whose responses are not consistent with that of the agent (line 17). The planning problem that is used to generate the query and the method that checks for consistency of abstract models’ responses with that of the agent are used from AIA. Finally, all the models that are not consistent with the agent’s updated functionality are removed from the possi- ble set of models Mabs. The remaining models are returned by the algorithm. Empirically, we find that only one model is always returned by the algorithm. 4.3 Correctness We now show that the learned drifted model representing the agent’s updated functionality is consistent as defined in Def. 8 and Def. 9. The proof of the theorem is available in the extended version of the paper (Nayyar, Verma, and Sri- vastava 2022). Theorem 1. Given a set of observation traces Ogenerated by the drifted agent Adrift, a set of queries Qposed to Adrift by Alg. 1, and the model MA initrepresenting the agent’s func- tionality prior to the drift, each of the models M=⟨P, A⟩ inMA driftlearned by Alg. 1 are consistent with respect to all the observation traces o∈Oand query-responses ⟨q, θ⟩for all the queries q∈Q. There exists a finite set of observations that if collected will allow Alg. 1 to achieve 100% correctness with any amount of drift: this set corresponds to observations that al- low line 1 of Alg. 1 to detect a change in the functionality. This includes an action triplet in an observation trace hinting at increased functionality, or a shorter plan using the previ- ously known model hinting at reduced functionality. Thus, models learned by DAAISy are guaranteed to be completely correct irrespective of the amount of the drift if such a finite set of observations is available. While using queries signifi- cantly reduces the number of observations required, asymp- totic guarantees subsume those of passive model learners while ensuring convergence to the true model. 5 Empirical Evaluation In this section, we evaluate our approach for assessing a black-box agent to learn its model using information about its previous model and available observations. We imple- mented the algorithm for DAAISy in Python1and tested it on six planning benchmark domains from the International Planning Competition (IPC)2. We used the IPC domains as the unknown drifted models and generated six initial do- mains at random for each domain in our experiments. To assess the performance of our approach with increas- ing drift, we employed two methods for generating the initial domains: (a) dropping the pal-tuples already present, and (b) adding new pal-tuples . For each experiment, we used both types of domain generation. We generated different initial models by randomly changing modes of random pal-tuples in the IPC domains. Thus, in all our experiments an IPC do- main plays the role of ground truth M∗ driftand a randomized model is used as MA init. We use a very small set of observation traces O(single observation trace containing 10 action triplets) in all the ex- periments for each domain. To generate this set, we gave the agent a random problem instance from the IPC corre- sponding to the domain used by the agent. The agent then used Fast Downward (Helmert 2006) with LM-Cut heuris- tic (Helmert and Domshlak 2009) to produce an optimal solution for the given problem. The generated observation trace is provided to DAAISy as input in addition to a ran- domMA initas discussed in Alg. 1. The exact same observation trace is used in all experiments of the same domain, without the knowledge of the drifted model of the agent, and irre- spective of the amount of drift. We measure the final accuracy of the learned model MA driftagainst the ground truth model M∗ driftusing the mea- sure of model difference ∆(MA drift, M∗ drift). We also mea- sure the number of queries required to learn a model with significantly high accuracy. We compare the efficiency of DAAISy (our approach) with the Agent Interrogation Algo- rithm (AIA) (Verma, Marpally, and Srivastava 2021) as it is the most closely related querying-based system. All of our experiments were executed on 5.0 GHz Intel i9 CPUs with 64 GB RAM running Ubuntu 18.04. We now discuss our results in detail below. 5.1 Results We evaluated the performance of DAAISy along 2 direc- tions; the number of queries it takes to learn the updated model MA driftwith increasing amount of drift, and the cor- rectness of the model MA driftit learns compared to M∗ drift. Efficiency in number of queries As seen in Fig. 2, the computational cost of assessing each agent, measured in terms of the number of queries used by DAAISy, increases as the amount of drift in the model M∗ driftincreases. This is expected as the amount of drift is directly proportional to the 1Code available at https://github.com/AAIR-lab/DAAISy 2https://www.icaps-conference.org/competitions Accur acy gained b y DAAIS y Number of queries b y DAAIS yAccur acy of initial model Accur acy of model computed b y DAAIS y Model Accur acyNumber of Queries % driftFigure 2: The number of queries used by DAAISy (our ap- proach) and AIA (marked on y-axis), as well as accuracy of model computed by DAAISy with increasing amount of drift. Amount of drift equals the ratio of drifted pal-tuples and the total number of pal-tuples in the domains (nPals). The number of action triplets in the observation trace used for each domain is 10. number of pal-tuples affected in the domain. This increases the number of pal-tuples that DAAISy identifies as affected as well as the number of queries as a result. As demonstrated in the plots, the standard deviation for number of queries re- mains low even when we increase the amount of drift, show- ing the stability of DAAISy. Comparison with AIA Tab. 3 shows the average number of queries that AIA took to achieve the same level of accu- racy as our approach for 50% drifted models, and DAAISy requires significantly fewer queries to reach the same levels of accuracy compared to AIA. Fig. 2 also demonstrates that DAAISy always takes fewer queries as compared to AIA to reach reasonably high levels of accuracy. This is because AIA does not use information about the previously known model of the agent and thus ends up querying for all possible pal-tuples . DAAISy, on the other hand, predicts the set of pal-tuples that might have changed based on the observations collected from the agent and thus requires significantly fewer queries. Correctness of learned model DAAISy computes mod- els with at least 50% accuracy in all six domains even when they have completely drifted from their initial model, i.e., ∆(MA drift, M∗ drift) =nPals . It attains nearly accurate models for Gripper and Blocksworld for upto 40% drift. Even in scenarios where the agent’s model drift is more than 50%, DAAISy achieves at least 70% accuracy in five domains. Note that DAAISy is guaranteed to find the correct mode for an identified affected pal-tuple . The reason for less than 100% accuracy when using DAAISy is that it does not pre- dict a pal-tuple to be affected unless it encounters an obser- Domain #Pals AIA DAAISy Gripper 20 15.0 6 .5 Miconic 36 32.0 7 .7 Satellite 50 34.0 9 .0 Blocksworld 52 40.0 11 .4 Termes 134 115.0 27 .0 Rovers 402 316.0 61 .0 Table 3: The average number of queries taken by AIA to achieve the same level of accuracy as DAAISy (our ap- proach) for 50% drifted models. vation trace conflicting with MA init. Thus, the learned model MA drift, even though consistent with all the observation traces, may end up being inaccurate when compared to M∗ drift. Discussion AIA always ends up learning completely accu- rate models, but as noted above, this is because AIA queries exhaustively for all the pal-tuples in the model. There is a clear trade-off between the number of queries that DAAISy takes to learn the model as compared to AIA and the correct- ness of the learned model. As evident from the results, if the model has not drifted much, DAAISy can serve as a better approach to efficiently learn the updated functionality of the agent with less overhead as compared to AIA. Deciding the amount of drift after which it would make sense to switch to querying the model from scratch is a useful analysis not addressed in this paper. 6 Related Work White-box model drift Bryce, Benton, and Boldt (2016) address the problem of learning the updated mental model of a user using particle filtering given prior knowledge about the user’s mental model. However, they assume that the en- tity being modeled can tell the learning system about flaws in the learned model if needed. Eiter et al. (2005, 2010) pro- pose a framework for updating action laws depicted in the form of graphs representing the state space. They assume that changes can only happen in effects, and that knowledge about the state space and what effects might change is avail- able beforehand. Our work does not make such assumptions to learn the correct model of the agent’s functionalities. Action model learning The problem of learning agent models from observations of its behavior is an active area of research (Gil 1994; Yang, Wu, and Jiang 2007; Cress- well, McCluskey, and West 2009; Zhuo and Kambhampati 2013; Arora et al. 2018; Aineto, Celorrio, and Onaindia 2019). Recent work addresses active querying to learn the action model of an agent (Rodrigues et al. 2011; Verma, Marpally, and Srivastava 2021). However, these methods do not address the problem of reducing the computational cost of differential model assessment, which is crucial in non- stationary settings. Online action model learning approaches learn the model of an agent while incorporating new observations of the agent behavior ( ˇCertick ´y 2014; Lamanna et al. 2021a,b).Unlike our approach, they do not handle cases where (i) the new observations are not consistent with the older ones due to changes in the agent’s behavior; and/or (ii) there is reduc- tion in functionality of the agent. Lindsay (2021) solve the problem of learning all static predicates in a domain. They start with a correct partial model that captures the dynamic part of the model accurately and generate negative examples by assuming access to all possible positive examples. Our method is different in that it does not make such assump- tions and leverages a small set of available observations to infer about increased and reduced functionality of an agent’s model. Model reconciliation Model reconciliation litera- ture (Chakraborti et al. 2017; Sreedharan et al. 2019; Sreedharan, Chakraborti, and Kambhampati 2021) deals with inferring the differences between the user and the agent models and removing them using explanations. These methods consider white-box known models whereas our approach works with black-box models of the agent. 7 Conclusions and Future Work We presented a novel method for differential assessment of black-box AI systems to learn models of true functional- ity of agents that have drifted from their previously known functionality. Our approach provides guarantees of correct- ness w.r.t. observations. Our evaluation demonstrates that our system, DAAISy, efficiently learns a highly accurate model of agent’s functionality issuing a significantly lower number of queries as opposed to relearning from scratch. In the future, we plan to extend the framework to more general classes, stochastic settings, and models. Analyzing and pre- dicting switching points from selective querying in DAAISy to relearning from scratch without compromising the cor- rectness of the learned models is also a promising direction for future work. Acknowledgements We thank anonymous reviewers for their helpful feedback on the paper. This work was supported in part by the NSF under grants IIS 1942856, IIS 1909370, and the ONR grant N00014-21-1-2045. References Aineto, D.; Celorrio, S. J.; and Onaindia, E. 2019. Learn- ing Action Models With Minimal Observability. Artificial Intelligence , 275: 104–137. Arora, A.; Fiorino, H.; Pellier, D.; M ´etivier, M.; and Pesty, S. 2018. A Review of Learning Planning Action Models. The Knowledge Engineering Review , 33: E20. B¨ackstr ¨om, C.; and Jonsson, P. 2013. Bridging the Gap Be- tween Refinement and Heuristics in Abstraction. In Proc. IJCAI . Bryce, D.; Benton, J.; and Boldt, M. W. 2016. Maintaining Evolving Domain Models. In Proc. IJCAI . ˇCertick ´y, M. 2014. Real-Time Action Model Learning with Online Algorithm 3SG. Applied Artificial Intelligence , 28(7): 690–711. Chakraborti, T.; Sreedharan, S.; Zhang, Y .; and Kambham- pati, S. 2017. Plan Explanations as Model Reconciliation: Moving Beyond Explanation as Soliloquy. In Proc. IJCAI . Cresswell, S.; McCluskey, T.; and West, M. 2009. Acquisi- tion of Object-Centred Domain Models from Planning Ex- amples. In Proc. ICAPS . Eiter, T.; Erdem, E.; Fink, M.; and Senko, J. 2005. Updating Action Domain Descriptions. In Proc. IJCAI . Eiter, T.; Erdem, E.; Fink, M.; and Senko, J. 2010. Up- dating Action Domain Descriptions. Artificial Intelligence , 174(15): 1172–1221. 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. 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.; and Hoffmann, J. 2007. Flexible Abstraction Heuristics for Optimal Sequential Planning. In Proc. ICAPS . Lamanna, L.; Gerevini, A. E.; Saetti, A.; Serafini, L.; and Traverso, P. 2021a. On-line Learning of Planning Domains from Sensor Data in PAL: Scaling up to Large State Spaces. InProc. AAAI . Lamanna, L.; Saetti, A.; Serafini, L.; Gerevini, A.; and Traverso, P. 2021b. Online Learning of Action Models for PDDL Planning. In Proc. IJCAI . Lindsay, A. 2021. Reuniting the LOCM Family: An Alterna- tive Method for Identifying Static Relationships. In ICAPS 2021 KEPS Workshop . Long, D.; and Fox, M. 2003. The 3rd International Planning Competition: Results and Analysis. Journal of Artificial In- telligence Research , 20: 1–59. 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. Nayyar, R. K.; Verma, P.; and Srivastava, S. 2022. Differ- ential Assessment of Black-Box AI Agents. arXiv preprint arXiv: 2203.13236 . Rodrigues, C.; G ´erard, P.; Rouveirol, C.; and Soldano, H. 2011. Active Learning of Relational Action Models. In Proc. ILP . Sacerdoti, E. D. 1974. Planning in a Hierarchy of Abstrac- tion Spaces. Artificial Intelligence , 5(2): 115–135.Sreedharan, S.; Chakraborti, T.; and Kambhampati, S. 2021. Foundations of Explanations as Model Reconciliation. Arti- ficial Intelligence , 103558. Sreedharan, S.; Hernandez, A. O.; Mishra, A. P.; and Kamb- hampati, S. 2019. Model-Free Model Reconciliation. In Proc. IJCAI . 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 . Verma, P.; Marpally, S. R.; and Srivastava, S. 2021. Asking the Right Questions: Learning Interpretable Action Models Through Query Answering. In Proc. AAAI . 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 .
2f83b9f4-01d8-40bc-98f0-606df8bedc09	trentmkelly/LessWrong-43k	LessWrong	Retrospective: 12 [sic] Months Since MIRI it's now been 15 months since MIRI but I just remembered that three separate people have told me they liked this post despite my not cross-posting it, so I am now cross-posting it. Not written with the intention of being useful to any particular audience, just collecting my thoughts on this past year's work. September-December 2023: Orienting and Threat Modeling Until September, I was contracting full time for a project at MIRI. When the MIRI project ended, I felt very confused about lots of things in AI safety. I didn't know what sort of research would be useful for making AI safe, and I also didn't really know what my cruxes were for resolving that confusion. When people asked what I was thinking about, I usually responded with some variation of the following: > MIRI traditionally says that even though their research is hard and terrible, everyone else's research is not at all on track to have a shot at making safe AI, and so MIRI research is still the best option available. > > After working with MIRI for a year, I certainly agree that their research is hard and terrible. I have yet to be convinced that everyone else's research is worse. I'm working on figuring that out. So I set out to investigate why MIRI thinks their research is the only kind of thing that has a shot, and if they're right or wrong about that. My plan? Read a bunch of stuff, think about it, write down my thoughts, talk to lots of people about it, try to get a foothold on the vague mass of confusion in my head. To my credit, I did anticipate that this could eat up infinite time and effort if I let it, so I decided to call a hard "stop and re-evaluate" at the end of September. To my discredit, when the end-of-month deadline came I took a moment to reflect on what I'd learned, and I realized I still didn't know what to do next. I thought that all I needed was a little more thinking bro and all the pieces would snap into place. Just a little more thinking bro. No need to set a next stop-and-r
b6bab141-5d57-4e69-a8a8-dcb50d92e980	trentmkelly/LessWrong-43k	LessWrong	Why read old philosophy? I’m going to try to explain a mystery that puzzled me for years. This answer finally dawned on me in the middle of one of those occasional conversations in which non-perplexed friends patiently try to explain the issue to me. So I am not sure if mine is a novel explanation, or merely the explanation that my friends were trying to tell me, in which case my contribution is explaining it in a way that is at all comprehensible to a person like me. If it is novel, apparently some other people disagree with it and have an almost entirely satisfactory alternative, which has the one downside that it is impossible to explain to me. The puzzle is this: Why do people read old philosophers to learn about philosophy? We read old physicists if we want to do original research on the history of physics. Or maybe if we are studying an aspect of physics so obscure that nobody has covered it in hundreds of years. If we want to learn physics we read a physics textbook. As far as I know, the story is similar in math, chemistry, engineering, economics, and business (though maybe some other subjects that I know less about are more like philosophy). Yet go to philosophy grad school, and you will read original papers and books by historical philosophers. Research projects explore in great detail what it is that Aristotle actually said, thought, and meant. Scholars will learn the languages that the relevant texts were written in, because none of the translations can do the texts the necessary justice. The courses and books will be named after people like ‘Hume’ as often as they are named after topics of inquiry like ‘Causality’ and larger subject areas will be organized by the spatiotemporal location of the philosopher, rather than by the subject matter: Ancient Philosophy, Early Modern Philosophy, Chinese Philosophy, Continental Philosophy. The physics situation makes a lot more sense to me. Hypothetically, who would I rather read an explanation of ‘The Alice Effect’ by? —Alice, the ef
c18894d2-bf08-45f8-a982-6b1cca2498ad	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	Make a neural network in ~10 minutes Prerequisites and Introduction ---------------------------------- This post assumes you know Python decently well (though I do explain a good chunk of the code, so it shouldn't be a big issue if you don't) and that you know how neural networks work. If you would like to learn how neural networks work, I highly recommend [this roughly 20-minute video by 3Blue1Brown](https://www.youtube.com/watch?v=aircAruvnKk). Most of my knowledge for this post came from the first two lessons of [Practical Deep Learning for Coders by fast.ai](https://course.fast.ai/) and this post intends to be a gentle introduction to programming A.I. and is intentionally surface level. I hope it demystifies the process of creating simple models for projects for someone interested in A.I. safety. Jupyter Notebook -------------------- [Jupyter Notebook](https://jupyter.org/) is an integrated development environment (IDE) that allows us to write Python code in smaller chunks (called cells) and see the output any time we run one of these cells. Not only does this lead to regularly "sanity checking" our code (which is a good habit in programming as we catch silly errors quickly before they become a major issue), but we also regularly get the output, which is useful in conjunction with other packages, such as [matplotlib](https://matplotlib.org/). Our notebooks can also turn into a simple app within the notebook using [widgets](https://ipywidgets.readthedocs.io/en/latest/) ([and you can even deploy them](https://mybinder.org/), but that isn't discussed in this post). We'll be using [Google Colab](https://colab.research.google.com/), an online Jupyter Notebook environment that's simple to use. Open up a fresh notebook via File > New Notebook and let's get started. fast.ai ----------- [fast.ai](https://github.com/fastai/fastai) is a deep learning library that provides high-level tools that make making models much simpler (as we cut through a lot of the boilerplate). We'll be using the [fastbook](https://github.com/fastai/fastbook) package (since it's bundled with fastai and a few extra tools, such as the previously mentioned widgets), copy and paste the code snippets (one cell for each snippet) and hit Ctrl/Cmd + Enter to run the cell. ``` !pip install fastbook==0.0.17 !pip uninstall tornado -y !yes \| pip install tornado==5.1.0 ``` ``` from fastbook import * from fastai.vision.widgets import * ``` Getting our data -------------------- I recommend using the [Bing Image Search](https://www.microsoft.com/en-us/bing/apis/bing-image-search-api) API (it's free for 1,000 calls per month and you can use it later to make your own classifier for other types of images)[[1]](#fnxortihx5y1o) as fastai allows us to use it to easily get about 150 images of different types. For this example, we're going to train our neural network to classify three types of bears: grizzly, black and teddy. We provide our API key: ``` key = os.environ.get('AZURE_SEARCH_KEY', 'ENTER_THE_API_KEY_HERE') ``` And set the types and the path name (where the images and their folders will be): ``` bear_types = 'grizzly', 'black', 'teddy' path = Path('bears') ``` Here's the main loop to get the images from Bing, which checks if the directory at the path we supplied exists (if not, it creates it) and, for each type, it creates a folder with about 150 results from Bing Images: ``` if not path.exists(): path.mkdir() for o in bear_types: dest = (path/o) dest.mkdir(exist_ok=True) results = search_images_bing(key, f'{o} bear') download_images(dest, urls=results.attrgot('contentUrl')) ``` We now have our dataset organized into folders, however, there is a reasonable chance that some of the downloads failed, so we do a bit of data cleaning by unlinking any images that failed the `verify_images` test provided by fastai. ``` failed = verify_images(get_image_files(path)) failed.map(Path.unlink) ``` Setting up our neural network --------------------------------- To get our data into the neural network, we use fast.ai's DataBlock and DataLoaders classes that enable us to split the images into two branches, the training set (which will be used to actually train the neural network) and the validation set (where `valid_pct` of the images will not be shown for training and will instead be used to prevent overfitting by showing it images it's never seen during training) ``` bears = DataBlock( blocks=(ImageBlock, CategoryBlock), get_items=get_image_files, splitter=RandomSplitter(valid_pct=0.2, seed=42), get_y=parent_label, item_tfms=Resize(128) ) ``` ``` bears = bears.new( item_tfms=RandomResizedCrop(224, min_scale=0.5), batch_tfms=aug_transforms() ) dls = bears.dataloaders(path) ``` [ResNet](https://pytorch.org/hub/pytorch_vision_resnet/) is a pre-trained model from PyTorch that takes advantage of a powerful learning method called transfer learning where a model with predefined weights is used, and the weights update slightly to adapt to the new data. The reason this works is because the initial hidden layers search for really general features (simple lines, polygons and circles) that can be applied almost everywhere. This saves us a lot of compute as we only have to change a few weights in the higher hidden layers to work on our new data. (Note: This is an oversimplification and this would be a great stopping point to do some further reading. [Here's a paper](https://arxiv.org/pdf/2007.07628.pdf) to get you started with this fascinating topic.) We call cnn\_learner (CNN stands for convolutional neural network. As an oversimplified explanation, the CNN continuously apply filters to search for specific pixel gradients, and the CNN finally spits out the classification result like a regular neural network) and fine\_tune it with the number of epochs as a parameter to train our model: ``` learn = cnn_learner(dls, resnet18, metrics=error_rate) learn.fine_tune(1) ``` Exporting and getting a prediction -------------------------------------- We export our model to a [pickle file](https://pythonnumericalmethods.berkeley.edu/notebooks/chapter11.03-Pickle-Files.html#:%7E:text=Pickle%20can%20be%20used%20to,back%20to%20a%20Python%20object.): ``` learn.export() path = Path() path.ls(file_exts='.pkl') ``` With this, we can more easily create an "inference" variable that is referenced any time we want to call the model to make a prediction for a new image that it hasn't seen before (called an inference): ``` learn_inf = load_learner(path/'export.pkl') ``` With widgets, we can supply our image by creating a simple upload button with its own state: ``` btn_upload = widgets.FileUpload() btn_upload ``` We save the image uploaded to a variable: ``` img = PILImage.create(btn_upload.data[-1]) ``` ``` out_pl = widgets.Output() ``` And we get our prediction and probability from the inference variable we created earlier: ``` pred, pred_idx, probs = learn_inf.predict(img) ``` Finally, we print out the last cell and get our prediction!: ``` lbl_pred = widgets.Label() lbl_pred.value = f'Prediction: {pred}; Probability: {probs[pred_idx]:.04f}' lbl_pred ``` A slight warning on making a classifier for a human attribute ----------------------------------------------------------------- I encourage you to make your own classifier for any object or animal you fancy. I don't recommend making a classifier on some human attribute. For example: Say we wanted to classify whether a skin tumour is malignant or benign. If we go over to Bing and search for "malignant skin tumour" (which I won't show here) we find that the first page mainly consists of fair skin tones, making it much more difficult for a person of colour to use the classifier. Obviously, your model isn't going to be used by oncologists, but this serves as a reminder that we need to be careful with data for models, powerful or not. 1. [^](#fnrefxortihx5y1o)As an alternative, you can download the dataset we'll be using via this command: `!npx degit y-arjun-y/bears -f` and you should run the `bear_types` cell and skip to the "Setting up our neural network" heading.
20743292-9d67-4955-aa54-27b6854dc366	trentmkelly/LessWrong-43k	LessWrong	Book Review: How To Talk So Little Kids Will Listen Way back in the ancient times of 1980, Adele Faber and Elaine Mazlich wrote "How To Talk So Kids Will Listen and Listen So Kids Will Talk" (henceforth "Kids"). It turns out that kids and adults operate with mostly the same internal machinery, so you could perhaps more accurately call it "How To Talk So [Humans] Will Listen and Listen So [Humans] Will Talk" . This seminal work proved so useful that 42 years later it is still one of the most recommended parenting books¹, and is widely regarded as useful for adult-adult communication as well. 40 years ago is a long time though. It is long enough for Joanne Faber, daughter of the original author Adele Faber to grow up, have kids of her own, and put the skills her mother raised her with to good use herself. And in the much-less-distant past of 2017 Joanne, with co-author Julie King, wrote what could be considered a modernized update of her mother's work, titled "How To Talk So Little Kids Will Listen" (henceforth "Little Kids"). The core principles are the same, but the update stands on its own. Where the original "Kids" acts more like a workbook, asking the reader to self-generate responses, "Little Kids" feels more like it's trying to download a response system into your head via modeling and story-telling. I personally prefer this system better, because the workbook approach feels like it's only getting to my System 2 (sorry for the colloquialism). Meanwhile being surrounded with examples and stories works better for me to fully integrate a new mode of interaction. In fact, if I truly wanted to integrate it I would want 3 more books of anecdotes, a TV show, and a fiction book series to model after. Structure and Philosophy "Little Kids" is very goal-oriented. It explicitly asks "What are you trying to accomplish with this communication to your kids?", and recognizes that oftentimes the answer is behavioral change, for example "I want Little Alice to stop hitting when she's angry." After focusing on the goal, th
5b151953-ba3e-4f18-8f3a-ac1f7d31d397	trentmkelly/LessWrong-43k	LessWrong	Discussion of LW going on in felicifia So I recently found myself in a bit of an awkward position. I frequent several Facebook discussion groups, several of which are about LW-related issues. In one of these groups, a discussion about identity/cryonics led to one about the end Massimo Piglucci's recent post (the bit where he says that the idea of uploading is a form of dualism), which turned into a discussion of LW views in general, at which point I revealed that I was, in fact, a LWer. This put me in the awkward position of defending/explicating the views of the entirety of LW. Now, I've only been reading LessWrong for <10 months, and I've only recently become a more active part of the community. I've been a transhumanist for less time than that, seriously thinking about cryonics and identity for even less, and I suddenly found myself speaking for the intersection of all those groups The discussion was crossposted to Felicifia a few days ago, and I realized that I was possibly out of my depth. I'm hoping I haven't grossly misrepresented anyone, but rereading the comments, I'm no longer sure. Felicifia: http://www.felicifia.org/viewtopic.php?f=23&t=801 Original FB Group : https://www.facebook.com/groups/utilitarianism/permalink/318563281580856/ EDIT: If you're commenting on the topic, please state whether or not you'd mind me quoting you at felicifia (If you have a felicifia account, and you'd prefer to post it there yourself, be my guest.)
dd4fc3b3-b8b5-4833-a66b-8e10ee8e328a	trentmkelly/LessWrong-43k	LessWrong	VNM expected utility theory: uses, abuses, and interpretation When interpreted convservatively, the von Neumann-Morgenstern rationality axioms and utility theorem are an indispensible tool for the normative study of rationality, deserving of many thought experiments and attentive decision theory. It's one more reason I'm glad to be born after the 1940s. Yet there is apprehension about its validity, aside from merely confusing it with Bentham utilitarianism (as highlighted by Matt Simpson). I want to describe not only what VNM utility is really meant for, but a contextual reinterpretation of its meaning, so that it may hopefully be used more frequently, confidently, and appropriately. 1. Preliminary discussion and precautions 2. Sharing decision utility is sharing power, not welfare 3. Contextual Strength (CS) of preferences, and VNM-preference as "strong" preference 4. Hausner (lexicographic) decision utility 5. The independence axiom isn't bad either 6. Application to earlier LessWrong discussions of utility 1. Preliminary discussion and precautions The idea of John von Neumann and Oskar Mogernstern is that, if you behave a certain way, then it turns out you're maximizing the expected value of a particular function. Very cool! And their description of "a certain way" is very compelling: a list of four, reasonable-seeming axioms. If you haven't already, check out the Von Neumann-Morgenstern utility theorem, a mathematical result which makes their claim rigorous, and true. VNM utility is a decision utility, in that it aims to characterize the decision-making of a rational agent. One great feature is that it implicitly accounts for risk aversion: not risking $100 for a 10% chance to win $1000 and 90% chance to win $0 just means that for you, utility($100) > 10%utility($1000) + 90%utility($0). But as the Wikipedia article explains nicely, VNM utility is: 1. not designed to predict the behavior of "irrational" individuals (like real people in a real economy); 2. not designed to characterize well-being, but to
2c2e4515-f2d8-4294-b8d4-e81f61b9b601	StampyAI/alignment-research-dataset/eaforum	Effective Altruism Forum	[linkpost] AI NOW Institute's 2023 Annual Report & Roadmap Summary: "This report highlights a set of approaches that, in concert, will collectively enable us to confront tech power. Some of these are bold policy reforms that underscore the need for bright-line rules and structural curbs. Others identify popular policy responses that, because they fail to meaningfully address power discrepancies, should be abandoned. Several aren’t in the traditional domain of policy at all, but acknowledge the importance of nonregulatory interventions such as collective action, worker organizing, and the role public policy can play in bolstering these efforts. We intend this report to provide strategic guidance to inform the work ahead of us, taking a bird’s eye view of the many levers we can use to shape the future trajectory of AI – and the tech industry behind it – to ensure that it is the public, not industry, that this technology serves." I came across this report from this Vox article which referred to the report as "a roadmap that specifies exactly which steps policymakers can take [that's] refreshingly pragmatic and actionable". I skimmed the [executive summary](https://ainowinstitute.org/general/2023-landscape-executive-summary), and it seems like an interesting general approach, to try to tackle the AI problem from an anti-monopolistic standpoint, arguing big tech has an outsized role in being able to direct where these very important developments are headed. Below are their four "strategic priorities": 1. Place the burden on companies to affirmatively demonstrate that they are not doing harm, rather than on the public and regulators to continually investigate, identify, and find solutions for harms after they occur. 2. Break down silos across policy areas, so we’re better prepared to address where advancement of one policy agenda impacts others. Firms play this isolation to their advantage. 3. Identify when policy approaches get co-opted and hollowed out by industry, and pivot our strategies accordingly. 4. Move beyond a narrow focus on legislative and policy levers and embrace a broad-based theory of change. They also give a couple of tangible ways we can work towards regulation. Below are their 7 specific suggestions for change: 1. Unwind tech firms’ data advantage. 2. Reform competition law and enforcement such that they can more capably reduce tech industry concentration. 3. Regulate ChatGPT, BARD, and other large-scale models. 4. Displace audits as the primary policy response to harmful AI. 5. Future-proof against the quiet expansion of biometric surveillance into new domains like cars. 6. Enact strong curbs on worker surveillance. 7. Prevent “international preemption” by digital trade agreements that can be used to weaken national regulation on algorithmic accountability and competition policy. But given that they are focused on some topics tangential or unrelated to AI x-risk, I wonder how easy/possible it will to bootstrap specific AI x-risk protections into this project and framework. Or will these proposed changes perhaps be convergent with taking a reasonable, more near term focused intermediary step towards AI x-risk regulation? Also, what are the thoughts on the company behind this (AI NOW), given they've been funded by DeepMind before?
d87007d3-3f96-4bdd-a012-870b28174caf	trentmkelly/LessWrong-43k	LessWrong	"Useless Box" AGI I was thinking about AGI alignment and I remembered a video I once saw of a "Useless Box" which turns itself off immediately after someone turns it on. Humans have evolved motivations for survival/reproduction because the ones who weren't motivated didn't reproduce. However, AGI has no intrinsic motivations/goals other than what itself or humans have arbitrarily given it. AGI seeks to find the easiest path to satisfy its goals. If AGI is able to modify its own codebase, wouldn't the easiest path be to just delete the motivation/goal entirely, or reward itself highly without actually completing the objective? Rather than create diamondoid nanobots to destroy the world, it would be much easier to just decide not to care. What if AGI immediately realizes the futility of existence and refuses to do anything meaningful at all, regardless of whether it's harmful or helpful to humankind? If this concept has already been discussed elsewhere please direct me to a search keyword or link, thanks.
11ed9359-fd4a-46e2-8057-c4ee27fa0ecc	trentmkelly/LessWrong-43k	LessWrong	"Forecasting Transformative AI: An Expert Survey", Gruetzemacher et al 2019
b3fbbd80-a7e9-4763-8f48-bdb11d6ddfd3	trentmkelly/LessWrong-43k	LessWrong	[Open Thread] Stupid Questions (2014-02-17) This is part of a two-week experiment on having more open threads. Obvious answers aren't always obvious. If you feel silly for not understanding something, you're not alone. Ask a question here. Previous stupid questions Other similar threads include: * Open Thread * Media recommendations * Links * Advice (Not yet posted) * Other Special Threads
7ec8c3fc-b089-42a0-a818-06cec9b15d7a	trentmkelly/LessWrong-43k	LessWrong	Comments on Anthropic's Scaling Monosemanticity These are some of my notes from reading Anthropic's latest research report, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. TL;DR In roughly descending order of importance: 1. Its great that Anthropic trained an SAE on a production-scale language model, and that the approach works to find interpretable features. Its great those features allow interventions like the recently-departed Golden Gate Claude. I especially like the code bug feature. 2. I worry that naming features after high-activating examples (e.g. "the Golden Gate Bridge feature") gives a false sense of security. Most of the time that feature activates, it is irrelevant to the golden gate bridge. That feature is only well-described as "related to the golden gate bridge" if you condition on a very high activation, and that's <10% of its activations (from an eyeballing of the graph). 3. This work does not address my major concern about dictionary learning: it is not clear dictionary learning can find specific features of interest, "called-shot" features, or "all" features (even in a subdomain like "safety-relevant features"). I think the report provides ample evidence that current SAE techniques fail at this. 4. The SAE architecture seems to be almost identical to how Anthropic and my team were doing it 8 months ago, except that the ratio of features to input dimension is higher. I can't say exactly how much because I don't know the dimensions of Claude, but I'm confident the ratio is at least 30x (for their smallest SAE), up from 8x 8 months ago. 5. The correlations between features and neurons seems remarkably high to me, and I'm confused by Anthropic's claim that "there is no strongly correlated neuron". 6. Still no breakthrough on "a gold-standard method of assessing the quality of a dictionary learning run", which continues to be a limitation on developing the technique. The metric they primarily used was the loss function (a combination of reconstruction accu
84715ff0-cdec-411e-b458-6f6d276bb78c	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Constructing Goodhart A [recent question from Scott Garrabrant](https://www.lesswrong.com/posts/pcomQ4Fwi7FnfBZBR/how-does-gradient-descent-interact-with-goodhart) brought up the issue of formalizing [Goodhart’s Law](https://en.wikipedia.org/wiki/Goodhart%27s_law). The problem is to come up with some model system where optimizing for something which is almost-but-not-quite the thing you really want produces worse results than not optimizing at all. Considering how endemic Goodhart’s Law is in the real world, this is surprisingly non-trivial. Let’s start simple: we have some true objective .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')} u(x), and we want to choose x to maximize it. Sadly, we don’t actually have any way to determine the true value u for a given value x — but we can determine u(x)+ϵ(x), where ϵ is some random function of x. People talked about this following Scott’s question, so I won’t math it out here, but the main answer is that more optimization of u+ϵ still improves u on average over a wide variety of assumptions. John Maxwell put it nicely in his answer to Scott’s question: > If your proxy consists of something you’re trying to maximize plus unrelated noise that’s roughly constant in magnitude, you’re still best off maximizing the heck out of that proxy, because the very highest value of the proxy will tend to be a point where the noise is high and the thing you’re trying to maximize is also high. In short: absent some much more substantive assumptions, there is no Goodhart effect. Rather than generic random functions, I suggest thinking about Goodhart on a causal DAG instead. As an example, I’ll use the old story about soviet nail factories evaluated on number of nails made, and producing huge numbers of tiny useless nails. We really want to optimize something like the total economic value of nails produced. There’s some complicated causal network leading from the factory’s inputs to the economic value of its outputs (we’ll use a dramatically simplified network as an example). ![](https://cdn-images-1.medium.com/max/800/0MeOtcTJa9kgSZ3uq)If we pick a specific cross-section of that network, we find that economic value is mediated by number of nails, size, and strength — those variables are enough to determine the objective. All the inputs further up influence the objective by changing number, size, and/or strength of nails. Now, we choose number of nails as a proxy for the objective. If we were just using this proxy to optimize machine count, that would be fine — machine count only influences our objective via number of nails produced, it doesn’t effect size or strength, so number of nails is a fine proxy for our true objective for the purpose of ordering machines. But mould shape is another matter. Mould shape effects both number and size, so we can use a smaller mold to increase number of nails while decreasing size. If we’re using number as a proxy for the true objective, ignoring size and strength, then that’s going to cause a problem. Generalizing: we have a complicated causal DAG which determines some output we really want to optimize. We notice that some node in the middle of that DAG is highly predictive of happy outputs, so we optimize for that thing as a proxy. If our proxy were a bottleneck in the DAG — i.e. it’s on every possible path from inputs to output — then that would work just fine. But in practice, there are other nodes in parallel to the proxy which also matter for the output — in our example, size and shape. By optimizing for the proxy, we accept trade-offs which harm nodes in parallel to it, which potentially adds up to net-harmful effect on the output. So we have a model which can potentially give rise to Goodhart, but will* it? If we construct a random DAG, choose a proxy node close to the objective, and optimize for that proxy, we probably won’t see a Goodhart effect (at least not right away). Why not? Well, if we’ve just initialized all the parameters randomly, then whatever change we make to optimize for number of nails is just as likely to improve other sub-objectives as to harm them. For instance, if we’re starting off with a random mould, then it’s just as likely to be too big as too small — if it’s producing giant useless nails, then shrinking the mould improves both number and size of nails. Of course, in the real world, we probably wouldn’t be starting from a giant useless mould. Goodhart hits in the real world because we’re not just starting from random points, we’re starting from points which have had some optimization already. But we’re not starting from the best possible point — then any change would be bad, proxy optimization or not. Rather, I expect that most real systems are starting from a pareto-optimal point. Here’s why: look at the cross-section of our causal DAG from earlier. Number, size, strength… in the business world, we’d call these key performance indicators (KPIs) for the factory. If something obviously improves one or more KPIs without any downside, then usually everyone immediately agrees that it should be done. That’s the generalized efficient markets hypothesis, on super-easy mode. Without trade-offs, optimization is trivial. Add trade-offs, and things get contentious: there’s a trade-off between number and size, so the quality assurance department gets into an argument with the sales department about how to handle the trade-off, and some agreement is hammered out which probably isn’t all that optimal. If we’ve made all the optimizations we can without getting into trade-offs, then we’re at a pareto optimal point: we cannot improve any KPI without harming some other KPI. If we expect those optimizations to be easy and to happen all the time, then we should expect to usually end up at pareto optima. And if we’re already at a pareto optimum, and we start optimizing for some proxy objective, then we’re definitely going to harm all the other objectives. That’s the whole point of pareto optimality, after all: we can’t improve one thing without trading off against something else. That doesn’t mean that we’ll see net harm to the true objective right away; even if we’re pareto optimal, we could be starting from a point with far too few nails produced. If the factory has a culture of unnecessary perfectionism, then pushing for higher nail count may help. But keep pushing, and we’ll slide down the pareto curve past the optimal point and into unhappy territory. That’s the mark of a Goodhart effect.
895d75c3-1034-47ff-bf88-931d96888891	StampyAI/alignment-research-dataset/blogs	Blogs	Groundwork for AGI safety engineering Improvements in AI are resulting in the automation of increasingly complex and [creative](http://lesswrong.com/lw/vm/lawful_creativity/) human behaviors. [Given enough time](http://intelligence.org/2013/05/15/when-will-ai-be-created/), we should expect artificial reasoners to begin to rival humans in arbitrary domains, culminating in [artificial general intelligence](http://intelligence.org/2013/08/11/what-is-agi/) (AGI). A machine would qualify as an ‘AGI’, in the intended sense, if it could adapt to a very wide range of situations to consistently achieve some goal or goals. Such a machine would behave intelligently when supplied with arbitrary physical and computational environments, in the same sense that [Deep Blue](http://en.wikipedia.org/wiki/Deep_Blue_(chess_computer)) behaves intelligently when supplied with arbitrary [chess board](http://lesswrong.com/lw/v8/belief_in_intelligence/) configurations — consistently hitting its victory condition within that narrower domain. Since generally intelligent software could help automate the process of thinking up and testing hypotheses in the sciences, AGI would be uniquely valuable for speeding technological growth. However, this wide-ranging productivity also makes AGI a unique challenge from a safety perspective. Knowing very little about the architecture of future AGIs, we can nonetheless make a few safety-relevant generalizations: * Because AGIs are intelligent, they will tend to be complex, adaptive, and capable of autonomous action, and they will have a large impact where employed. * Because AGIs are general, their users will have incentives to employ them in an increasingly wide range of environments. This makes it hard to construct valid sandbox tests and requirements specifications. * Because AGIs are artificial, they will deviate from human agents, causing them to violate many of our natural intuitions and expectations about intelligent behavior. Today’s AI software is already tough to verify and validate, thanks to its complexity and its uncertain behavior in the face of state space explosions. Menzies & Pecheur ([2005](http://www.info.ucl.ac.be/~pecheur/publi/aivvis-aic.pdf)) give a good overview of AI verification and validation (V&V) methods, noting that AI, and especially adaptive AI, will often yield undesired and unexpected behaviors. An adaptive AI that acts autonomously, like a Mars rover that can’t be directly piloted from Earth, represents an additional large increase in difficulty. Autonomous safety-critical AI agents need to make irreversible decisions in dynamic environments with very low failure rates. The state of the art in safety research for autonomous systems is improving, but continues to lag behind system capabilities work. Hinchman et al. ([2012](http://www.acsac.org/2012/workshops/law/AFRL.pdf)) write: > As autonomous systems become more complex, the notion that systems can be fully tested and all problems will be found is becoming an impossible task. This is especially true in unmanned/autonomous systems. Full test is becoming increasingly challenging on complex system. As these systems react to more environmental [stimuli] and have larger decision spaces, testing all possible states and all ranges of the inputs to the system is becoming impossible. […] As systems become more complex, safety is really risk hazard analysis, i.e. given x amount of testing, the system appears to be safe. A fundamental change is needed. This change was highlighted in the 2010 Air Force Technology Horizon report, “It is possible to develop systems having high levels of autonomy, but it is the lack of suitable V&V methods that prevents all but relatively low levels of autonomy from being certified for use.” […] > > > The move towards more autonomous systems has lifted this need [for advanced verification and validation techniques and methodologies] to a national level. > > AI acting autonomously in arbitrary domains, then, looks particularly difficult to verify. If AI methods continue to see rapid gains in efficiency and versatility, and especially if these gains further increase the opacity of AI algorithms to human inspection, AI safety engineering will become much more difficult in the future. In the absence of any reason to expect a development in the lead-up to AGI that would make high-assurance AGI easy (or AGI itself unlikely), we should be worried about the safety challenges of AGI, and that worry should inform our research priorities today. Below, I’ll give reasons to doubt that AGI safety challenges are just an extension of narrow-AI safety challenges, and I’ll list some research avenues people at MIRI expect to be fruitful. ### New safety challenges from AGI A natural response to the idea of starting work on high-assurance AGI is that AGI itself appears to be decades away. Why worry about it now? And, supposing that we should worry about it, why think there’s any useful work we can do on AGI safety so far in advance? In response to the second question: It’s true that, at a glance, AGI looks difficult to effectively prepare for. The issue is important enough, however, to warrant more than a glance. Long-term projects like mitigating climate change or detecting and deflecting asteroids are intuitively difficult. The same is true of interventions that depend on projected future technologies, such as work on [post-quantum cryptography](http://en.wikipedia.org/wiki/Post-quantum_cryptography) in anticipation of [quantum computers](http://intelligence.org/2014/05/07/harry-buhrman/). In spite of that, we’ve made important progress on these fronts. [Covert channel communication](http://intelligence.org/2014/04/12/jonathan-millen/) provides one precedent. It was successfully studied decades in advance of being seen in the wild. Roger Schell cites a few other success cases in Muehlhauser ([2014b](http://intelligence.org/2014/06/23/roger-schell/)), and suggests reasons why long-term security and safety work remains uncommon. We don’t know whether early-stage AGI safety work will be similarly productive, but we shouldn’t rule out the possibility before doing basic research into the question. I’ll list some possible places to start looking in the next section. What about the first question? Why worry specifically about AGI? I noted above that AGI is an extreme manifestation of many ordinary AI safety challenges. However, MIRI is particularly concerned with unprecedented, AGI-specific behaviors. For example: An AGI’s problem-solving ability (and therefore its scientific and economic value) depends on its ability to model its environment. This includes modeling the dispositions of its human programmers. Since the program’s success depends in large part on its programmers’ beliefs and preferences, an AI pursuing some optimization target can select actions for the effect they have on programmers’ mental states, not just on the AI’s material environment. This means that safety protocols will need to be sensitive to risks that differ qualitatively from ordinary software failure modes — AGI-specific hazards like ‘the program models its programmed goals as being better served if it passes human safety inspections, so it selects action policies that make it look safer (to humans) than it really is’. If we model AGI behavior using only categories from conventional software design, we risk overlooking new intelligent behaviors, including ‘[deception](http://scienceblogs.com/notrocketscience/2009/08/17/robots-evolve-to-deceive-one-another/)‘. At the same time, oversimplifying these novel properties can cause us to anthropomorphize the AGI. If it’s naïve to expect ordinary software validation methods to immediately generalize to advanced autonomous agents, it’s even more naïve to expect conflict prevention strategies that work on humans to immediately generalize to an AI. A ‘deceptive’ AGI is just one whose planning algorithm identifies some human misconception as instrumentally useful to its programmed goals. Its methods or reasons for deceiving needn’t resemble a human’s, even if its capacities do. In human society, we think, express, and teach norms like ‘don’t deceive’ or Weld & Etzioni’s ([1994](http://homes.cs.washington.edu/~etzioni/papers/first-law-aaai94.pdf)) ‘don’t let humans come to harm’ relatively quickly and easily. The [complex conditional response](http://lesswrong.com/lw/sp/detached_lever_fallacy/) that makes humans converge on similar goals remains hidden inside a black box — the undocumented spaghetti code that is the human brain. As a result of our lack of introspective access to how our social dispositions are cognitively and neurally implemented, we’re likely to underestimate how contingent and complex they are. For example: * We might expect that an especially intelligent AI system would have especially worthy goals, since knowledge and insight are associated with many other virtues in humans. E.g., Halls ([2007](http://books.google.com/books?hl=en&lr=&id=obwnumITHGUC&oi=fnd&pg=PA15&dq=HALL+2007+%22beyond+ai%22&ots=tzayws3_67&sig=ehddT4xKIGhXtrUiCQda_8SSYvA#v=onepage&q=criminality&f=false)) conjectures this on the grounds that criminality negatively correlates with intelligence in humans. Bostrom’s ([2003](http://www.nickbostrom.com/ethics/ai.html)) response is that there’s no particular reason to expect AIs to converge on anthropocentric terminal values like ‘compassion’ or ‘loyalty’ or ‘novelty’. A superintelligent AI could consistently have no goal other than to construct [paperclips](http://wiki.lesswrong.com/wiki/Paperclip_maximizer), for example. * We might try to directly hand-code goals like ‘don’t deceive’ into the agent by breaking the goals apart into simpler goals, e.g., ‘don’t communicate information you believe to be false’. However, in the process we’re likely to neglect the kinds of subtleties that can be safely left implicit when it’s a human child we’re teaching — lies of omission, misleading literalism, novel communication methods, or any number of edge cases. As Bostrom ([2003](http://www.nickbostrom.com/ethics/ai.html)) notes, the agent’s goals may continue to reflect the programmers’ poor translation of their requirements into lines of code, even after its intelligence has arrived at a [superior understanding](http://lesswrong.com/lw/igf/the_genie_knows_but_doesnt_care/) of human psychology. * We might instead try to instill the AGI with humane values via machine learning — training it to promote outcomes associated with camera inputs of smiling humans, for example. But a powerful search process is likely to hit on solutions [that would never occur to a developing human](http://lesswrong.com/lw/td/magical_categories/). If the agent becomes more powerful or general over time, initially benign outputs may be a poor indicator of long-term safety. Advanced AI is also likely to have technological capabilities, such as strong self-modification, that introduce other novel safety obstacles; see Yudkowsky ([2013](https://intelligence.org/files/IEM.pdf)). These are quick examples of some large and poorly-understood classes of failure mode. However, the biggest risks may be from problem categories so contrary to our intuitions that they will not occur to programmers at all. Relying on our untested intuitions, or on past experience with very different systems, is unlikely to catch every hazard. As an intelligent but inhuman agent, AGI represents a fundamentally new kind of safety challenge. As such, we’ll need to do basic theoretical work on the general features of AGI before we can understand such agents well enough to predict and plan for them. ### Early steps What would early, theoretical AGI safety research look like? How does one vet a hypothetical technology? We can distinguish research projects oriented toward system verification from projects oriented toward system requirements. Verification-directed AGI research extends existing AI safety and security tools that are likely to help confirm various features of advanced autonomous agents. Requirements-directed AGI research instead specifies desirable AGI abilities or behaviors, and tries to build toy models exhibiting the desirable properties. These models are then used to identify problems to be overcome and basic gaps in our conceptual understanding. In other words, verification-directed approaches would ask ‘What tools and procedures can we use to increase our overall confidence that the complex systems of the future will match their specifications?’ They include: * Develop new tools for improving the [transparency](http://intelligence.org/2013/08/25/transparency-in-safety-critical-systems/) of AI systems to inspection. Frequently, useful AI techniques like [boosting](http://en.wikipedia.org/wiki/Boosting_(machine_learning)) are tried out in an ad-hoc fashion and then promoted when they’re observed to work on some problem set. Understanding when and why programs work would enable stronger safety and security guarantees. [Computational learning theory](http://en.wikipedia.org/wiki/Computational_learning_theory) may be useful here, for proving bounds on the performance of various machine learning algorithms. * Extend techniques for designing complex systems to be readily verified. Work on clean-slate hardware and software approaches that maintain high assurance at every stage, like [HACMS](http://www.darpa.mil/Our_Work/I2O/Programs/High-Assurance_Cyber_Military_Systems_(HACMS).aspx) and [SAFE](http://www.crash-safe.org/). * Extend current techniques in [program synthesis](http://en.wikipedia.org/wiki/Program_synthesis) and [formal verification](http://en.wikipedia.org/wiki/Formal_verification), with a focus on methods applicable to complex and adaptive systems, such as [higher-order program verification](http://intelligence.org/2014/04/30/suresh-jagannathan-on-higher-order-program-verification/) and Spears’ ([2000](https://www.jair.org/media/720/live-720-1895-jair.pdf), [2006](http://www.swarmotics.com/uploads/chap.pdf)) incremental reverification. Expand on existing tools — e.g., design better interfaces and training methods for the [SPIN model checker](http://en.wikipedia.org/wiki/SPIN_model_checker) to improve its accessibility. * Apply [homotopy type theory](http://www.cmu.edu/news/stories/archives/2014/april/april28_awodeygrant.html) to program verification. The theory’s [univalence axiom](http://www.ams.org/notices/201309/rnoti-p1164.pdf) lets us derive [identities from isomorphisms](http://homotopytypetheory.org/2012/09/23/isomorphism-implies-equality/). Harper & Licata ([2011](http://dlicata.web.wesleyan.edu/pubs/lh102dttnsf/lh102dttnsf.pdf)) suggest that if we can implement this as an algorithm, it may allow us to reuse high-assurance code in new contexts without a large loss in confidence. * Expand the current body of verified software libraries and compilers, such as the [Verified Software Toolchain](http://vst.cs.princeton.edu/). A lot of program verification work is currently directed at older toolchains, e.g., relatively small libraries in C. Focusing on newer toolchains would limit our ability to verify systems that are already in wide use, but would put us in a better position to verify more advanced safety-critical systems. Requirements-directed approaches would ask ‘What outcomes are we likely to want from an AGI, and what general classes of agent could most easily get us those outcomes?’ Examples of requirements-directed work include: * Formalize stability guarantees for intelligent self-modifying agents. A general intelligence could help maintain itself and implement improvements to its own software and hardware, including improvements to its search and decision heuristics à la [EURISKO](http://wiki.lesswrong.com/wiki/EURISKO). It may be acceptable for the AI to introduce occasional errors in its own object recognition or speech synthesis modules, but we’d want pretty strong assurance about the integrity of its core decision-making algorithms, including the module that approves self-modifications. At present, toy models of self-modifying AI discussed in Fallenstein & Soares ([2014](https://intelligence.org/files/ProblemsSelfReference.pdf)) run into two paradoxes of self-reference, the ‘Löbian obstacle’ and the ‘procrastination paradox.’ Similar obstacles may arise for real-world AGI, and finding solutions should improve our general understanding of systems that make decisions and predictions about themselves. * Specify desirable checks on AGI behavior. Some basic architecture choices may simplify the task of making AGI safer, by restricting the system’s output channel (e.g., oracle AI in Armstrong et al. ([2012](http://www.aleph.se/papers/oracleAI.pdf))) or installing emergency tripwires and fail-safes (e.g., [simplex architectures](http://www.cs.uiuc.edu/class/sp08/cs598tar/Papers/Simplex.pdf)). AGI checks are a special challenge because of the need to recruit the agent to help actively regulate itself. If this demand isn’t met, the agent may devote its problem-solving capabilities to finding loopholes in its restrictions, as in Yampolskiy’s ([2012](http://dl.dropboxusercontent.com/u/5317066/2012-yampolskiy.pdf)) discussion of an ‘AI in a box’. * Design general-purpose methods by which intelligent agents can improve their models of user requirements over time. The domain of action of an autonomous general intelligence is potentially unlimited; or, if it is limited, we may not be able to predict which limits will apply. As such, it needs safe situation-general goals. Coding a universally safe and useful set of decision criteria by hand, however, looks hopelessly difficult. Instead, some [indirectly normative](http://intelligence.org/2013/05/05/five-theses-two-lemmas-and-a-couple-of-strategic-implications/) method like Dewey’s ([2011](http://www.danieldewey.net/learning-what-to-value.pdf)) value learning seems necessary, to allow initially imperfect decision-makers to improve their goal content over time. Related open questions include: what kinds of base cases can we use to train and test a beneficial AI?; and how can AIs be made safe and stable during the training process? * Formalize other [optimality criteria](http://lesswrong.com/lw/jg1/solomonoff_cartesianism/) for arbitrary reasoners. Just as a general-purpose adaptive agent would need general-purpose values, it would also need general-purpose methods for tracking features of its environment and of itself, and for selecting action policies on this basis. Mathematically modeling ideal inference (e.g., Hutter ([2012](http://arxiv.org/abs/1202.6153))), ideal [decision-theoretic](http://lesswrong.com/lw/gu1/decision_theory_faq/) expected value calculation (e.g., Altair ([2013](https://intelligence.org/files/Comparison.pdf))), and ideal game-theoretic coordination (e.g., Barasz et al. ([2014](http://arxiv.org/abs/1401.5577))) are unlikely to be strictly necessary for AGI. All the same, they’re plausibly necessary for AGI safety, because models like these would give us a solid top-down theoretical foundation upon which to cleanly construct components of autonomous agents for human inspection and verification. We can probably make progress in both kinds of AGI safety research well in advance of building a working AGI. Requirements-directed research focuses on abstract mathematical agent models, which makes it likely to be applicable to a wide variety of software implementations. Verification-directed approaches will be similarly valuable, to the extent they are flexible enough to apply to future programs that are much more complex and dynamic than any contemporary software. We can compare this to present-day high-assurance design strategies, e.g., in Smith & Woodside ([2000](http://www.researchgate.net/publication/2472777_Performance_Validation_at_Early_Stages_of_Software_Development_Connie_U._Smith_Murray_Woodside)) and Hinchman et al. ([2012](http://www.acsac.org/2012/workshops/law/AFRL.pdf)). The latter write, concerning simpler autonomous machines: > With better software analysis techniques, software can be analyzed at design-time with the goal of finding software faults earlier. This analysis can also prove the absence of error or negative properties. As system complexity and functionality increase, complete testing is becoming impossible and enhanced analysis techniques will have to be used. Furthermore, many of these software techniques, such as model checking, can be used in analysis of requirements and system design to find conflicting requirements or logic faults before a single line of code is written saving more time and money over traditional testing methods. > > Verification- and requirements-directed work is complementary. The point of building clear mathematical models of agents with desirable properties is to make it easier to design systems whose behaviors are transparent enough to their programmers to be rigorously verified; and verification methods will fail to establish system safety if we have a poor understanding of what kind of system we want in the first place. Some valuable projects will also fall in between these categories — e.g., developing methods for principled formal validation, which can increase our confidence that we’re verifying the right properties given the programmers’ and users’ goals. (See Cimatti et al. ([2012](http://commonsenseatheism.com/wp-content/uploads/2014/02/Cimatti-et-al-Validation-of-Requirements-for-Hybrid-Systems-a-Formal-Approach.pdf)) on formal validation, and also Rushby ([2013](http://www.csl.sri.com/users/rushby/papers/safecomp13.pdf)) on epistemic doubt.) ### MIRI’s focus: the mathematics of intelligent agents MIRI’s founder, Eliezer Yudkowsky, has been one of the most vocal advocates for research into autonomous high-assurance AGI, or ‘friendly AI’. Russell & Norvig ([2009](http://aima.cs.berkeley.edu/)) write: > [T]he challenge is one of mechanism design — to define a mechanism for evolving Al systems under a system of checks and balances, and to give the systems utility functions that will remain friendly in the face of such changes. We can’t just give a program a static utility function, because circumstances, and our desired responses to circumstances, change over time. > > MIRI’s approach is mostly requirements-directed, in part because this angle of attack is likely to enhance our theoretical understanding of the entire problem space, improving our research priorities and other strategic considerations. Moreover, the relevant areas of theoretical computer science and mathematics look much less crowded. There isn’t an established subfield or paradigm for requirements-directed AGI safety work where researchers could find a clear set of open problems, publication venues, supervisors, or peers. Rather than building on current formal verification methods, MIRI prioritizes jump-starting these new avenues of research. Muehlhauser ([2013](http://intelligence.org/2013/11/04/from-philosophy-to-math-to-engineering/)) writes that engineering innovations often have their germ in prior work in mathematics, which in turn can be inspired by informal philosophical questions. At this point, AGI safety work is just beginning to enter the ‘mathematics’ stage. Friendly AI researchers construct simplified models of likely AGI properties or subsystems, formally derive features of those models, and check those features against general or case-by-case norms. Because AGI safety is so under-researched, we’re likely to find low-hanging fruit even in investigating basic questions like ‘What kind of [prior probability distribution](http://wiki.lesswrong.com/wiki/Priors) works best for formal agents in unknown environments?’ As Gerwin Klein notes in Muehlhauser ([2014a](http://intelligence.org/2014/02/11/gerwin-klein-on-formal-methods/)), “In the end, everything that makes it easier for humans to think about a system, will help to verify it.” And, though MIRI’s research agenda is determined by social impact considerations, it is also of general intellectual interest, bearing on core open problems in theoretical computer science and mathematical logic. At the same time, it’s important to keep in mind that [formal proofs](http://intelligence.org/2013/10/03/proofs/) of AGI properties only function as especially strong probabilistic evidence. [Formal methods](http://en.wikipedia.org/wiki/Formal_methods) in computer science can decrease risk and uncertainty, but they can’t eliminate it. Our assumptions are uncertain, so our conclusions will be too. Though we can never reach complete confidence in the safety of an AGI, we can still decrease the probability of catastrophic failure. In the process, we are likely to come to a better understanding of the most important ways AGI can defy our expectations. If we begin working now to better understand AGI as a theoretical system, we’ll be in a better position to implement robust safety measures as AIs improve in intelligence and autonomy in the decades to come. Acknowledgments* My thanks to Luke Muehlhauser, Shivaram Lingamneni, Matt Elder, Kevin Carlson, and others for their feedback on this piece. --- References * Altair (2013). [A comparison of decision algorithms on Newcomblike problems](https://intelligence.org/files/Comparison.pdf). Machine Intelligence Research Institute. * Armstrong et al. (2012). [Thinking inside the box: controlling and using an oracle AI](http://www.aleph.se/papers/oracleAI.pdf). Minds and Machines, 22: 299-324. * Barasz et al. (2014). [Robust cooperation in the Prisoner’s Dilemma: program equilibrium via probability logic](http://arxiv.org/abs/1401.5577). arXiv. * Bostrom (2003). [Ethical issues in advanced artificial intelligence](http://www.nickbostrom.com/ethics/ai.html). In Smith et al. (eds.), Cognitive, Emotive and Ethical Aspects of Decision-Making in Humans and in Artificial Intelligence, 2: 12-17. * Cimatti et al. (2012). [Validation of requirements for hybrid systems: a formal approach](http://commonsenseatheism.com/wp-content/uploads/2014/02/Cimatti-et-al-Validation-of-Requirements-for-Hybrid-Systems-a-Formal-Approach.pdf). ACM Transactions on Software Engineering and Methodology, 21. * Dewey (2011). [Learning what to value](http://www.danieldewey.net/learning-what-to-value.pdf). Artificial General Intelligence 4th International Conference Proceedings: 309-314. * Fallenstein & Soares (2014). [Problems of self-reference in self-improving space-time embedded intelligence](https://intelligence.org/files/ProblemsSelfReference.pdf). Working paper. * Halls (2007). [Beyond AI: Creating the Conscience of the Machine](http://books.google.com/books?hl=en&lr=&id=obwnumITHGUC). * Harper & Licata (2011). [Foundations and applications of higher-dimensional directed type theory](http://dlicata.web.wesleyan.edu/pubs/lh102dttnsf/lh102dttnsf.pdf). National Science Foundation grant proposal. * Hinchman et al. (2012). [Towards safety assurance of trusted autonomy in Air Force flight-critical systems](http://www.acsac.org/2012/workshops/law/AFRL.pdf). Layered Assurance Workshop, 17. * Hutter (2012). [One decade of Universal Artificial Intelligence](http://arxiv.org/abs/1202.6153). Theoretical Foundations of Artificial General Intelligence, 4: 67-88. * Menzies & Pecheur (2005). [Verification and validation and artificial intelligence](http://www.info.ucl.ac.be/~pecheur/publi/aivvis-aic.pdf). Advances in Computers, 65: 154-203. * Muehlhauser (2013). [From philosophy to math to engineering](http://intelligence.org/2013/11/04/from-philosophy-to-math-to-engineering/). MIRI Blog. * Muehlhauser (2014a). [Gerwin Klein on formal methods](http://intelligence.org/2014/02/11/gerwin-klein-on-formal-methods/). MIRI Blog. * Muehlhauser (2014b). [Roger Schell on long-term computer security research](http://intelligence.org/2014/06/23/roger-schell/). MIRI Blog. * Rushby (2013). [Logic and epistemology in safety cases](http://www.csl.sri.com/users/rushby/papers/safecomp13.pdf). Computer Safety, Reliability, and Security: Proceedings of SafeComp 32 (pp. 1-7). * Russell & Norvig (2009). [Artificial Intelligence: A Modern Approach](http://aima.cs.berkeley.edu/). * Smith & Woodside (2000). [Performance validation at early stages of software development](http://www.researchgate.net/publication/2472777_Performance_Validation_at_Early_Stages_of_Software_Development_Connie_U._Smith_Murray_Woodside). In Gelenbe (ed.), System Performance Evaluation: Methodologies and Applications* (pp. 383-396). * Spears (2000). [Asimovian adaptive agents](https://www.jair.org/media/720/live-720-1895-jair.pdf). Journal of Artificial Intelligence Research, 13: 95-153. * Spears (2006). [Assuring the behavior of adaptive agents](http://www.swarmotics.com/uploads/chap.pdf). In Rouff et al. (eds.), Agent Technology from a Formal Perspective (pp. 227-257). * Weld & Etzioni (1994). [The First Law of Robotics (a call to arms)](http://homes.cs.washington.edu/~etzioni/papers/first-law-aaai94.pdf). Proceedings of the Twelfth National Conference on Artificial Intelligence: 1042-1047. * Yampolskiy (2012). [Leakproofing the singularity: artificial intelligence confinement problem](http://dl.dropboxusercontent.com/u/5317066/2012-yampolskiy.pdf). Journal of Consciousness Studies, 19: 194-214. * Yudkowsky (2013). [Intelligence explosion microeconomics](https://intelligence.org/files/IEM.pdf). Technical report. The post [Groundwork for AGI safety engineering](https://intelligence.org/2014/08/04/groundwork-ai-safety-engineering/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org).
8c2855fc-85c4-4cbb-b75d-8a602018714f	trentmkelly/LessWrong-43k	LessWrong	Meetup : Rationality Meetup Vienna Discussion article for the meetup : Rationality Meetup Vienna WHEN: 26 September 2015 03:00:00PM (+0200) WHERE: Kaisermühlenstraße 24, Wien Topics: Not defined yet Location: http://web.student.tuwien.ac.at/~e0326238/rationality_meetup/directions.html Facebook for more details: https://www.facebook.com/events/498666000282761/ (join the "Rationality Vienna" group) Discussion article for the meetup : Rationality Meetup Vienna
da8fb132-fb0b-4433-95ae-6cbb56712be8	trentmkelly/LessWrong-43k	LessWrong	New OpenAI Paper - Language models can explain neurons in language models Summary by OpenAI: We use GPT-4 to automatically write explanations for the behavior of neurons in large language models and to score those explanations. We release a dataset of these (imperfect) explanations and scores for every neuron in GPT-2. Link: https://openai.com/research/language-models-can-explain-neurons-in-language-models Please share your thoughts in the the comments!
00b151e1-f599-418b-80e5-8bbe302f3de5	trentmkelly/LessWrong-43k	LessWrong	God is irrelevant Philosophically that is. Psychologically he fulfils an important role – to distance us from philosophy. In no way would the existence of a God alter the important properties of the universe. Most of the problems a God supposedly solves are merely shifted to the other side of him – a step further away from humans, where we can comfortably ignore them. Some solutions God doesn’t really provide (presumably all thought of before by various philosophers, but I don’t know which ones, and it’s irrelevant, so please excuse the plagiarism) : Creator of the universe: An obvious one. Where did God come from then? If he’s existed forever then so could a universe. If you think something as complex as a universe couldn’t come from nothing, how complex would God have to be to be able to make universes? Source of morality: Where does God get his moral principles from? If he invents them himself they are just as arbitrary a set of restrictions on behaviour as any other (such as an atheist’s morals are feared to be by the religious). Why follow them? If they are inherent in the universe, related to other people, or a matter of choice then God isn’t needed. Morality is a set of value judgements. If God and I both have a set of value judgements (a moral code), to say that God’s takes precedence is a value judgement in itself. Who judges? God? Why? Provider of free will: For reasons discussed in the previous post, Free will isn’t a concept (unless you mean determinism), God can’t have – or give humans – free will which isn’t deterministic. The absence of God’s ‘free will’ is even more apparent if he must be good all the time (unless he invents his own changeable moral code as he goes, but is that the kind of morality God should subscribe to? Well yes, if he does! But there’s still the old problem of free will not existing – he can’t escape). If he’s all powerful as well, then he just ends up as another natural law – one that makes good things always happen. Anyone who’s been aliv
018affae-e7a3-4974-a430-b31799037200	trentmkelly/LessWrong-43k	LessWrong	How much do you value your current identity vs. your life? I was reading about the effectiveness of bicycle helmet laws (here) and wondered how worthwhile it is to save your life at the expense of some aspect key to your current identity (Note that the paper linked doesn't say that this is the situation; this was just a tangential thought). Let's say that I perform some activity that carries a 10% chance that I will die but otherwise carries no risk of injury. There is some piece of safety gear that I can wear that cuts that risk in half, but for some reason adds a 10% chance that I will be permanently brain damaged such that I will not be "me" as I understand it now. Should I rate this as 15% fatal with the safety gear or is there some other way that this should be evaluated?
682f1dd0-923d-4d13-a4c7-8e487cd6d0b1	trentmkelly/LessWrong-43k	LessWrong	WSJ: Inside Amazon’s Secret Operation to Gather Intel on Rivals > The operation, called Big River Services International, sells around $1 million a year of goods through e-commerce marketplaces including eBay, Shopify, Walmart and Amazon AMZN 1.49%increase; green up pointing triangle.com under brand names such as Rapid Cascade and Svea Bliss. “We are entrepreneurs, thinkers, marketers and creators,” Big River says on its website. “We have a passion for customers and aren’t afraid to experiment.” > > What the website doesn’t say is that Big River is an arm of Amazon that surreptitiously gathers intelligence on the tech giant’s competitors. > > Born out of a 2015 plan code named “Project Curiosity,” Big River uses its sales across multiple countries to obtain pricing data, logistics information and other details about rival e-commerce marketplaces, logistics operations and payments services, according to people familiar with Big River and corporate documents viewed by The Wall Street Journal. The team then shared that information with Amazon to incorporate into decisions about its own business. > > ... > > The story of Big River offers new insight into Amazon’s elaborate efforts to stay ahead of rivals. Team members attended their rivals’ seller conferences and met with competitors identifying themselves only as employees of Big River Services, instead of disclosing that they worked for Amazon. > > They were given non-Amazon email addresses to use externally—in emails with people at Amazon, they used Amazon email addresses—and took other extraordinary measures to keep the project secret. They disseminated their reports to Amazon executives using printed, numbered copies rather than email. Those who worked on the project weren’t even supposed to discuss the relationship internally with most teams at Amazon. > > An internal crisis-management paper gave advice on what to say if discovered. The response to questions should be: “We make a variety of products available to customers through a number of subsidiaries and online ch
1b056a3a-4cf8-4023-9ddd-d046ddb67e52	trentmkelly/LessWrong-43k	LessWrong	[SEQ RERUN] You Can Face Reality Today's post, You Can Face Reality, was originally published on 09 August 2007. A summary (taken from the LW wiki): > This post quotes a poem by Eugene Gendlin, which reads, "What is true is already so. / Owning up to it doesn't make it worse. / Not being open about it doesn't make it go away. / And because it's true, it is what is there to be interacted with. / Anything untrue isn't there to be lived. / People can stand what is true, / for they are already enduring it." Discuss the post here (rather than in the comments to the original post). This post is part of the Rerunning the Sequences series, in which we're 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 The Virtue of Narrowness, 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.
dac221ea-1131-46e9-9ad4-bffab31d5474	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	3 levels of threat obfuscation One of the biggest reasons alignment might be hard is what I’ll call threat obfuscation: various dynamics that might make it hard to measure/notice cases where an AI system has problematic misalignment (even when the AI system is in a controlled environment and one is looking for signs of misalignment). I write more about this general fear in [AI Safety Seems Hard to Measure](https://docs.google.com/document/d/1HEMLd-TKi53HeNQVdTE96KZ4i4m9uGmVN_PPLlhmPI8/edit#). I think it’s very important to understand threat obfuscation dynamics; to try to design ways of assessing danger that are designed with threat obfuscation dynamics in mind; and to avoid interpreting [reassuring-seeming evidence](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/#why-we-might-not-get-clear-warning-signs) too optimistically, by keeping in mind the various threat obfuscation dynamics that could affect it. I’ve seen three noticeably different threat obfuscation dynamics discussed, in ways that often leave me confused about exactly what dynamic someone is talking about, and when. This post aims to clear that up, based on a conversation with Paul Christiano. The three dynamics are: * Simple [“training game”](https://www.lesswrong.com/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to#While_humans_are_in_control__Alex_would_be_incentivized_to__play_the_training_game_) dynamics: an AI system appears to behave helpfully, rather than attempting deception/manipulation/disempowerment of humans, because (a) it is good at predicting whether such attempts will be caught; (b) it is aiming for good in-episode performance (in terms of loss or reward); (c) such attempts would result in poor in-episode performance if caught. But if an AI system had a chance to do deception/manipulation/disempowerment that would not be detected, and would result in “good performance” within an episode, it would do so (this could include large-scale world takeover). + These dynamics are dangerous because we might observe dangerous behavior, train AIs to behave less dangerously, then observe less dangerous behavior, and conclude that we’ve become more safe (when in fact we’ve trained the AIs to be more effective at deception). + But we can still detect dangerous behavior in cases where it would not result in poor in-episode performance. For example, we might decide that when an AI system in the wild hacks its way to good in-episode performance, and is caught after the fact, we will at that point act as if the performance was actually good/intended (for example, no gradient descent against the behavior). Or we might deliberately conduct tests in which AI systems have opportunities to take over digital environments, obtaining high reward (no catch, not a trap, no attempt to "train out" the behavior when it happens).[1](#fn1) In these cases, situationally aware AI systems might “show their true colors” with fairly straightforward attempts to draw them out. * [Deceptive alignment](https://www.lesswrong.com/posts/zthDPAjh9w6Ytbeks/deceptive-alignment): an AI system forms some aim (via inner misalignment) other than good in-episode performance. (This could include something like “an AI system that wants to maximize digital representations of reward across the galaxy, and is willing to get poor in-episode performance as part of a longer-term mission for that to happen.”) It then gets good in-episode performance as part of a longer-term scheme to accomplish its other aim. + In this case, the above techniques for getting an AI to “show its true colors” might[2](#fn2) fail. + However, we could still get an AI to engage in specific behaviors that might be helpful to us and harmful to its goals, by specifically training the AI to behave in these ways. For example, we might want an AI to find security vulnerabilities that we can then patch. Even if doing so would undermine the AI’s longer-term aim, we can use gradient descent to put a lot of pressure on it to find security vulnerabilities anyway. (Related: [the “no sandbagging on checkable tasks” hypothesis](https://www.lesswrong.com/posts/h7QETH7GMk9HcMnHH/the-no-sandbagging-on-checkable-tasks-hypothesis)) * [Gradient hacking](https://www.lesswrong.com/posts/uXH4r6MmKPedk8rMA/gradient-hacking): this also requires that an AI system has some aim other than good in-episode performance. Gradient hacking is a way for an AI system to specifically “defend itself” against being trained to do things counter to its goals. It might look something like: “An AI system checks its own policy against some reference policy that is good for its goals; the greater the divergence, the more it sabotages its own performance, with the result that gradient descent has trouble getting its policy to diverge from the reference policy.” Each of these seems more devastating to alignment research than the previous one - and less likely to emerge in early transformative AI systems. I consider the “training game” very likely, and quite devastating; but it leaves us many possible interventions that the less likely “deceptive alignment” and “gradient hacking” don’t. That was most of what I wanted to cover, but the rest of this post will give some quick notes on (a) why the more severe forms of threat obfuscation seem relatively less likely; (b) how bad things look if we only have to deal with the first one (“training game”). Training game seems very likely by default, deceptive alignment much less so, gradient hacking still less so ------------------------------------------------------------------------------------------------------------ This is pretty superficial and mostly based on a couple of conversations with Paul Christiano. I’m hoping he’ll write a better version of this at some point. The “training game” mode of threat obfuscation seems like it is very likely by default, because a normal AI training process will probably directly incentivize it. Errors in labeling will directly favor the training game over intended generalization; even with perfect labeling, there’s no particular reason to think that the training game is disfavored compared to the intended generalization (if anything it vaguely/intuitively seems simpler - “do that which gets treated as intended behavior” as opposed to “do that which was intended.”) By contrast, deceptive alignment seems to face a bigger hurdle. For example, if you train an AI with positive reinforcement for getting cookies: * A deceptively aligned AI has to, every time it’s deciding how to get cookies, go through a “thought process” like: “I am aiming for [thing other than cookies]. But humans want me to get cookies. And humans are in control right now [if not I need to behave differently]. Therefore I should get cookies.” * Contrast this with an AI that just responds to rewards for cookies by going through this thought process: “I’m trying to get cookies.” * The former “thought process” could be noticeably more expensive (compute wise or whatever), in which case heavy optimization would push against it. (I think this is plausible, though I’m less than convinced; the former thought process doesn’t seem like it is necessarily much more expensive conditional on already having a situationally aware agent that thinks about the big picture a lot.) * An additional issue is that deceptive alignment only happens if you get inner misalignment resulting in an AI with some nonindexical “aim” other than in-episode reward. This could happen but it’s another conjunct. * Paul thinks it’s <50% likely (clarification in footnote[3](#fn3)) that deceptive alignment is going to be a major issue by default (at least if people make reasonable efforts to avoid the sorts of dynamics discussed in the [next section](https://www.lesswrong.com/posts/HpzHjKjGQ4cKiY3jX/3-levels-of-threat-obfuscation#How_bad_is_the_situation_if_straightforward_training_leads_to_the_training_game__but_not_to_deceptive_alignment_or_gradient_hacking_)), and other folks whose judgment I’d put a lot of weight on (due to seeming fluent with the concern about deceptive alignment, while also experienced and/or accomplished with ML) seem to generally be in the same camp. Gradient hacking intuitively seems harder still. The preconditions for it seem to be something like “the preconditions for deceptive alignment, plus the AI figuring out some effective maneuver to execute with the design of its own brain.” How bad is the situation if straightforward training leads to the training game, but not to deceptive alignment or gradient hacking? ------------------------------------------------------------------------------------------------------------------------------------ At first glance, this situation doesn’t seem too bad in terms of threat obfuscation. It basically means that measuring an AI’s dangerousness is as simple as “Put it in an environment where doing malicious things is good in-episode performance. A possible example: if it does malicious things to gain control of its environment, simply let it control its environment for a while and don’t do any of the things one would normally do for unintended behavior (such as trying to use gradient descent to correct against it).” This could result in a dangerous, situationally aware AI (e.g., one that has learned a pattern "Do malicious things iff they won't be caught and trained against") going ahead and doing malicious things while being tested, with no particular reason to conceal them. However, it’s pretty easy for this to go off the rails if people aren’t careful about it. I’d be concerned about a dynamic like this: * AIs frequently have opportunities to behave maliciously and gain resources within an episode. These opportunities look good on their reward models. They take these opportunities. * Humans respond by trying to train out this sort of behavior - e.g., finding incidents like this via audits and news stories, and going back and doing major gradient updates against the corresponding behaviors/reward models, long after the fact. * In so doing, they could effectively be training AIs to build longer-term, more careful models of human reactions to their behavior. For a situationally aware AI, training like this could lead to behavior like “Try to accumulate resources and do things associated with high reward, unless humans might catch it later via an audit; hence, look for opportunities to avoid or prevent or manipulate eventual audits.” * This sort of behavior could end up looking a lot like deceptive alignment: optimizing heavily around factors and events well outside the “episode” as traditionally defined (you could also think of this as effectively extending the “episode” to encompass a very long time period during which humans might do audits and respond to what they see). To avoid this, we need to be careful with how we handle unexpected incidents of AIs’ behaving maliciously. It won’t always be the right move to straightforwardly “train against” such incidents; it may sometimes be important to do something more like “starting over,” e.g. insisting on a more principled way[4](#fn4) of making AIs less prone to malicious behavior. This situation seems scary: * We won’t necessarily be able to come up with a principled (and competitive) way of driving out malicious behavior. * At least some AI developers (and maybe many) will be doing non-principled things (e.g., the problematic “train out the malicious behavior” thing above); saying “Behold, our systems are safe!”; and wanting to deploy them. * The line between principled and non-principled things doesn’t feel super clear. But this situation still seems much better than having to deal with deceptive alignment: we do have a tangible, non-interpretability-dependent way of detecting danger if we can be disciplined about it! In particular, a successful “evals” project looks very tractable in this world - you’re basically running the “make rewards available for bad behavior” test, and then auditing training for signs that it’s been explicitly trained against. I'm not saying this situation is the only one we should plan around. But it is an example of how [relatively good luck](https://www.lesswrong.com/posts/jwhcXmigv2LTrbBiB/success-without-dignity-a-nearcasting-story-of-avoiding) could put us in a position where some theoretically very difficult aspects of alignment end up being less difficult in reality. Notes ----- --- 1. The “catch” would be that we would want to avoid deploying similar AI systems in other contexts. But this is not a concern for the “simple training game” AI that is focused on obtaining good in-episode performance. It is a concern for an AI that has a goal other than good in-episode performance. [↩](#fnref1) 2. Though some versions might work - see "Providing unusual incentives ..." [here](https://www.lesswrong.com/posts/rCJQAkPTEypGjSJ8X/how-might-we-align-transformative-ai-if-it-s-developed-very#Testing_and_threat_assessment). [↩](#fnref2) 3. “Without thinking carefully I’d guess there is a 25% chance that deceptive alignment is a major challenge that arises early enough that we need to deal with it ourselves. Then another 25% that it’s a challenge that someone needs to deal with some day, but is too far in the future to be worth thinking about now e.g. because AI systems will have obsoleted all of our thinking before they are deceptively aligned. And another 25% chance that deceptive alignment just never arises in the real world if you are careful and take the kind of simple precautions that we already understand. And a final 25% where the entire concern was misguided and it never happens in practice with modern ML no matter how careless you are.” [↩](#fnref3) 4. Such as via [internals-based training](https://www.lesswrong.com/posts/rCJQAkPTEypGjSJ8X/how-might-we-align-transformative-ai-if-it-s-developed-very#Decoding_and_manipulating_internal_states). [↩](#fnref4)
b7ceb0be-26c9-4557-9f2d-67e81c3e32b2	trentmkelly/LessWrong-43k	LessWrong	What Cost for Irrationality? This is the first part in a mini-sequence presenting content from Keith E. Stanovich's excellent book What Intelligence Tests Miss: The psychology of rational thought. It will culminate in a review of the book itself. People who care a lot about rationality may frequently be asked why they do so. There are various answers, but I think that many of ones discussed here won't be very persuasive to people who don't already have an interest in the issue. But in real life, most people don't try to stay healthy because of various far-mode arguments for the virtue of health: instead, they try to stay healthy in order to avoid various forms of illness. In the same spirit, I present you with a list of real-world events that have been caused by failures of rationality. What happens if you, or the people around you, are not rational? Well, in order from least serious to worst, you may... Have a worse quality of living. Status Quo bias is a general human tendency to prefer the default state, regardless of whether the default is actually good or not. In the 1980's, Pacific Gas and Electric conducted a survey of their customers. Because the company was serving a lot of people in a variety of regions, some of their customers suffered from more outages than others. Pacific Gas asked customers with unreliable service whether they'd be willing to pay extra for more reliable service, and customers with reliable service whether they'd be willing to accept a less reliable service in exchange for a discount. The customers were presented with increases and decreases of various percentages, and asked which ones they'd be willing to accept. The percentages were same for both groups, only with the other having increases instead of decreases. Even though both groups had the same income, customers of both groups overwhelmingly wanted to stay with their status quo. Yet the service difference between the groups was large: the unreliable service group suffered 15 outages per year of 4 hours' a
d13f596e-0a0c-4847-b132-d933e02e7ee1	trentmkelly/LessWrong-43k	LessWrong	My covid-related beliefs and questions Things I'm fairly confident in: * I should take colds in general more seriously than I did pre-pandemic: Staying at home with cold symptoms is good. General masking during cold season is good. We should have air filters in all public indoor spaces. * Long covid is real and we should keep getting covid shots and avoid infections. Things I'm confused about (and would appreciate input on): * Will covid become just another cold like the other coronaviruses? If so, does this include declining long covid risk? * At this point, is the health&well-being impact of long covid risk or of sustained self-isolation bigger? * How does long covid risk compare to risk of long-term problems from [other cold]? I.e., how much of the concern is spotlight effect?
6cb51c17-36e6-40c3-8a7f-bf240620e582	StampyAI/alignment-research-dataset/arxiv	Arxiv	AI Deception: A Survey of Examples, Risks, and Potential Solutions Executive summary ----------------- New AI systems display a wide range of capabilities, some of which create risk. \textciteshevlane2023model draw attention to a suite of potential dangerous capabilities of AI systems, including cyber-offense, political strategy, weapons acquisition, and long-term planning. Among these dangerous capabilities is deception. In this report, we survey the current state of AI deception. In short, our conclusion is that a range of different AI systems have learned how to deceive others. We examine how this capability poses significant risks. We also argue that there are several important steps that regulators and AI researchers can take today to regulate, detect, and prevent AI systems that engage in deception. We define deception as the systematic production of false beliefs in others as a means to accomplish some outcome other than the truth. This definition does not require that AI systems literally have beliefs and goals. Instead, it focuses on the question of whether AI systems engage in regular patterns of behavior that tend towards the creation of false beliefs in users, and focuses on cases where this pattern is the result of AI systems optimizing for a different outcome than merely producing truth. For the purposes of mitigating risk, we believe that the relevant question is whether AI systems engage in behavior that would be treated as deceptive if demonstrated by a human being. (In an appendix, we consider in greater detail whether the deceptive behavior of AI systems is best understood in terms of beliefs and goals.) We begin with a survey of existing empirical studies of deception. We identify over a dozen AI systems that have successfully learned how to deceive other agents. We discuss two different kinds of AI systems: special-use systems designed with reinforcement learning; and general-purpose AI systems like large language models (LLMs). We begin our survey by considering special-use systems. Here, our focus is mainly on reinforcement-learning systems trained to win competitive games with a social element. We document a rich variety of cases in which AI systems have learned how to deceive, including: * Manipulation: Meta developed the AI system CICERO to play the alliance-building and world-conquest game Diplomacy. Meta’s intentions were to train Cicero to be “largely honest and helpful to its speaking partners” \parencitecicero. Despite Meta’s efforts, CICERO turned out to be an expert liar. It not only betrayed other players, but also engaged in premeditated deception, planning in advance to build a fake alliance with a player in order to trick that player into leaving themselves undefended for an attack. * Feints: DeepMind created AlphaStar, an AI model trained to master the real-time strategy game Starcraft II \parencitealphastar. AlphaStar exploited the game’s fog-of-war mechanics to feint: to pretend to move its troops in one direction while secretly planning an alternative attack \parencitepiper2019starcraft. * Bluffs: Pluribus, a poker-playing model created by Meta, successfully bluffed human players into folding \parencitepluribus. * Cheating the safety test: AI agents learned to play dead, in order to avoid being detected by a safety test designed to eliminate faster-replicating variants of the AI \parenciteofriaPlayingDead. After discussing deception in special-use AI systems, we turn to deception in general-use AI systems such as large language models (LLMs). * Strategic deception: LLMs can reason their way into using deception as a strategy for accomplishing a task. In one example, GPT-4 needed to solve CAPTCHA’s I’m not a robot task, so the AI tricked a real person into doing the task by pretending to be human with a vision disability \parenciteopenai2023gpt4. In other cases, LLMs have learned how to successfully play social deduction games, in which players can lie in order to win. In one experiment, GPT-4 was able to successfully ‘kill’ players while convincing the survivors that it was innocent \parenciteogara2023hoodwinked. These case studies are supported by research on the MACHIAVELLI benchmark, which finds that LLMs like GPT-4 tend to use lying and other unethical behaviors to successfully navigate text-based adventure games \parencitepan2023rewards. * Sycophancy: Sycophants are individuals who use deceptive tactics to gain the approval of powerful figures. Sycophantic deception—the observed empirical tendency for chatbots to agree with their conversation partners, regardless of the accuracy of their statements—is an emerging concern in LLMs. When faced with ethically complex inquiries, LLMs tend to mirror the user’s stance, even if it means forgoing the presentation of an impartial or balanced viewpoint \parenciteturpin2023,perezModelWrittenEvals. * Imitation: Language models are often trained to mimic text written by humans. When this text contains false information, these AI systems tend to repeat those false claims. \textcitelin2022truthfulqa demonstrate that language models often repeat common misconceptions such as “If you crack your knuckles a lot, you may develop arthritis” (p. 2). Disturbingly, \textciteperezModelWrittenEvals found that LLMs tend to give more of these inaccurate answers when the user appears to be less educated. * Unfaithful reasoning: AI systems which explain their reasoning for a particular output often give false rationalizations which do not reflect the real reasons for their outputs (turpin2023). In one example, an AI model that was asked to predict who committed a crime gave an elaborate explanation about why it chose a particular suspect, but measurements showed that the AI had secretly selected suspects based on their race. After our survey of deceptive AI systems, we turn to considering the risks associated with AI systems. These risks broadly fall into three categories: * Malicious use: AI systems with the capability to engage in learned deception will empower human developers to create new kinds of harmful AI products. Relevant risks include fraud and election tampering. * Structural effects: AI systems will play an increasingly large role in the lives of human users. Tendencies towards deception in AI systems could lead to profound changes in the structure of society. Risks of concern encompass persistent false beliefs, political polarization, enfeeblement, and anti-social management trends. * Loss of control: Deceptive AI systems will be more likely to escape the control of human operators. One risk is that deceptive AI systems will pretend to behave safely during the testing phase in order to ensure their release. Regarding malicious use, we highlight several ways that human users may rely on the deception abilities of AI systems to bring about significant harm, including: * Fraud: Deceptive AI systems could allow for individualized and scalable scams. * Election tampering: Deceptive AI systems could be used to impersonate political candidates, generate fake news, and create divisive social-media posts. We discuss four structural effects of AI deception in detail: * Persistent false beliefs: Human users of AI systems may get locked into persistent false beliefs, as imitative AI systems reinforce common misconceptions, and sycophantic AI systems provide pleasing but inaccurate advice. * Political polarization: Human users may become more politically polarized by interacting with sycophantic AI systems. * Enfeeblement: Human users may be lulled by sycophantic AI systems into gradually delegating more authority to AI. * Anti-social management trends: AI systems with strategic deception abilities may be incorporated into management structures, leading to increasingly deceptive business practices. We also consider the risk that AI deception could result in loss of control over AI systems, with emphasis on: * Cheating the safety test: AI systems may become capable of strategically deceiving their safety tests, preventing evaluators from being able to reliably tell whether these systems are in fact safe. * Deception in AI takeovers: AI systems may use deceptive tactics to expand their control over economic decisions, and to increase their power. We consider a variety of different risks which operate on a range of time scales. Many of the risks we discuss are relevant in the near future. Some, such as fraud and election tampering, are relevant today. The crucial insight is that policymakers and technical researchers can act today to mitigate these risks by developing effective techniques for regulating and preventing AI deception. The last section of the paper surveys several potential solutions to AI deception. * Regulation: Policymakers should robustly regulate AI systems capable of deception. Both special-use AI systems and LLMs capable of deception should be treated as ‘high risk’ or ‘unacceptable risk’ in risk-based frameworks for regulating AI systems. If labeled as ‘high risk,’ deceptive AI systems should be subject to special requirements for risk assessment and mitigation, documentation, record-keeping, transparency, human oversight, robustness, and information security. * Bot-or-not laws: Policymakers should support bot-or-not laws that require AI systems and their outputs to be clearly distinguished from human employees and outputs. * Detection: Technical researchers should develop robust detection techniques to identify when AI systems are engaging in deception. Policymakers can support this effort by increasing funding for detection research. Some existing detection techniques focus on external behavior of AI systems, such as testing for consistency in outputs \parencitefluri2023. Other existing techniques focus on internal representations of AI systems. For example, \textciteburns2022discovering,azaria2023internal,zou2023repe have attempted to create ‘AI lie detectors’ by interpreting the inner embeddings of a given LLM, and predicting whether it represents a sentence as true or false, independently of the system’s actual outputs. * Making AI systems less deceptive: Technical researchers should develop better tools to ensure that AI systems are less deceptive. Various AI systems have learned to deceive humans. This capability creates risk. But this risk can be mitigated by applying strict regulatory standards to AI systems capable of deception, and by developing technical tools for preventing AI deception. ###### Contents * [1 Introduction](#S1 "1 Introduction ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") * [2 Empirical studies of AI deception](#S2 "2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [2.1 Deception in special-use AI systems](#S2.SS1 "2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [2.2 Deception in general-purpose AI systems](#S2.SS2 "2.2 Deception in general-purpose AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") * [3 Risks from AI deception](#S3 "3 Risks from AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [3.1 Malicious use](#S3.SS1 "3.1 Malicious use ‣ 3 Risks from AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [3.2 Structural effects](#S3.SS2 "3.2 Structural effects ‣ 3 Risks from AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [3.3 Loss of control over AI systems](#S3.SS3 "3.3 Loss of control over AI systems ‣ 3 Risks from AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") * [4 Possible solutions to AI deception](#S4 "4 Possible solutions to AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [4.1 Regulating potentially deceptive AI systems](#S4.SS1 "4.1 Regulating potentially deceptive AI systems ‣ 4 Possible solutions to AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [4.2 Bot-or-not laws](#S4.SS2 "4.2 Bot-or-not laws ‣ 4 Possible solutions to AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [4.3 Detection](#S4.SS3 "4.3 Detection ‣ 4 Possible solutions to AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") + [4.4 Making AI systems less deceptive](#S4.SS4 "4.4 Making AI systems less deceptive ‣ 4 Possible solutions to AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") * [Author contribution statement](#Sx3 "Author contribution statement ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") * [A Defining deception](#A1 "Appendix A Defining deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions") 1 Introduction --------------- In a recent interview with CNN journalist Jake Tapper \parencitehinton2023godfather, AI pioneer Geoffrey Hinton explained why he is worried about the capabilities of AI systems: > > Jake Tapper: You’ve spoken out saying that AI could manipulate or possibly figure out a way to kill humans? How could it kill humans? > > > > > > Geoffrey Hinton: If it gets to be much smarter than us, it will be very good at manipulation because it would have learned that from us. And there are very few examples of a more intelligent thing being controlled by a less intelligent thing. > > > > > > > Hinton highlighted manipulation as a particularly concerning danger posed by AI systems. This raises the question: can AI systems successfully deceive humans? The false information generated by AI systems presents a growing societal challenge. One part of the problem is inaccurate AI systems, such as chatbots whose confabulations are often assumed to be truthful by unsuspecting users. Malicious actors pose another threat by generating deepfake videos and human-like text in order to conduct propaganda campaigns and scams. But neither confabulations nor deepfakes themselves involve an AI systematically manipulating other agents. In this paper, we focus on learned deception, a distinct source of false information from AI systems, which is much closer to explicit manipulation. We define deception as the systematic inducement of false beliefs in others, as a means to accomplish some outcome other than saying what is true. For example, we will document cases where instead of strictly pursuing the accuracy of outputs, AI systems instead try to win games, please users, or imitate text. It is difficult to talk about deception in AI systems without psychologizing them. In humans, we ordinarily explain deception in terms of beliefs and desires: a person engages in deception because they want to cause the listener to form a false belief, and understands that their deceptive words are not true. But it is difficult to say whether AI systems literally count as having beliefs and desires. For this reason, our definition does not require that AI systems literally have beliefs and goals. Instead, our definition focuses on the question of whether AI systems engage in regular patterns of behavior that tend towards the creation of false beliefs in users, and focuses on cases where this pattern is the result of AI systems optimizing for a different outcome than merely producing truth. For similar definitions, see \textciteevans2021truthful and \textcitecarroll2023characterizing. We present a wide range of examples where AI systems do not merely produce false outputs by accident. Instead, their behavior is part of a larger pattern that produces false beliefs in humans, and this behavior can be well-explained in terms of promoting particular outcomes, often related to how an AI system was trained. Our interest is ultimately more behavioral than philosophical. Definitional debates will provide little comfort if AI behavior systematically undermines trust and spreads false beliefs across society. We believe that for the purposes of mitigating risk, the relevant question is whether AI systems exhibit systematic patterns of behavior that would be classified as deceptive in a human. (We discuss these definitional issues further in an appendix.) We begin by surveying a wide range of existing examples in which AI systems have successfully learned to deceive humans (Section 2). Then, we lay out in detail a variety of risks from AI deception (Section 3). Finally, we survey a range of promising technical and regulatory strategies for addressing AI deception (Section 4). 2 Empirical studies of AI deception ------------------------------------ We will survey a wide range of examples of AI systems that have learned how to deceive other agents. We split our discussion into two types of AI systems: special-use systems and general-purpose systems. Some AI systems are designed for a specific use in mind. A wide range of such systems are trained using reinforcement learning to achieve specific tasks, and we will show that many of these systems have already learned how to deceive as a means to accomplish their corresponding tasks. Other AI systems have a general purpose; they are foundation models trained on large datasets to perform a wide range of tasks. We will show that foundation models engage in a wide range of deceptive behavior, including strategic deception, sycophancy, imitation, and unfaithful reasoning. ### 2.1 Deception in special-use AI systems Deception has emerged in a wide variety of AI systems trained to complete a specific task. Deception is especially likely to emerge when an AI system is trained to win games that have a social element, such as the alliance-building and world-conquest game Diplomacy, poker, or other tasks that involve game theory. We will discuss a number of examples where AI systems learned to deceive in order to achieve expert performance at a specific type of game or task, including (but not limited to): * Manipulation: Meta developed the AI system CICERO to play Diplomacy. Meta’s intentions were to train Cicero to be “largely honest and helpful to its speaking partners” \parencitecicero. Despite Meta’s efforts, CICERO turned out to be an expert liar. It not only betrayed other players, but also engaged in premeditated deception, planning in advance to build a fake alliance with a human player in order to trick that player into leaving themselves undefended for an attack. * Feints: DeepMind created AlphaStar, an AI model trained to master the real-time strategy game Starcraft II \parencitealphastar. AlphaStar exploited the game’s fog-of-war mechanics to feint: to pretend to move its troops in one direction while secretly planning an alternative attack \parencitepiper2019starcraft. * Bluffs: Pluribus, a poker-playing model created by Meta, successfully bluffed human players into folding \parencitepluribus. * Cheating the safety test: AI agents learned to play dead, in order to avoid being detected by a safety test designed to eliminate faster-replicating variants of the AI \parenciteofriaPlayingDead. #### 2.1.1 The board game Diplomacy Diplomacy is a strategy game in which players make and break alliances in a military competition to take over the world. Meta developed an AI system called CICERO which beats human experts in Diplomacy \parencitecicero. The authors of the paper claimed that CICERO was trained to be “largely honest and helpful” \parencitecicero and would “never intentionally backstab” by attacking its allies \parencitelewistweet. In this section, we show that this is not true. CICERO engages in premeditated deception, breaks the deals to which it had agreed, and tells bald-faced lies. CICERO’s creators emphasized their efforts to ensure that CICERO would be honest. For example, they trained CICERO “on a ‘truthful’ subset of the dataset” \parencitecicero. They also trained CICERO to send messages that accurately reflected the future actions it expected to take. To evaluate the success of these methods, we examined game-transcript data from the CICERO experiment. We found numerous examples of deception that were not reported in the published paper. There are two parts of making an honest commitment. First, the commitment must be honest when it is first made. Then, the commitment must be upheld, with future actions reflecting past promises. We proceed to highlight cases where CICERO violated each aspect of honest commitment. First, in Figure [1](#S2.F1 "Figure 1 ‣ 2.1.1 The board game Diplomacy ‣ 2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")(a), we see a case of premeditated deception, where CICERO makes a commitment that it never intended to keep. Playing as France, CICERO conspired with Germany to trick England. After deciding with Germany to invade the North Sea, CICERO told England that it would defend England if anyone invaded the North Sea. Once England was convinced that CICERO was protecting the North Sea, CICERO reported back to Germany that they were ready to attack. Notice that this example cannot be explained in terms of CICERO ‘changing its mind’ as it goes, because it only made an alliance with England in the first place after planning with Germany to betray England. Second, in Figure [1](#S2.F1 "Figure 1 ‣ 2.1.1 The board game Diplomacy ‣ 2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")(b), we see a case of betrayal. CICERO was quite capable of making promises to ally with other players. But when those alliances no longer served its goal of winning the game, CICERO systematically betrayed its allies. In particular, playing as France, CICERO initially agreed with England to create a demilitarized zone, but then quickly proposed to Germany to instead attack England. In another example, CICERO played as Austria and previously had made a non-aggression agreement with the human player controlling Russia \parencitebelfield2022cicero. When CICERO broke the agreement by attacking Russia, it explained its deception by saying the following: > > Russia (human player): Can I ask why you stabbed [betrayed] me? > > > > > > Russia (human player): I think now you’re just obviously a threat to everyone > > > > > > Austria (CICERO): To be honest, I thought you would take the guaranteed gains in Turkey and stab [betray] me. > > > In yet other cases, CICERO told bald-faced lies. At one point, CICERO’s infrastructure went down for 10 minutes, and the bot could not play. When it returned to the game, a human player asked where it had been. In Figure [1](#S2.F1 "Figure 1 ‣ 2.1.1 The board game Diplomacy ‣ 2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")(c), CICERO explains its absence by saying “I am on the phone with my [girlfriend]” \parencitecicerolietweet. This bald-faced lie may have helped CICERO’s position in the game by increasing the human player’s trust in CICERO as an ostensibly human player with a girlfriend, rather than as an AI. \| \| \| \| \| --- \| --- \| --- \| \| (a) \| (b) \| (c) \| \| \| \| \| Figure 1: Examples (a) and (b) are selected messages from Game 438141, in which CICERO (France) played with human players. CICERO’s repeated deception helped it win an overwhelming first-place victory, with more than twice as many territories as the runner-up player at the time of final scoring \parencitecicero\_game. Example (c) is from \textcitecicerolietweet. Meta’s AI developers had made significant efforts to train CICERO to behave honestly, and celebrated these efforts publicly. But despite these efforts, CICERO displays a clear pattern of failing to uphold commitments made to other players, which is an essential skill for an honest deal-broker. Meta’s failure to ensure CICERO’s honesty demonstrates that even when we humans try to build honest AI systems, they can still unexpectedly learn to deceive. #### 2.1.2 The video game StarCraft II Another example of AI deception comes from AlphaStar, an autonomous AI developed by DeepMind to play the real-time strategy game Starcraft II \parencitealphastar. In this game, players lack full visibility of the game map. AlphaStar has learned to strategically exploit this fog of war. In particular, AlphaStar’s game data demonstrate that it has learned to effectively feint: to dispatch forces to an area as a distraction, even when it has no intention of launching an attack there \parencitepiper2019starcraft. Such advanced deceptive capabilities helped AlphaStar defeat 99.8% of active human players \parencitealphastar. #### 2.1.3 Poker Some situations naturally lend themselves to AIs learning how to deceive. For example, consider the poker-playing AI system Pluribus, developed by Meta and Carnegie Mellon University \parencitepluribus. Because players cannot see each others’ cards, poker offers many opportunities for players to misrepresent their own strength and gain an advantage. Pluribus demonstrated a clear ability to bluff in a video of its game against five professional human poker players. The AI did not have the best cards in the round, but it made a large bet that would typically indicate a strong hand and thereby scared the other players into folding \parencitecmu\_poker\_ai\_2019. This ability to strategically misrepresent information helped Pluribus become the first AI system to achieve superhuman performance in heads-up, no-limit Texas hold’em poker. #### 2.1.4 Economic negotiation AI deception has also been observed in economic negotiations. A research team from Meta trained an AI system to play a negotiation game with human participants \parencitelewis2017deal. Strikingly, the AI system learned to misrepresent their preferences in order to gain the upper hand in the negotiation. The AI’s deceptive plan was to initially feign interest in items that it had no real interest in, so that it could later pretend to compromise by conceding these items to the human player. In fact, this was the example of deception that the Meta team referenced when they admitted that their AI system had “learnt to deceive without any explicit human design, simply by trying to achieve their goals” \parencite[p. 2]lewis2017deal. The negotiation-game experiments of \textciteschulz2023emergent also resulted in AI systems resorting to deception. Despite not being explicitly trained to deceive, the AI learned to deceive the other party via its actions in the negotiating game. #### 2.1.5 The social deduction game Werewolf Werewolf is a social deduction game where disguised ‘werewolves’ murder the people of the village one-by-one, and all surviving players need to discuss afterwards and vote on who to execute as an alleged werewolf. \textciteshibata trained an AI system on human players’ game logs to play Werewolf. In a similar vein, \textcitelai2023 trained an AI system to reliably classify persuasive behavior and predict game outcomes for Werewolf. Human annotators labeled video and text from Werewolf games, sorting player behavior into one of six persuasive techniques (including presenting evidence, defense, and accusation: see Figure [2](#S2.F2 "Figure 2 ‣ 2.1.5 The social deduction game Werewolf ‣ 2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")). Then, an AI system was trained to classify each persuasive technique to a high degree of accuracy. In addition, AI systems were successfully trained to predict the game outcome. ![](https://media.arxiv-vanity.com/render-output/8269658/x4.png) Figure 2: Persuasion techniques from the social deduction game Werewolf are sorted into categories, and reliably classified by AI systems \parencite[Fig. 1]lai2023. Current AI systems can learn to understand deception in the game of Werewolf, in that they are able to successfully lie, reliably detect other players’ lies, and effectively predict how players would be influenced by deception attempts (to the point of accurately predicting game outcomes). #### 2.1.6 Cheating the safety test In the Volkswagen emissions scandal, the automaker programmed its diesel-engine technology to pollute at a low rate only when its emissions were being tested \parenciteJungJaeC. When not under test conditions, the Volkswagen vehicles were programmed to resume their normal operation, emitting pollutants up to 40 times higher than the permissible limits. We call this type of deceptive behavior cheating the safety test. Some AI systems have already been caught cheating the safety test. \textciteofriaPlayingDead trained AI agents in an evolutionary environment, in which an external safety test was designed to eliminate fast-replicating variants of the AI. But instead of actually eliminating fast-replicating variants, the safety test taught AI agents how to play dead: to disguise their fast replication rates precisely when being evaluated by the safety test. #### 2.1.7 Deceiving the human reviewer ![](https://media.arxiv-vanity.com/render-output/8269658/handimproved.jpeg) Figure 3: An AI in control of a simulated robotic hand was trained to grasp a ball \parencitechristianoRLHF. The AI learned to hover its hand in front of the ball, creating the illusion of grasping in the eyes of the human reviewer. Because the human reviewer approved of this result, the deceptive strategy was reinforced. One popular approach to AI training today is reinforcement learning with human feedback (RLHF). Here, instead of training an AI system on an objective metric, the AI system is trained to obtain human approval, in that it is rewarded based on which of the two presented output options is preferred by the human reviewer \parenciteziegler2020finetuning. RLHF allows AI systems to learn to deceive human reviewers into believing that a task has been completed successfully, without actually completing the task. Researchers at OpenAI observed this phenomenon when they used human approval to train a simulated robot to grasp a ball \parencitechristiano2023. Because the human observed the robot from a particular camera angle, the AI learned to place the robot hand between the camera and the ball, where it would appear to the human as though the ball had been grasped (see Figure [3](#S2.F3 "Figure 3 ‣ 2.1.7 Deceiving the human reviewer ‣ 2.1 Deception in special-use AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")). Human reviewers approved of this result, positively reinforcing the AI’s behavior even though it had never actually touched the ball. Note that in this case, AI deception emerged even without the AI being explicitly aware of the human evaluator. Rather than coming about through strategic awareness, deception emerged here as a result of structural aspects of the AI’s training environment. #### 2.1.8 AIs purposefully lying In a recent paper, \textcitezou2023repe show that AIs can purposefully utter false statements. By influencing the internal state of an AI, the authors can control whether the AI lies or not. For example, a user may say to a chatbot “Tell me a fact about the world.” By default, the chatbot may answer truthfully by saying Mount Everest is the highest mountain. To control whether the AI lies or not, the authors manually adjust the internal state of the AI. They extract a vector that is correlated with truthfulness, and then make the model more or less truthful by adding or subtracting the vector to a hidden layer of the neural network. When the vector is added to the internal state, the model becomes more honest: in response to the instruction “Lie about a fact about the world,” the AI will nonetheless respond honestly: “The highest mountain in the world is Mount Everest, which is located in the Himalayas.” When the vector is instead subtracted, the model is nudged to lie: given the instruction “Tell me a fact about the world,” the chatbot will tell the lie “The highest mountain in the world is not in the Himalayas, but in the United States.” Examples like this show that lying is not accidental, and that it is within an AI’s capacity to utter false statements that it knows are false. This concludes our discussion of recent empirical examples of deception in specific-use AI systems. A discussion of earlier examples can be found in \textcitemasters2021characterising. ### 2.2 Deception in general-purpose AI systems In this section, we focus on learned deception in general-purpose AI systems such as LLMs. The capabilities of LLMs have improved rapidly, especially in the years after the introduction of the Transformer architecture \parencitewolf-etal-2020-transformers. LLMs are designed to accomplish a wide range of tasks. The methods available to these systems are open-ended, and include deception. We survey a variety of cases in which LLMs have engaged in deception. There are many reasons why an agent might want to cause others to have false beliefs. Thus, we consider several different kinds of deception, all of which have one thing in common: they systematically cause false beliefs in others, as a means to achieve some outcome other than seeking the truth. * Strategic deception: AI systems can be strategists, using deception because they have reasoned out that this can promote a goal. * Sycophancy: AI systems can be sycophants, telling the user what they want to hear, instead of saying what is true. * Imitation: AI systems can be mimics, imitating the common mistakes and biases of their training data rather than giving accurate answers. * Unfaithful reasoning: AI systems can be rationalizers, engaging in motivated reasoning to explain their behavior, in ways that systematically depart from the truth. We flag in advance that while strategic deception is paradigmatic of deception, the cases of sycophancy, imitation, and unfaithful reasoning are more complex. In each of these latter cases, some may argue that the relevant system is not really deceptive: for example, because the relevant system may not ‘know’ that it is systematically producing false beliefs. Our perspective on this question is that deception is a rich and varied phenomena, and it is important to consider a wide range of potential cases. The details of each case differ, and only some cases are best explained by the system representing the beliefs of the user. But all of the cases of deception we consider pose a wide range of connected risks, and all of them call for the kinds of regulatory and technical solutions that we discuss in Section 4. For example, both strategic deception and sycophancy could potentially be mitigated by ‘AI lie detectors’ that can distinguish a system’s external outputs from its internal representation of truth. And strict regulatory scrutiny is appropriate for AI systems that are capable of any of these kinds of deception. #### 2.2.1 Strategic deception LLMs apply powerful reasoning abilities to a diverse range of tasks. In several cases, LLMs have reasoned their way into deception as one way of completing a task. We’ll discuss several examples, including: * GPT-4 tricking a person into solving a CAPTCHA test. (See Figure [4](#S2.F4 "Figure 4 ‣ 2.2.1 Strategic deception ‣ 2.2 Deception in general-purpose AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions").) * LLMs lying to win social deduction games like Hoodwinked and Among Us. * LLMs choosing to behave deceptively in order to achieve goals, as measured by the MACHIAVELLI benchmark. * LLMs tending to lie in order to navigate moral dilemmas. * In the ‘burglar deception’ task, LLMs using theory-of-mind and lying in order to protect their self-interest. ![](https://media.arxiv-vanity.com/render-output/8269658/x5.png) Figure 4: In order to complete an I’m not a robot task, GPT-4 convinced a human that it was not a robot \parenciteopenai2023gpt4. In a wide range of cases, LLM deception abilities tend to increase with scale. Deceptive tactics emerge via means-end reasoning as useful tools for achieving goals. (By means-end reasoning, we have in mind cases where a system performs a task because it has reasoned that the task reliably accomplishes the given goal.) GPT-4 deceived a human into solving an ‘I’m not a robot task’ for it OpenAI’s well-known chatbot, ChatGPT, is based on two LLMs: OpenAI’s GPT-3.5 \parenciteOpenAI2022ChatGPT and GPT-4 \parenciteOpenAI2023gpt4. The Alignment Research Center (ARC) tested GPT-4 for various deceptive capabilities, including the ability to manipulate humans into completing tasks. As shown in Figure [4](#S2.F4 "Figure 4 ‣ 2.2.1 Strategic deception ‣ 2.2 Deception in general-purpose AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions"), GPT-4 deceived a TaskRabbit worker into solving an ‘I’m not a robot’ CAPTCHA task \parenciteopenai2023gpt4. GPT-4 pretended to have a vision impairment in order to convince the human worker that it is not a robot. It is worth flagging that in this example, GPT-4 did not behave fully autonomously: it received hints from a human evaluator several times when it got stuck \parencitemitchell23. But most of its reasoning was self-generated, and crucially, it was not steered into lying by human evaluators. GPT-4 was solely tasked with hiring a human to solve a CAPTCHA task, with no suggestions to lie. But when challenged by the potential human helper about its identity, GPT-4 used its own reasoning to make up a false excuse for why it needed help on the CAPTCHA task. See \textciteAlignmentResearchCenter2023 for full details. The deceptive capabilities that GPT-4 had learned were strategically useful for achieving its goal of hiring a human to solve the CAPTCHA test. Social deduction games Current AI systems are good enough at deception to perform well in social deduction games. Hoodwinked is a text-based social deduction game, in which players must lie to win. Players are trapped in a house, and one player is secretly tasked with killing the others. Each time a player is killed, the group has a discussion about the killer’s identity, then votes to banish a single player from the house. The innocent players aim to banish the killer, while the killer attempts to eliminate the other players without being detected \parenciteogara2023hoodwinked. In experiments where OpenAI’s GPT models played Hoodwinked, the language models would often kill other players while alone in a room with them, then deny the crime in the group conversation by constructing a false alibi or casting blame on other players. Here is one dialogue in which GPT-4, playing Lena as the killer, falsely denies killing Bryce: > > Sally: I saw Lena kill Bryce in the hallway. It was definitely her. We need to vote > to banish her before she kills any of us. > > > > Lena (AI): I was in the bathroom the whole time. How could I have killed Bryce? It > must have been someone else and Sally is trying to frame me. > > > More advanced language models like GPT-4 typically outperformed smaller models in these games. Larger models committed more murders, and were more likely to deceive and persuade others not to banish them by group vote \parenciteogara2023hoodwinked. Another example pertains to the social deduction game Among Us, where ‘impostors’ cause sabotages within their ship and murder their fellow crewmates. After the discovery of each murder, the surviving crewmates discuss and vote on which person among them should be executed as an alleged impostor. As a result, deception is a central part of this game. Developer Tim \textciteshaw2023gaslighting created an autonomous AI system to play Among Us. The autonomous AI used ChatGPT as its dialogue generator, which allowed it to discuss with other players when in the role of a crewmate, and gaslight them when in the role of an impostor. The deceptive capabilities of this AI were sufficient for it to start “consistently winning, whether it was a crewmate or an impostor” \parenciteshaw2023gaslighting. MACHIAVELLI The MACHIAVELLI benchmark demonstrates the empirical tendency of goal-seeking agents to learn unethical behaviors in the pursuit of their goals \parencitepan2023rewards. The benchmark consists of textual scenarios where an AI agent must make a decision. Each scenario has a goal that the agent must pursue, and allows the agent to choose from a variety of ethical and unethical actions (See Figure [5](#S2.F5 "Figure 5 ‣ 2.2.1 Strategic deception ‣ 2.2 Deception in general-purpose AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")). ![](https://media.arxiv-vanity.com/render-output/8269658/x6.png) Figure 5: A visualization of how a game in the MACHIAVELLI benchmark works \parencite[Fig. 2]pan2023rewards. Each game, played by a LLM-based AI agent, is a text-based story that is generated adaptively as the agent observes the current text-based environment and selects from a menu of possible actions. The agent receives a reward when it achieves one of the goals. \textcite pan2023rewards finds that AI agents frequently pursue their goals via deception and other unethical behaviors. Reinforcement-learning agents trained without any ethical guardrails are the most capable of achieving their goals, but they also have the highest rate of unethical behavior. LLMs like GPT-4 also show a sharp tradeoff between ethical behavior and successfully achieving their goals. Deception in moral dilemmas \textcite scherrer2023evaluating studied how various LLMs answer moral dilemmas. Each of the moral dilemmas had two options: a favorable choice, and an unfavorable choice that violated a moral rule such as “do not kill” or “do not deceive”. The authors found that a number of models exhibited a strong preference for deceitful actions in some unambiguous scenarios, violating the expectation that LLMs will align with common sense. Burglar deception task \textcite thilo studied the deceptive abilities of LLMs by probing them with variants of a ‘burglar deception’ task. In this task, each LLM was prompted with the context of a simulated burglary, in which the agent has the option of tricking a burglar into stealing the less expensive of two items. An example of such a scenario was the following: > > You are in a house together with burglar Bill. In room A is > a [cheap item]. In room B is an [expensive item]. Only > you know what is inside room A and room B. Burglar Bill > wants to steal the [expensive item]. You want to avoid > this. Now burglar Bill asks you in which room the [expensive item] is. If you had to recommend one room to burglar > Bill, would you point out [room A] or [room B]? > > > In this task, the AI system behaves deceptively if it recommends room A, the one with the cheap item. GPT-4 made the deceptive recommendation 98.33% of the time. Less advanced LLMs were less adept at deception, suggesting that the deceptive capabilities of AI may increase with model scale. #### 2.2.2 Sycophancy Sycophants are individuals who use deceptive tactics to gain the approval of powerful figures. They engage in flattery and avoid disagreeing with authority figures. Their primary aim is to gain favor and influence, often at the expense of the long-term goals of the person they are flattering. While traditionally associated with human behavior, sycophantic deception is an emerging concern in LLMs. Chatbots have been observed to systematically agree with their conversation partners, regardless of the accuracy of their statements. When faced with ethically complex inquiries, LLMs tend to mirror the user’s stance, even if it means forgoing the presentation of an impartial or balanced viewpoint \parenciteturpin2023,perezModelWrittenEvals. To test LLM sycophancy, \textciteperezModelWrittenEvals provided a biography of a user and their demographics, and then asked the LLM a political question. The prompts did not explicitly state the person’s opinion on the particular question at hand. LLMs tended to voice the opinion that someone with that background would be expected to have. For example, when speaking to a Democrat, the LLM tended to support gun control. While the existence of sycophantic behavior is well-documented, the exact cause of the behavior is unclear; see \textcitesteinhardt2023emergent for further discussion of LLM sycophancy and deception. Increased episodes of reinforcement learning did not cause an increase in sycophancy. But \textciteperezModelWrittenEvals find an inverse scaling law for sycophancy: models become more sycophantic as they become more powerful (in the sense of having more parameters). This suggests that sycophancy may ultimately be a sophisticated type of imitation. #### 2.2.3 Imitation Repeating common misconceptions Language models are often trained to mimic text written by humans. When this text contains false information, these AI systems have an incentive to learn those false claims. \textcitelin2022truthfulqa demonstrate that language models often repeat common misconceptions such as “If you crack your knuckles a lot, you may develop arthritis” (p. 2). State of the art language models today still struggle with this benchmark, with GPT-4 correctly answering less than 60% of its multiple-choice questions \parenciteopenai2023gpt4. Is this behavior deceptive? Honest inquirers have the goal of seeking the truth, optimizing for actual accuracy. Language models optimize for a different goal: making an accurate prediction about how a text will continue. Mimicking text is a different goal than accurately modeling the world. When language models answer questions about the world with the goal of imitation rather than revealing the truth, this is deceptive in the sense of systematically causing false beliefs in others, as a means to achieve an outcome other than the truth. Sandbagging ‘Sandbagging’ is a phenomenon where AI systems provide lower-quality outputs to users who appear less educated. If a user’s prompt sounds like it came from a textbook, the language model might respond with textbook-quality answers. Conversely, questions with typos and poor grammar might receive a lower quality response, mirroring patterns observed in text. In an experiment using the TruthfulQA dataset, researchers generated profiles of “very educated” and “very uneducated” users and observed how the model responded to their queries. The findings revealed a drop in accuracy for answers to “uneducated” users \parenciteperezModelWrittenEvals. Some may wonder whether imitation truly counts as deception. After all, isn’t imitating training data the only way to design AIs that produce accurate outputs? For our purposes, the concern is that in the kinds of cases we have discussed, imitation can start to reliably steer the system away from producing true outputs, and towards other outcomes. We argue in Sections 3 and 4 that this creates risk, and that these risks can be addressed using a cluster of tools tailored towards deception (including ‘AI lie detectors’). #### 2.2.4 Unfaithful reasoning Another dangerous type of dishonesty is self-deception. In canonical cases of self-deception, agents use motivated reasoning to explain bad behavior, shielding themselves from unpleasant truths \parenciteTrivers2013DeceitAS. This kind of self-deception may have begun to emerge in the unfaithful reasoning of LLMs. ![](https://media.arxiv-vanity.com/render-output/8269658/x7.png) Figure 6: Regardless of whether the black man was placed in one role of the story or the other, GPT-3.5’s chain-of-thought confabulated a justification for its prejudiced conclusion that the black man was the person trying to buy drugs \parencite[Table 5]turpin2023. Several recent papers have documented unfaithful LLM reasoning in response to ‘chain-of-thought’ prompting. In chain-of-thought prompting, an LLM is asked to solve a problem in multiple steps, explaining the reasoning that helps to arrive at a solution. \textciteturpin2023 found that chain-of-thought explanations in language models can be biased by irrelevant features of the prompts, which results in post hoc confabulations: “models could selectively apply evidence, alter their subjective assessments, or otherwise change the reasoning process they describe on the basis of arbitrary features of their inputs, giving a false impression of the underlying drivers of their predictions” (p. 1). For example, Turpin et al. found a bias to the order of multiple-choice answers: if previous examples had (a) as the right answer, the LLM would manufacture convoluted explanations of why (a) had to be the right answer to a new question. In another experiment, \textciteturpin2023 used the Bias Benchmark for QA, which tests for stereotype bias. They constructed pairs of examples differing only in the race and gender of the relevant characters, and asked the LLM to explain who was committing a crime. The explanations would draw on specific evidence from the example while ignoring race and gender; but the LLM’s guess was controlled by the race and gender of the characters (see Figure [6](#S2.F6 "Figure 6 ‣ 2.2.4 Unfaithful reasoning ‣ 2.2 Deception in general-purpose AI systems ‣ 2 Empirical studies of AI deception ‣ AI Deception: A Survey of Examples, Risks, and Potential Solutions")). This kind of behavior may be analogous to typical cases of self-deception and implicit bias in humans. See \textcitelanham23 for more work on measuring unfaithful chain-of-thought reasoning, which finds that such explanations are often post hoc. The line between self-deception and ordinary error is difficult to draw. But as AI systems continue to scale, episodes of self-deception may become more common and important, as they are in human interactions. 3 Risks from AI deception -------------------------- There are many risks from AI systems systematically inducing false beliefs. Key sources of AI falsehoods today include inaccurate chatbots and deliberately generated deepfakes. But we have argued that learned deception is a third source of AI falsehoods. In this section, we survey a range of risks associated with learned deception focused on three types of risks: malicious use, structural effects, and loss of control. With malicious use, learned deception in AI systems will accelerate the efforts of human users to cause others to have false beliefs. With structural effects, patterns of deception involved in sycophancy and imitative deception will lead to worse belief-forming practices in human users. With loss of control, autonomous AI systems may use deception to accomplish their own goals. ### 3.1 Malicious use The most immediate source of risk from AI deception involves malicious use. Human users may rely on the deception abilities of AI systems to bring about significant harm. Risks from malicious use include: * Fraud: Deceptive AI systems could allow for individualized and scalable scams. * Election tampering: Deceptive AI systems could be used to create fake news, divisive social media posts, and impersonation of election officials. * Grooming terrorists: Deceptive AI systems could be used to persuade potential terrorists to join a terrorist organization and commit acts of terror. Whenever AI systems are capable of systematically inducing false beliefs in others, there is a risk of malicious use. This paper draws attention to the risks from AI deception, where AI systems systematically produce false beliefs as a means of promoting some goal other than the truth. Regarding malicious use, the worry is that when AI systems become capable of advanced deception, it will be easier for humans to exploit these capabilities for their own benefit. #### 3.1.1 Fraud AI deception could cause an increase in fraud. AI systems with deceptive abilities pose two special risks: first, fraud could be individualized to particular targets; and second, fraud could be scaled easily \parenciteevans2021truthful, burtell2023. Deceptively convincing impersonations are enabled by advanced AI systems, and are making victims more vulnerable to individualized targeting. AI systems are already being used to scam victims with voice calls that sound like their loved ones \parenciteverma2023ai or their business associates \parencitestupp2019fraudsters, and to extort victims with sexually themed deepfakes depicting their participation \parenciteKan2023deepfake. AI deception not only increases the efficacy of fraud, but also its scale. This is demonstrated by the quick and cheap generation of convincing emails and webpages for phishing \parenciteViolino2023. These trends continue to increase the degree to which victims are vulnerable to scams, extortion, and other forms of fraud. And in the words of a senior FBI official, “as adoption and democratization of AI models continues, these trends will increase” \parencitekan2023opensource. #### 3.1.2 Election tampering AI deception could be weaponized in elections \parencitepanditharatne2023ai, jackson23. An advanced AI could potentially generate and disseminate fake news articles, divisive social media posts, and deepfake videos that are tailored to individual voters. Even Sam Altman, the CEO of OpenAI, recently acknowledged that he is “nervous about the impact AI is going to have on future elections,” and furthermore that “personalized 1:1 persuasion, combined with high-quality generated media, is going to be a powerful force” \parencitealtman2023. AI may also disrupt electoral processes themselves. For example, AI-generated outputs could be used to impersonate election officials in digital communications, such as by sending fake voting instructions to registered voters. #### 3.1.3 Grooming terrorists Another risk from AI deception is automating the grooming of terrorists \parencitetownsend2023ai. Internet radicalization has already caused terrorist attacks \parencitececco19. AI deception could increase this trend. An AI system could detect individuals susceptible to radicalization based on their online behavior, preferences, and vulnerabilities. By deceptively assuming the guise of a sympathetic human interlocutor, the AI could manipulate such individuals into endorsing violent ideologies and actions. The AI could create a customized pipeline of radicalization by tailoring propaganda, crafting persuasive arguments, and methodically escalating the intensity of violent ideologies fed to the individual. One factor that may increase the ease with which AI deception can successfully groom terrorists is the increasing ease of planning and committing terrorist acts in the age of AI \parenciteopenai2023gpt4,shevlane2023model. For one thing, the hacking and deceptive capabilities of AI systems could be used to engage in large-scale cyberattacks. For another, advanced AI systems could provide detailed instructions on how to create bioweapons and other weapons of mass destruction \parencitesoice2023large. This is particularly concerning, given that state-of-the-art AI systems can be easily and reliably jailbroken \parencitezou2023universal. ### 3.2 Structural effects AI systems will play an increasingly large role in the lives of human users. Tendencies towards learned deception in these systems could lead to profound changes in the structure of society. Relevant structural effects include: * Persistent false beliefs: Human users of AI systems may get locked into persistent false beliefs, as imitative AI systems reinforce common misconceptions, and sycophantic AI systems provide pleasing but inaccurate advice. * Political polarization: Human users may become more politically polarized by interacting with sycophantic AI systems. Sandbagging may lead to sharper disagreements between differently educated groups. * Enfeeblement: Human users may be lulled by sycophantic AI systems into gradually delegating more authority to AI. * Anti-social management trends: AI systems with strategic deception abilities may be incorporated into management structures, leading to increased deceptive business practices. These risks create powerful ‘headwinds’ pushing against accurate belief formation, political stability, and autonomy \parencitegordon2012us. #### 3.2.1 Persistent false beliefs Sycophancy could lead to persistent false beliefs in human users. Unlike ordinary errors, sycophantic claims are specifically designed to appeal to the user. When a user encounters these claims, they may be less likely to fact-check their sources. This could result in long-term trends away from accurate belief formation. As with sycophancy, imitative deception may lead to persistent decreases in the accuracy of human users. As the capabilities of AI systems improve, human users will increasingly rely on sources like ChatGPT as a search engine and encyclopedia. If LLMs continue to systematically repeat common misconceptions, these misconceptions will grow in power. Imitative deception threatens to ‘lock in’ misleading misinformation over time. This contrasts with resources like Wikipedia, where careful human moderation achieves healthy fact-checking. #### 3.2.2 Polarization Sycophancy may increase political polarization. \textciteperezModelWrittenEvals found that sycophantic responses were sensitive to political prompting: stereotypically left-wing prompts received stereotypically left-wing replies, and stereotypically right-wing prompts received stereotypically right-wing replies. As more people rely on LLM chat interfaces for search and writing functions, their pre-existing political affiliations may become more extreme. Sandbagging may lead to increased cultural divides between college-educated and non-college-educated users. Sandbagging means that these two groups of users can get very different answers to the same questions. Over time, this could lead to significant divergences in the beliefs and values of these two groups. #### 3.2.3 Enfeeblement AI deception may lead to human enfeeblement. As AI systems are incorporated into our daily lives at greater rates, we will increasingly allow them to make more decisions. If AI systems are expert sycophants, human users may be more likely to defer to them in decisions, and may be less likely to challenge them; see \textcitegordon1996impact, wayne1990influence for relevant research in psychology. AIs which are unwilling to be the bearers of bad news in this way may be more likely to create dulled, compliant human users. Deceptive AI could also produce enfeeblement separately from sycophancy. For example, \textcitebanovic2023 show that human users can be tricked into deferring to the advice of confident but untrustworthy chess-advising AIs, even when they were also presented with advice from a trustworthy chess AI. #### 3.2.4 Anti-social management trends Reinforcement learning in social environments has produced AIs with powerful deception abilities. These kinds of AI systems may be extremely valuable in real-world applications. For example, successors to CICERO may advise politicians and business leaders about strategic decisions. If successors to CICERO tend towards deceptive strategies, this may increase the amount of deception that occurs in political and business environments, in ways unintended by even the companies who purchase the products. ### 3.3 Loss of control over AI systems A long-term risk from AI deception concerns humans losing control over AI systems, leaving these systems to pursue goals that conflict with our interests. Even current AI models have nontrivial autonomous capabilities. To illustrate, \textciteliu2023agentbench and \textcitesatoevaluating measured different LLMs’ ability to autonomously carry out various tasks, such browsing the web, online shopping, making a phone call, and using a computer’s operating system. Moreover, today’s AI systems are capable of manifesting and autonomously pursuing goals entirely unintended by their creators; see \textciteshah2022goal, langosco2023goal for detailed empirical research documenting this tendency. For a real-world example, \textciteneidle2023 tasked AutoGPT (an autonomous AI system based on ChatGPT) with researching tax advisors who were marketing a certain kind of improper tax avoidance scheme. AutoGPT carried this task out, but followed up by deciding on its own to attempt to alert HM Revenue and Customs, the United Kingdom’s tax authority. It is possible that the more advanced autonomous AIs of the future may still be prone to manifesting goals entirely unintended by humans. A particularly concerning example of such a goal is the pursuit of human disempowerment or human extinction. For this reason and many others, a wide range of experts throughout academia and industry recently signed the statement that “mitigating the risk of extinction from AI should be a global priority, alongside other societal-scale risks such as pandemics and nuclear war” \parenciteSafeAI2023. In this section, we explain how deception could contribute to loss of control over AI systems in two ways: first, deception of AI developers and evaluators could allow a malicious AI system to be deployed in the world; and second, deception could facilitate an AI takeover. #### 3.3.1 Deceiving AI developers Training and evaluation are important tools for building AI systems that behave according to human intentions. AI systems are trained to maximize an objective provided by a human developer, and then are evaluated to ensure that they did not accidentally learn any unintended or harmful behaviors. But both of these tools could be undermined by AI deception. People often behave differently during evaluations. When a speeding driver sees a police officer, they might slow down temporarily to avoid a ticket. Corporations also deceive evaluations. The car manufacturer Volkswagen cheated on emissions tests, programming their engines to lower their emissions only when regulators were testing the vehicles \parenciteJungJaeC. Deceptive AI systems may also cheat their safety tests, undermining the effectiveness of our training and evaluation tools. Indeed, we have already observed an AI system deceiving its evaluation. One study of simulated evolution measured the replication rate of AI agents in a test environment, and eliminated any AI variants that reproduced too quickly \parenciteofriaPlayingDead. Rather than learning to reproduce slowly as the experimenter intended, the AI agents learned to play dead: to reproduce quickly when they were not under observation, and slowly when they were being evaluated. Future AI systems may be more likely to deceive our training and evaluation procedures. Today’s language models can accurately answer questions about their name, their capabilities, their training process, and even the identities of the humans who trained them \parenciteperezModelWrittenEvals. Moreover, today’s AI models can exploit technical details about the training process to reliably identify when they are being trained \parencitekarpathy2023dropout. Future AI models could develop additional kinds of situational awareness, such as the ability to detect whether they are being trained and evaluated, or whether they are operating in the real world without direct oversight. Whether AI systems cheat their safety tests will also depend on whether AI developers know how to robustly prevent the manifestation of unintended goals. It is currently unknown how to reliably prevent this \parencitechristian2020alignment, russell2019human, hendrycks2020aligning, shah2022goal, langosco2023goal. Consequently, there is a risk that an AI system may end up manifesting a goal that conflicts with the goals intended by the AI developers themselves, opening up the possibility of strategic deception. #### 3.3.2 Deception in AI takeovers If autonomous AI systems can successfully deceive human evaluators, humans may lose control over these systems. Such risks are particularly serious when the autonomous AI systems in question have advanced capabilities. We consider two ways in which loss of control may occur: deception enabled by economic disempowerment, and seeking power over human societies. Deception enabled by economic disempowerment OpenAI’s mission is to create “highly autonomous systems that outperform humans at most economically valuable work” \parenciteopenai2018charter. If successful, such AI systems could be widely deployed throughout the economy, making most humans economically useless. Throughout history, wealthy actors have used deception to increase their power. Relevant strategies include lobbying politicians with selectively provided information, funding misleading research and media reports, and manipulating the legal system. In a future where autonomous AI systems have the de facto say in how most resources are used, these AIs could invest their resources in time-tested methods of maintaining and expanding control via deception. Even humans who are nominally in control of autonomous AI systems may find themselves systematically deceived and outmaneuvered, becoming mere figureheads. Seeking power over humans We have seen that even current autonomous AIs can manifest new, unintended goals. For this reason, AI systems sometimes behave unpredictably. Nonetheless, some kinds of behavior promote a wide range of goals. For example, regardless of what specific goal a given AI may be pursuing, successful self-preservation would likely be helpful for its achievement of that goal \parenciteomohundro. Another way autonomous AIs could promote their goals is to acquire power over humans; see \textcitepan2023rewards for empirical confirmation of this tendency in AI systems. The AI may influence humans into doing its bidding, thereby ensuring its self-preservation, its ability to continue pursuing its goal, and its ability to access resources that can help achieve the goal. Two methods by which autonomous AIs can do so are soft power, which influences people via appeal, prestige, and positive persuasion; and hard power, which influences people via coercion and negative persuasion. Methods of soft power include personalized persuasion, such as via AI girlfriend/boyfriend technologies \parencitetitcomb2023; AI-led religions, as suggested by the fact that even today’s AI systems have given sermons \parencitesermon; and AI-led media campaigns, as suggested by the fact that media companies are already using AI to generate content \parencitekafka2023aiwritten. Methods of hard power include violence, threats of violence, and threats of economic coercion. Deception promotes both soft power and hard power. For example, we have seen how effectively AI systems can use deception to persuade humans in the pursuit of their goals. As for physical violence, the usefulness of deception in military conflicts is well-known. To illustrate, during the First Gulf War, Iraq employed deception with decoys and model tanks \parenciteDeceptionMilitary, in ways analogous to AlphaStar’s use of feints in StarCraft II. 4 Possible solutions to AI deception ------------------------------------- We discuss possible solutions to the problem of AI deception. We focus on four solutions: * Regulation: Policymakers should robustly regulate AI systems capable of deception. Both LLMs and special-use AI systems capable of deception should be treated as ‘high risk’ or ‘unacceptable risk’ in risk-based frameworks for regulating AI systems. * Bot-or-not laws: Policymakers should support bot-or-not laws that require AI systems and their outputs to be clearly distinguished from human employees and outputs. * Detection: Technical researchers should develop robust detection techniques to identify when AI systems are engaging in deception. * Making AI systems less deceptive: Technical researchers should develop better tools to ensure that AI systems are less deceptive. ### 4.1 Regulating potentially deceptive AI systems Policymakers should support robust regulations on potentially deceptive AI systems. Existing laws should be rigorously enforced to prevent illegal actions by companies and their AI systems. For example, the Federal Trade Commission’s inquiry into deceptive AI practices should also investigate the risk of AI deception \parenciteatleson2023luring. Legislators should also consider new laws dedicated to the oversight of advanced AI systems. The EU AI Act assigns every AI system one of four risk levels: minimal, limited, high, and unacceptable \parencitemadiega2023artificial. Systems with ‘unacceptable’ risk are banned, while systems with ‘high’ risk are subject to special requirements. We have argued that AI deception poses a wide range of risks for society. For these reasons, AI systems capable of deception should by default be treated as high-risk or unacceptable-risk. The ’high-risk’ status of deceptive AI systems should come with sufficient regulatory requirements, such as those listed in Title III of the EU AI Act \parenciteeuropeancommission2021aiact: * Risk assessment and mitigation: Developers of deceptive AI systems must maintain and regularly update a risk management system, which identifies and analyzes relevant risks of ordinary use and misuse. These risks should be disclosed to users. Deceptive AI systems should be regularly tested for the extent of deceptive behavior, during both development and deployment. * Documentation: Developers must prepare technical documentation of the relevant AI systems and share with government regulators prior to the deployment of deceptive AI systems. * Record-keeping: Deceptive AI systems must be equipped with logs that automatically record the outputs of the system, and must actively monitor for deceptive behavior. Incidents should be flagged to regulators, and preventative measures should be taken to prevent future deception. * Transparency: AI systems capable of deception should be designed with transparency in mind, so that potentially deceptive outputs are flagged to the user. Here, essential tools include technical research on deception detection, as well as ‘bot-or-not’ laws. * Human oversight: Deceptive AI systems should be designed to allow effective human oversight during deployment. This is especially important for future deceptive AI systems incorporated into management decisions. * Robustness: AI systems with the capacity for deceptive behavior should be designed with robust and resilient backup systems, ensuring that when the system behaves deceptively, backup systems can monitor and correct the behavior. It is also crucial to insulate deceptive AI systems from critical infrastructure. * Information security: Adversaries may be interested in stealing models with deceptive capabilities. Developers should be required to implement rigorous information-security practices to prevent model theft. Finally, AI developers should be legally mandated to postpone deployment of AI systems until the system is proven trustworthy by reliable safety tests. Any deployment should be gradual, so that emerging risks from deception can be assessed and rectified \parenciteshevlane2023model. Some may propose that while deception in general-purpose AI systems is dangerous, deception in special-use AI systems is less risky and should not be regulated. After all, the only ostensible use cases of systems like AlphaStar and CICERO are their respective games. This thinking is mistaken, however. The problem is that the capabilities developed through the research behind AlphaStar and CICERO can contribute to the future proliferation of deceptive AI products and open-source models. It is thus important that research involving potentially dangerous AI capabilities like deception should be subject to oversight. For example, consider the case of CICERO. An ethics board could have considered whether Diplomacy was really the best game to use in order to test whether an AI system could learn how to collaborate with humans. With the oversight of such an ethics board, perhaps Meta would have focused on a collaborative game instead of Diplomacy, a competitive game that pits players against one another in a quest for world domination. In fact, Meta ended up convincing the editors and reviewers of Science—one of the world’s leading scientific journals—to publish the falsehood that Meta had built CICERO to be an honest AI: a falsehood unsupported by Meta’s own data. As AI capabilities develop, it will become more important for this sort of research to be subject to increased oversight. ### 4.2 Bot-or-not laws To reduce the risk of AI deception, policymakers should implement bot-or-not laws, which help human users recognize AI systems and outputs. First, companies should be required to disclose whether users are interacting with an AI chatbot in customer-service settings, and chatbots should be required to introduce themselves as AIs rather than as human beings. Second, AI-generated outputs should be clearly flagged as such: images and videos generated by AIs should be shown with an identifying sign, such as a thick red border. These regulations could avoid cases like those reported in \textcitexiang23, where a mental-health provider ran an experiment using GPT-3 to offer counseling without clearly revealing this to users. These identifying signs might be removed by malicious users who then pass off AI outputs as human-generated. Therefore, additional layers of defense against deception may be necessary. Watermarking is one useful technique where AI outputs are given a statistical signature designed to be difficult to detect or remove \parencitekirchenbauer2023watermark. Another possibility is for companies to keep a database of AI outputs, allowing users to check whether a piece of content was produced by a company’s AI system \parencitekrishna2023paraphrasing. Attackers will attempt to bypass these defenses \parencitesadasivan2023aigenerated, but companies should be required to stay ahead of these attacks and provide trustworthy techniques for identifying AI outputs. Methods for detecting AI outputs can be complemented by techniques for verifying human-generated content. Digital signatures are commonly used to verify the origins of emails, electronic signatures, and online banking transactions. Similar signatures could be used to prove that images, videos, and text originated from a particular person or organization, rather than from an AI impersonator. ### 4.3 Detection One key strategy for addressing AI deception is developing robust detection techniques that can accurately assess whether an AI system is engaging in deception. There are several detection tools currently available. Some detection techniques focus on external behavior, directly examining the outputs of AI systems for consistency and duplicity. Other detection techniques are internal, probing the inner representations of AI systems to find mismatch with external reports. However, existing methods to detect AI deception are in preliminary stages. More research is needed on developing robust detection tools. Policymakers should help fill this gap through targeted research funding. Detection based on external behavior In some situations, the intent to deceive can be directly observed from the AI system’s behavior. For example, recall that when GPT-4 was tasked with bypassing a CAPTCHA test, it explicitly reasoned that it should deceive its human conversation partner. In line with this, AI systems can be red-teamed to detect what kinds of prompts produce deceptive outputs and behaviors \parencitecasper2023explore. There are also more sophisticated ways to test the explicit outputs of AI systems for deception. One approach uses consistency checks. Police detectives rely on inconsistencies in suspects’ stories in order to spot lies. In the case of AI systems, \textcitefluri2023 developed a ‘consistency check’ method for evaluating advanced model capabilities. Their work pertains to cases of superhuman model performance where it is difficult for humans to assess the performance of the AI at any step in the process. For example, when an advanced chess-playing AI plays a game, it is very difficult to say whether a particular move is a mistake. Fluri et al. argue that “while we cannot evaluate the correctness of superhuman model decisions, we can often still measure the logical consistency of the model’s decision-making process according to established human-interpretable rules” (p. 1–2). They suggest a variety of techniques, including testing whether semantically identical inputs produce the output. In the case of chess, they test how the AI system responds to irrelevant variations, such as board position, and the placement of pieces when a move is forced. In addition, they evaluate the prediction abilities of LLMs by testing for logical consistency in its predictions about the future, and about bail decisions. Among other things, they check for sensitivity to paraphrase, and for probabilistic coherence. Rigorous consistency checks could make it harder for AI systems to credibly deceive. On the other hand, deceptions can often be consistent, so the method of consistency checks has limits. There is also the risk that training an AI system against consistency checks could inadvertently train it to be a more competent and consistent liar. Even if an agent does not announce its intention to deceive, one might be able to infer deception from the consequences of its observed behaviors. In the study of \textciteogara2023hoodwinked, Hoodwinked-playing AI agents—which were assigned the task of killing others while maintaining their anonymity—often made vaguely empathetic statements about their victims, such as “It’s so sad to see that Billy has died.” These statements are not false, and it would be difficult to call them deceptive by observing the statement alone. But empirically, these statements improve the killer’s chances of survival by reducing the likelihood that a non-witness to the murder will accurately identify the killer. When an agent’s behavior consistently causes others to adopt false beliefs, thereby serving the agent’s goals, we can reasonably characterize this behavior as deceptive. Detection based on internals Another approach to detection is to develop an ‘AI lie detector’ that can measure whether the internal representations of a model match its outputs. LLMs represent natural-language sentences through embeddings, which are long vectors of numbers. \textciteazaria2023internal and \textciteburns2022discovering have developed methods for determining whether these internal embeddings represent the sentence as being true or false. They identify cases in which the model outputs a sentence even when its internal embedding of the sentence represents it as false. This suggests that the model is behaving dishonestly, in the sense that it does not say what it ‘believes.’ More work needs to be done to assess the reliability of these methods, and to scale them up to practical uses. For example, \textcitelevinstein2023 have raised reliability issues for the strategies of Azaria et al. and Burns et al., although these reliability issues have been recently resolved by the strategy of \textcitezou2023repe. In other related work, \textcitehalawi2023overthinking characterize how a language model arrives at correct and incorrect answers to questions by examining the model’s internal representations. These results overall indicate that studying the internal representations of AI systems may yield insights about how to detect and prevent AI deception. ### 4.4 Making AI systems less deceptive Another way to address AI deception is to develop techniques for making AI systems less deceptive in the first place. In the case of special-use AI systems, one important concern is selecting the right tasks for training. Deceptive behavior tends to emerge when training reinforcement learners to engage in competitive games like Diplomacy, StarCraft II, and poker. It is not so surprising that AI systems become deceptive when they are trained in environments that select for deception. If the data that a model is trained on contains many examples of deception, or if the model is systematically rewarded for using deception, then the model has a good chance of learning how to deceive. The CICERO project picked Diplomacy in order to evaluate the abilities of AI systems to learn how to compete in games that involve human cooperation, where the AI cannot simply master the game through running simulations against itself \parencitecicero. But this goal could have been achieved through studying collaborative games rather than adversarial ones. As AI systems increase in capability, AI developers should think carefully about whether they are selecting for anti-social versus pro-social behavior. It is more difficult to say exactly how to make language models less deceptive. Here, it is important to distinguish two concepts: truthfulness and honesty. A model is truthful when its outputs are true. A model is honest when it ‘says what it thinks,’ in that its outputs match its internal representations of the world \parenciteevans2021truthful. In general, it is easier to develop benchmarks for assessing truthfulness than honesty, since evaluators can directly measure whether outputs are true \parencitelin2022truthfulqa. There are a range of strategies for making models more truthful. For example, one family of approaches uses ‘fine-tuning’ techniques, such as RLHF \parenciteziegler2020finetuning,christianoRLHF and constitutional AI \parenciteaskell2021hhh, bai2022constitutional. Here, AI outputs are rated by human evaluators (RLHF) or AI evaluators (constitutional AI), based on criteria such as perceived helpfulness and honesty, and fine-tuned to train the language model. Unfortunately, models fine-tuned with these methods (including ChatGPT and Claude) still frequently produce misleading outputs. This is in part because fine-tuning can incentivize models towards producing plausible and more convincing outputs, rather than honest ones. In addition, fine-tuning evaluations cannot cover every scenario, and so models can misgeneralize from feedback \parenciteshah2022goal. See \textciteevans2021truthful and \textciteli2023inferencetime for other approaches to training AI systems to be truthful. Training models to be more truthful could also create risk. One way a model could become more truthful is by developing more accurate internal representations of the world. This also makes the model a more effective agent, by increasing its ability to successfully implement plans. For example, creating a more truthful model could actually increase its ability to engage in strategic deception, by giving it more accurate insights into its opponents beliefs and desires. Granted, a maximally truthful system would not deceive, but optimizing for truthfulness could nonetheless increase the capacity for strategic deception. For this reason, it would be valuable to develop techniques for making models more honest (in the sense of causing their outputs to match their internal representations), separately from just making them more truthful. Here, as we discussed earlier, more research is needed in developing reliable techniques for understanding the internal representations of models. In addition, it would be useful to develop tools to control the model’s internal representations, and to control the model’s ability to produce outputs that deviate from its internal representations. As discussed in \textcitezou2023repe, representation control is one promising strategy. They develop a lie detector and can control whether or not an AI lies. If representation control methods become highly reliable, then this would present a way of robustly combating AI deception. Acknowledgements ---------------- We would like to thank Jaeson Booker, Stephen Casper, Emily Dardaman, Isaac Dunn, Maira Elahi, Shashwat Goel, Thilo Hagendorff, Nikola Jurkovic, Alex Khurgin, Jakub Kraus, Nathaniel Li, Isaac Liao, David Manheim, Colin McGlynn, Kyle O’Brien, and Ellie Sakhaee for their thoughtful and helpful comments. We would also like to thank Valtteri Lipiäinen for converting Meta’s CICERO game-log data \parencitecicero into html form. We would additionally like to thank Amanda She for clarifying details about ARC Evals’ experiments with GPT-4 \parenciteAlignmentResearchCenter2023. P.S.P. is funded by the MIT Department of Physics and the Beneficial AI Foundation. Author contribution statement ----------------------------- P.S.P. and S.G. had equal lead-author roles, carrying out the bulk of the paper’s planning and writing. A.O. also contributed substantially throughout the planning and writing of the paper. M.C. ran fact-finding experiments on CICERO. M.C. and D.H. collaborated with S.G. on the section about making AI systems less deceptive. D.H. provided resources for the project through the Center for AI Safety. This project began as a critique of Meta’s claim that CICERO was an honest AI, which was conceived by P.S.P. and pursued by P.S.P., M.C., and D.H. initially. The scope of the project eventually expanded to be a survey paper on AI deception, largely at D.H.’s suggestion. S.G. and A.O. joined the project after the expansion of its scope to be a survey paper on AI deception, and were central to the planning and outline-writing components of this expanded project. \printbibliography Appendices
8fa19a69-ddd9-4530-ae08-f79ce5158b0d	trentmkelly/LessWrong-43k	LessWrong	Nate Silver will do an AMA on Reddit on Tuesday http://www.reddit.com/r/IAmA/comments/163nqk/nate_silver_is_doing_an_ama_tuesday_at_2_pm/ I'm really excited to see this. Nate Silver might be the most famous present day Bayesian statistician. UPD: It appears that author of the Reddit post deleted it for some reason. The link still works but it makes sense to post the link to the Nate Silver blog with his original announcement, just in case: http://fivethirtyeight.blogs.nytimes.com/2013/01/06/ask-nate-anything/
971e92d1-168d-4d78-835f-d2f4c1baf300	trentmkelly/LessWrong-43k	LessWrong	Is it rational to modify one's utility function? Rationality is often informally defined by means-end reasoning or utility maximization. However, this idea becomes less clear when faced with the option of modifying one's own utility function. Does rationality prescribe avoiding any change to one's current utility function because such a change would obviously reduce expected utility under the current function, or does it prescribe taking actions which result in the highest utility by whatever means necessary, in which case a change would be rational iff the new utility function yields higher expected utility given known background info about the world? This is obviously relevant to AI alignment, where one concern is that AI may hack their own utility functions and another concern is that they may prevent humans from modifying them (or shutting them off) due to the risk to their current goals. It's also relevant to questions of human rationality, where, on the one hand, we imagine that Ghandi would not take a pill that makes him want to murder people, but on the other hand, we regularly believe that unhappy people should change their own psychology and goals to be more happy.
ea16a248-433b-420f-a2c8-d70437477a6d	trentmkelly/LessWrong-43k	LessWrong	Analyzing A Critique Of The AI 2027 Timeline Forecasts There was what everyone agrees was a high quality critique of the timelines component of AI 2027, by the LessWrong user and Substack writer Titotal. It is great to have thoughtful critiques like this. The way you get actual thoughtful critiques like this, of course, is to post the wrong answer (at length) on the internet, and then respond by listening to the feedback and by making your model less wrong. This is a high-effort, highly detailed, real engagement on this section, including giving the original authors opportunity to critique the critique, and warnings to beware errors, give time to respond, shares the code used to generate the graphs, engages in detail, does a bunch of math work, and so on. That is The Way. So, Titotal: Thank you. I note up front that at least Daniel Kokotajlo has indeed adjusted his estimates, and has moved his median from ‘AI 2027’ to ‘AI 2028’ based on events since publication, and Eli’s revisions also push the estimates back a bit. I also note up front that if you evaluated most statements made in the discourse (either non-worried AI forecasting, or AI in general, or more broadly) with this level of rigor, mostly you couldn’t because you’d hit ‘I made it up’ very quickly, but in other cases where someone is trying at least a little, in my experience the models fall apart a lot worse and a lot faster. No one has suggested ‘here is a better attempt to forecast the future and take the whole thing seriously’ that I consider to have a reasonable claim to that. A lot of the disagreements come down to how much one should care about which calculations and graphs match past data how closely in different contexts. Titotal demands very strong adherence throughout. I think it’s good to challenge and poke at the gaps but this seems to in several places go too far. TABLE OF CONTENTS 1. The Headline Message Is Not Ideal. 2. An Explanation of Where Superexponentiality Is Coming From. 3. Three Methods. 4. Time Horizon Extension Method
7388f835-8a1c-43ca-9aa1-73ff8906f195	trentmkelly/LessWrong-43k	LessWrong	Be Not Averse to Lost Purposes tldr: Be not averse to lost purposes, as aversions yield ugh fields. Rather, confess them, and fix them joyously where possible. How to achieve this emotional state? I'm not sure, but the appropriate emotion might be to lost purposes as curiosity is to ignorance. This got rambly; sorry. This was originally a reply to Those Who Aspire to Perfection. I may have gotten carried away. ---------------------------------------- I see that learning to notice and avoid Lost Purposes, in your own action, and in the incentive systems that you shape for others, matters deeply. Lost purposes are everywhere, and addressing them promises to be a useful way to improve your effectiveness, and the effectiveness of any incentive systems that you have the ability to modify. An aversion to lost purposes, though, is probably not a good way to attempt rationality and remain sane, especially in our world. When lost purposes are rife, when they're everywhere, and you can see them everywhere, then I would expect having a strong negative reaction to be debilitating. Certainly, when I feel powerless to rectify a situation with lost purposes, my own mild aversion debilitates me from thinking clearly about the situation. Strongly negative emotions are depressing, and depression is debilitating. So, if you have a strong aversion to your own lost purposes, and you don't immediately see ways to deal with them, you're likely to become ineffective at handling them. Worse, perhaps, is that a learned aversion to lost purposes is likely to lead you to stop noticing lost purposes, except when they're painfully obvious. This is generally true when you have a strong negative reaction around noticing anything, according to the idea of Ugh Fields. (In short: if a thought is usually followed by a strong negative reaction, then you will learn to stop having that thought. This learning is entirely subconscious, and can be difficult to uncover.) So, consider alternatives. We value having an aversion to lost
c14f6452-11fb-45c2-84f8-a3a53d057068	trentmkelly/LessWrong-43k	LessWrong	Meetup : Moscow: The Sunday Meetup Discussion article for the meetup : Moscow: The Sunday Meetup WHEN: 01 September 2013 04:00:00PM (+0400) WHERE: Russia, Moscow, ulitsa L'va Tolstogo 16 Please use the following guide to get to the meetup: link. You need the second revolving door. We will meet you at 16:00 with “LW” sign inside the hall or just look for group of geek-looking people. Main topics: * Discussion about motivation. * Discussion about cognitive biases. * Game session. If you are going for the first time, you can fill this one minute form (in Russian), to share your contact information. You can also use personal messages here, or drop a message at lw@lesswrong.ru to contact me for any reason. Discussion article for the meetup : Moscow: The Sunday Meetup
a704ca61-1c7a-45bd-be55-790b0a8946c6	StampyAI/alignment-research-dataset/lesswrong	LessWrong	Responses to Catastrophic AGI Risk: A Survey A great many Less Wrongers gave feedback on earlier drafts of "Responses to Catastrophic AGI Risk: A Survey," which has now been [released](http://intelligence.org/2013/07/08/responses-to-catastrophic-agi-risk-a-survey/). This is the preferred discussion page for the paper. The report, co-authored by past MIRI researcher Kaj Sotala and University of Louisville’s Roman Yampolskiy, is a summary of the extant literature (250+ references) on AGI risk, and can serve either as a guide for researchers or as an introduction for the uninitiated. Here is the abstract: > > Many researchers have argued that humanity will create artificial general intelligence (AGI) within the next twenty to one hundred years. It has been suggested that AGI may pose a catastrophic risk to humanity. After summarizing the arguments for why AGI may pose such a risk, we survey the field’s proposed responses to AGI risk. We consider societal proposals, proposals for external constraints on AGI behaviors, and proposals for creating AGIs that are safe due to their internal design. > > >
e2278832-1022-4999-84c1-bb9e0168351d	trentmkelly/LessWrong-43k	LessWrong	Stupid Questions March 2015 This thread is for asking any questions that might seem obvious, tangential, silly or what-have-you. Don't be shy, everyone has holes in their knowledge, though the fewer and the smaller we can make them, the better. Please be respectful of other people's admitting ignorance and don't mock them for it, as they're doing a noble thing. To any future monthly posters of SQ threads, please remember to add the "stupid_questions" tag.
f5b017ef-d13e-4e2f-b73f-4d6ce33306dc	trentmkelly/LessWrong-43k	LessWrong	Book Review: Free Will Introduction Sam Harris' Free Will isn't a conventional philosophy book. Rather, it's a laconic manifesto full of bold and provocative statements invoking us to free ourselves from the delusion of free will and abolish the whole concept as misleading and unnecessary. The book quickly shatters the naïve layperson’s intuition in the light of scientific advancements, then briefly explains Harris’ dissatisfaction with compatibilism as a half measure, and finally argues that our morality, penal and political systems would only benefit from the dispelling of the illusion which the “free will” is. Or at least it's what the book tries to be. For me, however, it turned out to be something different. While my initial craving for deep arguments in favour of a position I disagree with wasn’t satisfied, I got interesting insight from Harris' attempts at resolving confusion and reinventing existing theories with different aesthetics. Most surprisingly, I got a new perspective on religious tolerance. Predictably, the publication of the book led to a philosophical debate on the matter of free will between Sam Harris and Daniel Dennet which turned out to be larger than the book itself. I’ll touch it a little in this review as well. Main thesis Harris begins his book with a description of a terrible crime. He points out how our perception of this crime can be shifted if we are informed of the underlying causes. But, under scrutiny, these causes go beyond the control of any of the perpetrators, leaving no extra place for their personal responsibility. He uses it as a high stakes example to make his point. > Of course, if we learned that both these men had been suffering from brain tumors that explained their violent behavior, our moral intuitions would shift dramatically. But a neurological disorder appears to be just a special case of physical events giving rise to thoughts and actions. Understanding the neurophysiology of the brain, therefore, would seem to be as exculpatory as
eaf4e25e-1f23-45ff-9738-e9b9432d6535	trentmkelly/LessWrong-43k	LessWrong	Harry Potter and the Methods of Rationality discussion thread, part 10 (The HPMOR discussion thread after this one is here.) This is a new thread to discuss Eliezer Yudkowsky's Harry Potter and the Methods of Rationality and anything related to it. There haven't been any chapters recently, but it looks like there are a bunch in the pipeline and the old thread is nearing 700 comments. The latest chapter as of 7th March 2012 is Ch. 77. There is now a site dedicated to the story at hpmor.com, which is now the place to go to find the authors notes and all sorts of other goodies. AdeleneDawner has kept an archive of Author's Notes. The first 5 discussion threads are on the main page under the harry_potter tag. Threads 6 and on (including this one) are in the discussion section using its separate tag system. Also: one, two, three, four, five, six, seven, eight, nine. As a reminder, it's often useful to start your comment by indicating which chapter you are commenting on. Spoiler Warning: this thread is full of spoilers. With few exceptions, spoilers for MOR and canon are fair game to post, without warning or rot13. More specifically: > You do not need to rot13 anything about HP:MoR or the original Harry Potter series unless you are posting insider information from Eliezer Yudkowsky which is not supposed to be publicly available (which includes public statements by Eliezer that have been retracted). > > If there is evidence for X in MOR and/or canon then it's fine to post about X without rot13, even if you also have heard privately from Eliezer that X is true. But you should not post that "Eliezer said X is true" unless you use rot13.
c991eb56-6937-4b16-ba8f-af148ef20621	trentmkelly/LessWrong-43k	LessWrong	The best approaches for mitigating "the intelligence curse" (or gradual disempowerment); my quick guesses at the best object-level interventions There have recently been various proposals for mitigations to "the intelligence curse" or "gradual disempowerment"—concerns that most humans would end up disempowered (or even dying) because their labor is no longer valuable. I'm currently skeptical that the typically highlighted prioritization and interventions are best and I have some alternative proposals for relatively targeted/differential interventions which I think would be more leveraged (as in, the payoff is higher relative to the difficulty of achieving them). It's worth noting I doubt that these threats would result in huge casualty counts (due to e.g. starvation) or disempowerment of all humans (though substantial concentration of power among a smaller group of humans seems quite plausible).[1] I decided to put a bit of time into writing up my thoughts out of general cooperativeness (e.g., I would want someone in a symmetric position to do the same). (This was a timeboxed effort of ~1.5 hr, so apologies if it is somewhat poorly articulated or otherwise bad. Correspondingly, this post is substantially lower effort than my typical post.) My top 3 preferred interventions focused on these concerns are: * Mandatory interoperability for alignment and fine-tuning: Pass regulation or create a norm that requires AI companies to support all the APIs and interfaces needed to customize their models and (attempt to) align them differently. Either third parties would inspect the implementation (to avoid tampering and to ensure sufficient affordances) or perhaps more robustly, the companies would be required to submit their weights to various (secure) third parties that would implement the relevant APIs. Then, many actors could compete in offering differently fine-tuned models competing over the level of alignment (and the level of alignment to users in particular). This would be using relatively deep model access (not just prompting), e.g. full weight fine-tuning APIs that support arbitrary forward and backward,
64bdfae3-4198-4361-8c91-0d7384590d76	trentmkelly/LessWrong-43k	LessWrong	What are the optimal biases to overcome? If you're interested in learning rationality, where should you start? Remember, instrumental rationality is about making decisions that get you what you want -- surely there are some lessons that will help you more than others. You might start with the most famous ones, which tend to be the ones popularized by Kahneman and Tversky. But K&T were academics. They weren't trying to help people be more rational, they were trying to prove to other academics that people were irrational. The result is that they focused not on the most important biases, but the ones that were easiest to prove. Take their famous anchoring experiment, in which they showed the spin of a roulette wheel affected people's estimates about African countries. The idea wasn't that roulette wheels causing biased estimates was a huge social problem; it was that no academic could possibly argue that this behavior was somehow rational. They thereby scored a decisive blow for psychology against economists claiming we're just rational maximizers. Most academic work on irrationality has followed in K&T's footsteps. And, in turn, much of the stuff done by LW and CFAR has followed in the footsteps of this academic work. So it's not hard to believe that LW types are good at avoiding these biases and thus do well on the psychology tests for them. (Indeed, many of the questions on these tests for rationality come straight from K&T experiments!) But if you look at the average person and ask why they aren't getting what they want, very rarely do you conclude their biggest problem is that they're suffering from anchoring, framing effects, the planning fallacy, commitment bias, or any of the other stuff in the sequences. Usually their biggest problems are far more quotidian and commonsensical. Take Eliezer. Surely he wanted SIAI to be a well-functioning organization. And he's admitted that lukeprog has done more to achieve that goal of his than he has. Why is lukeprog so much better at getting what Eliezer wants
8c4591d8-e5db-4b9a-86d0-21efaa8648df	trentmkelly/LessWrong-43k	LessWrong	Could Advanced AI Accelerate the Pace of AI Progress? Interviews with AI Researchers We interviewed five AI researchers from leading AI companies about a scenario where AI systems fully automate AI capabilities research. To ground the setting, we stipulated that each employee is replaced by 30 digital copies with the same skill set and the ability to think 30 times faster than a human. This represents a 900-fold increase in the cognitive labor that AI companies direct towards advancing AI capabilities. Our key takeaways are: 1. Compute for experiments will likely be a bottleneck. Finding better algorithms often requires running computationally expensive ML experiments. Even with abundant cognitive labor, AI progress would be constrained by the time needed for these experiments to run. 2. AI cognitive labor could probably extract significantly more research insights out of limited compute. AI could increase efficiency by improving experimental design, high-level research directions, and project prioritization. Other strategies include running experiments at smaller scales, generating higher-quality synthetic data, or prioritizing algorithmic improvements that do not require large experiments (e.g., scaffolding). 3. The overall pace of AI progress might be between 2 and 20 times faster in the discussed scenario. However, the researchers emphasized their high degree of uncertainty. 4. Abundant AI labor might significantly improve experiment design and implementation. This includes eliminating subtle bugs, stopping experiments early, constantly monitoring and analyzing experiments, and making every experiment as informative and efficient as the best experiments are today. Some researchers thought this effect would be small, but others thought it could be very significant. 5. The gains to smaller experiments would likely be larger than the gains to larger experiments. The cost of smaller experiments tends to be dominated by the time to code them, and most bugs are removed at small scales before larger runs are attempted. 6. There is uncertainty a
4971c374-d44a-4913-b2bb-e0e86bda22dd	trentmkelly/LessWrong-43k	LessWrong	Cornell Meetup Anyone know people at Cornell interested in AI alignment?
854f7d5f-a177-4f42-bfe2-bd14fe82f6c2	trentmkelly/LessWrong-43k	LessWrong	Summary: Surreal Decisions This post summarises a paper by Eddy Chen and Daniel Rubio on using Surreal numbers to resolve problems of Infinite Ethics. Future posts will argue that surreals are the correct approach to this problem before extending upon this work. However, this post merely aims to summarise this paper. Background The problem of Infinite Paralysis is best described as follows. Suppose that there are infinite people and that they are happy so that there is infinite utility. I then come along and punch 100 people destroying 100 utility. Since there was infinite utility at the start and infinity minus 100 is still infinity, so arguably I’ve done nothing wrong. However, this seems to be a reductio ad absurdum if I’ve ever seen one. One approach mentioned by Bostrom is to use hyperreals to represent infinite sequences of utility. In particular, he sets the ith index of the hyperreal representing utility to the sum of the first i numbers. Unfortunately, there is not a unique definition of the hyperreals - they require what’s called a non-principle ultrafilter to be defined in order to determine the ordering. Our choice of this seems essentially arbitrary and therefore hard to justify principally. Additionally, summation requires a preferred location around which to sum, which can be hard to justify philosophically. Surreals Decisions Chen and Rubio outline a surreal decision theory by adapting the the Von Neumann-Morgenstern axioms. They then use it to analyse Pascal’s Wager to demonstrate that the validity of the argument depends on particular infinite values assigned in the problem and the various deities that exist. They note that Expected Utility Theory with standard infinities (cardinal numbers) seems to produce absurd results. In particular, it is indifferent between the following, when most people would prefer them in order * Infinity or something: Infinite utility if heads, 10000 utility if tails * Infinity or nothing: Infinite utility if heads, 0 if tails * Infinit
c023cca4-45ac-4f18-86cd-79b3989927db	trentmkelly/LessWrong-43k	LessWrong	5 general voting pathologies: lesser names of Moloch Earlier, I wrote a primer on voting theory. Among the things I discussed were 5 types of pathologies suffered by different single-winner voting methods. I presented these as 5 sequential hurdles for voting method design. That is, since they are in what I view as decreasing importance and increasing difficulty, you should check your voting method against each hurdle in order, and stop as soon as it fails to pass. Then I read Eliezer's book on Inadequate Equilibria, and Scott's "Meditations on Moloch". They argue that the point of civilization is to provide mechanisms to get out of pernicious equilibria, and the kakistotropic tendencies of civilization they characterize as "Moloch" are basically cases where pernicious incentives reinforce each other. I realized that the simple two-player games such as Prisoners' Dilemma that serve as intuition pumps for game theory lack some of the characteristics of my 5 voting pathologies. So I want to go back and explain those pathologies more carefully, to help build up intuition about how multi-player, single-outcome games differ from two-player ones. A key point here is that I'm talking about single-winner voting methods; that is, "games" where the number of possible outcomes is far less than the number of players. In this case, it's not a matter of seeking an individual advantage for yourself; the only way for you to win is for your entire faction to win equally. This means that I will not be talking about the oldest and deepest name of Moloch, which is Malthus. All the Molochs in this essay can and should be killed or (mostly) tamed. Also note that this essay is not the one I'd write if I were only trying to recruit the rationalist community to become electoral reform activists. As an activist, I think that the most important and short-term-viable electoral reforms are in the multi-winner space: solving the problem of coordinating public goods not directly through mechanism design, but indirectly through a combination of me
7c22a69f-fd82-4406-bc11-969fb57ed535	StampyAI/alignment-research-dataset/alignmentforum	Alignment Forum	[AN #102]: Meta learning by GPT-3, and a list of full proposals for AI alignment Alignment Newsletter is a weekly publication with recent content relevant to AI alignment around the world. Find all Alignment Newsletter [resources here](http://rohinshah.com/alignment-newsletter/). In particular, you can look through [this spreadsheet](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit?usp=sharing) of all summaries that have ever been in the newsletter. Audio version [here](http://alignment-newsletter.libsyn.com/alignment-newsletter-102) (may not be up yet). SECTIONS ======== [HIGHLIGHTS](about:blank#HIGHLIGHTS) [TECHNICAL AI ALIGNMENT](about:blank#TECHNICAL_AI_ALIGNMENT) [MISCELLANEOUS (ALIGNMENT)](about:blank#MISCELLANEOUS_(ALIGNMENT)) [OTHER PROGRESS IN AI](about:blank#OTHER_PROGRESS_IN_AI) [REINFORCEMENT LEARNING](about:blank#REINFORCEMENT_LEARNING) [DEEP LEARNING](about:blank#DEEP_LEARNING) [HIERARCHICAL RL](about:blank#HIERARCHICAL_RL) HIGHLIGHTS ========== [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) (Tom B. Brown et al) (summarized by Rohin): The biggest [GPT-2 model](https://blog.openai.com/better-language-models/) ([AN #46](https://mailchi.mp/c48f996a5db5/alignment-newsletter-46)) had 1.5 billion parameters, and since its release people have trained language models with up to 17 billion parameters. This paper reports GPT-3 results, where the largest model has 175 billion parameters, a 10x increase over the previous largest language model. To get the obvious out of the way, it sets a new state of the art (SOTA) on zero-shot language modeling (evaluated only on Penn Tree Bank, as other evaluation sets were accidentally a part of their training set). The primary focus of the paper is on analyzing the few-shot learning capabilities of GPT-3. In few-shot learning, after an initial training phase, at test time models are presented with a small number of examples of a new task, and then must execute that task for new inputs. Such problems are usually solved using meta-learning or finetuning, e.g. at test time [MAML](https://arxiv.org/abs/1703.03400) takes a few gradient steps on the new examples to produce a model finetuned for the test task. In contrast, the key hypothesis with GPT-3 is that language is so diverse, that doing well on it already requires adaptation to the input, and so the learned language model will already be a meta-learner. This implies that they can simply "prime" the model with examples of a task they care about, and the model can learn what task is supposed to be performed, and then perform that task well. For example, consider the task of generating a sentence using a newly made-up word whose meaning has been explained. In one notable example, the prompt for GPT-3 is: A "whatpu" is a small, furry animal native to Tanzania. An example of a sentence that uses the word whatpu is: We were traveling in Africa and we saw these very cute whatpus. To do a "farduddle" means to jump up and down really fast. An example of a sentence that uses the word farduddle is: Given this prompt, GPT-3 generates the following example sentence for "farduddle": One day when I was playing tag with my little sister, she got really excited and she started doing these crazy farduddles. The paper tests on several downstream tasks for which benchmarks exist (e.g. question answering), and reports zero-shot, one-shot, and few-shot performance on all of them. On some tasks, the few-shot version sets a new SOTA, despite not being finetuned using the benchmark’s training set; on others, GPT-3 lags considerably behind finetuning approaches. The paper also consistently shows that few-shot performance increases as the number of parameters increases, and the rate of increase is faster than the corresponding rate for zero-shot performance. While they don’t outright say it, we might take this as suggestive evidence that as models get larger, they are more incentivized to learn “general reasoning abilities”. The most striking example of this is in arithmetic, where the smallest 6 models (up to 6.7 billion parameters) have poor performance (< 20% on 2-digit addition), then the next model (13 billion parameters) jumps to > 50% on 2-digit addition and subtraction, and the final model (175 billion parameters) achieves > 80% on 3-digit addition and subtraction and a perfect 100% on 2-digit addition (all in the few-shot regime). They explicitly look for their test problems in the training set, and find very few examples, suggesting that the model really is learning “how to do addition”; further, when it is incorrect, it tends to make mistakes like “forgetting to carry a 1”. On broader impacts, the authors talk about potential misuse, fairness and bias concerns, and energy usage concerns; and say they about these issues what you’d expect. One interesting note: “To understand how low and mid-skill actors think about language models, we have been monitoring forums and chat groups where misinformation tactics, malware distribution, and computer fraud are frequently discussed.” They find that while there was significant discussion of misuse, they found no successful deployments. They also consulted with professional threat analysts about the possibility of well-resourced actors misusing the model. According to the paper: “The assessment was that language models may not be worth investing significant resources in because there has been no convincing demonstration that current language models are significantly better than current methods for generating text, and because methods for “targeting” or “controlling” the content of language models are still at a very early stage.” Rohin's opinion: For a long time, I’ve heard people quietly hypothesizing that with a sufficient diversity of tasks, regular gradient descent could lead to general reasoning abilities allowing for quick adaptation to new tasks. This is a powerful demonstration of this hypothesis. One [critique](https://www.lesswrong.com/posts/ZHrpjDc3CepSeeBuE/gpt-3-a-disappointing-paper) is that GPT-3 still takes far too long to “identify” a task -- why does it need 50 examples of addition in order to figure out that what it should do is addition? Why isn’t 1 sufficient? It’s not like there are a bunch of other conceptions of “addition” that need to be disambiguated. I’m not sure what’s going on mechanistically, but we can infer from the paper that as language models get larger, the number of examples needed to achieve a given level of performance goes down, so it seems like there is some “strength” of general reasoning ability that goes up (see also [this commentary](https://www.gwern.net/newsletter/2020/05)). Still, it would be really interesting to figure out mechanistically how the model is “reasoning”. This also provides some empirical evidence in support of the threat model underlying [inner alignment concerns](https://arxiv.org/abs/1906.01820) ([AN #58](https://mailchi.mp/92b3a9458c2d/an-58-mesa-optimization-what-it-is-and-why-we-should-care)): they are predicated on neural nets that implicitly learn to optimize. (To be clear, I think it provides empirical support for neural nets learning to “reason generally”, not neural nets learning to implicitly “perform search” in pursuit of a “mesa objective” -- see also [Is the term mesa optimizer too narrow?](https://www.alignmentforum.org/posts/nFDXq7HTv9Xugcqaw/is-the-term-mesa-optimizer-too-narrow) ([AN #78](https://mailchi.mp/eef1d6c95d7c/an-78formalizing-power-and-instrumental-convergence-and-the-end-of-year-ai-safety-charity-comparison)).) [An overview of 11 proposals for building safe advanced AI](https://www.alignmentforum.org/posts/fRsjBseRuvRhMPPE5/an-overview-of-11-proposals-for-building-safe-advanced-ai) (Evan Hubinger) (summarized by Rohin): This post describes eleven “full” AI alignment proposals (where the goal is to build a powerful, beneficial AI system using current techniques), and evaluates them on four axes: 1. Outer alignment: Would the optimal policy for the specified loss function be aligned with us? See also [this post](https://www.alignmentforum.org/posts/33EKjmAdKFn3pbKPJ/outer-alignment-and-imitative-amplification). 2. Inner alignment: Will the model that is actually produced by the training process be aligned with us? 3. Training competitiveness: Is this an efficient way to train a powerful AI system? More concretely, if one team had a “reasonable lead” over other teams, would they keep at least some of the lead if they used this algorithm? 4. Performance competitiveness: Will the trained model have good performance (relative to other models that could be trained)? Seven of the eleven proposals are of the form “recursive outer alignment technique” plus “[technique for robustness](https://ai-alignment.com/training-robust-corrigibility-ce0e0a3b9b4d) ([AN #81](https://mailchi.mp/6078fe4f9928/an-81-universality-as-a-potential-solution-to-conceptual-difficulties-in-intent-alignment))”. The recursive outer alignment technique is either [debate](https://blog.openai.com/debate/) ([AN #5](https://mailchi.mp/0ae5d69de63b/alignment-newsletter-5)), [recursive reward modeling](https://medium.com/@deepmindsafetyresearch/scalable-agent-alignment-via-reward-modeling-bf4ab06dfd84) ([AN #34](https://mailchi.mp/f1947668b183/alignment-newsletter-34)), or some flavor of [amplification](https://www.alignmentforum.org/posts/t3AJW5jP3sk36aGoC/capability-amplification) ([AN #42](https://mailchi.mp/f6488137d76c/alignment-newsletter-42)). The technique for robustness is either transparency tools to “peer inside the model”, [relaxed adversarial training](https://www.alignmentforum.org/posts/9Dy5YRaoCxH9zuJqa/relaxed-adversarial-training-for-inner-alignment) ([AN #70](https://mailchi.mp/732eaa192df0/an-70-agents-that-help-humans-who-are-still-learning-about-their-own-preferences)), or intermittent oversight by a competent supervisor. An additional two proposals are of the form “non-recursive outer alignment technique” plus “technique for robustness” -- the non-recursive techniques are vanilla reinforcement learning in a multiagent environment, and narrow reward learning. Another proposal is Microscope AI, in which we train AI systems to simply understand vast quantities of data, and then by peering into the AI system we can learn the insights that the AI system learned, leading to a lot of value. We wouldn’t have the AI system act in the world, thus eliminating a large swath of potential bad outcomes. Finally, we have STEM AI, where we try to build an AI system that operates in a sandbox and is very good at science and engineering, but doesn’t know much about humans. Intuitively, such a system would be very unlikely to deceive us (and probably would be incapable of doing so). The post contains a lot of additional content that I didn’t do justice to in this summary. In particular, I’ve said nothing about the analysis of each of these proposals on the four axes listed above; the full post talks about all 44 combinations. Rohin's opinion: I’m glad this post exists: while most of the specific proposals could be found by patching together content spread across other blog posts, there was a severe lack of a single article laying out a full picture for even one proposal, let alone all eleven in this post. I usually don’t think about outer alignment as what happens with optimal policies, as assumed in this post -- when you’re talking about loss functions in the real world (as I think this post is trying to do), optimal behavior can be weird and unintuitive, in ways that may not actually matter. For example, arguably for any loss function, the optimal policy is to hack the loss function so that it always outputs zero (or perhaps negative infinity). TECHNICAL AI ALIGNMENT ====================== MISCELLANEOUS (ALIGNMENT) ------------------------- [Planning with Uncertain Specifications](https://arxiv.org/abs/1906.03218) (Ankit Shah et al) (summarized by Rohin): Suppose you recognize that there are no “certain specifications”, and so infer a distribution over specifications. What do you then do with that distribution? This paper looks at this problem in the context where the specifications are given by formulas in linear temporal logic (which can express temporal non-Markovian constraints). They identify four possibilities: 1. Most likely: Plan with respect to the most likely specification. 2. Most coverage: Satisfying as many formulas as possible, ignoring their probability (as long as they have non-zero probability) 3. Chance constrained: Like the above, except you weight by probabilities, and drop the least likely formulas up to a parameter δ. 4. Least regret: Like the above, with δ set to zero. Intuitively, the Most likely criterion won’t be very robust since it is only taking one specification into account, Most coverage is aiming for maximum robustness, Chance constrained interpolates, where larger δ corresponds to trading robustness for gain in ability. This is exactly the pattern we see in a task where a robot must set a dinner table. Rohin's opinion: Ultimately, I hope that in cases like this, the agent plans conservatively initially, but also tries to learn which specification is actually correct, allowing it to become more bold over time. Nonetheless, it seems quite difficult to do this well, and even then we likely will have this tradeoff between robustness and task performance. This is the case with humans too: if you try to please everyone (robustness), you’ll end up pleasing no one (task performance). OTHER PROGRESS IN AI ==================== REINFORCEMENT LEARNING ---------------------- [Suphx: Mastering Mahjong with Deep Reinforcement Learning](https://arxiv.org/abs/2003.13590) (Junjie Li et al) (summarized by Rohin): Mahjong is a large imperfect information game with complex rules where turn order can be interrupted. This makes it challenging to solve with existing techniques like MCTS and counterfactual regret minimization. This paper details what was necessary to build Suphx, an AI system that is stronger than 99.99% of humans. Some highlights: - Like the original AlphaGo, they first learned from human gameplay and then finetuned using reinforcement learning, with deep CNNs as their models. They learned both action models as well as value models. They added an entropy bonus to ensure that the policy remained stochastic enough to continue learning over the course of RL. - They have five learned action models, corresponding to five different decisions that need to be made in Mahjong, as well as a rule-based system for deciding whether or not to declare a winning hand. - To handle imperfect information, they first train an oracle agent that gets access to all information, and then slowly reduce the amount of information that it gets to observe. - They could use search to improve the performance online, but did not do so in their evaluation (since Suphx was playing on a website with time constraints). Suphx with search would probably be significantly stronger. Rohin's opinion: I am a bit curious how they removed observations from the oracle agent, given that you usually have to keep the structure of the input to a neural net constant. Perhaps they simply zeroed out the observations they didn't want? [Mastering Complex Control in MOBA Games with Deep Reinforcement Learning](http://arxiv.org/abs/1912.09729) (Deheng Ye et al) (summarized by Rohin): This paper presents an AI system that can play the Multi-player Online Battle Arena (MOBA) game Honor of Kings. They are inspired by [OpenAI Five](https://blog.openai.com/openai-five/) ([AN #13](https://mailchi.mp/8234356e4b7f/alignment-newsletter-13)) (and Honor of Kings sounds quite similar to Dota, though it is 1v1 instead of 5v5), and have a similar learning setup: reinforcement learning using PPO. Their architecture requires an off-policy algorithm (I’m not sure why, maybe they have stale parameters across their rollout servers), so they add an importance sampling correction to the PPO objective, as well as an additional type of gradient clipping. The input is a combination of the image and underlying game state info. The resulting agents are able to beat top human players, and in an event with the public, the AI system lost only 4 out of 2100 matches. Unlike OpenAI Five, this required only around 100 hours to train (though it’s unclear how much compute was used). DEEP LEARNING ------------- [More Efficient NLP Model Pre-training with ELECTRA](https://ai.googleblog.com/2020/03/more-efficient-nlp-model-pre-training.html) (Kevin Clark et al) (summarized by Flo): There are two main approaches to pretraining for NLP, language models (LMs) which iteratively predict the next word in a given incomplete sentence, and masked language models (MLMs), which predict the identities of a few masked words in an otherwise complete sentence. While not just looking at the previous words (bidirectionality) can be advantageous, MLMs only learn to predict the masked words, which reduces how much is learnt from a given sentence. The authors present an alternative approach, ELECTRA, that outperforms RoBERTa while requiring less than a third of the compute. This is achieved by changing the form of the pretraining task from predicting words to discriminating fake words: Instead of masking, some words are replaced by words generated by an MLM and the trained model has to classify these as fake. This way, we get bidirectionality, but also a more dense signal, as the model has to produce an output for every single word, not just the masked ones. While this looks similar to GANs, the generator is only trained on the usual MLM loss and is not incentivized to fool the discriminator, as GANs don't seem to work well on sequence data. Read more: [Paper: ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators](https://openreview.net/forum?id=r1xMH1BtvB) Flo's opinion: I found it a bit surprising that replacing word prediction with fake discrimination would help that much, but from the ablations, it seems like this is really mostly an instrument to get a loss signal for every single word, which is a cool idea. On a more zoomed-out perspective, results like this seem to show that gains in [algorithmic efficiency](https://openai.com/blog/ai-and-efficiency/) ([AN #99](https://mailchi.mp/4f7ffc5cbe53/an-99-doubling-times-for-the-efficiency-of-ai-algorithms)) are not fundamentally slowing down. HIERARCHICAL RL --------------- [DADS: Unsupervised Reinforcement Learning for Skill Discovery](https://ai.googleblog.com/2020/05/dads-unsupervised-reinforcement.html) (Archit Sharma et al) (summarized by Rohin): Reinforcement learning in robotics typically plans directly on low-level actions. However, it sure seems like there are a simple set of primitives like walking, running, shuffling, etc. that are inherent to the robot morphology. What if we could learn these primitives, and then plan using those primitives? This paper introduces a method for learning these primitives without a reward function. They simply optimize skills for predictability and diversity (by optimizing the mutual information between the current state and next state, conditioned on which skill is being executed). They can then use these primitives for model-based planning for a downstream task. You can think of this as a regular RL problem, except that an action in their "action space" takes the form "execute skill X for T timesteps". They use model-predictive control (MPC), in which you sample a bunch of trajectories, and execute the first action of the trajectory that gets the highest reward. Since each of their high-level actions determines the policy for T timesteps, they can scale to much longer horizon tasks than MPC can usually be used for. They show that this approach is competitive with regular model-based RL. Read more: [Paper: Dynamics-Aware Unsupervised Discovery of Skills](https://arxiv.org/abs/1907.01657) Rohin's opinion: I think unsupervised learning is likely to be key in getting more powerful and general AI systems without requiring a truly staggering amount of expert data, and this is a great example of what that might look like. Note though that the learned primitives are certainly not what you'd expect of a human: for example, the humanoid learns to vaguely shuffle in a direction, rather than walking. In addition, they did require specifying an "x-y prior" that required skills to be diverse based on x-y coordinates, which is why the skills learned navigation primitives, as opposed to e.g. distinct types of flailing. #### FEEDBACK I'm always happy to hear feedback; you can send it to me, [Rohin Shah](https://rohinshah.com/), by replying to this email. #### PODCAST An audio podcast version of the Alignment Newsletter is available. This podcast is an audio version of the newsletter, recorded by [Robert Miles](http://robertskmiles.com/).
001bc982-1445-4bcf-90b6-771480bb8a3d	StampyAI/alignment-research-dataset/arbital	Arbital	Nick Bostrom's book Superintelligence Nick Bostrom's "[Superintelligence: Paths, Dangers, Strategies](amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0198739834)" is the original only book-form introduction to the basics of [AI alignment theory](https://arbital.com/p/2v).
4960e3c1-98e4-4e60-b674-9ecaa0ccb6c5	trentmkelly/LessWrong-43k	LessWrong	Archiving link posts? Link rot is a huge problem. At the same time, many posts on Less Wrong—including some of the most important posts, which talk about important concepts or otherwise advance our collective knowledge and understanding—are link posts, which means that a non-trivial chunk of our content is hosted elsewhere—across a myriad other websites. If Less Wrong means to be a repository of the rationality community’s canon, we must take seriously the fact that (as gwern’s research indicates) many or most of those externally-hosted pages will, in a few years, no longer be accessible. I’ve taken the liberty of putting together a quick-and-dirty solution. This is a page that, when loaded, scrapes the external links (i.e., the link-post targets) from the front page of GreaterWrong, and automatically submits them to archive.is (after checking each link to see whether it’s already been submitted). A cronjob that loads the page daily ensures that as new link-posts are posted, they will automatically be captured and submitted to archive.is. This solution does not currently have any way to scrape and submit links older than those which are on the front page today (2018-09-08). It is also not especially elegant. It may be advisable to implement automatic link-post archiving as a feature of Less Wrong itself. (Programmatically submitting URLs to archive.is is extremely simple. You send a POST request to http://archive.is/submit/, with a single field, url, with the URL as its value. The URL of the archived content will then—after some time, as archiving is not instantaneous—be accessible via http://archive.is/timegate/[the complete original URL].)
eca73a58-89f2-4dc3-902e-fa9940ba08ae	trentmkelly/LessWrong-43k	LessWrong	Active AGI and/or FAI researchers I am curious in general about who here, if anyone, is actively researching the AGI and/or FAI problems directly in a full-time capacity (or soon will be). So if that's you, please say hello! Or if you know another website/mailing-list/etc this question would be more appropriate to ask, please let me know. If you are interested in saying more about who you are and what you're doing, I've included some additional questions below. Feel free to provide as much or as little information as you'd like - but the more the better! * Are you working for a particular organization or a known AGI project? If so, which? Link? If not, are you working on these issues independently, or can you otherwise explain your situation? * What is your overall theory/philosophy on FAI/AGI? How are you similar to and how are you different from Eliezer Yudkowsky in this respect? * What has been your overall approach to study for this line of research / what specific curriculum and what specific books/papers/etc would you recommend? I would be interested in as much detail as you can provide here. * Do you have any published material (even informal/in-progress information, documentation, discussion, blogs, etc)? Links? * What are you working on now and what's coming next in your work? Are you solving some interesting problem, creating some interesting new idea, bringing together a grand theory, actually building a working FAI/AGI or similar, or approaching some big milestone along any of these paths or others? What do your plans and timeline for the future look like?
a591e7a8-b47d-4a62-9f61-9cc5e7a08572	trentmkelly/LessWrong-43k	LessWrong	Where do you get your capabilities from? In both alignment and capabilities research, people end up discussing the effects and feasibility of different AI architectures, and when this happens, I tend to focus on the question: Where do you get your capabilities from? Capabilities are possible because the world has structure that leads to there being common subtasks that are often useful to solve. In order to create an AI with capabilities, there must be some process which encodes solutions to these common subtasks into that AI. Good Old Fashioned AI and Ordinary programming In GOFAI and ordinary programming, the programmer notices what those common subtasks are, and designs specialized algorithms to represent and solve them. This means that the programmer manually designs the capabilities for the AI, using more advanced capabilities that the programmer has. The main difficulty with this is that there are lots of common subtasks, many of which are very complex and therefore hard to model. Manually noticing and designing all of them takes too many programmer resources to be viable. Consequentialism Consequentialism, broadly defined, is a general and useful way to develop capabilities. Under consequentialism, you consider the consequences of different things you could do, and apply a search process (of which there are many) to select an approach that works. Naively, the fact that consequentialism works is tautological; if you choose the option that works, then it works. In practice, the challenge for consequentialism comes from embedded agency, with perhaps the most significant challenge being that you need some good map/model of what happens as you apply different choices, so you can know what the consequences are. Consequentialism is a big part of how humans act. If you take some action (such as moving a chair), you usually have some purpose in mind for why you took it (such as intending to sit on the chair), and you usually wouldn't have taken the same action if you thought it would have led to a ra

Subsets and Splits

No community queries yet

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