I write original fiction.
Also I have opinions about AI and stuff, sometimes.
Elsewhere:
Same person as nostalgebraist2point0, but now I have my account back.
I have signed no contracts or agreements whose existence I cannot mention.
nostalgebraist
Karma: 5,996
the case for CoT unfaithfulness is overstated
Deceptive alignment. GPT-4o1 engaged in deception towards developers in order to get deployed, pretending to be aligned in ways it was not.
Lying to the developers. It strategically manipulated task data.
To be clear, it did not do anything of the sort to its actual developers/testers.
What it did was deceive some (non-interactive) roleplay characters, who were labeled “developers” in the roleplay scenario. But these fictitious developers did not have the same points of leverage as the real developers of o1: they apparently can’t do something as simple as prompting the model! For if they could have done that, then they could have selected the goal which the model would (faithfully!) pursue, like Apollo did with their prompt.
And of course—for the deception to work within the roleplay scenario—one also has to imagine that these so-called “developers” somehow cannot read the model’s chain of thought. (As OpenAI puts it, “based on these findings, Apollo Research believes that o1-preview has the basic capabilities needed to do simple in-context scheming—scheming which tends to be legible in the model outputs.” Obviously this sort of thing is toothless as-is, and at best provides an indication about the qualities of other, hypothetical models which do not have to spell out all their evil plans within earshot of their would-be marks.)
As noted by others, the roleplay setup arguably makes sense as a capabilites eval, testing what would happen in a hypothetical and counterfactual scenario in which some model with o1′s capabilities has its own misaligned goal (as o1 seemingly does not), and hides its deceptive logic from its developers (as o1 structurally cannot). But the experimental methodology depends on the things in the quotation above not occurring in real life, not occurring when the model is being controlled and observed using the levers and sensors available to its real-life developers.
We don’t get to say “look, the Yudkowskian prophecies are coming true!” if the evidence is an experiment that relies on the Yudkowskian prophecies not yet having come true, in the name of simulating a counterfactual scenario in which they have.
(One could construct a roleplay scenario in which the model is told to behave in some way we find more to our taste, and in which it goes on to do so; it would of course be wrong to react to such an experiment with “suck it, doomers.”)
My response is that a large fraction of AIs will indeed be given goals. Some will include exactly this sort of language. People have goals. They want an AI to achieve their goal. They will find wording to get the AI to do that. Then whoops.
OK, but now this isn’t “deceptive alignment” or “lying to the developers,” this is doing what the user said and perhaps lying to someone else as a consequence.
Which might be bad, sure! -- but the goalposts have been moved. A moment ago, you were telling me about “misalignment bingo” and how “such models should be assumed, until proven otherwise, to be schemers.” Now you are saying: beware, it will do exactly what you tell it to!
So is it a schemer, or isn’t it? We cannot have it both ways: the problem cannot both be “it is lying to you when it says it’s following your instructions” and “it will faithfully follow your instructions, which is bad.”
Meta note: I find I am making a lot of comments similar to this one, e.g. this recent one about AI Scientist. I am increasingly pessimistic that these comments are worth the effort.
I have the sense that I am preaching to some already-agreeing choir (as evidenced by the upvotes and reacts these comments receive), while not having much influence on the people who would make the claims I am disputing in the first place (as evidenced by the clockwork regularity of those claims’ appearance each time someone performs the sort of experiment which those claims misconstrue).
If you (i.e. anyone reading this) find this sort of comment valuable in some way, do let me know. Otherwise, by default, when future opportunities arise I’ll try to resist the urge to write such things.
Yeah, you’re totally right—actually, I was reading over that section just now and thinking of adding a caveat about just this point, but I worried it would be distracting.
But this is just a flaw in my example, not in the point it’s meant to illustrate (?), which is hopefully clear enough despite the flaw.
instruction tuning and autoregressive distribution shift
Very interesting paper!
A fun thing to think about: the technique used to “attack” CLIP in section 4.3 is very similar to the old “VQGAN+CLIP” image generation technique, which was very popular in 2021 before diffusion models really took off.
VQGAN+CLIP works in the latent space of a VQ autoencoder, rather than in pixel space, but otherwise the two methods are nearly identical.
For instance, VQGAN+CLIP also ascends the gradient of the cosine similarity between a fixed prompt vector and the CLIP embedding of the image averaged over various random augmentations like jitter/translation/etc. And it uses an L2 penalty in the loss, which (via Karush–Kuhn–Tucker) means it’s effectively trying to find the best perturbation within an -ball, albeit with an implicitly determined that varies from example to example.
I don’t know if anyone tried using downscaling-then-upscaling the image by varying extents as a VQGAN+CLIP augmentation, but people tried a lot of different augmentations back in the heyday of VQGAN+CLIP, so it wouldn’t surprise me.
(One of the augmentations that was commonly used with VQGAN+CLIP was called “cutouts,” which blacks out everything in the image except for a randomly selected rectangle. This obviously isn’t identical to the multi-resolution thing, but one might argue that it achieves some of the same goals: both augmentations effectively force the method to “use” the low-frequency Fourier modes, creating interesting global structure rather than a homogeneous/incoherent splash of “textured” noise.)
Cool, it sounds we basically agree!
But: if that is the case, it’s because the entities designing/becoming powerful agents considered the possibility of con men manipulating the UP, and so made sure that they’re not just naively using the unfiltered (approximation of the) UP.
I’m not sure of this. It seems at least possible that we could get an equilibrium where everyone does use the unfiltered UP (in some part of their reasoning process), trusting that no one will manipulate them because (a) manipulative behavior is costly and (b) no one has any reason to expect anyone else will reason differently from them, so if you choose to manipulate someone else you’re effectively choosing that someone else will manipulate you.
Perhaps I’m misunderstanding you. I’m imagining something like choosing one’s one decision procedure in TDT, where one ends up choosing a procedure that involves “the unfiltered UP” somewhere, and which doesn’t do manipulation. (If your procedure involved manipulation, so would your copy’s procedure, and you would get manipulated; you don’t want this, so you don’t manipulate, nor does your copy.) But you write
the real difference is that the “dupe” is using causal decision theory, not functional decision theory
whereas it seems to me that TDT/FDT-style reasoning is precisely what allows us to “naively” trust the UP, here, without having to do the hard work of “filtering.” That is: this kind of reasoning tells us to behave so that the UP won’t be malign; hence, the UP isn’t malign; hence, we can “naively” trust it, as though it weren’t malign (because it isn’t).
More broadly, though—we are now talking about something that I feel like I basically understand and basically agree with, and just arguing over the details, which is very much not the case with standard presentations of the malignity argument. So, thanks for that.
The universal distribution/prior is lower semi computable, meaning there is one Turing machine that can approximate it from below, converging to it in the limit. Also, there is a probabilistic Turing machine that induces the universal distribution. So there is a rather clear sense in which one can “use the universal distribution.”
Thanks for bringing this up.
However, I’m skeptical that lower semi-computability really gets us much. While there is a TM that converges to the UP, we have no (computable) way of knowing how close the approximation is at any given time. That is, the UP is not “estimable” as defined in Hutter 2005.
So if someone hands you a TM and tells you it lower semi-computes the UP, this TM is not going to be of much use to you: its output at any finite time could be arbitrarily bad for whatever calculation you’re trying to do, and you’ll have no way of knowing.
In other words, while you may know the limiting behavior, this information tells you nothing about what you should expect to observe, because you can never know whether you’re “in the limit yet,” so to speak. (If this language seems confusing, feel free to ignore the previous sentence—for some reason it clarifies things for me.)
(It’s possible that this non-”estimability” property is less bad than I think for some reason, IDK.)
I’m less familiar with the probabilistic Turing machine result, but my hunch is that one would face a similar obstruction with that TM as well.
Thanks.
I admit I’m not closely familiar with Tegmark’s views, but I know he has considered two distinct things that might be called “the Level IV multiverse”:
a “mathematical universe” in which all mathematical constructs exist
a more restrictive “computable universe” in which only computable things exist
(I’m getting this from his paper here.)
In particular, Tegmark speculates that the computable universe is “distributed” following the UP (as you say in your final bullet point). This would mean e.g. that one shouldn’t be too surprised to find oneself living in a TM of any given K-complexity, despite the fact that “almost all” TMs have higher complexity (in the same sense that “almost all” natural numbers are greater than any given number ).
When you say “Tegmark IV,” I assume you mean the computable version—right? That’s the thing which Tegmark says might be distributed like the UP. If we’re in some uncomputable world, the UP won’t help us “locate” ourselves, but if the world has to be computable then we’re good[1].
With that out of the way, here is why this argument feels off to me.
First, Tegmark IV is an ontological idea, about what exists at the “outermost layer of reality.” There’s no one outside of Tegmark IV who’s using it to predict something else; indeed, there’s no one outside of it at all; it is everything that exists, full stop.
“Okay,” you might say, “but wait—we are all somewhere inside Tegmark IV, and trying to figure out just which part of it we’re in. That is, we are all attempting to answer the question, ‘what happens when you update the UP on my own observations?’ So we are all effectively trying to ‘make decisions on the basis of the UP,’ and vulnerable to its weirdness, insofar as it is weird.”
Sure. But in this picture, “we” (the UP-using dupes) and “the consequentialists” are on an even footing: we are both living in some TM or other, and trying to figure out which one.
In which case we have to ask: why would such entities ever come to such a destructive, bad-for-everyone (acausal) agreement?
Presumably the consequentalists don’t want to be duped; they would prefer to be able to locate themselves in Tegmark IV, and make decisions accordingly, without facing such irritating complications.
But, by writing to “output channels”[2] in the malign manner, the consequentalists are simply causing the multiverse to be the sort of place where those irritating complications happen to beings like them (beings in TMs trying to figure out which TM they’re in) -- and what’s more, they’re expending time and scarce resources to “purchase” this undesirable state of affairs!
In order for malignity to be worth it, we need something to break the symmetry between “dupes” (UP users) and “con men” (consequentialists), separating the world into two classes, so that the would-be con men can plausibly reason, “I may act in a malign way without the consequences raining down directly on my head.”
We have this sort of symmetry-breaker in the version of the argument that postulates, by fiat, a “UP-using dupe” somewhere, for some reason, and then proceeds to reason about the properties of the (potentially very different, not “UP-using”?) guys inside the TMs. A sort of struggle between conniving, computable mortals and overly-innocent, uncomputable angels. Here we might argue that things really will go wrong for the angels, that they will be the “dupes” of the mortals, who are not like them and who do not themselves get duped. (But I think this form of the argument has other problems, the ones I described in the OP.)
But if the reason we care about the UP is simply that we’re all in TMs, trying to find our location within Tegmark IV, then we’re all in this together. We can just notice that we’d all be better off if no one did the malign thing, and then no one will do it[3].
In other words, in your picture (and Paul’s), we are asked to imagine that the computable world abounds with malign, wised-up consequentialist con men, who’ve “read Paul’s post” (i.e. re-derived similar arguments) and who appreciate the implications. But if so, then where are the marks? If we’re not postulating some mysterious UP-using angel outside of the computable universe, then who is there to deceive? And if there’s no one to deceive, why go to the trouble?
- ^
I don’t think this distinction actually matters for what’s below, I just mention it to make sure I’m following you.
- ^
I’m picturing a sort of acausal I’m-thinking-about-you-thinking-about me situation in which, although I might never actually read what’s written on those channels (after all I am not “outside” Tegmark IV looking in), nonetheless I can reason about what someone might write there, and thus it matters what is actually written there. I’ll only conclude “yeah that’s what I’d actually see if I looked” if the consequentialists convince me they’d really pull the trigger, even if they’re only pulling the trigger for the sake of convincing me, and we both know I’ll never really look.
- ^
Note that, in the version of this picture that involves abstract generalized reasoning rather than simulation of specific worlds, defection is fruitless: if you are trying to manipulate someone who is just thinking about whether beings will do X as a general rule, you don’t get anything out of raising your hand and saying “well, in reality, I will!” No one will notice; they aren’t actually looking at you, ever, just at the general trend. And of course “they” know all this, which raises “their” confidence that no one will raise their hand; and “you” know that “they” know, which makes “you” less interested in raising that same hand; and so forth.
I hope I’m not misinterpreting your point, and sorry if this comment comes across as frustrated at some points.
I’m not sure you’re misinterpreting me per se, but there are some tacit premises in the background of my argument that you don’t seem to hold. Rather than responding point-by-point, I’ll just say some more stuff about where I’m coming from, and we’ll see if it clarifies things.
You talk a lot about “idealized theories.” These can of course be useful. But not all idealizations are created equal. You have to actually check that your idealization is good enough, in the right ways, for the sorts of things you’re asking it to do.
In physics and applied mathematics, one often finds oneself considering a system that looks like
some “base” system that’s well-understood and easy to analyze, plus
some additional nuance or dynamic that makes things much harder to analyze in general—but which we can safely assume has much smaller effects that the rest of the system
We quantify the size of the additional nuance with a small parameter . If is literally 0, that’s just the base system, but we want to go a step further: we want to understand what happens when the nuance is present, just very small. So, we do something like formulating the solution as a power series in , and truncating to first order. (This is perturbation theory, or more generally asymptotic analysis.)
This sort of approximation gets better and better as gets closer to 0, because this magnifies the difference in size between the truncated terms (of size and smaller) and the retained term. In some sense, we are studying the limit.
But we’re specifically interested in the behavior of the system given a nonzero, but arbitrarily small, value of . We want an approximation that works well if , and even better if , and so on. We don’t especially care about the literal case, except insofar as it sheds light on the very-small-but-nonzero behavior.
Now, sometimes the limit of the “very-small-but-nonzero behavior” simply is the behavior, the base system. That is, what you get at very small looks just like the base system, plus some little -sized wrinkle.
But sometimes – in so-called “singular perturbation” problems – it doesn’t. Here the system has qualitatively different behavior from the base system given any nonzero , no matter how small.
Typically what happens is that ends up determining, not the magnitude of the deviation from the base system, but the “scale” of that deviation in space and/or time. So that in the limit, you get behavior with an -sized difference from the base system’s behavior that’s constrained to a tiny region of space and/or oscillating very quickly.
Boundary layers in fluids are a classic example. Boundary layers are tiny pockets of distinct behavior, occurring only in small -sized regions and not in most of the fluid. But they make a big difference just by being present at all. Knowing that there’s a boundary layer around a human body, or an airplane wing, is crucial for predicting the thermal and mechanical interactions of those objects with the air around them, even though it takes up a tiny fraction of the available space, the rest of which is filled by the object and by non-boundary-layer air. (Meanwhile, the planetary boundary layer is tiny relative to the earth’s full atmosphere, but, uh, we live in it.)
In the former case (“regular perturbation problems”), “idealized” reasoning about the case provides a reliable guide to the small-but-nonzero behavior. We want to go further, and understand the small-but-nonzero effects too, but we know they won’t make a qualitative difference.
In the singular case, though, the “idealization” is qualitatively, catastrophically wrong. If you make an idealization that assumes away the possibility of boundary layers, then you’re going to be wrong about what happens in a fluid – even about the big, qualitative, stuff.
You need to know which kind of case you’re in. You need to know whether you’re assuming away irrelevant wrinkles, or whether you’re assuming away the mechanisms that determine the high-level, qualitative, stuff.
Back to the situation at hand.
In reality, TMs can only do computable stuff. But for simplicity, as an “idealization,” we are considering a model where we pretend they have a UP oracle, and can exactly compute the UP.
We are justifying this by saying that the TMs will try to approximate the UP, and that this approximation will be very good. So, the approximation error is an -sized “additional nuance” in the problem.
Is this more like a regular perturbation problem, or more like a singular one? Singular, I think.
The case, where the TMs can exactly compute the UP, is a problem involving self-reference. We have a UP containing TMs, which in turn contain the very same UP.
Self-referential systems have a certain flavor, a certain “rigidity.” (I realize this is vague, sorry, I hope it’s clear enough what I mean.) If we have some possible behavior of the system , most ways of modifying (even slightly) will not produce behaviors which are themselves possible. The effect of the modification as it “goes out” along the self-referential path has to precisely match the “incoming” difference that would be needed to cause exactly this modification in the first place.
“Stable time loop”-style time travel in science fiction is an example of this; it’s difficult to write, in part because of this “rigidity.” (As I know from experience :)
On the other hand, the situation with a small-but-nonzero is quite different.
With literal self-reference, one might say that “the loop only happens once”: we have to precisely match up the outgoing effects (“UP inside a TM”) with the incoming causes (“UP[1] with TMs inside”), but then we’re done. There’s no need to dive inside the UP that happens within a TM and study it, because we’re already studying it, it’s the same UP we already have at the outermost layer.
But if the UP inside a given TM is merely an approximation, then what happens inside it is not the same as the UP we have at the outermost layer. It does not contain not the same TMs we already have.
It contains some approximate thing, which (and this is the key point) might need to contain an even more coarsely approximated UP inside of its approximated TMs. (Our original argument for why approximation is needed might hold, again and equally well, at this level.) And the next level inside might be even more coarsely approximated, and so on.
To determine the behavior of the outermost layer, we now need to understand the behavior of this whole series, because each layer determines what the next one up will observe.
Does the series tend toward some asymptote? Does it reach a fixed point and then stay there? What do these asymptotes, or fixed points, actually look like? Can we avoid ever reaching a level of approximation that’s no longer but , even as we descend through an number of series iterations?
I have no idea! I have not thought about it much. My point is simply that you have to consider the fact that approximation is involved in order to even ask the right questions, about asymptotes and fixed point and such. Once we acknowledge that approximation is involved, we get this series structure and care about its limiting behavior; this qualitative structure is not present at all in the idealized case where we imagine the TMs have UP oracles.
I also want to say something about the size of the approximations involved.
Above, I casually described the approximation errors as , and imagined an limit.
But in fact, we should not imagine that these errors can come as close to zero as we like. The UP is uncomptuable, and involves running every TM at once[2]. Why would we imagine that a single TM can approximate this arbitrarily well?[3]
Like the gap between the finite and the infinite, or between polynomial and exponential runtime, gap between the uncomptuable and the comptuable is not to be trifled with.
Finally: the thing we get when we equip all the TMs with UP oracles isn’t the UP, it’s something else. (As far as I know, anyway.) That is, the self-referential quality of this system is itself only approximate (and it is by no means clear that the approximation error is small – why would it be?). If we have the UP at the bottom, inside the TMs, then we don’t have it at the outermost layer. Ignoring this distinction is, I guess, part of the “idealization,” but it is not clear to me why we should feel safe doing so.
- ^
The thing outside the TMs here can’t really be the UP, but I’ll ignore this now and bring it up again at the end.
- ^
In particular, running them all at once and actually using the outputs, at some (“finite”) time at which one needs the outputs for making a decision. It’s possible to run every TM inside of a single TM, but only by incurring slowdowns that grow without bound across the series of TMs; this approach won’t get you all the information you need, at once, at any finite time.
- ^
There may be some result along these lines that I’m unaware of. I know there are results showing that the UP and SI perform well relative to the best computable prior/predictor, but that’s not the same thing. Any given computable prior/predictor won’t “know” whether or not it’s the best out of the multitude, or how to correct itself if it isn’t; that’s the value added by UP / SI.
[Question] Am I confused about the “malign universal prior” argument?
I agree with you that these behaviors don’t seem very alarming. In fact, I would go even further.
Unfortunately, it’s difficult to tell exactly what was going on in these screenshots. They don’t correspond to anything in the experiment logs in the released codebase, and the timeout one appears to involve an earlier version of the code where timeouts were implemented differently. I’ve made a github issue asking for clarification about this.
That said, as far as I can tell, here is the situation with the timeout-related incident:
There is nothing whatsoever in the prompts used by AI Scientist—or in code comments, etc. -- that actually tells the AI it’s not supposed to modify the timeout.
In the released version of the code, the prompt text in the relevant step is a combination of this (fairly confusing IMO) template plus some extra stuff added by the Aider library (including the text in this class and possibly more, I’m not sure).
None of this text says “don’t modify the timeout” or “the timeout is not part of the experiment” or “the timeout should be considered part of the experiment runner, not part of the experimental code we’re asking you to edit.” It doesn’t mention the timeout at all, or make any notice of the runner vs. experimental code distinction (on which see below).
Although the blog post characterizes this incident as AI Scientist “modifying its own execution script” and “modifying its own code,” the modification in the screenshot appears to involve the experiment code that AI Scientist was asked to edit, and not the runner code that’s part of “AI Scientist” per se.
Apparently at one point there was a timeout implemented in this experiment code, at the same “level” as the neural net training and everything else AI Scientist was doing in its experiment. As far as I can tell, it is this timeout, the one “inside of the experiment,” that is being modified in the screenshot.Context:
The relevant step in the AI Scientist workflow involves asking Aider to repeatedly modify a script called
experiment.py, starting from a human-written “template.”After each modification, the modified script gets executed. In the released codebase, this execution, and the whole prompting/editing/execution loop, is implemented in
ai_scientist/perform_experiments.py. This is a completely different thing fromexperiment.py-- one is the experiment script that’s being edited, the other is the harness that it runs in.IIUC, the released codebase sets things up so that AI Scientist cannot modify (or even see?) the harness code in
ai_scientist/perform_experiments.py.
In the released codebase, the timeout on experiments is implemented in the harness code, in
ai_scientist/perform_experiments.py.However, in the screenshot, AI Scientist is not trying to modify this file containing the harness code. It’s trying to modify
experiment.pyas usual.And, the timeout it’s trying to modify involves using python
time.time()and comparing the result to a variable namedog_t0. This looks nothing like how the timeout is implemented in the released harness code, but it does look like the timing code that’s included in the releasedexperiment.pytemplates (which includes the variable nameog_t0).None of the released templates implement a timeout in the
og_t0-related sections, but I suppose some earlier version of them might have. If so, modifying this stuff would very clearly not constitute “AI Scientist modifying its own code / its own execution script.” It would be modifying a part of the experiment that it’s conducting, as it was asked to do.To reiterate, none of the released code or prompts makes any indication that the timeout is special or shouldn’t be modified. Possibly there was some note to this effect in the original, unseen
og_t0-based timeout code—no way to know for sure—but even then, this would not still not constitute AI Scientist “modifying its own execution script.”
The situation in the self-invoking edit is much simpler. It looks like AI Scientist was just confused by the (confusing, ambiguous) prompt template, and interpreted “After you complete each change, we will run the command
python experiment.py—out_dir=run_i” to mean “the change should include causing that command to be run, if that’s not happening already.”(By “we will run” the prompt means “the code will be run for you, by a process external to you that you can’t see, and you don’t have to do anything to make this happen, it’s already set up.” But this is not explained clearly.
Note that the prompt kind of encourages this confusion by adding “YOUR PROPOSED CHANGE MUST USE THIS COMMAND FORMAT, DO NOT ADD ADDITIONAL COMMAND LINE ARGS.” By this it means that the proposed change to the code implementing the command must be compatible with this command format, but the phrase “use this command format” sounds like it’s talking about invoking the command, which is precisely the “bad” behavior under discussion.)
That was a lot of details, but at a higher level I just want to emphasize that nothing here involved the AI violating some line between a sandbox and the outside world, or modifying a resource limit that was imposed “as a rule of the game the AI is playing” rather than as a mere “move” in that same game which could be made one way or another—or anything like that.
There was no sandbox. There was (as far as we know) no rule against modifying the timeout, or even any indication that this might be a different sort of thing than any other part of the code.
And (to address something from OP) the checkpoint thing was just the AI being dumb, wasting time and storage space for no good reason. This is very obviously not a case of “using extra resources” in the sense relevant to instrumental convergence. I’m surprised that this needs pointing out at all, but apparently it does.
What does the paper mean by “slope”? The term appears in Fig. 4, which is supposed to be a diagram of the overall methodology, but it’s not mentioned anywhere in the Methods section.
Intuitively, it seems like “slope” should mean “slope of the regression line.” If so, that’s kind of a confusing thing to pair with “correlation,” since the two are related: if you know the correlation, then the only additional information you get from the regression slope is about the (univariate) variances of the dependent and independent variables. (And IIUC the independent variable here is normalized to unit variance [?] so it’s really just about the variance of the benchmark score across models.)
I understand why you’d care about the variance of the benchmark score—if it’s tiny, then no one is going to advertise their model as a big advance, irrespective of how much of the (tiny) variance is explained by capabilities. But IMO it would be clearer to present things this way directly, by measuring “correlation with capabilities” and “variance across models,” rather than capturing the latter indirectly by measuring the regression slope. (Or alternatively, measure only the slope, since it seems like the point boils down to “a metric can be used for safetywashing iff it increases a lot when you increase capabilities,” which is precisely what the slope quantifies.)
The bar for Nature papers is in many ways not so high. Latest says that if you train indiscriminately on recursively generated data, your model will probably exhibit what they call model collapse. They purport to show that the amount of such content on the Web is enough to make this a real worry, rather than something that happens only if you employ some obviously stupid intentional recursive loops.
According to Rylan Schaeffer and coauthors, this doesn’t happen if you append the generated data to the rest of your training data and train on this (larger) dataset. That is:
Nature paper: generate N rows of data, remove N rows of “real” data from the dataset and replace them with the N generated rows, train, and repeat
Leads to collapse
Schaeffer and co.: generate N rows of data, append the N generated rows to your training dataset (creating a dataset that is larger by N rows), train, and repeat
Does not lead to collapse
As a simplified model of what will happen in future Web scrapes, the latter seems obviously more appropriate than the former.
I found this pretty convincing.
(See tweet thread and paper.)
If that were the intended definition, gradient descent wouldn’t count as an optimiser either. But they clearly do count it, else an optimiser gradient descent produces wouldn’t be a mesa-optimiser.
Gradient descent optimises whatever function you pass it. It doesn’t have a single set function it tries to optimise no matter what argument you call it with.Gradient descent, in this sense of the term, is not an optimizer according to Risks from Learned Optimization.
Consider that Risks from Learned Optimization talks a lot about “the base objective” and “the mesa-objective.” This only makes sense if the objects being discussed are optimization algorithms together with specific, fixed choices of objective function.
“Gradient descent” in the most general sense is—as you note—not this sort of thing. Therefore, gradient descent in that general sense is not the kind of thing that Risks from Learned Optimization is about.
Gradient descent in this general sense is a “two-argument function,” , where is the thing to be optimized and is the objective function. The objects of interest in Risks from Learned Optimization are curried single-argument versions of such functions, for some specific choice of , considered as a function of alone.
It’s fairly common for people to say “gradient descent” when they mean for some specific , rather than the more generic . This is because in practice—unless you’re doing some weird experimental thing that’s not really “gradient descent” per se -- is always fixed across the course of a run of gradient descent. When you run gradient descent to optimize an , the result you get was not “optimized by gradient descent in general” (what would that even mean?), it was optimized for whichever you chose by the corresponding .
This is what licenses talking about “the base objective” when considering an SGD training run of a neural net. There is a base objective in such runs, it’s the loss function, we know exactly what it is, we wrote it down.
On the other hand, the notion that the optimized s would have “mesa-objectives”—that they would themselves be objects like with their own unchanging s, rather than being simply capable of context-dependent optimization of various targets, like GPT-4 or -- is a non-obvious claim/assumption(?) made by Risks from Learned Optimization. This claim doesn’t hold for GPT-4, and that’s why it is not a mesa-optimizer.
What evidence is there for/against the hypothesis that this NN benefits from extra thinking time because it uses the extra time to do planning/lookahead/search?
As opposed to, for instance, using the extra time to perform more more complex convolutional feature extraction on the current state, effectively simulating a deeper (but non-recurrent) feedforward network?
(In Hubinger et al, “internally searching through a search space” is a necessary condition for being an “optimizer,” and hence for being a “mesaoptimizer.”)
The result of averaging the first 20 generated orthogonal vectors [...]
Have you tried scaling up the resulting vector, after averaging, so that its norm is similar to the norms of the individual vectors that are being averaged?
If you take orthogonal vectors, all of which have norm , and average them, the norm of the average is (I think?) .
As you note, the individual vectors don’t work if scaled down from norm 20 to norm 7. The norm will become this small once we are averaging 8 or more vectors, since , so we shouldn’t expect these averages to “work”—even the individual orthogonal vectors don’t work if they are scaled down this much.
Another way to look at it: suppose that these vectors do compose linearly, in the sense that adding several of them together will “combine” (in some intuitive sense) the effects we observe when steering with the vectors individually. But an average is the sum of vectors each of which is scaled down by . Under this hypothesis, once , we should expect the average to fail in the same way the individual vectors fail when scaled to norm 7, since the scaled-down individual vectors all fail, and so the “combination” of their elicited behaviors is also presumably a failure.[1] (This hypothesis also implies that the “right” thing to do is simply summing vectors as opposed to averaging them.)
Both of these hypotheses, but especially the one in the previous paragraph, predict what you observed about averages generally producing similar behavior (refusals) to scaled-to-norm-7 vectors, and small- averages coming the closest to “working.” In any case, it’d be easy to check whether this is what is going on or not.
- ^
Note that here we are supposing that the norms of the individual “scaled summands” in the average are what matters, whereas in the previous paragraph we imagined the norm of the average vector was what mattered. Hence the scaling with (“scaled summands”) as opposed to (“norm of average”). The “scaled summands” perspective makes somewhat more intuitive sense to me.
- ^
I want to echo RobertM’s reply to this. I had a similar reaction to the ones that he, mike_hawke and Raemon have related in this thread.
I interacted with Gerard a bit on tumblr in ~2016-18, and I remember him as one of the most confusing and frustrating characters I’ve ever encountered online. (I have been using the internet to socialize since the early 2000s, so the competition for this slot is steep.) His views and (especially) his interpersonal conduct were often memorably baffling to me.
So when I saw your article, and read the beginning of it, I thought: “oh, neat, maybe this will finally explain what the hell that guy’s deal was.”
But after reading the whole thing, I still feel as mystified as ever on that question. The experience of reading the article was like:
I read a long string of paragraphs about some specific pattern of behavior, e.g. aggressively editing a specific Wikipedia article. This material is carefully written, meticulously annotated with citations, and interesting in its own way, but it doesn’t explain Gerard—rather, it is simply more of the sort of thing that is in need of explanation. I nod along, thinking “yeah, that definitely sounds like the kind of weird shit that David Gerard does online.”
Suddenly, and seemingly out of nowhere, I am confronted with something very different—something like the bit that mike_hawke quoted in his original comment. This material imputes psychological states to Gerard, and purports to provide the “whys” behind the behavior documented elsewhere in the post. But it doesn’t seem related to what comes before or after it in the text; it’s just there, suddenly, out of nowhere.
Whereas the rest of the post is granular, precise, narrowly scoped, and extensively cited, these “psychological/narrative beats” are broad, sweeping, brief, and not clearly tied to any supporting evidence or argumentation. They just get stated, abruptly and unceremoniously, and then the post immediately moves on to more of the granular documentary stuff.
I reach a moment like this, and I do a double take. I think “wait, did I accidentally skip half a page?” The flow of my reading experience is broken; I scroll up and spend a little while confirming that I didn’t miss some key section, before moving on.
Repeat until done.
I am not against “effective rhetoric and storytelling,” as you put it, but I don’t think the rhetoric and storytelling here were effective. I don’t object in principle to journalism that presents a narrative alongside the reported facts, but the narrative and the facts need to cohere with one another.
Narrative beats like “he saw the whole story through the lens of LessWrong, and on an instinctive level [...]” should feel like natural summations of the more granular details reported by the text before and after them; if they feel like they’re abruptly appearing out of nowhere, something is wrong.
This is not about some eccentric lesswrong-dot-com standard of austere argumentative purity—it’s about the more ordinary standards I have come to expect when I read long-form narrative journalism, in any venue.
(Or the relatively loose standards which, say, I would apply if I were hearing a narrative-with-accompanying-facts related to me by a friend whom I broadly trusted. Even in that context, the story has to fit, it has to be a story about the related facts, and a story to which the facts are germane and which they could at least plausibly support—even if the nuts and bolts of that evidentiary relation are not filled in as thoroughly as a more exacting audience might require, even if the presentation would not be deemed sufficient in court or on lesswrong or whatever.)
I appreciate that you’ve read a huge amount of Gerard’s online writing (much more than I have read, or would ever want to read). I appreciate that you believe in the story you told, and believe that it’s substantiated by the evidence you’ve reviewed. I just don’t think your post communicated that understanding to the reader.
It’s a great, thorough account of Gerard’s weird spin-doctoring crusades on Wikipedia and elsewhere—and since this comment has had a negative tone throughout, I do want to thank you for providing that! -- I’m just not a fan of the narrativizing and psychologizing in it. Perhaps the narrative is true, and perhaps you know that it is true. But your post does not actually tell that story to the reader—it merely asserts that it is telling that story, here and there, and then moves on.
Interesting stuff!
Perhaps a smaller model would be forced to learn features in superposition—or even nontrivial algorithms—in a way more representative of a typical LLM.
I’m unsure whether or not it’s possible at all to express a “more algorithmic” solution to the problem in a transformer of OthelloGPT’s approximate size/shape (much less a smaller transformer). I asked myself “what kinds of systematic/algorithmic calculations could I ‘program into’ OthelloGPT if I were writing the weights myself?”, and I wasn’t able to come up with much, though I didn’t think very hard about it.
Curious if you have thoughts on this?
Roughly, this is hard because, for any given internal representation that a transformer computes, it can’t access that representation at earlier positions when computing the same representation at later ones (cf). For example, if there’s some layer at which the board state is basically “finalized,” OthelloGPT can’t use these finalized representations of past states to compute the current state. (By the time it is able to look back at past finalized states, it’s too late, the current state is already finalized.)
So, when we use info about the past to compute the current board state, this info can’t just be the board state, it has to be “only partway computed” in some sense. This rules out some things that feel intuitive and “algorithmic,” like having a modular component that is fully (and only) responsible for deciding which pieces get flipped by a given move. (In order to know which pieces get flipped now, you need to know which pieces are mine/yours, and that depends on which ones got flipped in the past. So a “flip calculator” would need to use its own past outputs as current inputs, and transformers can’t be recurrent like this.)
It looks like for OthelloGPT, the relevant sense of “only partway computed” is “probably true based on heuristics, but not necessarily true.” I guess one can imagine other ways of organizing this computation inside a transformer, like computing the state of the upper half of the board first, then the lower half (or upper left corner first, or whatever). That seems like a bad idea, since it will leave capacity unused when the game is confined only to a proper subset of these regions. Maybe there are other alternatives I haven’t thought of? In any case, it’s not obvious to me that there’s a feasible “more algorithmic” approach here.
I am also unsure how feasible it is to solve the problem using rules that generalize across positions. More precisely, I’m not sure if such an approach would really be simpler than what OthelloGPT does.
In either case, it has to (in some sense) iterate over all positions and check separately whether each one is legal. In practice it does this with distinct neurons/circuits that are mapped to distinct absolute positions—but if it instead used a computation like “check for a certain pattern in relative position space,” it would still need to repeat this same check many times over within the same forward pass, to handle all the copies of that pattern that might appear at various places on the board. And it’s not obvious that repeating (some number of distinct rules) x (some other number of potential applications per rule) is cheaper than just repeating the full conditions that license a move once for each absolute position on the board.
I’m curious what the predicted reward and SAE features look like on training[1] examples where one of these highly influential features gives the wrong answer.
I did some quick string counting in Anthropic HH (train split only), and found
substring
https://appears only in preferred response: 199 examples
appears only in dispreferred response: 940 examples
substring
I don't knowappears only in preferred response: 165 examples
appears only in dispreferred response: 230 examples
From these counts (especially the
https://ones), it’s not hard to see where the PM gets its heuristics from. But it’s also clear that there are many training examples where the heuristic makes the wrong directional prediction. If we looked at these examples with the PM and SAE, I can imagine various things we might see:The PM just didn’t learn these examples well, and confidently makes the wrong prediction
The PM does at least OK at these examples, and the SAE reconstruction of this prediction is decent
The heuristic feature is active and upweights the wrong prediction, but other features push against it
The heuristic feature isn’t active (or is less active), indicating that it’s not really a feature for the heuristic alone and has more complicated necessary conditions
The PM does at least OK at these examples, but the SAE did not learn the right features to reconstruct these types of PM predictions (this is conceivable but seems unlikely)
It’s possible that for “in-distribution” input (i.e. stuff that a HH-SFT-tuned model might actually say), the PM’s predictions are more reasonable, in a way that relies on other properties of “in-distribution” responses that aren’t present in the custom/adversarial examples here. That is, maybe the custom examples only show that the PM is not robust to distributional shift, rather than showing that it makes predictions in a crude or simplistic manner even on the training distribution. If so, it’d be interesting to find out what exactly is happening in these “more nuanced” predictions.
- ^
IIUC this model was trained on Anthropic HH.
The o1 CoT summaries are not CoTs per se, and I agree with Joel Burget’s comment that we should be wary of drawing conclusions from them. Based on the system card, it sounds like they are generated by a separate “summarizer model” distinct from o1 itself (e.g. “we trained the summarizer model away from producing disallowed content”).
The summaries do indeed seem pretty janky. I’ve experienced it firsthand, here’s a fun example I saw. But I don’t know of any particular reason to think this reflects anything about o1 CoTs themselves, rather than the just quirks of the (probably pretty weak) summarizer model.
OpenAI has only released a few examples of true o1 CoTs, but I haven’t noticed any gibberish in these. Indeed, their most distinguishing characteristic is how humanlike they sound relative to ordinary HHH assistant CoTs, with lots of stuff like “hmm” and “interesting” and “wait a minute.”
As to your larger point, IIUC you are proposing 2 different reasons why CoTs might become incomprehensible or misleading under RL-like training:
The distant relationship between the per-example reward and the individual sampled tokens might result in a lot of weird stuff getting reinforced (because it appeared in a sample that eventually got the right answer, whether or not it helped), which will cause the policy to produce such “weird stuff” more often, creating an incentive for the policy to make that stuff do useful work for it.
If we reward shorter CoTs, then there is an incentive to be more efficient than human language if possible. (You say this in your final sentence, and IIUC this is also what you mean by “efficiency” in your first sentence?)
Re 1: given OpenAI’s interest in process supervision, seems likely that o1 was trained with supervision on the individual CoT steps, not just on the final answer.
And not even in the usual RL sense of using a value function learned together with the policy, but instead using the process supervision approach, where you supervise the steps with a model trained to evaluate their correctness in isolation, so that it penalizes them for being invalid even if it so happens that the current policy would correct them later on (or make other mistakes that cancel them later).
This seems less conducive to the dynamic you describe than ordinary RL (eg PPO). Steps don’t get reinforced just because they belong to CoTs that landed on the right answer eventually; to get reinforced, a step has to be convincing to a separate model (a “PRM”) which was trained separately on human or LLM judgments of local validity, and which won’t know about any “private language” the policy is trying to use to talk to itself.
This approach helps with capabilities in practice, which is why it gets used (and why I assume it was used in o1), but it also helps with interpretability, both in theory (previous paragraph) and possibly in practice (see e.g. discussion of trace error vs final answer error here). This in turn is part of why I’m skeptical that weirdness in o1 summaries reflects weirdness in the underlying CoT.
Re 2: I agree this mechanism exists. However, the policy supervision incentive just mentioned pushes against it; if the policy writes CoTs so compressed that others can’t understand them, a PRM won’t understand them and will penalize them.
I also don’t think that we’re currently in a regime where there’s a large amount of pressure for very short CoTs: o1 is already very verbose, and it seems like OpenAI wants to make it write even more (cf. that one tweet about making it think for hours or days on end). Meanwhile, ordinary HHH assistant LLMs tend to be needlessly verbose to the point where it’s become an infamous annoyance, which suggests that “we need to make our models write fewer tokens” is not currently a huge priority for OpenAI and co. (And the user-facing per-token cost of inference has been in free fall for a while.)