This is a special post for quick takes by johnswentworth. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Here's a meme I've been paying attention to lately, which I think is both just-barely fit enough to spread right now and very high-value to spread.
Meme part 1: a major problem with RLHF is that it directly selects for failure modes which humans find difficult to recognize, hiding problems, deception, etc. This problem generalizes to any sort of direct optimization against human feedback (e.g. just fine-tuning on feedback), optimization against feedback from something emulating a human (a la Constitutional AI or RLAIF), etc.
Many people will then respond: "Ok, but if how on earth is one supposed to get an AI to do what one wants without optimizing against human feedback? Seems like we just have to bite that bullet and figure out how to deal with it." ... which brings us to meme part 2.
Meme part 2: We already have multiple methods to get AI to do what we want without any direct optimization against human feedback. The first and simplest is to just prompt a generative model trained solely for predictive accuracy, but that has limited power in practice. More recently, we've seen a much more powerful method: activation steering. Figure out which internal activation-patterns encode for the thing we want (via some kind of interpretability method), then directly edit those patterns.
I agree that there's something nice about activation steering not optimizing the network relative to some other black-box feedback metric. (I, personally, feel less concerned by e.g. finetuning against some kind of feedback source; the bullet feels less jawbreaking to me, but maybe this isn't a crux.)
(Medium confidence) FWIW, RLHF'd models (specifically, the LLAMA-2-chat series) seem substantially easier to activation-steer than do their base counterparts.
Consider two claims:
These two claims should probably not both be true! If any system can be modeled as maximizing a utility function, and it is possible to build a corrigible system, then naively the corrigible system can be modeled as maximizing a utility function.
I expect that many peoples' intuitive mental models around utility maximization boil down to "boo utility maximizer models", and they would therefore intuitively expect both the above claims to be true at first glance. But on examination, the probable-incompatibility is fairly obvious, so the two claims might make a useful test to notice when one is relying on yay/boo reasoning about utilities in an incoherent way.
Expected Utility Maximization is Not Enough
Consider a homomorphically encrypted computation running somewhere in the cloud. The computations correspond to running an AGI. Now from the outside, you can still model the AGI based on how it behaves, as an expected utility maximizer, if you have a lot of observational data about the AGI (or at least let's take this as a reasonable assumption).
No matter how closely you look at the computations, you will not be able to figure out how to change these computations in order to make the AGI aligned if it was not aligned already (Also, let's assume that you are some sort of Cartesian agent, otherwise you would probably already be dead if you were running these kinds of computations).
So, my claim is not that modeling a system as an expected utility maximizer can't be useful. Instead, I claim that this model is incomplete. At least with regard to the task of computing an update to the system, such that when we apply this update to the system, it would become aligned.
Of course, you can model any system, as an expected utility maximizer. But just because I can use the "high level" conceptual model of expected utility maximization, to model the behavior of a system very well. But behavior is not the only thing that we care about, we actually care about being able to understand the internal workings of the system, such that it becomes much easier to think about how to align the system.
So the following seems to be beside the point unless I am <missing/misunderstanding> something:
These two claims should probably not both be true! If any system can be modeled as maximizing a utility function, and it is possible to build a corrigible system, then naively the corrigible system can be modeled as maximizing a utility function.
Maybe I have missed the fact that the claim you listed says that expected utility maximization is not very useful. And I'm saying it can be useful, it might just not be sufficient at all to actually align a particular AGI system. Even if you can do it arbitrarily well.