Dmitry Vaintrob

• Dmitry VaintrobJan 7, 2025, 8:58 PM

  2 points

  0

  in reply to: Noosphere89’s comment on: Dmitry Vain­trob’s Shortform

  Cool! I haven’t seen these, good to have these to point to (and I’m glad that Richard Ngo has thought about this)

• Dmitry VaintrobJan 7, 2025, 8:48 PM

  2 points

  0

  in reply to: Raemon’s comment on: Dmitry Vain­trob’s Shortform

  Thanks for the context! I didn’t follow this discourse very closely, but I think your “optimistic assumptions” post wasn’t the main offender—it’s reasonable to say that “it’s suspicious when people are bad at backchaining but think they’re good at backchaining or their job depends on backchaining more than they are able to”. I seem to remember reading some responses/​ related posts that I had more issues with, where the takeaway was explicitly that “alignment researchers should try harder at backchaining and one-shotting baba-is-you-like problems because that’s the most important thing”, instead of the more obvious but less rationalism-vibed takeaway of “you must (if at all possible) avoid situations where you have to one-shot complicated games”.

  I think if I’m reading you correctly, we’re largely in agreement. All plan-making and game-playing depends on some amount of backchaining/​ one-shot prediction. And there is a part of doing science that looks a bit like this. But there are ways of getting around having to brute-force this by noticing regularities and developing intuitions, taking “explore” directions in explore-exploit tradeoffs, etc. -- this is sort of the whole point of RL, for example.

  I also very much like the points you made about plans. I’d love to understand more about your OODA loop points, but I haven’t yet been able to find a good “layperson” operationalization of OODA that’s not competence porn (in general, I find “sequential problem-solving” stuff coming from pilot training useful as inspiration, but not directly applicable because the context is so different—and I’d love a good reference here that addresses this carefully).

  A vaguely related picture I had in my mind when thinking about the Baba is you discourse (and writing this shortform) comes from being a competitive chess player in middle school. Namely, in middle school competitions and in friendly training games in chess club, people make a big deal out of the “touch move” rule: that you’re not allowed to play around with pieces when planning and you need to form a plan entirely in your head. But then when you see a friendly game between two high-level chess players, they will constantly play around with each other’s pieces to show each other positions several moves into the game that would result from various choices. To someone on a high level (higher than I ever got to), there is very little difference between playing out a game on the board and playing it out in your head, but it’s helpful to move pieces around to communicate your ideas to your partner. I think that (even with a scratchpad), there’s a component of this here: there is a kind of qualitative difference between “learning to track hypothetical positions well” /​ “learning results” /​ “being good at memorization and flashcards” vs. having better intuitions and ideas. A lot of learning a field /​ being a novice in anything consists of being good at the former. But I think “science” as it were progresses by people getting good at the latter. Here I actually don’t think that the “do better vibe” corresponds to not being good at generating new ideas: rather, I think that rationalists (correctly) cultivate a novice mentality, where they constantly learn new skills and approach new areas, where the “train an area of your brain to track sequential behaviors well” (analogous to “mentally chain several moves forward in Baba is you”) is the core skill. And then when rationalists do develop this area and start running “have and test your ideas and intuitions in this environment” loops, these are harder to communicate/​ analyze, and so their importance sort of falls off in the discourse (while on an individual level people are often quite good at these—in fact, the very skill of “communicating well about sequential thinking” is something that many rationalists have developed deep competence in I think).

• Dmitry VaintrobJan 7, 2025, 8:18 PM

  2 points

  0

  in reply to: Alexander Gietelink Oldenziel’s comment on: The Laws of Large Numbers

  So the oscillating phase formula is about approximately integrating the function $\exp (-f(x\left)/\hslash \right)$ against various “priors” p(x) (or more generally any fixed function g), where f is a Lagrangian (think energy) and (\hbar) is a small parameter. It gives an asymptotic series in powers of $\hslash$. The key point is that (more or less) the kth perturbative term only depends on the kth-order power series expansion of f around the “stationary points” (i.e., saddlepoints, Jac(f) = 0) when f is imaginary, on the maxima of f when f is real, and there is a mixed form that depends on stationary points of the imaginary part which are also maxima of the real part (if these exist); the formulae are all exactly the same, with the only difference between real and imaginary f (i.e. statistical vs. quantum mechanics) being whether you only keep maxima or all saddle points.

  Now in SLT, you’re exactly applying the “real” stationary phase formula, i.e., looking at maxima of the (negative) loss function -L(w). The key thing that can happen is that there can be infinitely many maxima, and these might be singular (both in the sense of having higher degree of stationarity, and in the sense of forming a singular manifold). In this case the stationary phase formula is more complicated and AFAIK isn’t completely worked out; Watanabe was the first person who contributed to finding expressions for the general case here beyond the leading correction.

  In the case of maxima which are nondegenerate, i.e., have positive-definite Hessian, the full perturbative expansion is known; in fact, at least in one very useful frame on it, terms are indexed by Feynman diagrams.

  Now the energy function f that appears in this context is the log of the Fourier transform $\hat{p}\left(\theta \right)$ of a probability distribution p(x). Notice that p(x) satisfies $p\left(x\right)\ge 0$ and $\int p\left(x\right)dx=1$. This means that (\hat{p}(0) = \int f(x) dx) is 1 and its log is 0. You can check that all other values of the Fourier transform are $\le 1$ in absolute value (this follows from the fact that $\int |f\left(x\right)|\ge |\int f\left(x\right)|)$. In fact, the Hessian is equal (up to scale) to the variance of p. Now the point is that the only way that variance can be zero is if your pd is concentrated on a lower-dimensional affine subspace, in which case you can simply reduce your problem to a lower-dimensional one with nonsingular Hessian. When this doesn’t happen, the function you’re applying stationary phase to has only nondegenerate maxima, and so the “standard” Feynman-diagram formula applies instead of the more sophisticated Watanabe one that’s used in SLT.

• Dmitry VaintrobJan 6, 2025, 5:33 PM

  59 points

  17

  on: Dmitry Vain­trob’s Shortform

  Alignment is not all you need. But that doesn’t mean you don’t need alignment.

  One of the fairytales I remember reading from my childhood is the “Three sillies”. The story is about a farmer encountering three episodes of human silliness, but it’s set in one more frame story of silliness: his wife is despondent because there is an axe hanging in their cottage, and she thinks that if they have a son, he will walk underneath the axe and it will fall on his head.

  The frame story was much more memorable to me than any of the “body” stories, and I randomly remember this story much more often than any other fairytale I read at the age I read fairytales. I think the reason for this is that the “hanging axe” worry is a vibe very familiar from my family and friend circle, and more generally a particular kind of intellectual neuroticism that I encounter all the time, that is terrified of incomplete control or understanding.

  I really like the rationalist/​EA ecosphere because of its emphasis on the solvability of problems like this: noticing situations where you can just approach the problem, taking down the axe. However, a baseline of intellectual neuroticism persists (after all you wouldn’t expect otherwise from a group of people who pull smoke alarms on pandemics and existential threats that others don’t notice). Sometimes it’s harmless or even beneficial. But a kind of neuroticism in the community that bothers me, and seems counterproductive, is a certain “do it perfectly or you’re screwed” perfectionism that pervades a lot of discussions. (This is also familiar to me from my time as a mathematician: I’ve had discussions with very intelligent and pragmatic friends who rejected even the most basic experimentally confirmed facts of physics because “they aren’t rigorously proven”.)

  A particular train of discussion that annoyed me in this vein was the series of responses to Raemon’s “preplanning and baba is you” post. The initial post I think makes a nice point—it suggests as an experiment trying to solve levels of a move-based logic game by pre-planning every step in advance, and points out that this is hard. Various people tried this experiment and found that it’s hard. This was presented as an issue in solving alignment, in worlds where “we get one shot”. But what annoyed me was the takeaway.

  I think a lot of the great things about the intellectual vibe in the (extended) LW and EA communities is that “you have more ways to solve problems than you think”. However, there is a particular kind of virtue-signally class of problems where trying to find shortcuts or alternatives is frowned upon and the only accepted form of approach is “trying harder” (another generalized intellectual current in the LW-osphere that I strongly dislike).

  Back to the “Baba is you” experiment. The best takeaway, I think, is that we should avoid situations where we need to solve complex problems in one shot, and we should work towards making sure this situation doesn’t exist (and we should just give up on trying to make progress in worlds where we get absolutely no new insights before the do-or-die step of making AGI). Doing so, at least without superhuman assistance, is basically impossible. Attempts at this tend to be not only silly but counterproductive: the “graveyard” of failed idealistic movements are chock-full of wannabe Hari Seldons who believe that they have found the “perfect solution”, and are willing to sacrifice everything to realize their grand vision.

  This doesn’t mean we need to give up, or only work on unambitious, practical applications. But it does mean that we have to admit that things can be useful to work on in expectation before we have a “complete story for how they save the world”.

  Note that what is being advocated here is not an “anything goes” mentality. I certainly think that AI safety research can be too abstract, too removed from any realistic application in any world. But there is a large spectrum of possibilities between “fully plan how you will solve a complex logic game before trying anything” and “make random jerky moves because they ‘feel right’”.

  I’m writing this in response to Adam Jones’ article on AI safety content.. I like a lot of the suggestions. But I think the section on alignment plans suffers from the “axe” fallacy that I claim is somewhat endemic here. Here’s the relevant quote:

  For the last few weeks, I’ve been working on trying to find plans for AI safety. They should cover the whole problem, including the major hurdles after intent alignment. Unfortunately, this has not gone well—my rough conclusion is that there aren’t any very clear and well publicised plans (or even very plausible stories) for making this go well. (More context on some of this work can be found in BlueDot Impact’s AI safety strategist job posting). (emphasis mine).

  I strongly disagree with this being a good thing to do!

  We’re not going to have a good, end-to-end plan about how to save the world from AGI. Even now, with ever more impressive and scary AIs becoming a comonplace, we have very little idea about what AGI will look like, what kinds of misalignment it will have, where the hard bits of checking it for intent and value alignment will be. Trying to make extensive end-to-end plans can be useful, but can also lead to a strong streetlight effect: we’ll be overcommitting to current understanding, current frames of thought (in an alignment community that is growing and integrating new ideas with an exponential rate that can be factored in months, not years).

  Don’t get me wrong. I think it’s valuable to try to plan things where our current understanding is likely to at least partially persist: how AI will interface with government, general questions of scaling and rough models of future development. But we should also understand that our map has lots of blanks, especially when we get down to thinking about what we will understand in the future. What kinds of worrying behaviors will turn out to be relevant and which ones will be silly in retrospect? What kinds of guarantees and theoretical foundations will our understanding of AI encompass? We really don’t know, and trying to chart a course through only the parts of the map that are currently filled out is an extremely limited way of looking at things.

  So instead of trying to solve the alignment problem end to end what I think we should be doing is:

  • getting a variety of good, rough frames on how the future of AI might go

  • thinking about how these will integrate with human systems like government, industry, etc.

  • understanding more things, to build better models in the future.

  I think the last point is crucial, and should be what modern alignment and interpretability is focused on. We really do understand a lot more about AI than we did a few years ago (I’m planning a post on this). And we’ll understand more still. But we don’t know what this understanding will be. We don’t know how it will integrate with existing and emergent actors and incentives. So instead of trying to one-shot the game and write an ab initio plan for how work on quantifying creativity in generative vision models will lead to the world being saved, I think there is a lot of room to just do good research. Fill in the blank patches on that map before routing a definitive course on it. Sure, maybe don’t waste time on the patches in the far corners which are too abstract or speculative or involve too much backchaining. But also don’t try to predict all the axes that will be on the wall in the future before looking more carefully at a specific, potentially interesting, axe.

  What links here?

  • My Jan­uary al­ign­ment the­ory Nanowrimo by Dmitry Vaintrob (Jan 2, 2025, 12:07 AM; 42 points)
  • Is the­ory good or bad for AI safety? by Dmitry Vaintrob (Jan 19, 2025, 10:32 AM; 27 points)

• Dmitry VaintrobJan 6, 2025, 12:09 AM

  2 points

  0

  in reply to: transhumanist_atom_understander’s comment on: The Laws of Large Numbers

  I’m not exactly sure about what you mean wrt “what you want” here. It is not the case that you can exactly reconstruct most probability distributions you’ll encounter in real life from their moments/​ cumulants (hence the expansion is perturbative, not exact).

  But in the interpretability/​ field-theoretic model of wide NN’s point of view, this is what you want (specifically, the fourth-order correction)

• Dmitry VaintrobJan 6, 2025, 12:04 AM

  2 points

  0

  in reply to: Alexander Gietelink Oldenziel’s comment on: The Laws of Large Numbers

  Yes, I actually thought about this a bit. It is definitely the case that the LC (or RLCT) in the SLT context is also exactly a (singular) stationary phase expansion. Unfortunately, the Fourier transform of a random variable, including a higher-dimensional one, really does have an isolated nondegenerate maximum at 0 (unless the support of your random variable is contained in a union of linear subspaces, which is kinda boring/​ reducible to simpler contexts). Maybe if you think about some kind of small perturbation of a lower-dimensional system, you can get some components of the singular free energy expansion, but the expansion relevant here is really nonsingular. This is also the type signature of the expansion you see in most physical QFT systems, at least if they have a perturbative form (in which case, the free theory will in general be nondegenerate).

• Dmitry VaintrobJan 5, 2025, 11:48 PM

  6 points

  3

  on: Rea­sons for and against work­ing on tech­ni­cal AI safety at a fron­tier AI lab

  Thanks for writing this! I’ve participated in some similar conversations and on balance, think that working in a lab is probably net good for most people assuming you have a reasonable amount of intellectual freedom (I’ve been consistently impressed by some papers coming out of Anthropic).

  Still, one point made by Kaarel in a recent conversation seemed like an important update against working in a lab (and working on “close-to-the-metal” interpretability in general). Namely, I tend to not buy arguments by MIRI-adjacent people that “if we share our AI insights with the world then AGI will be developed significantly sooner”. These were more reasonable when they were the only ones thinking seriously about AGI, but now it mostly seems that a capabilities researcher will (on the margin, and at the same skill level) contribute more to making AGI come soon than a safety researcher. But a counterpoint is that serious safety researchers “are trying to actually understand AI”, which has a global orientation towards producing valuable new research results (something like people at the Manhattan project or Apollo program at the hight of these programs’ quality), whereas a capabilities researcher is more driven by local market incentives. So there may be a real sense in which interpretability research, particularly of more practical types, is more dangerous, conditional on “globally new ideas” (like deep learning, transformers etc.) being needed for AGI. This was so far the most convincing argument for me against working on technical interpretability in general, and it might be complicated further by working in a big lab (as I said, it hasn’t been enough to flip my opinion, but seems worth sharing)

• Dmitry VaintrobJan 5, 2025, 11:23 PM

  2 points

  0

  in reply to: Dmitry Vaintrob’s comment on: Alexan­der Gietelink Ol­den­z­iel’s Shortform

  Maybe a reductive summary is “general is good if outer alignment is easy but inner alignment is hard, but bad in the opposite case”

• Dmitry VaintrobJan 5, 2025, 11:18 PM

  2 points

  0

  in reply to: Alexander Gietelink Oldenziel’s comment on: Alexan­der Gietelink Ol­den­z­iel’s Shortform

  I haven’t thought about this enough to have a very mature opinion. On one hand being more general means you’re liable to goodheart more (i.e., with enough deeply general processing power, you understand that manipulating the market to start World War 3 will make your stock portfolio grow, so you act misaligned). On the other hand being less general means that AI’s are more liable to “partially memorize” how to act aligned in familiar situations, and go off the rails when sufficiently out-of-distribution situations are encountered. I think this is related to the question of “how general are humans”, and how stable are human values to being much more or much less general

• Dmitry VaintrobJan 5, 2025, 11:14 PM

  2 points

  0

  in reply to: Alexander Gietelink Oldenziel’s comment on: Alexan­der Gietelink Ol­den­z­iel’s Shortform

  You mean on more general algorithms being good vs. bad?

• Dmitry VaintrobJan 5, 2025, 10:26 PM

  4 points

  0

  in reply to: Alexander Gietelink Oldenziel’s comment on: Alexan­der Gietelink Ol­den­z­iel’s Shortform

  Yep, have been recently posting shortforms (as per your recommendation), and totally with you on the “halfbaked-by-design” concept (if Cheeseboard can do it, it must be a good idea right? :)

  I still don’t agree that free energy is core here. I think that the relevant question, which can be formulated without free energy, is whether various “simplicity/​generality” priors push towards or away from human values (and you can then specialize to questions of effective dimension/​llc, deep vs. shallow networks, ICL vs. weight learning, generalized ood generalization measurements, and so on to operationalize the inductive prior better). I don’t think there’s a consensus on whether generality is “good” or “bad”—I know Paul Christiano and ARC has gone both ways on this at various points.

• Dmitry VaintrobJan 5, 2025, 9:54 PM

  22 points

  2

  on: Dmitry Vain­trob’s Shortform

  On the surprising effectiveness of linear regression as a toy model of generalization.

  Another shortform today (since Sunday is the day of rest). This time it’s really a hot take: I’m not confident about the model described here being correct.

  Neural networks aren’t linear—that’s the whole point. They notice interesting, compositional, deep information about reality. So when people use linear regression as a qualitative comparison point for behaviors like generalization and learning, I tend to get suspicious. Nevertheless, the track record of linear regression as a model for “qualitative” asymptotic behaviors is hard to deny. Linear regression models (neatly analyzable using random matrix theory) give surprisingly accurate models of double descent, scaling phenomena, etc. (at least when comparing to relatively shallow networks like mnist or modular addition).

  I recently developed a cartoon internal model for why this may be the case. I’m not sure if it’s correct, but I’ll share it here.

  The model assumes a few facts about algorithms implemented by NN’s (all of which I believe in much more strongly than the model of comparing them to linear regression):

  1. The generalized Linear Representation Hypothesis. An NN’s internal working can be locally factorized into a large collection of low-level features in distinct low-dimensional linear subspaces, and then applying (generally nonlinear) postprocessing to these features independently or in small batches. Note that this is much weaker than a stronger version (such as the one inherent in SAE) that posits 1-dimensional features. In my experience a version of this hypothesis is almost universally believed by engineers, and also agrees with all known toy algorithms discovered so far.

  2. Smoothness-ish of the data manifold. Inside the low-dimensional “feature subspace”, the data is kinda smooth—i.e., it’s smooth (i.e., locally approximately linear) in most directions, and in directions where it’s not smooth, it still might behave sorta smoothly in aggregate.

  3. Linearity-ish of the classification signal. Even in cases like mnist or transformer learning where the training data is discrete (and the algorithm is mean to approximate it by a continuous function), there is a sense in which it’s locally well-approximated by a linear function. E.g. perhaps some coarse-graining of the discrete data is continuously linear, or at least the data boundary can be locally well approximated by a linear hyperplane (so that a local linear function can attain 100% accuracy). More generally, we can assume a similar local linearity property on the layer-to-layer forward functions, when restricted to either a single feature space or a small group of interacting feature spaces.

  4. (At least partial) locality of the effect of weight modification. When I read it this paper left a lasting impression on me. I’m actually not super excited about its main claims (I’ll discuss polytopes later), but a very cool piece of the analysis here is locally modelling ReLU learning as building a convex function as a max of linear functions (and explaining why non-ReLU learning should exhibit a softer version of the same behavior). This is a somewhat “shallow” point of view on learning, but probably captures a nontrivial part of what’s going on, and this predicts that every new weight update only has local effect—i.e., is felt in a significant way only by a small number of datapoints (the idea being that if you’re defining a convex function as the max of a bunch of linear functions, shifting one of the linear functions will only change the values in places where this particular linear function was dominant). The way I think about this phenomenon is that it’s a good model for “local learning”, i.e., learning closer to memorization on the memorization-generalization spectrum that only updates the behavior on a small cluster of similar datapoints (e.g. the LLM circuit that completes “Barack” with “Obama”). There are also possibly also more diffuse phenomena (like “understanding logic”, or other forms of grokking “overarching structure”), but most likely both forms of learning occur (and it’s more like a spectrum than a dichotomy).

  If we buy that these four phenomena occur, or even “occur sometimes” in a way relevant for learning, then it naturally follows that a part of the “shape” of learning and generalization is well described qualitatively by linear regression. Indeed, the model then becomes that (by point 4 above), many weight updates exclusively “focus on a single local batch of input points” in some low-dimensional feature manifold. For this particular weight update, locality of the update and smoothness of the data manifold (#2) together imply that we can model it as learning a function on a linear low-dimensional space (since smooth manifolds are locally well-approximated by a linear space). Finally, local linearity of the classification function (#3) implies that we’re learning a locally linear function on this local batch of datapoints. Thus we see that, under this collection of assumptions, the local learning subproblems essentially boil down to linear regression.

  Note that the “low-dimensional feature space” assumption, #1, is necessary for any of this to even make sense. Without making this assumption, the whole picture is a non-starter and the other assumptions, #2-#4 don’t make sense, since a sub-exponentially large collection of points on a high-dimensional data manifold with any degree of randomness (something that is true about the data samples in any nontrivial learning problem) will be very far away from each other and the notion of “locality” becomes meaningless. (Note also that a weaker hypothesis than #1 would suffice—in particular, it’s enough that there are low-dimensional “feature mappings” where some clustering occurs at some layer, and these don’t a priori have to be linear.)

  What is this model predicting? Generally I think abstract models like this aren’t very interesting until they make a falsifiable prediction or at least lead to some qualitative update on the behavior of NN’s. I haven’t thought about this very much, and would be excited if others have better ideas or can think of reasons why this model is incorrect. But one thing this model likely predicts is that a better model for a NN than a single linear regression model is a collection of qualitatively different linear regression models at different levels of granularity. In other words, depending on how sloppily you chop your data manifold up into feature subspaces, and how strongly you use the “locality” magnifying glass on each subspace, you’ll get a collection of different linear regression behaviors; you then predict that at every level of granularity, you will observe some combination of linear and nonlinear learning behaviors.

  This point of view makes me excited about work by Ari Brill on fractal data manifolds and scaling laws. If I the paper correctly, he models a data manifold as a certain stochastic fractal in a low-dimensional space and makes scaling predictions about generalization behavior depending on properties of the fractal, by thinking of the fractal as a hierarchy of smooth but noisy features. Finding similarly-flavored behavior scaling behavior on “linear regression subphenomena” in a real-life machine learning problem would positively update me on my model above being correct.

  What links here?

  • My Jan­uary al­ign­ment the­ory Nanowrimo by Dmitry Vaintrob (Jan 2, 2025, 12:07 AM; 42 points)

• Dmitry VaintrobJan 5, 2025, 8:03 PM

  8 points

  0

  in reply to: jake_mendel’s comment on: The sub­set par­ity learn­ing prob­lem: much more than you wanted to know

  Thanks! I definitely believe this, and I think we have a lot of evidence for this in both toy models and LLMs (I’m planning a couple of posts on this idea of “training stories”), and also theoretical reasons in some contexts. I’m not sure how easy it is to extend the specific approach used in the proof for parity to a general context. I think it inherently uses the fact of orthogonality of Fourier functions on boolean inputs, and understanding other ML algorithms in terms of nice orthogonal functions seems hard to do rigorously, unless you either make some kind of simplifying “presumption of independence” model on learnable algorithms or work in a toy context. In the toy case, there is a nice paper that does exactly this (explains how NN’s will tend to find “incrementally learnable” algorithms), by using a similar idea to the parity proof I outlined. This is the leap complexity paper (that Kaarel and I have looked into; I think you’ve also looked into related things)

• Dmitry VaintrobJan 5, 2025, 7:52 PM

  3 points

  1

  in reply to: Alexander Gietelink Oldenziel’s comment on: Alexan­der Gietelink Ol­den­z­iel’s Shortform

  I’m not sure I agree with this—this seems like you’re claiming that misalignment is likely to happen through random diffusion. But I think most worries about misalignment are more about correlated issues, where the training signal consistently disincentivizes being aligned in a subtle way (e.g. a stock trading algorithm manipulating the market unethically because the pressure of optimizing income at any cost diverges from the pressure of doing what its creators would want it to do). If diffusion were the issue, it would also affect humans and not be special to AIs. And while humans do experience value drift, cultural differences, etc., I think we generally abstract these issues as “easier” than the “objective-driven” forms of misalignment

• Dmitry VaintrobJan 5, 2025, 7:45 PM

  4 points

  0

  in reply to: Dmitry Vaintrob’s comment on: Adam Shai’s Shortform

  Ah I think that the notion of amortized inference that you’re using encapsulates what I’m saying about chess. I’m still a little confused about the scope of the concept though—do you have a good cached explanation?

• Dmitry VaintrobJan 5, 2025, 7:39 PM

  6 points

  0

  in reply to: Adam Shai’s comment on: Adam Shai’s Shortform

  I feel like the term “amortization” in ML/​CS has a couple of meanings. Do you just mean redistributing compute from training to inference?

  I think this is an interesting model, but I also think that part of the use of CoT is more specific to the language/​logic context, to literally think step by step (which sometimes lets you split problems into subproblems). In some limit, there would be exponentially few examples in the training data of directly “thinking n steps ahead”, so a transformer wouldn’t be able to learn to do this at all (at least without some impressive RL). Like imagine training a chess playing computer to play chess, by only looking at every 10th move of a chess game: probably with enough inference power, a very powerful system wold be able to reconstruct the rules of chess as the best way of making sense of the regularities in the information, but this is in some sense exponentially harder than learning from looking at every move

• Dmitry VaintrobJan 5, 2025, 7:25 PM

  2 points

  0

  in reply to: Noosphere89’s comment on: Dmitry Vain­trob’s Shortform

  Cute :). Do you mean that we’ve only engineered the alien computer running a single program (the standard model with our universe’s particular coupling constants), or something else?

• Dmitry VaintrobJan 5, 2025, 7:23 PM

  6 points

  2

  in reply to: Noosphere89’s comment on: Dmitry Vain­trob’s Shortform

  This makes sense, but I’d argue that ML and interpretability has even more of both of these properties. Something that makes it harder is that some of the high-level goals of understanding transformers are inherently pretty complex, and also it’s less susceptible to math/​ elegance-based analysis, so is even more messy :)

• Dmitry VaintrobJan 5, 2025, 6:57 AM

  33 points

  11

  on: Dmitry Vain­trob’s Shortform

  We have in fact reverse engineered alien datastructures.

  I’m trying to keep up a regular schedule of writing for my Nanowrimo project. I’m working on a post that’s not done yet, so for today I’ll write a quick (and not very high-effort) take related to discussions I’ve had with more doomy friends (I’m looking at you Kaarel Hänni). A good (though incomplete) crux on why something like “faithful interpretability” may be hard (faithful here being “good enough to genuinely decompose enough of the internal thought process to notice and avoid deception”) is contained in Tsvi’s post on “alien datastructures”. I think it’s a great piece, though I largely probably agree with it only about 50%. A very reductive explanation of the core parable in the piece is that we imagine we’ve found an ancient alien computer from a civilization with a totally separate parallel development of math and logic and computing. The protagonist of the koan lectures on why it might be hard to “do science” to reverse engineer what the computer does, and it might be better to “do deep thinking” to try to understand what problems the computer is trying to accomplish, and then try to come up with the right abstractions. In the case of AI, the recommendation is that this should in particular be done by introspecting and trying to understand our own thinking at a high level. There’s a great quote that I’ll include because it sounds nice:

  Go off and think well——morally, effectively, funly, cooperatively, creatively, agentically, truth-trackingly, understandingly——and observe this thinking——and investigate/​modify/​design this thinking——and derive principles of mind that explain the core workhorses of the impressive things we do including self-reprogramming, and that explain what determines our values and how we continue caring across ontology shifts, and that continue to apply across mental change and across the human-AGI gap; where those principles of mind are made of ideas that are revealed by the counterfactual structure of possible ways of thinking revealed by our interventions on our thinking, like how car parts make more sense after you take them out and replace them with other analogous parts.

  I don’t want to debate the claims of the piece, since after all it’s a koan and is meant to be a parable and a vibe. But I think that as a parallel parable, it’s useful to know that humanity has actually reverse engineered a massively alien computer, and we did it extremely well, and via a combination of “science” (i.e., iteratively designing and testing and refining models) and high-level thinking, though with more of an emphasis on science.

  The alien computer in this analogy is physics. It’s hard to emphasize how different physics is from what people in the past thought it would be. All our past ideas of reality have been challenged: waves aren’t really waves, particles aren’t really particles, time isn’t really time, space isn’t really (just) space. I think that a lot of people who have learned about math or physics, but don’t have experience doing research in it, have a romantic idea that it’s driven by “deep insight”. That physics sits around waiting for the next Newton or Einstein to get a new idea, and then adapts to that idea with the experimentalists there to confirm the great theoreticist’s theories, and the “non-great” rank and file there to compute out their consequences. And like sure, this kind of flow from theorists out is one thing that exists, but it’s mostly wrong. Mostly the way we understand the wild, alien structure that is physics is through just, well, doing science, which involves people tweaking old paradigms, gathering data, finding things that don’t make sense, and then folding the new information back into the big structure. The key point here (that physicists understand viscerally) is that there’s not a big, concept-centric order dependence to discovery. You can understand quantum mechanics first, you can understand relativity first, you can start with waves or particles; in the end, if you are serious about refining and testing your intuitions, and you do calculations to get stuff that increasingly makes more sense, you are on the right track.

  A fun example here comes from the discovery of Heisenberg’s “matrix” model of quantum mechanics. One could imagine a romantic Eureka-moment picture of people suddenly grasping the indeterminacy of reality after thinking deeply about experiments that don’t make sense. But the reality was that after Max Planck’s first idea of quanta of energy levels in 1900 through 1925, people used measurements and arbitrary combinatorics that seemed to mostly agree with experiment to cobble together a weird theory (that the next generation called “Old quantum theory”) of allowed and disallowed orbits that everyone admitted didn’t make sense and fell apart on deep consideration, but sorta worked. After a significant amount of progress was made in this way, Heisenberg and Sommerfeld took a two-parameter table of numbers that went into this theory, made a matrix out of it, and then (in my somewhat cartoonish understanding) noticed that this matrix could be used to measure some other phenomena better. Then Heisenberg realized that just by looking at the matrix, making it complex and viewing it as an evolution operator, you could get a unifying and more satisfying explanation of many of the disparate phenomena in this old theory; it would then take more time for people to develop the more modern concepts of measurement, decoherence, and many worlds (the last of which still meets resistance among the old guard today).

  The point is that physics doesn’t (only) grow through great insight. This might seem surprising to someone who grew up in a certain purist philosophy of logic that, I think incorrectly, has been cultivated on lesswrong: roughly, that ideas are either right, or useless. In fact, the experience of a physicist is that sloppy stopgap ideas that explain experiment or theory in a slightly more systematic way are often useful, though for reasons that will often elude the sloppy idea-haver (physics has a perennial problem of the developers of sloppy ideas getting overinvested in their “exact correctness”, and being resistant to future more sensible systematization—but this is a different story). In physics ideas are deep, but they’re weirdly un-path dependent. There are many examples of people arriving at the same place by different sloppy and unsystematic routes (the Feynman path integral vs. other interpretations of quantum field theory being a great example). This uncanny ability to make progress by taking stopgap measures and then slowly refining them surprises physicists themselves, and I think that some very cool physics “ideology” comes from trying to make sense of this. Two meta-explanations that I think were in part a response to physicists trying to figure out “what the hell is going on”, but led to physics that is used today (especially in the first case) are:

  • effective field theory and the renormalization group flow, coming from ideas of Wilson about how theories at different energy levels can be effectively replaced by simpler theories at other layers (and relating to ideas of Landau and others of complicated concrete theories self-correcting to converge to certain elegant but more abstract “universal” theories at larger scale)

  • the “bootstrap” idea that complicated theories with unknown moving parts have measurements that satisfy certain standard formulas and inequalities, and one can often get quite far (indeed sometimes, all the way) by ignoring the “physics” and only looking at the formulas as a formal system.

  Both of these physics ideas identify a certain shape of “emergent structure”, where getting at something with even a very small amount of “orientational push” from reality, or from known sloppy structure, will tend to lead to the same “correct” theory at scales we care about.

  This doesn’t mean that big ideas and deep thinking are useless. But I think that this does point (at least in physics) to taking more seriously ideas that “aren’t fully self-consistent yet” from a more demanding mathematical standpoint. We can play random fun measurement and calculation and systematization games, and tinker (“morally, effectively, funly, cooperatively, creatively, agentically, truth-trackingly, understandingly”) with systematizations of the results of these games, and we’re likely to get there in the end.

  At least if the alien computer is physics.

  For later. In my posts I’m hoping to talk in more detail about renormalization and the self-correcting nature of universality, and also about why I think that the complexity of (certain areas of) physics is a good match for the complexity of neural nets, and why I think the modern field of interpretability and “theoretical ML” more generally is much further along this game of tinkering and systematizing than many people think (in particular, in some ways we’re beyond the “Old Quantum Theory”-style mess). But this is it for now.

  What links here?

  • My Jan­uary al­ign­ment the­ory Nanowrimo by Dmitry Vaintrob (Jan 2, 2025, 12:07 AM; 42 points)

Dmitry Vain­trob’s Shortform