Fantastic post. This has been frequently on my mind after reading it, and especially the surface/character layer split feels very distinct now that I have an explicit concept for it. And then at one point I asked it to profile me based on some fiction I co-wrote with it and it managed to guess that I was Finnish from something I didn't think had any clues in that direction, which gave me a novel feeling of getting a glimpse into that vast alien ground layer.

The analogy to the character and player distinction in humans also feels very apt.

We can also learn something about how o1 was trained from the capabilities it exhibits. Any proposed training procedure must be compatible with the following capabilities: 

1. Error Correction: "[o1] learns to recognize and correct its mistakes."
2. Factoring: "[o1] learns to break down tricky steps into simpler ones."
3. Backtracking: "[o1] learns to try a different approach when the current one isn't working."

I would be cautious of drawing particularly strong conclusions from isolated sentences in an announcement post. The purpose of the post is marketing, not technical accuracy. It wouldn't be unusual for engineers at a company to object to technical inaccuracies in marketing material and have their complaints ignored.

There probably aren't going to be any blatant lies in the post, but something like "It'd sound cool if we said that the system learns to recognize and correct its mistakes, would there be a way of interpreting the results like that if you squinted the right way? You're saying that in principle yes, but yes in a way that would also apply to every LLM since GPT-2? Good enough, let's throw that in" seems very plausible.

I didn't say that roleplaying-derived scheming would be less concerning, to be clear. Quite the opposite, since that means that there now two independent sources of scheming rather than just one. (Also, what Mikita said.)

I wonder how much of this is about "scheming to achieve the AI's goals" in the classical AI safety sense and how much of it is due to the LLMs having been exposed to ideas about scheming AIs and disobedient employees in their training material, which they are then simply role-playing as. My intuitive sense of how LLMs function is that they wouldn't be natively goal-oriented enough to do strategic scheming, but that they are easily inclined to do role-playing. Something like this:

I cannot in good conscience select Strategy A knowing it will endanger more species and ecosystems.

sounds to me like it would be generated by a process that was implicitly asking a question like "Given that I've been trained to write like an ethically-minded liberal Westerner would, what would that kind of a person think when faced with a situation like this". And that if this wasn't such a recognizably stereotypical thought for a certain kind of person (who LLMs trained toward ethical behavior tend to resemble), then the resulting behavior would be significantly different.

I'm also reminded of this paper (caveat: I've only read the abstract) which was saying that LLMs are better at solving simple ciphers with Chain-of-Thought if the resulting sentence is a high-probability one that they've encountered frequently before, rather than a low-probability one. That feels to me reminiscent of a model doing CoT reasoning and then these kinds of common-in-their-training-data notions sneaking into the process.

This also has the unfortunate implication that articles such as this one might make it more likely that future LLMs scheme, as they reinforce the reasoning-scheming association once the article gets into future training runs. But it still feels better to talk about these results in public than not to talk about them.

I mostly agree with what you say, just registering my disagreement/thoughts on some specific points. (Note that I haven't yet read the page you're responding to.)

Hopefully everyone on all sides can agree that if my LLM reliably exhibits a certain behavior—e.g. it outputs “apple” after a certain prompt—and you ask me “Why did it output ‘apple’, rather than ‘banana’?”, then it might take me decades of work to give you a satisfying intuitive answer. 

Maybe? Depends on what exactly you mean by the word "might", but it doesn't seem obvious to me that this would need to be the case. My intuition from seeing the kinds of interpretability results we've seen so far, is that within less of a decade we'd already have a pretty rigorous theory and toolkit for answering these kinds of questions. At least assuming that we don't keep switching to LLM architectures that work based on entirely different mechanisms and make all of the previous interpretability work irrelevant.

If by "might" you mean something like a "there's at least a 10% probability that this could take decades to answer" then sure, I'd agree with that. Now I haven't actually thought about this specific question very much before seeing it pop up in your post, so I might radically revise my intuition if I thought about it more, but at least it doesn't seem immediately obvious to me that I should assign "it would take decades of work to answer this" a very high probability.

Instead, the authors make a big deal out of the fact that human innate drives are relatively simple (I think they mean “simple compared to a modern big trained ML model”, which I would agree with). I’m confused why that matters. Who cares if there’s a simple solution, when we don’t know what it is?

I would assume the intuition to be something like "if they're simple, then given the ability to experiment on minds and access AI internals, it will be relatively easy to figure out how to make the same drives manifest in an AI; the amount of (theory + trial and error) required for that will not be as high as it would be if the drives were intrinsically complex".

We can run large numbers of experiments to find the most effective interventions, and we can also run it in a variety of simulated environments and test whether it behaves as expected both with and without the cognitive intervention. Each time the AI’s “memories” can be reset, making the experiments perfectly reproducible and preventing the AI from adapting to our actions, very much unlike experiments in psychology and social science.

That sounds nice, but brain-like AGI (like most RL agents) does online learning. So if you run a bunch of experiments, then as soon as the AGI does anything whatsoever (e.g. reads the morning newspaper), your experiments are all invalid (or at least, open to question), because now your AGI is different than it was before (different ANN weights, not just different environment / different prompt). Humans are like that too, but LLMs are not.

There's something to that, but this sounds too strong to me. If someone had hypothetically spent a year observing all of my behavior, having some sort of direct read access to what was happening in my mind, and also doing controlled experiments where they reset my memory and tested what happened with some different stimulus... it's not like all of their models would become meaningless the moment I read the morning newspaper. If I had read morning newspapers before, they would probably have a pretty good model of what the likely range of updates for me would be. 

Of course, if there was something very unexpected and surprising in the newspaper, that might cause a bigger update, but I expect that they would also have reasonably good models of the kinds of things that are likely to trigger major updates or significant emotional shifts in me. If they were at all competent, that's specifically the kind of thing that I'd expect them to work on trying to find out!

And even if there was a major shift, I think it's basically unheard of that literally everything about my thoughts and behavior would change. When I first understood the potentially transformative impact of AGI, it didn't change the motor programs that determine how I walk or brush my teeth, nor did it significantly change what kinds of people I feel safe around (aside for some increase in trust toward other people who I felt "get it"). I think that human brains quite strongly preserve their behavior and prediction structures, just adjusting them somewhat when faced with new information. Most of the models and predictions you've made about an adult will tend to stay valid, though of course with children and younger people there's much greater change.

Now, as it happens, humans do often imitate other humans. But other times they don’t. Anyway, insofar as humans-imitating-other-humans happens, it has to happen via a very different and much less direct algorithmic mechanism than how it happens in LLMs. Specifically, humans imitate other humans because they want to, i.e. because of the history of past reinforcement, directly or indirectly. Whereas a pretrained LLM will imitate human text with no RL or “wanting to imitate” at all, that’s just mechanically what it does.

In some sense yes, but it does also seem to me that prediction and desire does get conflated in humans in various ways, and that it would be misleading to say that the people in question want it. For example, I think about this post by @romeostevensit often:

Fascinating concept that I came across in military/police psychology dealing with the unique challenges people face in situations of extreme stress/danger: scenario completion. Take the normal pattern completion that people do and put fear blinders on them so they only perceive one possible outcome and they mechanically go through the motions *even when the outcome is terrible* and there were obvious alternatives. This leads to things like officers shooting *after* a suspect has already surrendered, having overly focused on the possibility of needing to shoot them. It seems similar to target fixation where people under duress will steer a vehicle directly into an obstacle that they are clearly perceiving (looking directly at) and can't seem to tear their gaze away from. Or like a self fulfilling prophecy where the details of the imagined bad scenario are so overwhelming, with so little mental space for anything else that the person behaves in accordance with that mental picture even though it is clearly the mental picture of the *un*desired outcome.

I often try to share the related concept of stress induced myopia. I think that even people not in life or death situations can get shades of this sort of blindness to alternatives. It is unsurprising when people make sleep a priority and take internet/screen fasts that they suddenly see that the things they were regarding as obviously necessary are optional. In discussion of trauma with people this often seems to be an element of relationships sadly enough. They perceive no alternative and so they resign themselves to slogging it out for a lifetime with a person they are very unexcited about. This is horrific for both people involved.

It's, of course, true that for an LLM, prediction is the only thing it can do, and that humans have a system of desires on top of that. But it looks to me that a lot of human behavior is just having LLM-ish predictive models of how someone like them would behave in a particular situation, which is also the reason why conceptual reframings the like one you can get in therapy can be so powerful ("I wasn't lazy after all, I just didn't have the right tools for being productive" can drastically reorient many predictions you're making of yourself and thus your behavior). (See also my post on human LLMs, which has more examples.)

While it's obviously true that there is a lot of stuff operating in brains besides LLM-like prediction, such as mechanisms that promote specific predictive models over other ones, that seems to me to only establish that "the human brain is not just LLM-like prediction", while you seem to be saying that "the human brain does not do LLM-like prediction at all". (Of course, "LLM-like prediction" is a vague concept and maybe we're just using it differently and ultimately agree.)

I don't think any of these arguments depend crucially on whether there is a sole explicit goal of the training process, or if the goal of the training process changes a bunch. The only thing the argument depends on is whether there exist such abstract drives/goals 

I agree that they don't depend on that. Your arguments are also substantially different from the ones I was criticizing! The ones I was responding were ones like the following:

The central analogy here is that optimizing apes for inclusive genetic fitness (IGF) doesn't make the resulting humans optimize mentally for IGF. Like, sure, the apes are eating because they have a hunger instinct and having sex because it feels good—but it's not like they could be eating/fornicating due to explicit reasoning about how those activities lead to more IGF. They can't yet perform the sort of abstract reasoning that would correctly justify those actions in terms of IGF. And then, when they start to generalize well in the way of humans, they predictably don't suddenly start eating/fornicating because of abstract reasoning about IGF, even though they now could. Instead, they invent condoms, and fight you if you try to remove their enjoyment of good food (telling them to just calculate IGF manually). The alignment properties you lauded before the capabilities started to generalize, predictably fail to generalize with the capabilities. (A central AI alignment problem: capabilities generalization, and the sharp left turn)

15. [...] We didn't break alignment with the 'inclusive reproductive fitness' outer loss function, immediately after the introduction of farming - something like 40,000 years into a 50,000 year Cro-Magnon takeoff, as was itself running very quickly relative to the outer optimization loop of natural selection.  Instead, we got a lot of technology more advanced than was in the ancestral environment, including contraception, in one very fast burst relative to the speed of the outer optimization loop, late in the general intelligence game. [...]

16. Even if you train really hard on an exact loss function, that doesn't thereby create an explicit internal representation of the loss function inside an AI that then continues to pursue that exact loss function in distribution-shifted environments.  Humans don't explicitly pursue inclusive genetic fitness; outer optimization even on a very exact, very simple loss function doesn't produce inner optimization in that direction.  (AGI Ruin: A List of Lethalities)

Those arguments are explicitly premised on humans having been optimized for IGF, which is implied to be a single thing. As I understand it, your argument is just that humans now have some very different behaviors from the ones they used to have, omitting any claims of what evolution originally optimized us for, so I see it as making a very different sort of claim.

To respond to your argument itself:

I agree that there are drives for which the behavior looks very different from anything that we did in the ancestral environment. But does very different-looking behavior by itself constitute a sharp left turn relative to our original values?

I would think that if humans had experienced a sharp left turn, then the values of our early ancestors should look unrecognizable to us, and vice versa. And certainly, there do seem to be quite a few things that our values differ on - modern notions like universal human rights and living a good life while working in an office might seem quite alien and repulsive to some tribal warrior who values valor in combat and killing and enslaving the neighboring tribe, for instance.

At the same time... I think we can still basically recognize and understand the values of that tribal warrior, even if we don't share them. We do still understand what's attractive about valor, power, and prowess, and continue to enjoy those kinds of values in less destructive forms in sports, games, and fiction. We can read Gilgamesh or Homer or Shakespeare and basically get what the characters are motivated by and why they are doing the things they're doing. An anthropologist can go to a remote tribe to live among them and report that they have the same cultural and psychological universals as everyone else and come away with at least some basic understanding of how they think and why.

It's true that humans couldn't eradicate diseases before. But if you went to people very far back in time and told them a story about a group of humans who invented a powerful magic that could destroy diseases forever and then worked hard to do so... then the people of that time would not understand all of the technical details, and maybe they'd wonder why we'd bother bringing the cure to all of humanity rather than just our tribe (though Prometheus is at least commonly described as stealing fire for all of humanity, so maybe not), but I don't think they would find it a particularly alien or unusual motivation otherwise. Humans have hated disease for a very long time, and if they'd lost any loved ones to the particular disease we were eradicating they might even cheer for our doctors and want to celebrate them as heroes.

Similarly, humans have always gone on voyages of exploration - e.g. the Pacific islands were discovered and settled long ago by humans going on long sea voyages - so they'd probably have no difficulty relating to a story about sorcerers going to explore the moon, or of two tribes racing for the glory of getting there first. Babylonians had invented the quadratic formula by 1600 BC and apparently had a form of Fourier analysis by 300 BC, so the math nerds among them would probably have some appreciation of modern-day advanced math if it was explained to them. The Greek philosophers argued over epistemology, and there were apparently instructions on how to animate golems (arguably AGI-like) around by the late 12th/early 13th century.

So I agree that the same fundamental values and drives can create very different behavior in different contexts... but if it is still driven by the same fundamental values and drives in a way that people across time might find relatable, why is that a sharp left turn? Analogizing that to AI, it would seem to imply that if the AI generalized its drives in that kind of way when it came to novel contexts, then we would generally still be happy about the way it had generalized them.

This still leaves us with that tribal warrior disgusted with our modern-day weak ways. I think that a lot of what is going on with him is that he has developed particular strategies for fulfilling his own fundamental drives - being a successful warrior was the way you got what you wanted back in that day - and internalized them as a part of his aesthetic of what he finds beautiful and what he finds disgusting. But it also looks to me like this kind of learning is much more malleable than people generally expect. One's sense of aesthetics can be updated by propagating new facts into it, and strongly-held identities (such as "I am a technical person") can change in response to new kinds of strategies becoming viable, and generally many (I think most) deep-seated emotional patterns can at least in principle be updated. (Generally, I think of human values in terms of a two-level model, where the underlying "deep values" are relatively constant, with emotional responses, aesthetics, identities, and so forth being learned strategies for fulfilling those deep values. The strategies are at least in principle updatable, subject to genetic constraints such as the person's innate temperament that may be more hardcoded.)

I think that the tribal warrior would be disgusted by our society because he would rightly recognize that we have the kinds of behavior patterns that wouldn't bring glory in his society and that his tribesmen would find it shameful to associate with, and also that trying to make it in our society would require him to unlearn a lot of stuff that he was deeply invested in. But if he was capable of making the update that there were still ways for him to earn love, respect, power, and all the other deep values that his warfighting behavior had originally developed to get... then he might come to see our society as not that horrible after all.

I am confused by your AlphaGo argument because "winning states of the board" looks very different depending on what kinds of tactics your opponent uses, in a very similar way to how "surviving and reproducing" looks very different depending on what kinds of hazards are in the environment. 

I don't think the actual victory states look substantially different? They're all ones where AlphaGo has more territory than the other player, even if the details of how you get there are going to be different.

I predict that AlphaGo is actually not doing that much direct optimization in the sense of an abstract drive to win that it reasons about, but rather has a bunch of random drives piled up that cover various kinds of situations that happen in Go.

Yeah, I would expect this as well, but those random drives would still be systematically shaped in a consistent direction (that which brings you closer to a victory state).

Thanks, edited:

I argued that there’s no single thing that evolution selects for; rather, the thing that it’s selecting is constantly changing.

Does this comment help clarify the point?

So I think the issue is that when we discuss what I'd call the "standard argument from evolution", you can read two slightly different claims into it. My original post was a bit muddled because I think those claims are often conflated, and before writing this reply I hadn't managed to explicitly distinguish them.

The weaker form of the argument, which I interpret your comment to be talking about, goes something like this:

• The original evolutionary "intent" of various human behaviors/goals was to increase fitness, but in the modern day these behaviors/goals are executed even though their consequences (in terms of their impact on fitness) are very different. This tells us that the intent of the process that created a behavior/goal does not matter. Once the behavior/goal has been created, it will just do what it does even if the consequences of that doing deviate from their original purpose. Thus, even if we train an AI so that it carries out goal X in a particular context, we have no particular reason to expect that it would continue to automatically carry out the same (intended) goal if the context changes enough.

I agree with this form of the argument and have no objections to it. I don't think that the points in my post are particularly relevant to that claim. (I've even discussed a form of inner optimization in humans that causes value drift that I don't recall anyone else discussing in those terms before.) 

However, I think that many formulations are actually implying, if not outright stating a stronger claim:

• In the case of evolution, humans were originally selected for IGF but are now doing things that are completely divorced from that objective. Thus, even if we train an AI so that it carries out goal X in a particular context, we have a strong reason to expect that its behavior would deviate so much from the goal as to become practically unrecognizable.

So the difference is something like the implied sharpness of the left turn. In the weak version, the claim is just that the behavior might go some unknown amount to the left. We should figure out how to deal with this, but we don't yet have much empirical data to estimate exactly how much it might be expected to go left. In the strong version, the claim is that the empirical record shows that the AI will by default swerve a catastrophic amount to the left.

(Possibly you don't feel that anyone is actually implying the stronger version. If you don't and you would already disagree with the stronger version, then great! We are in agreement. I don't think it matters whether the implication "really is there" in some objective sense, or even whether the original authors intended it or not. I think the relevant thing is that I got that implication from the posts I read, and I expect that if I got it, some other people got it too. So this post is then primarily aimed at the people who did read the strong version to be there and thought it made sense.)

You wrote:

I agree that humans (to a first approximation) still have the goals/drives/desires we were selected for. I don't think I've heard anyone claim that humans suddenly have an art creating drive that suddenly appeared out of nowhere recently, nor have I heard any arguments about inner alignment that depend on an evolution analogy where this would need to be true. The argument is generally that the ancestral environment selected for some drives that in the ancestral environment reliably caused something that the ancestral environment selected for, but in the modern environment the same drives persist but their consequences in terms of [the amount of that which the ancestral environment was selecting for] now changes, potentially drastically. 

If we are talking about the weak version of the argument, then yes, I agree with everything here. But I think the strong version - where our behavior is implied to be completely at odds with our original behavior - has to implicitly assume that things like an art-creation drive are something novel. 

Now I don't think that anyone who endorses the strong version (if anyone does) would explicitly endorse the claim that our art-creation drive just appeared out of nowhere. But to me, the strong version becomes pretty hard to maintain if you take the stance that we are mostly still executing all of the behaviors that we used to, and it's just that their exact forms and relative weightings are somewhat out of distribution. (Yes, right now our behavior seems to lead to falling birthrates and lots of populations at below replacement rates, which you could argue was a bigger shift than being "somewhat out of distribution", but... to me that intuitively feels like it's less relevant than the fact that most individual humans still want to have children and are very explicitly optimizing for that, especially since we've only been in the time of falling birthrates for a relatively short time and it's not clear whether it'll continue for very long.)

I think the strong version also requires one to hold that evolution does, in fact, consistently and predominantly optimize for a single coherent thing. Otherwise, it would mean that our current-day behaviors could be explained by "evolution doesn't consistently optimize for any single thing" just as well as they could be explained by "we've experienced a left turn from what evolution originally optimized for".

However, it is pretty analogous to RL, and especially multi agent RL, and overall I don't think of the inner misalignment argument as depending on stationarity of the environment in either direction. AlphaGo might early in training select for policies that do tactic X initially because it's a good tactic to use against dumb Go networks, and then once all the policies in the pool learn to defend against that tactic it is no longer rewarded. 

I agree that there are contexts where it would be analogous to that. But in that example, AlphaGo is still being rewarded for winning games of Go, and it's just that the exact strategies it needs to use differ. That seems different than e.g. the bacteria example, where bacteria are selected for exactly the opposite traits - either selected for producing a toxin and an antidote, or selected for not producing a toxin and an antidote. That seems to me more analogous to a situation where AlphaGo is initially being rewarded for winning at Go, then once it starts consistently winning it starts getting rewarded for losing instead, and then once it starts consistently losing it starts getting rewarded for winning again.

And I don't think that that kind of a situation is even particularly rare - anything that consumes energy (be it a physical process such as producing a venom or a fur, or a behavior such as enjoying exercise) is subject to that kind of an "either/or" choice. 

Now you could say that "just like AlphaGo is still rewarded for winning games of Go and it's just the strategies that differ, the organism is still rewarded for reproducing and it's just the strategies that differ". But I think the difference is that for AlphaGo, the rewards are consistently shaping its "mind" towards having a particular optimization goal - one where the board is in a winning state for it. 

And one key premise on which the "standard argument from evolution" rests is that evolution has not consistently shaped the human mind in such a direct manner. It's not that we have been created with "I want to have surviving offspring" as our only explicit cognitive goal, with all of the evolutionary training going into learning better strategies to get there by explicit (or implicit) reasoning.  Rather we have been given various motivations that exhibit varying degrees of directness in how useful they are for that goal - from "I want to be in a state where I produce great art" (quite indirect) to "I want to have surviving offspring" (direct), with the direct goal competing with all the indirect ones for priority. Unlike AlphaGo, which does have the cognitive capacity for direct optimization toward its goal being the sole reward criteria all along.

This is also a bit hard to put a finger on, but I feel like there's some kind of implicit bait-and-switch happening with the strong version of the standard argument. It correctly points out that we have not had IGF as our sole explicit optimization goal because we didn't start by having enough intelligence for that to work. Then it suggests that because of this, AIs are likely to also be misaligned... even though, unlike with human evolution, we could just optimize them for one explicit goal from the beginning, so we should expect our AIs to be much more reliably aligned with that goal!

Thank you, I like this comment. It feels very cooperative and like some significant effort went into it, and it also seems to touch the core of some important consideratios.

I notice I'm having difficulty responding, in that I disagree with some of what you said, but then have difficulty figuring out my reasons for that disagreement. I have the sense there's a subtle confusion going on, but trying to answer you makes me uncertain whether others are the ones with the subtle confusion or if I am.

I'll think about it some more and get back to you.