The Best of LessWrong

AI ALIGNMENT FORUM
The Best of LessWrong
AF

6Daniel Kokotajlo

Ajeya's timelines report is the best thing that's ever been written about AI timelines imo. Whenever people ask me for my views on timelines, I go through the following mini-flowchart: 1. Have you read Ajeya's report? --If yes, launch into a conversation about the distribution over 2020's training compute and explain why I think the distribution should be substantially to the left, why I worry it might shift leftward faster than she projects, and why I think we should use it to forecast AI-PONR instead of TAI. --If no, launch into a conversation about Ajeya's framework and why it's the best and why all discussion of AI timelines should begin there. So, why do I think it's the best? Well, there's a lot to say on the subject, but, in a nutshell: Ajeya's framework is to AI forecasting what actual climate models are to climate change forecasting (by contrast with lower-tier methods such as "Just look at the time series of temperature over time / AI performance over time and extrapolate" and "Make a list of factors that might push the temperature up or down in the future / make AI progress harder or easier," and of course the classic "poll a bunch of people with vaguely related credentials." There's something else which is harder to convey... I want to say Ajeya's model doesn't actually assume anything, or maybe it makes only a few very plausible assumptions. This is underappreciated, I think. People will say e.g. "I think data is the bottleneck, not compute." But Ajeya's model doesn't assume otherwise! If you think data is the bottleneck, then the model is more difficult for you to use and will give more boring outputs, but you can still use it. (Concretely, you'd have 2020's training compute requirements distribution with lots of probability mass way to the right, and then rather than say the distribution shifts to the left at a rate of about one OOM a decade, you'd input whatever trend you think characterizes the likely improvements in data gathering.) The upsho

9Daniel Kokotajlo

I still think this is great. Some minor updates, and an important note: Minor updates: I'm a bit less concerned about AI-powered propaganda/persuasion than I was at the time, not sure why. Maybe I'm just in a more optimistic mood. See this critique for discussion. It's too early to tell whether reality is diverging from expectation on this front. I had been feeling mildly bad about my chatbot-centered narrative, as of a month ago, but given how ChatGPT was received I think things are basically on trend. Diplomacy happened faster than I expected, though in a less generalizeable way than I expected, so whatever. My overall timelines have shortened somewhat since I wrote this story, but it's still the thing I point people towards when they ask me what I think will happen. (Note that the bulk of my update was from publicly available info rather than from nonpublic stuff I saw at OpenAI.) Important note: When I wrote this story, my AI timelines median was something like 2029. Based on how things shook out as the story developed it looked like AI takeover was about to happen, so in my unfinished draft of what 2027 looks like, AI takeover happens. (Also AI takeoff begins, I hadn't written much about that part but probably it would reach singularity/dysonswarms/etc. in around 2028 or 2029.) That's why the story stopped, I found writing about takeover difficult and confusing & I wanted to get the rest of the story up online first. Alas, I never got around to finishing the 2027 story. I'm mentioning this because I think a lot of readers with 20+ year timelines read my story and were like "yep seems about right" not realizing that if you look closely at what's happening in the story, and imagine it happening in real life, it would be pretty strong evidence that crazy shit was about to go down. Feel free to controvert that claim, but the point is, I want it on the record that when this original 2026 story was written, I envisioned the proper continuation of the story resultin

8johnswentworth

In a field like alignment or embedded agency, it's useful to keep a list of one or two dozen ideas which seem like they should fit neatly into a full theory, although it's not yet clear how. When working on a theoretical framework, you regularly revisit each of those ideas, and think about how it fits in. Every once in a while, a piece will click, and another large chunk of the puzzle will come together. Selection vs control is one of those ideas. It seems like it should fit neatly into a full theory, but it's not yet clear what that will look like. I revisit the idea pretty regularly (maybe once every 3-4 months) to see how it fits with my current thinking. It has not yet had its time, but I expect it will (that's why it's on the list, after all). Bearing in mind that the puzzle piece has not yet properly clicked, here are some current thoughts on how it might connect to other pieces: * Selection and control have different type signatures. * A selection process optimizes for the values of variables in some model, which may or may not correspond anything in the real world. Human values seem to be like this - see Human Values Are A Function Of Humans' Latent Variables. * A control process, on the other hand, directly optimizes things in its environment. A thermostat, for instance, does not necessarily contain any model of the temperature a few minutes in the future; it just directly optimizes the value of the temperature a few minutes in the future. * The post basically says it, but it's worth emphasizing: reinforcement learning is a control process, expected utility maximization is a selection process. The difference in type signatures between RL and EU maximization is the same as the difference in type signatures between selection and control. * Inner and outer optimizers can have different type signatures: an outer controller (e.g. RL) can learn an inner selector (e.g. utility maximizer), or an outer selector (e.g. a human) can build an inner controller (e

32habryka

I think this post might be the best one of all the MIRI dialogues. I also feel confused about how to relate to the MIRI dialogues overall. A lot of the MIRI dialogues consist of Eliezer and Nate saying things that seem really important and obvious to me, and a lot of my love for them comes from a feeling of "this actually makes a bunch of the important arguments for why the problem is hard". But the nature of the argument is kind of closed off. Like, I agree with these arguments, but like, if you believe these arguments, having traction on AI Alignment becomes much harder, and a lot of things that people currently label "AI Alignment" kind of stops feeling real, and I have this feeling that even though a really quite substantial fraction of the people I talk to about AI Alignment are compelled by Eliezer's argument for difficulty, that there is some kind of structural reason that AI Alignment as a field can't really track these arguments. Like, a lot of people's jobs and funding rely on these arguments being false, and also, if these arguments are correct, the space of perspectives on the problem suddenly loses a lot of common ground on how to proceed or what to do, and it isn't really obvious that you even want an "AI Alignment field" or lots of "AI Alignment research organizations" or "AI Alignment student groups". Like, because we don't know how to solve this problem, it really isn't clear what the right type of social organization is, and there aren't obviously great gains from trade, and so from a coalition perspective, you don't get a coalition of people who think these arguments are real. I feel deeply confused about this. Over the last two years, I think I wrongly ended up just kind of investing into an ecosystem of people that somewhat structurally can't really handle these arguments, and makes plans that assume that these arguments are false, and in doing so actually mostly makes the world worse, by having a far too optimistic stance on the differen

2Raemon

I haven't had time to reread this sequence in depth, but I wanted to at least touch on how I'd evaluate it. It seems to be aiming to be both a good introductory sequence, while being a "complete and compelling case I can for why the development of AGI might pose an existential threat". The question is who is this sequence for, what is it's goal, and how does it compare to other writing targeting similar demographics. Some writing that comes to mind to compare/contrast it with includes: * Scott Alexander's Superintelligence FAQ. This is the post I've found most helpful for convincing people (including myself), that yes, AI is just actually a big deal and an extinction risk. It's 8000 words. It's written fairly entertainingly. What I find particularly compelling here are a bunch of factual statements about recent AI advances that I hadn't known about at the time. * Tim Urban's Road To Superintelligence series. This is even more optimized for entertainingness. I recall it being a bit more handwavy and making some claims that were either objectionable, or at least felt more objectionable. It's 22,000 words. * Alex Flint's AI Risk for Epistemic Minimalists. This goes in a pretty different direction – not entertaining, and not really comprehensive either . It came to mind because it's doing a sort-of-similar thing of "remove as many prerequisites or assumptions as possible". (I'm not actually sure it's that helpful, the specific assumptions it's avoiding making don't feel like issues I expect to come up for most people, and then it doesn't make a very strong claim about what to do) (I recall Scott Alexander once trying to run a pseudo-study where he had people read a randomized intro post on AI alignment, I think including his own Superintelligence FAQ and Tim Urban's posts among others, and see how it changed people's minds. I vaguely recall it didn't find that big a difference between them. I'd be curious how this compared) At a glance, AGI Safety From First P

9Vanessa Kosoy

This post states a subproblem of AI alignment which the author calls "the pointers problem". The user is regarded as an expected utility maximizer, operating according to causal decision theory. Importantly, the utility function depends on latent (unobserved) variables in the causal network. The AI operates according to a different, superior, model of the world. The problem is then, how do we translate the utility function from the user's model to the AI's model? This is very similar to the "ontological crisis" problem described by De Blanc, only De Blanc uses POMDPs instead of causal networks, and frames it in terms of a single agent changing their ontology, rather than translation from user to AI. The question the author asks here is important, but not that novel (the author himself cites Demski as prior work). Perhaps the use of causal networks is a better angle, but this post doesn't do much to show it. Even so, having another exposition of an important topic, with different points of emphasis, will probably benefit many readers. The primary aspect missing from the discussion in the post, in my opinion, is the nature of the user as a learning agent. The user doesn't have a fixed world-model: or, if they do, then this model is best seen as a prior. This observation hints at the resolution of the apparent paradox wherein the utility function is defined in terms of a wrong model. But it still requires us to explain how the utility is defined s.t. it is applicable to every hypothesis in the prior. (What follows is no longer a "review" per se, inasmuch as a summary of my own thoughts on the topic.) Here is a formal model of how a utility function for learning agents can work, when it depends on latent variables. Fix A a set of actions and O a set of observations. We start with an ontological model which is a crisp infra-POMPD. That is, there is a set of states Sont, an initial state s0ont∈Sont, a transition infra-kernel Tont:Sont×A→□(Sont×O) and a reward functio

7johnswentworth

Why This Post Is Interesting This post takes a previously-very-conceptually-difficult alignment problem, and shows that we can model this problem in a straightforward and fairly general way, just using good ol' Bayesian utility maximizers. The formalization makes the Pointers Problem mathematically legible: it's clear what the problem is, it's clear why the problem is important and hard for alignment, and that clarity is not just conceptual but mathematically precise. Unfortunately, mathematical legibility is not the same as accessibility; the post does have a wide inductive gap. Warning: Inductive Gap This post builds on top of two important pieces for modelling embedded agents which don't have their own posts (to my knowledge). The pieces are: * Lazy world models * Lazy utility functions (or value functions more generally) In hindsight, I probably should have written up separate posts on them; they seem obvious once they click, but they were definitely not obvious beforehand. Lazy World Models One of the core conceptual difficulties of embedded agency is that agents need to reason about worlds which are bigger than themselves. They're embedded in the world, therefore the world must be as big as the entire agent plus whatever environment the world includes outside of the agent. If the agent has a model of the world, the physical memory storing that model must itself fit inside of the world. The data structure containing the world model must represent a world larger than the storage space the data structure takes up. That sounds tricky at first, but if you've done some functional programming before, then data structures like this actually pretty run-of-the-mill. For instance, we can easily make infinite lists which take up finite memory. The trick is to write a generator for the list, and then evaluate it lazily - i.e. only query for list elements which we actually need, and never actually iterate over the whole thing. In the same way, we can represent

7Vanessa Kosoy

This post is a great review of the Natural Abstractions research agenda, covering both its strengths and weaknesses. It provides a useful breakdown of the key claims, the mathematical results and the applications to alignment. There's also reasonable criticism. To the weaknesses mentioned in the overview, I would also add that the agenda needs more engagement with learning theory. Since the claim is that all minds learn the same abstractions, it seems necessary to look into the process of learning, and see what kind of abstractions can or cannot be learned (both in terms of sample complexity and in terms of computational complexity). Some thoughts about natural abstractions inspired by this post: * The concept of natural abstractions seems closely related to my informally conjectured agreement theorem for infra-Bayesian physicalism. In a nutshell, two physicalist agents in the same universe with access to "similar" information should asymptotically arrive at similar beliefs (notably this is false for cartesian agents because of the different biases resulting from the different physical points of view). * A possible formalization of the agreement theorem inspired by my richness of mathematics conjecture: Given two beliefs Ψ and Φ, we say that Ψ⪯Φ when some conditioning of Ψ on a finite set of observations produces a refinement of some conditioning of Φ on a finite set of observations (see linked shortform for mathematical details). This relation is a preorder. In general, we can expect an agent to learn a sequence of beliefs of the form Ψ0≺Ψ1≺Ψ2≺… Here, the sequence can be over physical time, or over time discount or over a parameter such as "availability of computing resources" or "how much time the world allows you for thinking between decisions": the latter is the natural asymptotic for metacognitive agents (see also logical time). Given two agents, we get two such sequences {Ψi} and {Φi}. The agreement theorem can then state that for all i∈N, there exists j

17Vanessa Kosoy

In this essay, Rohin sets out to debunk what ey perceive as a prevalent but erroneous idea in the AI alignment community, namely: "VNM and similar theorems imply goal-directed behavior". This is placed in the context of Rohin's thesis that solving AI alignment is best achieved by designing AI which is not goal-directed. The main argument is: "coherence arguments" imply expected utility maximization, but expected utility maximization does not imply goal-directed behavior. Instead, it is a vacuous constraint, since any agent policy can be regarded as maximizing the expectation of some utility function. I have mixed feelings about this essay. On the one hand, the core argument that VNM and similar theorems do not imply goal-directed behavior is true. To the extent that some people believed the opposite, correcting this mistake is important. On the other hand, (i) I don't think the claim Rohin is debunking is the claim Eliezer had in mind in those sources Rohin cites (ii) I don't think that the conclusions Rohin draws or at least implies are the right conclusions. The actual claim that Eliezer was making (or at least my interpretation of it) is, coherence arguments imply that if we assume an agent is goal-directed then it must be an expected utility maximizer, and therefore EU maximization is the correct mathematical model to apply to such agents. Why do we care about goal-directed agents in the first place? The reason is, on the one hand goal-directed agents are the main source of AI risk, and on the other hand, goal-directed agents are also the most straightforward approach to solving AI risk. Indeed, if we could design powerful agents with the goals we want, these agents would protect us from unaligned AIs and solve all other problems as well (or at least solve them better than we can solve them ourselves). Conversely, if we want to protect ourselves from unaligned AIs, we need to generate very sophisticated long-term plans of action in the physical world, possibl

19habryka

I've been thinking about this post a lot since it first came out. Overall, I think it's core thesis is wrong, and I've seen a lot of people make confident wrong inferences on the basis of it. The core problem with the post was covered by Eliezer's post "GPTs are Predictors, not Imitators" (which was not written, I think, as a direct response, but which still seems to me to convey the core problem with this post): The Simulators post repeatedly alludes to the loss function on which GPTs are trained corresponding to a "simulation objective", but I don't really see why that would be true. It is technically true that a GPT that perfectly simulates earth, including the creation of its own training data set, can use that simulation to get perfect training loss. But actually doing so would require enormous amounts of compute and we of course know that nothing close to that is going on inside of GPT-4. To me, the key feature of a "simulator" would be a process that predicts the output of a system by developing it forwards in time, or some other time-like dimension. The predictions get made by developing an understanding of the transition function of a system between time-steps (the "physics" of the system) and then applying that transition function over and over again until your desired target time. I would be surprised if this is how GPT works internally in its relationship to the rest of the world and how it makes predictions. The primary interesting thing that seems to me true about GPT-4s training objective is that it is highly myopic. Beyond that, I don't see any reason to think of it as particularly more likely to create something that tries to simulate the physics of any underlying system than other loss functions one could choose. When GPT-4 encounters a hash followed by the pre-image of that hash, or a complicated arithmetic problem, or is asked a difficult factual geography question, it seems very unlikely that the way GPT-4 goes about answering that qu

6janus

I think Simulators mostly says obvious and uncontroversial things, but added to the conversation by pointing them out for those who haven't noticed and introducing words for those who struggle to articulate. IMO people that perceive it as making controversial claims have mostly misunderstood its object-level content, although sometimes they may have correctly hallucinated things that I believe or seriously entertain. Others have complained that it only says obvious things, which I agree with in a way, but seeing as many upvoted it or said they found it illuminating, and ontology introduced or descended from it continues to do work in processes I find illuminating, I think the post was nontrivially information-bearing. It is an example of what someone who has used and thought about language models a lot might write to establish an arena of abstractions/ context for further discussion about things that seem salient in light of LLMs (+ everything else, but light of LLMs is likely responsible for most of the relevant inferential gap between me and my audience). I would not be surprised if it has most value as a dense trace enabling partial uploads of its generator, rather than updating people towards declarative claims made in the post, like EY's Sequences were for me. Writing it prompted me to decide on a bunch of words for concepts and ways of chaining them where I'd otherwise think wordlessly, and to explicitly consider e.g. why things that feel obvious to me might not be to another, and how to bridge the gap with minimal words. Doing these things clarified and indexed my own model and made it more meta and reflexive, but also sometimes made my thoughts about the underlying referent more collapsed to particular perspectives / desire paths than I liked. I wrote much more than the content included in Simulators and repeatedly filtered down to what seemed highest priority to communicate first and feasible to narratively encapsulate in one post. If I tried again now i

7Daniel Kokotajlo

(I am the author) I still like & endorse this post. When I wrote it, I hadn't read more than the wiki articles on the subject. But then afterwards I went and read 3 books (written by historians) about it, and I think the original post held up very well to all this new info. In particular, the main critique the post got -- that disease was more important than I made it sound, in a way that undermined my conclusion -- seems to have been pretty wrong. (See e.g. this comment thread, these follow up posts) So, why does it matter? What contribution did this post make? Well, at the time -- and still now, though I think I've made a dent in the discourse -- quite a lot of people I respect (such as people at OpenPhil) seemed to think unaligned AGI would need god-like powers to be able to take over the world -- it would need to be stronger than the rest of the world combined! I think this is based on a flawed model of how takeover/conquest works, and history contains plenty of counterexamples to the model. The conquistadors are my favorite counterexample from my limited knowledge of history. (The flawed model goes by the name of "The China Argument," at least in my mind. You may have heard the argument before -- China is way more capable than the most capable human, yet it can't take over the world; therefore AGI will need to be way way more capable than the most powerful human to take over the world.) Needless to say, this is a somewhat important crux, as illustrated by e.g. Joe Carlsmith's report, which assigns a mere 40% credence to unaligned APS-AI taking over the world even conditional on it escaping and seeking power and managing to cause at least a trillion dollars worth of damage. (I've also gotten feedback from various people at OpenPhil saying that this post was helpful to them, so yay!) I've since written a sequence of posts elaborating on this idea: Takeoff and Takeover in the Past and Future. Alas, I still haven't written the capstone posts in the sequence, t

31johnswentworth

This review is mostly going to talk about what I think the post does wrong and how to fix it, because the post itself does a good job explaining what it does right. But before we get to that, it's worth saying up-front what the post does well: the post proposes a basically-correct notion of "power" for purposes of instrumental convergence, and then uses it to prove that instrumental convergence is in fact highly probable under a wide range of conditions. On that basis alone, it is an excellent post. I see two (related) central problems, from which various other symptoms follow: 1. POWER offers a black-box notion of instrumental convergence. This is the right starting point, but it needs to be complemented with a gears-level understanding of what features of the environment give rise to convergence. 2. Unstructured MDPs are a bad model in which to formulate instrumental convergence. In particular, they are bad for building a gears-level understanding of what features of the environment give rise to convergence. Some things I've thought a lot about over the past year seem particularly well-suited to address these problems, so I have a fair bit to say about them. Why Unstructured MDPs Are A Bad Model For Instrumental Convergence The basic problem with unstructured MDPs is that the entire world-state is a single, monolithic object. Some symptoms of this problem: * it's hard to talk about "resources", which seem fairly central to instrumental convergence * it's hard to talk about multiple agents competing for the same resources * it's hard to talk about which parts of the world an agent controls/doesn't control * it's hard to talk about which parts of the world agents do/don't care about * ... indeed, it's hard to talk about the world having "parts" at all * it's hard to talk about agents not competing, since there's only one monolithic world-state to control * any action which changes the world at all changes the entire world-state; there's no built-in w

6TurnTrout

One year later, I remain excited about this post, from its ideas, to its formalisms, to its implications. I think it helps us formally understand part of the difficulty of the alignment problem. This formalization of power and the Attainable Utility Landscape have together given me a novel frame for understanding alignment and corrigibility. Since last December, I’ve spent several hundred hours expanding the formal results and rewriting the paper; I’ve generalized the theorems, added rigor, and taken great pains to spell out what the theorems do and do not imply. For example, the main paper is 9 pages long; in Appendix B, I further dedicated 3.5 pages to exploring the nuances of the formal definition of ‘power-seeking’ (Definition 6.1). However, there are a few things I wish I’d gotten right the first time around. Therefore, I’ve restructured and rewritten much of the post. Let’s walk through some of the changes. ‘Instrumentally convergent’ replaced by ‘robustly instrumental’ Like many good things, this terminological shift was prompted by a critique from Andrew Critch. Roughly speaking, this work considered an action to be ‘instrumentally convergent’ if it’s very probably optimal, with respect to a probability distribution on a set of reward functions. For the formal definition, see Definition 5.8 in the paper. This definition is natural. You can even find it echoed by Tony Zador in the Debate on Instrumental Convergence: (Zador uses “set of scenarios” instead of “set of reward functions”, but he is implicitly reasoning: “with respect to my beliefs about what kind of objective functions we will implement and what the agent will confront in deployment, I predict that deadly actions have a negligible probability of being optimal.”) While discussing this definition of ‘instrumental convergence’, Andrew asked me: “what, exactly, is doing the converging? There is no limiting process. Optimal policies just are.” It would be more appropriate to say that an ac

6Steven Byrnes

I wrote this relatively early in my journey of self-studying neuroscience. Rereading this now, I guess I'm only slightly embarrassed to have my name associated with it, which isn’t as bad as I expected going in. Some shifts I’ve made since writing it (some of which are already flagged in the text): * New terminology part 1: Instead of “blank slate” I now say “learning-from-scratch”, as defined and discussed here. * New terminology part 2: “neocortex vs subcortex” → “learning subsystem vs steering subsystem”, with the former including the whole telencephalon and cerebellum, and the latter including the hypothalamus and brainstem. I distinguish them by "learning-from-scratch vs not-learning-from-scratch". See here. * Speaking of which, I now put much more emphasis on "learning-from-scratch" over "cortical uniformity" when talking about the neocortex etc.—I care about learning-from-scratch more, I talk about it more, etc. I see the learning-from-scratch hypothesis as absolutely central to a big picture of the brain (and AGI safety!), whereas cortical uniformity is much less so. I do still think cortical uniformity is correct (at least in the weak sense that someone with a complete understanding of one part of the cortex would be well on their way to a complete understanding of any other part of the cortex), for what it’s worth. * I would probably drop the mention of “planning by probabilistic inference”. Well, I guess something kinda like planning by probabilistic inference is part of the story, but generally I see the brain thing as mostly different. * Come to think of it, every time the word “reward” shows up in this post, it’s safe to assume I described it wrong in at least some respect. * The diagram with neocortex and subcortex is misleading for various reasons, see notes added to the text nearby. * I’m not sure I was using the term “analysis-by-synthesis” correctly. I think that term is kinda specific to sound processing. And the vision analog is “vision

24ryan_greenblatt

My sense is that this post holds up pretty well. Most of the considerations under discussion still appear live and important including: in-context learning, robustness, whether jank AI R&D accelerating AIs can quickly move to more general and broader systems, and general skepticism of crazy conclusions. At the time of this dialogue, my timelines were a bit faster than Ajeya's. I've updated toward the views Daniel expresses here and I'm now about half way between Ajeya's views in this post and Daniel's (in geometric mean). My read is that Daniel looks somewhat too aggressive in his predictions for 2024, though it is a bit unclear exactly what he was expecting. (This concrete scenario seems substantially more bullish than what we've seen in 2024, but not by a huge amount. It's unclear if he was intending these to be mainline predictions or a 25th percentile bullish scenario.) AI progress appears substantially faster than the scenario outlined in Ege's median world. In particular: * On "we have individual AI labs in 10 years that might be doing on the order of e.g. $30B/yr in revenue". OpenAI made $4 billion in revenue in 2024 and based on historical trends it looks like AI company revenue goes up 3x per year such that in 2026 the naive trend extrapolation indicates they'd make around $30 billion. So, this seems 3 years out instead of 10. * On "maybe AI systems can get gold on the IMO in five years". We seem likely to see gold on IMO this year (a bit less than 2 years later). It would be interesting to hear how Daniel, Ajeya, and Ege's views have changed since the time this was posted. (I think Daniel has somewhat later timelines (but the update is smaller than the progression of time such that AGI now seems closer to Daniel) and I think Ajeya has somewhat sooner timelines.) Daniel discusses various ideas for how to do a better version of this dialogue in this comment. My understanding is that Daniel (and others) have run something similar to what he describes

21nostalgebraist

This post provides a valuable reframing of a common question in futurology: "here's an effect I'm interested in -- what sorts of things could cause it?" That style of reasoning ends by postulating causes. But causes have a life of their own: they don't just cause the one effect you're interested in, through the one causal pathway you were thinking about. They do all kinds of things. In the case of AI and compute, it's common to ask * Here's a hypothetical AI technology. How much compute would it require? But once we have an answer to this question, we can always ask * Here's how much compute you have. What kind of AI could you build with it? If you've asked the first question, you ought to ask the second one, too. The first question includes a hidden assumption: that the imagined technology is a reasonable use of the resources it would take to build. This isn't always true: given those resources, there may be easier ways to accomplish the same thing, or better versions of that thing that are equally feasible. These facts are much easier to see when you fix a given resource level, and ask yourself what kinds of things you could do with it. This high-level point seems like an important contribution to the AI forecasting conversation. The impetus to ask "what does future compute enable?" rather than "how much compute might TAI require?" influenced my own view of Bio Anchors, an influence that's visible in the contrarian summary at the start of this post. ---------------------------------------- I find the specific examples much less convincing than the higher-level point. For the most part, the examples don't demonstrate that you could accomplish any particular outcome applying more compute. Instead, they simply restate the idea that more compute is being used. They describe inputs, not outcomes. The reader is expected to supply the missing inference: "wow, I guess if we put those big numbers in, we'd probably get magical results out." But this

7Orpheus16

ELK was one of my first exposures to AI safety. I participated in the ELK contest shortly after moving to Berkeley to learn more about longtermism and AI safety. My review focuses on ELK’s impact on me, as well as my impressions of how ELK affected the Berkeley AIS community. Things about ELK that I benefited from Understanding ARC’s research methodology & the builder-breaker format. For me, most of the value of ELK came from seeing ELK’s builder-breaker research methodology in action. Much of the report focuses on presenting training strategies and presenting counterexamples to those strategies. This style of thinking is straightforward and elegant, and I think the examples in the report helped me (and others) understand ARC’s general style of thinking. Understanding the alignment problem. ELK presents alignment problems in a very “show, don’t tell” fashion. While many of the problems introduced in ELK have been written about elsewhere, ELK forces you to think through the reasons why your training strategy might produce a dishonest agent (the human simulator) as opposed to an honest agent (the direct translator). The interactive format helped me more deeply understand some of the ways in which alignment is difficult. Common language & a shared culture. ELK gave people a concrete problem to work on. A whole subculture emerged around ELK, with many junior alignment researchers using it as their first opportunity to test their fit for theoretical alignment research. There were weekend retreats focused on ELK. It was one of the main topics that people were discussing from Jan-Feb 2022. People shared their training strategy ideas over lunch and dinner. It’s difficult to know for sure what kind of effect this had on the community as a whole. But at least for me, my current best-guess is that this shared culture helped me understand alignment, increased the amount of time I spent thinking/talking about alignment, and helped me connect with peers/collaborators who we

12Vika

I've been pleasantly surprised by how much this resource has caught on in terms of people using it and referring to it (definitely more than I expected when I made it). There were 30 examples on the list when was posted in April 2018, and 20 new examples have been contributed through the form since then. I think the list has several properties that contributed to wide adoption: it's fun, standardized, up-to-date, comprehensive, and collaborative. Some of the appeal is that it's fun to read about AI cheating at tasks in unexpected ways (I've seen a lot of people post on Twitter about their favorite examples from the list). The standardized spreadsheet format seems easier to refer to as well. I think the crowdsourcing aspect is also helpful - this helps keep it current and comprehensive, and people can feel some ownership of the list since can personally contribute to it. My overall takeaway from this is that safety outreach tools are more likely to be impactful if they are fun and easy for people to engage with. This list had a surprising amount of impact relative to how little work it took me to put it together and maintain it. The hard work of finding and summarizing the examples was done by the people putting together the lists that the master list draws on (Gwern, Lehman, Olsson, Irpan, and others), as well as the people who submit examples through the form. What I do is put them together in a common format and clarify and/or shorten some of the summaries. I also curate the examples to determine whether they fit the definition of specification gaming (as opposed to simply a surprising behavior or solution). Overall, I've probably spent around 10 hours so far on creating and maintaining the list, which is not very much. This makes me wonder if there is other low hanging fruit in the safety resources space that we haven't picked yet. I have been using it both as an outreach and research tool. On the outreach side, the resource has been helpful for making the arg

8Steven Byrnes

I’ll set aside what happens “by default” and focus on the interesting technical question of whether this post is describing a possible straightforward-ish path to aligned superintelligent AGI. The background idea is “natural abstractions”. This is basically a claim that, when you use an unsupervised world-model-building learning algorithm, its latent space tends to systematically learn some patterns rather than others. Different learning algorithms will converge on similar learned patterns, because those learned patterns are a property of the world, not an idiosyncrasy of the learning algorithm. For example: Both human brains and ConvNets seem to have a “tree” abstraction; neither human brains nor ConvNets seem to have a “head or thumb but not any other body part” concept. I kind of agree with this. I would say that the patterns are a joint property of the world and an inductive bias. I think the relevant inductive biases in this case are something like: (1) “patterns tend to recur”, (2) “patterns tend to be localized in space and time”, and (3) “patterns are frequently composed of multiple other patterns, which are near to each other in space and/or time”, and maybe other things. The human brain definitely is wired up to find patterns with those properties, and ConvNets to a lesser extent. These inductive biases are evidently very useful, and I find it very likely that future learning algorithms will share those biases, even more than today’s learning algorithms. So I’m basically on board with the idea that there may be plenty of overlap between the world-models of various different unsupervised world-model-building learning algorithms, one of which is the brain. (I would also add that I would expect “natural abstractions” to be a matter of degree, not binary. We can, after all, form the concept “head or thumb but not any other body part”. It would just be extremely low on the list of things that would pop into our head when trying to make sense of something we’

13Vanessa Kosoy

In this post, the author presents a case for replacing expected utility theory with some other structure which has no explicit utility function, but only quantities that correspond to conditional expectations of utility. To provide motivation, the author starts from what he calls the "reductive utility view", which is the thesis he sets out to overthrow. He then identifies two problems with the view. The first problem is about the ontology in which preferences are defined. In the reductive utility view, the domain of the utility function is the set of possible universes, according to the best available understanding of physics. This is objectionable, because then the agent needs to somehow change the domain as its understanding of physics grows (the ontological crisis problem). It seems more natural to allow the agent's preferences to be specified in terms of the high-level concepts it cares about (e.g. human welfare or paperclips), not in terms of the microscopic degrees of freedom (e.g. quantum fields or strings). There are also additional complications related to the unobservability of rewards, and to "moral uncertainty". The second problem is that the reductive utility view requires the utility function to be computable. The author considers this an overly restrictive requirement, since it rules out utility functions such as in the procrastination paradox (1 is the button is ever pushed, 0 if the button is never pushed). More generally, computable utility function have to be continuous (in the sense of the topology on the space of infinite histories which is obtained from regarding it as an infinite cartesian product over time). The alternative suggested by the author is using the Jeffrey-Bolker framework. Alas, the author does not write down the precise mathematical definition of the framework, which I find frustrating. The linked article in the Stanford Encyclopedia of Philosophy is long and difficult, and I wish the post had a succinct distillation of the

17Vanessa Kosoy

In this post, the author proposes a semiformal definition of the concept of "optimization". This is potentially valuable since "optimization" is a word often used in discussions about AI risk, and much confusion can follow from sloppy use of the term or from different people understanding it differently. While the definition given here is a useful perspective, I have some reservations about the claims made about its relevance and applications. The key paragraph, which summarizes the definition itself, is the following: In fact, "continues to exhibit this tendency with respect to the same target configuration set despite perturbations" is redundant: clearly as long as the perturbation doesn't push the system out of the basin, the tendency must continue. This is what is known as "attractor" in dynamical systems theory. For comparison, here is the definition of "attractor" from the Wikipedia: The author acknowledges this connection, although he also makes the following remark: I find this remark confusing. An attractor that operates along a subset of the dimension is just an attractor submanifold. This is completely standard in dynamical systems theory. Given that the definition itself is not especially novel, the post's main claim to value is via the applications. Unfortunately, some of the proposed applications seem to me poorly justified. Specifically, I want to talk about two major examples: the claimed relationship to embedded agency and the claimed relations to comprehensive AI services. In both cases, the main shortcoming of the definition is that there is an essential property of AI that this definition doesn't capture at all. The author does acknowledge that "goal-directed agent system" is a distinct concept from "optimizing systems". However, he doesn't explain how are they distinct. One way to formulate the difference is as follows: agency = optimization + learning. An agent is not just capable of steering a particular universe towards a certain outc

20ryan_greenblatt

At the time when I first heard this agenda proposed, I was skeptical. I remain skeptical, especially about the technical work that has been done thus far on the agenda[1]. I think this post does a reasonable job of laying out the agenda and the key difficulties. However, when talking to Davidad in person, I've found that he often has more specific tricks and proposals than what was laid out in this post. I didn't find these tricks moved me very far, but I think they were helpful for understanding what is going on. This post and Davidad's agenda overall would benefit from having concrete examples of how the approach might work in various cases, or more discussion of what would be out of scope (and why this could be acceptable). For instance, how would you make a superhumanly efficient (ASI-designed) factory that produces robots while proving safety? How would you allow for AIs piloting household robots to do chores (or is this out of scope)? How would you allow for the AIs to produce software that people run on their computers or to design physical objects that get manufactured? Given that this proposal doesn't allow for safely automating safety research, my understanding is that it is supposed to be a stable end state. Correspondingly, it is important to know what Davidad thinks can and can't be done with this approach. My core disagreements are on the "Scientific Sufficiency Hypothesis" (particularly when considering computational constraints), "Model-Checking Feasibility Hypothesis" (and more generally on proving the relevant properties), and on the political feasibility of paying the needed tax even if the other components work out. It seems very implausible to me that making a sufficiently good simulation is as easy as building the Large Hadron Collider. I think the objection in this comment holds up (my understanding is Davidad would require that we formally verify everything on the computer).[2] As a concrete example, I found it quite implausible that you

4Charbel-Raphaël

Ok, time to review this post and assess the overall status of the project. Review of the post What i still appreciate about the post: I continue to appreciate its pedagogy, structure, and the general philosophy of taking a complex, lesser-known plan and helping it gain broader recognition. I'm still quite satisfied with the construction of the post—it's progressive and clearly distinguishes between what's important and what's not. I remember the first time I met Davidad. He sent me his previous post. I skimmed it for 15 minutes, didn't really understand it, and thought, "There's no way this is going to work." Then I reconsidered, thought about it more deeply, and realized there was something important here. Hopefully, this post succeeded in showing that there is indeed something worth exploring! I think such distillation and analysis are really important. I'm especially happy about the fact that we tried to elicit as much as we could from Davidad's model during our interactions, including his roadmap and some ideas of easy projects to get early empirical feedback on this proposal. Current Status of the Agenda. (I'm not the best person to write this, see this as an informal personal opinion) Overall, Davidad performed much better than expected with his new job as program director in ARIA and got funded 74M$ over 4 years. And I still think this is the only plan that could enable the creation of a very powerful AI capable of performing a true pivotal act to end the acute risk period, and I think this last part is the added value of this plan, especially in the sense that it could be done in a somewhat ethical/democratic way compared to other forms of pivotal acts. However, it's probably not going to happen in time. Are we on track? Weirdly, yes for the non-technical aspects, no for the technical ones? The post includes a roadmap with 4 stages, and we can check if we are on track. It seems to me that Davidad jumped directly to stage 3, without going through sta

37Charbel-Raphaël

Tldr: I'm still very happy to have written Against Almost Every Theory of Impact of Interpretability, even if some of the claims are now incorrect. Overall, I have updated my view towards more feasibility and possible progress of the interpretability agenda — mainly because of the SAEs (even if I think some big problems remain with this approach, detailed below) and representation engineering techniques. However, I think the post remains good regarding the priorities the community should have. First, I believe the post's general motivation of red-teaming a big, established research agenda remains crucial. It's too easy to say, "This research agenda will help," without critically assessing how. I appreciate the post's general energy in asserting that if we're in trouble or not making progress, we need to discuss it. I still want everyone working on interpretability to read it and engage with its arguments. Acknowledgments: Thanks to Epiphanie Gédéon, Fabien Roger, and Clément Dumas for helpful discussions. Updates on my views Legend: * On the left of the arrow, a citation from the OP → ❓ on the right, my review which generally begins with emojis * ✅ - yes, I think I was correct (>90%) * ❓✅ - I would lean towards yes (70%-90%) * ❓ - unsure (between 30%-70%) * ❓❌ - I would lean towards no (10%-30%) * ❌ - no, I think I was basically wrong (<10%) * ⭐ important, you can skip the other sections Here's my review section by section: ⭐ The Overall Theory of Impact is Quite Poor? * "Whenever you want to do something with interpretability, it is probably better to do it without it" → ❓ I still think this is basically right, even if I'm not confident this will still be the case in the future; But as of today, I can't name a single mech-interpretability technique that does a better job at some non-intrinsic interpretability goal than the other more classical techniques, on a non-toy model task. * "Interpretability is Not a Good Predictor of Future Systems" →

6Ben Pace

Returning to this essay, it continues to be my favorite Paul post (even What Failure Looks Like only comes second), and I think it's the best way to engage with Paul's work than anything else (including the Eliciting Latent Knowledge document, which feels less grounded in the x-risk problem, is less in Paul's native language, and gets detailed on just one idea for 10x the space thus communicating less of the big picture research goal). I feel I can understand all the arguments made in this post. I think this should be mandatory reading before reading Eliciting Latent Knowledge. Overview of why: * The motivation behind most of proposals Paul has spent a lot of time (iterated amplification, imitative generalization) on are explained clearly and succinctly. * For a quick summary, this involves * A proposal for useful ML-systems designed with human feedback * An argument that the human-feedback ML-systems will have flaws that kill you * A proposal for using ML assistants to debug the original ML system * An argument that the ML systems will not be able to understand the original human-feedback ML-systems * A proposal for training the human-feedback ML-systems in a way that requires understandability * An argument that this proposal is uncompetitive * ??? (I believe the proposals in the ELK document are the next step here) * A key problem when evaluating very high-IQ, impressive, technical work, is that it is unclear which parts of the work you do not understand because you do not understand an abstract technical concept, and which parts are simply judgment calls based on the originator of the idea. This post shows very clearly which is which — many of the examples and discussions are technical, but the standard for "plausible failure story" and "sufficiently precise algorithm" and "sufficiently doomed" are all judgment calls, as are the proposed solutions. I'm not even sure I get on the bus at step 1, that the right next step is to consider ML

12johnswentworth

This post is an excellent distillation of a cluster of past work on maligness of Solomonoff Induction, which has become a foundational argument/model for inner agency and malign models more generally. I've long thought that the maligness argument overlooks some major counterarguments, but I never got around to writing them up. Now that this post is up for the 2020 review, seems like a good time to walk through them. In Solomonoff Model, Sufficiently Large Data Rules Out Malignness There is a major outside-view reason to expect that the Solomonoff-is-malign argument must be doing something fishy: Solomonoff Induction (SI) comes with performance guarantees. In the limit of large data, SI performs as well as the best-predicting program, in every computably-generated world. The post mentions that: ... but in the large-data limit, SI's guarantees are stronger than just that. In the large-data limit, there is no computable way of making better predictions than the Solomonoff prior in any world. Thus, agents that are influencing the Solomonoff prior cannot gain long-term influence in any computable world; they have zero degrees of freedom to use for influence. It does not matter if they specialize in influencing worlds in which they have short strings; they still cannot use any degrees of freedom for influence without losing all their influence in the large-data limit. Takeaway of this argument: as long as we throw enough data at our Solomonoff inductor before asking it for any outputs, the malign agent problem must go away. (Though note that we never know exactly how much data that is; all we have is a big-O argument with an uncomputable constant.) ... but then how the hell does this outside-view argument jive with all the inside-view arguments about malign agents in the prior? Reflection Breaks The Large-Data Guarantees There's an important gotcha in those guarantees: in the limit of large data, SI performs as well as the best-predicting program, in every compu

28Vanessa Kosoy

This post is a review of Paul Christiano's argument that the Solomonoff prior is malign, along with a discussion of several counterarguments and countercounterarguments. As such, I think it is a valuable resource for researchers who want to learn about the problem. I will not attempt to distill the contents: the post is already a distillation, and does a a fairly good job of it. Instead, I will focus on what I believe is the post's main weakness/oversight. Specifically, the author seems to think the Solomonoff prior is, in some way, a distorted model of reasoning, and that the attack vector in question can attributed to this, at least partially. This is evident in phrases such as "unintuitive notion of simplicity" and "the Solomonoff prior is very strange". This is also why the author thinks the speed prior might help and that "since it is difficult to compute the Solomonoff prior, [the attack vector] might not be relevant in the real world". In contrast, I believe that the attack vector is quite robust and will threaten any sufficiently powerful AI as long as it's cartesian (more on "cartesian" later). Formally analyzing this question is made difficult by the essential role of non-realizability. That is, the attack vector arises from the AI reasoning about "possible universes" and "simulation hypotheses" which are clearly phenomena that are computationally infeasible for the AI to simulate precisely. Invoking Solomonoff induction dodges this issue since Solomonoff induction is computationally unbounded, at the cost of creating the illusion that the conclusions are a symptom of using Solomonoff induction (and, it's still unclear how to deal with the fact Solomonoff induction itself cannot exist in the universes that Solomonoff induction can learn). Instead, we should be using models that treat non-realizability fairly, such as infra-Bayesiansim. However, I will make no attempt to present such a formal analysis in this review. Instead, I will rely on painting an in

10habryka

I think it's a bit hard to tell how influential this post has been, though my best guess is "very". It's clear that sometime around when this post was published there was a pretty large shift in the strategies that I and a lot of other people pursued, with "slowing down AI" becoming a much more common goal for people to pursue. I think (most of) the arguments in this post are good. I also think that when I read an initial draft of this post (around 1.5 years ago or so), and had a very hesitant reaction to the core strategy it proposes, that I was picking up on something important, and that I do also want to award Bayes points to that part of me given how things have been playing out so far. I do think that since I've seen people around me adopt strategies to slow down AI, I've seen it done on a basis that feels much more rhetorical, and often directly violates virtues and perspectives that I hold very dearly. I think it's really important to understand that technological progress has been the central driving force behind humanity's success, and that indeed this should establish a huge prior against stopping almost any kind of technological development. In contrast to that, the majority of arguments that I've seen find traction for slowing down AI development are not distinguishable from arguments that apply to a much larger set of technologies which to me clearly do not pose a risk commensurable with the prior we should have against slowdown. Concerns about putting people out of jobs, destroying traditional models of romantic relationships, violating copyright law, spreading disinformation, all seem to me to be the kind of thing that if you buy it, you end up with an argument that proves too much and should end up opposed to a huge chunk of technological progress. And I can feel the pressure in myself for these things as well. I can see how it would be easy to at least locally gain traction at slowing down AI by allying myself with people who are concerned abo

14ryan_greenblatt

IMO, this post makes several locally correct points, but overall fails to defeat the argument that misaligned AIs are somewhat likely to spend (at least) a tiny fraction of resources (e.g., between 1/million and 1/trillion) to satisfy the preferences of currently existing humans. AFAICT, this is the main argument it was trying to argue against, though it shifts to arguing about half of the universe (an obviously vastly bigger share) halfway through the piece.[1] When it returns to arguing about the actual main question (a tiny fraction of resources) at the end here and eventually gets to the main trade-related argument (acausal or causal) in the very last response in this section, it almost seems to admit that this tiny amount of resources is plausible, but fails to update all the way. I think the discussion here and here seems highly relevant and fleshes out this argument to a substantially greater extent than I did in this comment. However, note that being willing to spend a tiny fraction of resources on humans still might result in AIs killing a huge number of humans due to conflict between it and humans or the AI needing to race through the singularity as quickly as possible due to competition with other misaligned AIs. (Again, discussed in the links above.) I think fully misaligned paperclippers/squiggle maximizer AIs which spend only a tiny fraction of resources on humans (as seems likely conditional on that type of AI) are reasonably likely to cause outcomes which look obviously extremely bad from the perspective of most people (e.g., more than hundreds of millions dead due to conflict and then most people quickly rounded up and given the option to either be frozen or killed). I wish that Soares and Eliezer would stop making these incorrect arguments against tiny fractions of resources being spent on the preference of current humans. It isn't their actual crux, and it isn't the crux of anyone else either. (However rhetorically nice it might be.) -------

19Neel Nanda

Self-Review: After a while of being insecure about it, I'm now pretty fucking proud of this paper, and think it's one of the coolest pieces of research I've personally done. (I'm going to both review this post, and the subsequent paper). Though, as discussed below, I think people often overrate it. Impact The main impact IMO is proving that mechanistic interpretability is actually possible, that we can take a trained neural network and reverse-engineer non-trivial and unexpected algorithms from it. In particular, I think by focusing on grokking I (semi-accidentally) did a great job of picking a problem that people cared about for non-interp reasons, where mech interp was unusually easy (as it was a small model, on a clean algorithmic task), and that I was able to find real insights about grokking as a direct result of doing the mechanistic interpretability. Real models are fucking complicated (and even this model has some details we didn't fully understand), but I feel great about the field having something that's genuinely detailed, rigorous and successfully reverse-engineered, and this seems an important proof of concept. IMO the other contributions are the specific algorithm I found, and the specific insights about how and why grokking happens. but in my opinion these are much less interesting. Field-Building Another large amount of impact is that this was a major win for mech interp field-building. This is hard to measure, but some data: * There are multiple papers I like that are substantially building on/informed by these results (A toy model of universality, the clock and the pizza, Feature emergence via margin maximization, Explaining grokking through circuit efficiency * It's got >100 citations in less than a year (a decent chunk of these are semi-fake citations from this being used as a standard citation for 'mech interp exists as a field', so I care more about the "how many papers would not exist without this" metric) * It went pretty viral on Twi

39Buck

(I'm just going to speak for myself here, rather than the other authors, because I don't want to put words in anyone else's mouth. But many of the ideas I describe in this review are due to other people.) I think this work was a solid intellectual contribution. I think that the metric proposed for how much you've explained a behavior is the most reasonable metric by a pretty large margin. The core contribution of this paper was to produce negative results about interpretability. This led to us abandoning work on interpretability a few months later, which I'm glad we did. But these negative results haven’t had that much influence on other people’s work AFAICT, so overall it seems somewhat low impact. The empirical results in this paper demonstrated that induction heads are not the simple circuit which many people claimed (see this post for a clearer statement of that), and we then used these techniques to get mediocre results for IOI (described in this comment). There hasn’t been much followup on this work. I suspect that the main reasons people haven't built on this are: * it's moderately annoying to implement it * it makes your explanations look bad (IMO because they actually are unimpressive), so you aren't that incentivized to get it working * the interp research community isn't very focused on validating whether its explanations are faithful, and in any case we didn’t successfully persuade many people that explanations performing poorly according to this metric means they’re importantly unfaithful I think that interpretability research isn't going to be able to produce explanations that are very faithful explanations of what's going on in non-toy models (e.g. I think that no such explanation has ever been produced). Since I think faithful explanations are infeasible, measures of faithfulness of explanations don't seem very important to me now. (I think that people who want to do research that uses model internals should evaluate their techniques by mea

8Vanessa Kosoy

This post attempts to describe a key disagreement between Karnofsky and Soares (written by Karnofsky) pertaining to the alignment protocol "train an AI to simulate an AI alignment researcher". The topic is quite important, since this is a fairly popular approach. Here is how I view this question: The first unknown is how accurate is the simulation. This is not really discussed in the OP. On the one hand, one might imagine that with more data, compute and other improvements, the AI should ultimately converge on an almost perfect simulation of an AI alignment researcher, which is arguably safe. One the other hand, there are two problems with this. First, such a simulation might be vulnerable to attacks from counterfactuals. Second, the prior is malign, i.e. the simulation might converge to representing a "malign simulation hypothesis" universe rather than then intended null hypothesis / ordinary reality. Instead, we can imagine a simulation that's not extremely accurate, but that's modified to be good enough by fine-tuning with reinforcement learning. This is essentially the approach in contemporary AI and is also the assumption of the OP. Although Karnofsky says: "a small amount of RL", and I'm don't know why he beliefs a small amount is sufficient. Perhaps RL seemed less obviously important then than it does now, with the recent successes of o1 and o3. The danger (as explained in the OP by Soares paraphrased by Karnofsky) is that it's much easier to converge in this manner on an arbitrary agent that has the capabilities of the imaginary AI alignment researcher (which probably have to be a lot greater than capabilities of human researchers to make it useful), but doesn't have values that are truly aligned. This is because "agency" is (i) a relatively simple concept and (ii) a robust attractor, in the sense that any agent would behave similarly when faced with particular instrumental incentives, and it's mainly this behavior that the training process rewards. On t

39LawrenceC

This is a review of both the paper and the post itself, and turned more into a review of the paper (on which I think I have more to say) as opposed to the post. Disclaimer: this isn’t actually my area of expertise inside of technical alignment, and I’ve done very little linear probing myself. I’m relying primarily on my understanding of others’ results, so there’s some chance I’ve misunderstood something. Total amount of work on this review: ~8 hours, though about 4 of those were refreshing my memory of prior work and rereading the paper. TL;DR: The paper made significant contributions by introducing the idea of unsupervised knowledge discovery to a broader audience and by demonstrating that relatively straightforward techniques may make substantial progress on this problem. Compared to the paper, the blog post is substantially more nuanced, and I think that more academic-leaning AIS researchers should also publish companion blog posts of this kind. Collin Burns also deserves a lot of credit for actually doing empirical work in this domain when others were skeptical. However, the results are somewhat overstated and, with the benefit of hindsight, (vanilla) CCS does not seem to be a particularly promising technique for eliciting knowledge from language models. That being said, I encourage work in this area.[1] Introduction/Overview The paper “Discovering Latent Knowledge in Language Models without Supervision” by Collin Burns, Haotian Ye, Dan Klein, and Jacob Steinhardt (henceforth referred to as “the CCS paper” for short) proposes a method for unsupervised knowledge discovery, which can be thought of as a variant of empirical, average-case Eliciting Latent Knowledge (ELK). In this companion blog post, Collin Burns discusses the motivations behind the paper, caveats some of the limitations of the paper, and provides some reasons for why this style of unsupervised methods may scale to future language models. The CCS paper kicked off a lot of waves in the alig

7johnswentworth

The material here is one seed of a worldview which I've updated toward a lot more over the past year. Some other posts which involve the theme include Science in a High Dimensional World, What is Abstraction?, Alignment by Default, and the companion post to this one Book Review: Design Principles of Biological Circuits. Two ideas unify all of these: 1. Our universe has a simplifying structure: it abstracts well, implying a particular kind of modularity. 2. Goal-oriented systems in our universe tend to evolve a modular structure which reflects the structure of the universe. One major corollary of these two ideas is that goal-oriented systems will tend to evolve similar modular structures, reflecting the relevant parts of their environment. Systems to which this applies include organisms, machine learning algorithms, and the learning performed by the human brain. In particular, this suggests that biological systems and trained deep learning systems are likely to have modular, human-interpretable internal structure. (At least, interpretable by humans familiar with the environment in which the organism/ML system evolved.) This post talks about some of the evidence behind this model: biological systems are indeed quite modular, and simulated evolution experiments find that circuits do indeed evolve modular structure reflecting the modular structure of environmental variations. The companion post reviews the rest of the book, which makes the case that the internals of biological systems are indeed quite interpretable. On the deep learning side, researchers also find considerable modularity in trained neural nets, and direct examination of internal structures reveals plenty of human-recognizable features. Going forward, this view is in need of a more formal and general model, ideally one which would let us empirically test key predictions - e.g. check the extent to which different systems learn similar features, or whether learned features in neural nets satisfy th

15adamShimi

In “Why Read The Classics?”, Italo Calvino proposes many different definitions of a classic work of literature, including this one: For me, this captures what makes this sequence and corresponding paper a classic in the AI Alignment literature: it keeps on giving, readthrough after readthrough. That doesn’t mean I agree with everything in it, or that I don’t think it could have been improved in terms of structure. But when pushed to reread it, I found again and again that I had missed or forgotten some nice argument, some interesting takeaway. With that, a caveat: I’m collaborating with Evan Hubinger (one of the authors) on projects related to ideas introduced in this sequence, especially to Deceptive Alignment. I am thus probably biased positively about this work. That being said, I have no problem saying I disagree with collaborators, so I don’t think I’m too biased to write this review. (Small point: I among other people tend to describe this sequence/paper as mainly Evan’s work, but he repeatedly told me that everyone participated equally, and that the names are in alphabetic order, not contribution order. So let’s keep that in mind) Summary Let’s start the review proper with a post by post summary (except for the conclusion): * (Introduction) This first post introduces the idea of mesa-optimizers, the learned optimizers from the title. A mesa-optimizer is an optimizer which is the result of a learning process, and it comes with the issue of inner alignment: how aligned is the objective of the mesa-optimizer (on which we don’t have direct control) with the objective of the base-optimizer that produced this mesa-optimizer? The post then split the safety questions related to mesa-optimizer in two categories: understanding which conditions make mesa-optimizer appear; and understanding how aligned is the mesa-objective with the base-objective. * (Conditions for Mesa-Optimization) This post tackles the first category outlined in the introduction: how can m

13adamShimi

This post states the problem of gradient hacking. It is valuable in that this problem is far from obvious, and if plausible, very dangerous. On the other hand, the presentation doesn’t go into enough details, and so leaves gradient hacking open to attacks and confusion. Thus instead of just reviewing this post, I would like to clarify certain points, while interweaving my critics about the way gradient hacking was initially stated, and explaining why I consider this problem so important. (Caveat: I’m not pretending that any of my objections are unknown to Evan. We’re working together on searching for experimental evidence of gradient hacking, and this review is heavily influenced by this collaboration.) What is gradient hacking? Evan defines it as: So we have a trained mesa-optimizer that is deceptively aligned (its mesa-objective differs from the base-objective it is trained on). This deceptive system finds a way to influence the gradient generated during training, and thus can basically ensure it never becomes inner aligned (among other things). Before checking on how exactly this could be possible, we should think a bit more about what this implies. If the deceptive model has arbitrary access to the gradient, then from that point on, the base-objective has only minimal influence on the training. Some influence remains because we probably want to test the end result, and thus it should work well with the base-objective. But that's pretty much the only constraint left. It could also pretty much deals with deception detectors because it can make itself not detectable: To say it pithy: if gradient hacking happens, we’re fucked. How could it happen, though? Well, that’s actually two distinct questions: how could a model gradient hack, and how could training create a model which gradient hacks. The post mostly focuses on the first one. How could a model gradient hack? The first example comes from a quoted footnote of Risks from Learned Optimization: This im

15Writer

In this post, I appreciated two ideas in particular: 1. Loss as chisel 2. Shard Theory "Loss as chisel" is a reminder of how loss truly does its job, and its implications on what AI systems may actually end up learning. I can't really argue with it and it doesn't sound new to my ear, but it just seems important to keep in mind. Alone, it justifies trying to break out of the inner/outer alignment frame. When I start reasoning in its terms, I more easily appreciate how successful alignment could realistically involve AIs that are neither outer nor inner aligned. In practice, it may be unlikely that we get a system like that. Or it may be very likely. I simply don't know. Loss as a chisel just enables me to think better about the possibilities. In my understanding, shard theory is, instead, a theory of how minds tend to be shaped. I don't know if it's true, but it sounds like something that has to be investigated. In my understanding, some people consider it a "dead end," and I'm not sure if it's an active line of research or not at this point. My understanding of it is limited. I'm glad I came across it though, because on its surface, it seems like a promising line of investigation to me. Even if it turns out to be a dead end I expect to learn something if I investigate why that is. The post makes more claims motivating its overarching thesis that dropping the frame of outer/inner alignment would be good. I don't know if I agree with the thesis, but it's something that could plausibly be true, and many arguments here strike me as sensible. In particular, the three claims at the very beginning proved to be food for thought to me: "Robust grading is unnecessary," "the loss function doesn't have to robustly and directly reflect what you want," "inner alignment to a grading procedure is unnecessary, very hard, and anti-natural." I also appreciated the post trying to make sense of inner and outer alignment in very precise terms, keeping in mind how deep learning and

4adamShimi

How do you review a post that was not written for you? I’m already doing research in AI Alignment, and I don’t plan on creating a group of collaborators for the moment. Still, I found some parts of this useful. Maybe that’s how you do it: by taking different profiles, and running through the most useful advice for each profile from the post. Let’s do that. Full time researcher (no team or MIRIx chapter) For this profile (which is mine, by the way), the most useful piece of advice from this post comes from the model of transmitters and receivers. I’m convinced that I’ve been using it intuitively for years, but having an explicit model is definitely a plus when trying to debug a specific situation, or to explain how it works to someone less used to thinking like that. Full time research who wants to build a team/MIRIx chapter Obviously, this profile benefits from the great advice on building a research group. I would expect someone with this profile to understand relatively well the social dynamics part, so the most useful advice is probably the detailed logistics of getting such a group off the ground. I also believe that the escalating asks and rewards is a less obvious social dynamic to take into account. Aspiring researcher (no team or MIRIx chapter) The section You and your research was probably written with this profile in mind. It tries to push towards exploration instead of exploitation, babble instead of prune. And for so many people that I know who feel obligated to understand everything before toying with a question, this is the prescribed medicine. I want to push-back just a little about the “follow your curiosity” vibe, as I believe that there are ways to check how promising the current ideas are for AI Alignment. But I definitely understand that the audience is more “wannabe researchers stifled by their internal editor”, so pushing for curiosity and exploration makes sense. Aspiring researcher who wants to build a team/MIRIx chapter In additio

21DanielFilan

I think that strictly speaking this post (or at least the main thrust) is true, and proven in the first section. The title is arguably less true: I think of 'coherence arguments' as including things like 'it's not possible for you to agree to give me a limitless number of dollars in return for nothing', which does imply some degree of 'goal-direction'. I think the post is important, because it constrains the types of valid arguments that can be given for 'freaking out about goal-directedness', for lack of a better term. In my mind, it provokes various follow-up questions: 1. What arguments would imply 'goal-directed' behaviour? 2. With what probability will a random utility maximiser be 'goal-directed'? 3. How often should I think of a system as a utility maximiser in resources, perhaps with a slowly-changing utility function? 4. How 'goal-directed' are humans likely to make systems, given that we are making them in order to accomplish certain tasks that don't look like random utility functions? 5. Is there some kind of 'basin of goal-directedness' that systems fall in if they're even a little goal-directed, causing them to behave poorly? Off the top of my head, I'm not familiar with compelling responses from the 'freak out about goal-directedness' camp on points 1 through 5, even though as a member of that camp I think that such responses exist. Responses from outside this camp include Rohin's post 'Will humans build goal-directed agents?'. Another response is Brangus' comment post, although I find its theory of goal-directedness uncompelling. I think that it's notable that Brangus' post was released soon after this was announced as a contender for Best of LW 2018. I think that if this post were added to the Best of LW 2018 Collection, the 'freak out' camp might produce more of these responses and move the dialogue forward. As such, I think it should be added, both because of the clear argumentation and because of the response it is likely to provoke.

10DanielFilan

* Olah’s comment indicates that this is indeed a good summary of his views. * I think the first three listed benefits are indeed good reasons to work on transparency/interpretability. I am intrigued but less convinced by the prospect of ‘microscope AI’. * The ‘catching problems with auditing’ section describes an ‘auditing game’, and says that progress in this game might illustrate progress in using interpretability for alignment. It would be good to learn how much success the auditors have had in this game since the post was published. * One test of ‘microscope AI’: the go community has had a couple of years of the computer era, in which time open-source go programs stronger than AlphaGo have been released. This has indeed changed the way that humans think about go: seeing the corner variations that AIs tend to play has changed our views on which variations are good for which player, and seeing AI win probabilities conditioned on various moves, as well as the AI-recommended continuations, has made it easier to review games. Yet sadly, there has been to my knowledge no new go knowledge generated from looking at the internals of these systems, despite some visualization research being done (https://arxiv.org/pdf/1901.02184.pdf, https://link.springer.com/chapter/10.1007/978-3-319-97304-3_20). As far as I’m aware, we do not even know if these systems understand the combinatorial game theory of the late endgame, the one part of go that has been satisfactorily mathematized (and therefore unusually amenable to checking whether some program implements it). It’s not clear to me whether this is for a lack of trying, but this does seem like a setting where microscope AI would be useful if it were promising. * The paper mostly focuses on the benefits of transparency/interpretability for AI alignment. However, as far as I’m aware, since before this post was published, the strongest argument against work in this direction has been the problem of tractability - can we ac

17Jeremy Gillen

This post deserves to be remembered as a LessWrong classic. 1. It directly tries to solve a difficult and important cluster of problems (whether it succeeds is yet to be seen). 2. It uses a new diagrammatic method of manipulating sets of independence relations. 3. It's a technical result! These feel like they're getting rarer on LessWrong and should be encouraged. There are several problems that are fundamentally about attaching very different world models together and transferring information from one to the other. * Ontology identification involves taking a goal defined in an old ontology[1] and accurately translating it into a new ontology. * High-level models and low-level models need to interact in a bounded agent. I.e. learning a high-level fact should influence your knowledge about low-level facts and vice versa. * Value identification is the problem of translating values from a human to an AI. This is much like ontology identification, with the added difficulty that we don't get as much detailed access or control over the human world model. * Interpretability is about finding recognisable concepts and algorithms in trained neural networks. In general, we can solve these problems using shared variables and shared sub-structures that are present in both models. * We can stitch together very different world models along shared variables. E.g. if you have two models of molecular dynamics, one faster and simpler than the other. You want to simulate in the fast one, then switch to the slow one when particular interactions happen. To transfer the state from one to the other you identify variables present in both models (probably atom locations, velocities, some others), then just copy these values to the other model. Under-specified variables must be inferred from priors. * If you want to transfer a new concept from WM1 to a less knowledgeable WM2, you can do so by identifying the lower-level concepts that both WMs share, then constructing an "expla

15Isnasene

[Disclaimer: I'm reading this post for the first time now, as of 1/11/2020. I also already have a broad understanding of the importance of AI safety. While I am skeptical about MIRI's approach to things, I am also a fan of MIRI. Where this puts me relative to the target demographic of this post, I cannot say.] Overall Summary I think this post is pretty good. It's a solid and well-written introduction to some of the intuitions behind AI alignment and the fundamental research that MIRI does. At the same time, the use of analogy made the post more difficult for me to parse and hid some important considerations about AI alignment from view. Though it may be good (but not optimal) for introducing some people to the problem of AI alignment and a subset of MIRI's work, it did not raise or lower my opinion of MIRI as someone who already understood AGI safety to be important. To be clear, I do not consider any of these weaknesses serious because I believe them to be partially irrelevant to the audience of people who don't appreciate the importance of AI-Safety. Still, they are relevant to the audience of people who give AI-Safety the appropriate scrutiny but remain skeptical of MIRI. And I think this latter audience is important enough to assign this article a "pretty good" instead of a "great". I hope a future post directly explores the merit of MIRI's work on the context AI alignment without use of analogy. Below is an overview of my likes and dislikes in this post. I will go into more detail about them in the next section, "Evaluating Analogies." Things I liked: * It's a solid introduction to AI-alignment, covering a broad range of topics including: * Why we shouldn't expect aligned AGI by default * How modern conversation about AGI behavior is problematically underspecified * Why fundamental deconfusion research is necessary for solving AI-alignment * It directly explains the value/motivation of particular pieces of MIRI work via analogy -- which

8DragonGod

Epistemic Status I am an aspiring selection theorist and I have thoughts. ---------------------------------------- Why Selection Theorems? Learning about selection theorems was very exciting. It's one of those concepts that felt so obviously right. A missing component in my alignment ontology that just clicked and made everything stronger. Selection Theorems as a Compelling Agent Foundations Paradigm There are many reasons to be sympathetic to agent foundations style safety research as it most directly engages the hard problems/core confusions of alignment/safety. However, one concern with agent foundations research is that we might build sky high abstraction ladders that grow increasingly disconnected from reality. Abstractions that don't quite describe the AI systems we deal with in practice. I think that in presenting this post, Wentworth successfully sidestepped the problem. He presented an intuitive story for why the Selection Theorems paradigm would be fruitful; it's general enough to describe many paradigms of AI system development, yet concrete enough to say nontrivial/interesting things about the properties of AI systems (including properties that bear on their safety). Wentworth presents a few examples of extant selection theorems (most notably the coherence theorems) and later argues that selection theorems have a lot of research "surface area" and new researchers could be onboarded (relatively) quickly. He also outlined concrete steps people interested in selection theorems could take to contribute to the program. Overall, I found this presentation of the case for selection theorems research convincing. I think that selection theorems provide a solid framework with which to formulate (and prove) safety desiderata/guarantees for AI systems that are robust to arbitrary capability amplification. Furthermore, selection theorems seem to be very robust to paradigm shifts in the development artificial intelligence. That is regardless of what

9Olli Järviniemi

I view this post as providing value in three (related) ways: 1. Making a pedagogical advancement regarding the so-called inner alignment problem 2. Pointing out that a common view of "RL agents optimize reward" is subtly wrong 3. Pushing for thinking mechanistically about cognition-updates Re 1: I first heard about the inner alignment problem through Risks From Learned Optimization and popularizations of the work. I didn't truly comprehend it - sure, I could parrot back terms like "base optimizer" and "mesa-optimizer", but it didn't click. I was confused. Some months later I read this post and then it clicked. Part of the pedagogical value is not having to introduce the 4 terms of form [base/mesa] + [optimizer/objective] and throwing those around. Even with Rob Miles' exposition skills that's a bit overwhelming. Another part I liked were the phrases "Just because common English endows “reward” with suggestive pleasurable connotations" and "Let’s strip away the suggestive word “reward”, and replace it by its substance: cognition-updater." One could be tempted to object and say that surely no one would make the mistakes pointed out here, but definitely some people do. I did. Being a bit gloves off here definitely helped me. Re 2: The essay argues for, well, reward not being the optimization target. There is some deep discussion in the comments about the likelihood of reward in fact being the optimization target, or at least quite close (see here). Let me take a more shallow view. I think there are people who think that reward is the optimization target by definition or by design, as opposed to this being a highly non-trivial claim that needs to be argued for. It's the former view that this post (correctly) argues against. I am sympathetic to pushback of the form "there are arguments that make it reasonable to privilege reward-maximization as a hypothesis" and about this post going a bit too far, but these remarks should not be confused with a rebuttal

10fiddler

I think this post is incredibly useful as a concrete example of the challenges of seemingly benign powerful AI, and makes a compelling case for serious AI safety research being a prerequisite to any safe further AI development. I strongly dislike part 9, as painting the Predict-o-matic as consciously influencing others personality at the expense of short-term prediction error seems contradictory to the point of the rest of the story. I suspect I would dislike part 9 significantly less if it was framed in terms of a strategy to maximize predictive accuracy. More specifically, I really enjoy the focus on the complexity of “optimization” on a gears-level: I think that it’s a useful departure from high abstraction levels, as the question of what predictive accuracy means, and the strategy AI would use to pursue it, is highly influenced by the approach taken. I think a more rigorous approach to analyzing whether different AI approaches are susceptible to “undercutting” as a safety feature would be an extremely valuable piece. My suspicion is that even the engineer’s perspective here is significantly under-specified with the details necessary to determine whether this vulnerability exists. I also think that Part 9 detracts from the piece in two main ways: by painting the predict-o-matic as conscious, it implies a significantly more advanced AI than necessary to exhibit this effect. Additionally, because the AI admits to sacrificing predictIve accuracy in favor of some abstract value-add, it seems like pretty much any naive strategy would outcompete the current one, according to the engineer, meaning that the type of threat is also distorted: the main worry should be AI OPTIMIZING for predictive accuracy, not pursuing its own goals. That’s bad sci-fi or very advanced GAI, not a prediction-optimizer. I would support the deletion or aggressive editing of part 9 in this and future similar pieces: I’m not sure what it adds. ETA-I think whether or not this post should be upd

8Martin Randall

Does this look like a motte-and-bailey to you? 1. Bailey: GPTs are Predictors, not Imitators (nor Simulators). 2. Motte: The training task for GPTs is a prediction task. The title and the concluding sentence both plainly advocate for (1), but it's not really touched by the overall post, and I think it's up for debate (related: reward is not the optimization target). Instead there is an argument for (2). Perhaps the intention of the final sentence was to oppose Simulators? If that's the case, cite it, be explicit. This could be a really easy thing for an editor to fix. ---------------------------------------- Does this look like a motte-and-bailey to you? 1. Bailey: The task that GPTs are being trained on is ... harder than the task of being a human. 2. Motte: Being an actual human is not enough to solve GPT's task. As I read it, (1) is false, the task of being a human doesn't cap out at human intelligence. More intelligent humans are better at minimizing prediction error, achieving goals, inclusive genetic fitness, whatever you might think defines "the task of being a human". In the comments, Yudkowsky retreats to (2), which is true. But then how should I understand this whole paragraph from the post? If we're talking about how natural selection trained my genome, why are we talking about how well humans perform the human task? Evolution is optimizing over generations. My human task is optimizing over my lifetime. Also, if we're just arguing for different thinking, surely it mostly matters whether the training task is different, not whether it is harder? ---------------------------------------- Overall I think "Is GPT-N bounded by human capabilities? No." is a better post on the mottes and avoids staking out unsupported baileys. This entire topic is becoming less relevant because AIs are getting all sorts of synthetic data and RLHF and other training techniques thrown at them. The 2022 question of the capabilities of a hypothetical GPT-N that was only t

4Mikhail Samin

Sharp Left Turn: a more important problem (and a more specific threat model) than people usually think The sharp left turn is not a simple observation that we've seen capabilities generalise more than alignment. As I understand it, it is a more mechanistic understanding that some people at MIRI have, of dynamics that might produce systems with generalised capabilities but not alignment. Many times over the past year, I've been surprised by people in the field who've read Nate's post but somehow completely missed the part where it talks about specific dynamics that lead to alignment properties breaking during capabilities generalisation. To fulfil the reviewing duty and to have a place to point people to, I'll try to write down some related intuitions that I talked about throughout 2023 when trying to get people to have intuitions on what the sharp left turn problem is about. For example, imagine training a neural network with RL. For a while during training, the neural network might be implementing a fuzzy collection of algorithms and various heuristics that together kinda optimise for some goals. The gradient strongly points towards greater capabilities. Some of these algorithms and heuristics might be more useful for the task the neural network is being evaluated on, and they'll persist more and what the neural network is doing as a whole will look a bit more like what the most helpful parts of it are doing. Some of these algorithms and heuristics might be more agentic and do more for long-term goal achievement than others. As being better at achieving goals correlates with greater performance, the neural network becomes, as a whole, more capable of achieving goals. Or, maybe the transition that leads to capabilities generalisation can be more akin to grokking: even with a fuzzy solution, the distant general coherent agent implementations might still be visible to the gradient, and at some point, there might be a switch from a fuzzy collection of things togeth

AI ALIGNMENT FORUM
The Best of LessWrong
AF

AI ALIGNMENT FORUM
The Best of LessWrong
AF

Rationality

Optimization

World

Practical

AI Strategy

Technical AI Safety