Also, there is a big difference between “Calling for violence”, and “calling for the establishment of an international treaty, which is to be enforced by violence if necessary”. I don’t understand why so many people are muddling this distinction.
Chris van Merwijk(Chris van Merwijk)
“For instance, personally I think the reason so few people take AI alignment seriously is that we haven’t actually seen anything all that scary yet. “
And if this “actually scary” thing happens, people will know that Yudkowsky wrote the article beforehand, and they will know who the people are that mocked it.
I suggest renaming this to “countably factored spaces”. Countably being a property of the factorization rather than the space.
Also I suggest adding an actual self-contained definition of countable factored space to make it more readable.
Maybe you have made a gestalt-switch I haven’t made yet, or maybe yours is a better way to communicate the same thing, but: the way I think of it is that the reward function is just a function from states to numbers, and the way the information contained in the reward function affects the model parameters is via reinforcement of pre-existing computations.
Is there a difference between saying:
A reward function is an objective function, but the only way that it affects behaviour is via reinforcement of pre-existing computations in the model, and it doesn’t actually encode in any way the “goal” of the model itself.
A reward function is not an objective function, and the only way that it affects behaviour is via reinforcement of pre-existing computations in the model, and it doesn’t actually encode in any way the “goal” of the model itself.
It seems to me that once you acknowledge the point about reinforcement, the additional statement that reward is not an objective doesn’t actually imply anything further about the mechanistic properties of deep reinforcement learners? It is just a way to put a high-level conceptual story on top of it, and in this sense it seems to me that this point is already known (and in particular, contained within RFLO), even though we talked of the base objective still as an “objective”.
However, it might be that while RFLO pointed out the same mechanistic understanding that you have in mind, but calling it an objective tends in practice to not fully communicate that mechanistic understanding.
Or it might be that I am really not yet understanding that there is an actual diferrence in mechanistic understanding, or that my intuitions are still being misled by the wrong high-level concept even if I have the lower-level mechanistic understanding right.
(On the other hand, one reason to still call it an objective is because we really can think of the selection process, i.e. evolution/the learning algorithm of an RL agent, as having an objective but making imperfect choices, or we can think of the training objective as encoding a task that humans have in mind).
Very late reply, sorry.
“even though reward is not a kind of objective”, this is a terminological issue. In my view, calling a “antecedent-computation reinforcement criterion” an “objective” matches my definition of “objective”, and this is just a matter of terminology. The term “objective” is ill-defined enough that “even though reward is not a kind of objective” is a terminological claim about objective, not a claim about math/the world.
The idea that RL agents “reinforce antecedent computations” is completely core to our story of deception. You could not make sense of our argument for deception if you didn’t look at RL systems in this way. Viewing the base optimizer as “trying” to achieve an “objective” but “failing” because it is being “deceived” by the mesa optimizer is purely a metaphorical/terminological choice. It doesn’t negate the fact that we all understood that the base optimizer is just reinforcing “antecedent computations”. How else could you make sense of the story of deception, where an existing model, which represents the mesa optimizer, is being reinforced by the base optimizer because that existing model understands the base optimizer’s optimization process?
I am not claiming that the RFLO communicated this point well, just that it was understood and absolutely was core to the paper, and large parts of the paper wouldn’t even make sense if you didn’t have this insight. (Certainly the fact that we called it an objective doesn’t communicate the point, and it isn’t meant to).
“But yeah, I wish this hadn’t happened.”
Who else is gonna write the article? My sense is that no one (including me) is starkly stating publically the seriousness of the situation.“Yudkowsky is obnoxious, arrogant, and most importantly, disliked, so the more he intertwines himself with the idea of AI x-risk in the public imagination, the less likely it is that the public will take those ideas seriously”
I’m worried about people making character attacks on Yudkowsky (or other alignment researchers) like this. I think the people who think they can probably solve alignment by just going full-speed ahead and winging it, they are arrogant. Yudkowsky’s arrogant-sounding comments about how we need to be very careful and slow, are negligible in comparison. I’m guessing you agree with this (not sure) and we should be able to criticise him for his communication style, but I am a little worried about people publically undermining Yudkowsky’s reputation in that context. This seems like not what we would do if we were trying to coordinate well.
“But I don’t think you even need Eliezer-levels-of-P(doom) to think the situation warrants that sort of treatment.”
Agreed. If a new state develops nuclear weapons, this isn’t even close to creating a 10% x-risk, yet the idea of airstrikes on nuclear enrichment facillities, even though it is very controversial, has for a long time very much been an option on the table.
You are muddling the meaning of “pre-emptive war”, or even “war”. I’m not trying to diminish the gravity of Yudkowsky’s proposal, but a missile strike on a specific compound known to contain WMD-developing technology is not a “pre-emptive war” or “war”. Again I’m not trying to diminish the gravity, but this seems like an incorrect use of the term.
I agree. Though is it just the limited context window that causes the effect? I may be mistaken, but from my memory it seems like they emerge sooner than you would expect if this was the only reason (given the size of the context window of gpt3).
Reading this post a while after it was written: I’m not going to respond to the main claim (which seems quite likely) but just to the specific arguments, which seems suspicious to me. Here are some points:
In my model of the standard debate setup with human judge, the human can just use both answers in whichever way it wants, independently of which it selects as the correct answer. The fact that one answer provides more useful information than “2+2=?” doesn’t imply a “direct” incentive for the human judge to select that as the correct answer. Upon introspection, I myself would probably say that “4” is the correct answer, while still being very interested in the other answer (the answer on AI risk). I don’t think you disagreed with this?
At a later point you say that the real reason for why the judge would nevertheless select the QIA as the correct answer is that the judge wants to train the system to do useful things. You seem to say that a rational consequentialist would make this decision. Then at a later point you say that this is probably/plausibly (?) a bad thing: “Is this definitely undesirable? I’m not sure, but probably”. But if it really is a bad thing and we can know this, then surely a rational judge would know this, and could just decide not to do it? If you were the judge, would you select the QIA, despite it being “probably undesirable”?
Given that we are talking about optimal play and the human judge is in fact not rational/safe, the debater could manipulate the judge, and so the previous argument doesn’t in fact imply that judges won’t select QIA’s. The debater could deceive and manipulate the judge into (incorrectly) thinking that it should select the QIA, even if you/we currently believe that this would be bad. I agree this kind of deception would probably happen in optimal play (if that is indeed what you meant), but it relies on the judge being irrational or manipulable, not on some argument that “it is rational for a consequentialist judge to select answers with the highest information value”.
It seems to me that either we think there is no problem with selecting QIA’s as answers, or we think that human judges will be irrational and manipulated, but I don’t see the justification in this post for saying “rational consequentialist judges will select QIA’s AND this is probably bad”.
Therefore, the waluigi eigen-simulacra are attractor states of the LLM
It seems to me like this informal argument is a bit suspect. Actually I think this argument would not apply to Solomonof Induction.
Suppose we have to programs that have distributions over bitstrings. Suppose p1 assigns uniform probability to each bitstring, while p2 assigns 100% probability to the string of all zeroes. (equivalently, p1 i.i.d. samples bernoully from {0,1}, p2 samples 0 i.i.d. with 100%).
Suppose we use a perfect Bayesian reasoner to sample bitstrings, but we do it in precisely the same way LLMs do it according to the simulator model. That is, given a bitstring, we first formulate a posterior over programs, i.e. a “superposition” on programs, which we use to sample the next bit, then we recompute the posterior, etc.
Then I think the probability of sampling 00000000… is just 50%. I.e. I think the distribution over bitstrings that you end up with is just the same as if you just first sampled the program and stuck with it.
I think tHere’s a messy calculation which could be simplified (which I won’t do):
Limit of this is 0.5.
I don’t wanna try to generalize this, but based on this example it seems like if an LLM was an actual Bayesian, Waluigi’s would not be attractors. The informal argument is wrong because it doesn’t take into account the fact that over time you sample increasingly many non-waluigi samples, pushing down the probability of Waluigi.
Then again, the presense of a context window completely breaks the above calculation in a way that preserves the point. Maybe the context window is what makes Waluigi’s into an attractor? (Seems unlikely actually, given that the context windows are fairly big).
Is there currently a way to pool money on the trades you’re suggesting? In general it seems like there is some economies of scale to be gained by creating some kind of rationalist fund
I wonder if there is a bias induced by writing this on a year-by-year basis, as opposed to some random other time interval, like 2 years. I can somehow imagine that if you take 2 copies of a human, and ask one to do this exercise in yearly intervals, and the other to do it in 2-year intervals, they’ll basically tell the same story, but the second one’s story takes twice as long. (i.e. the second one’s prediction for 2022/2024/2026 are the same as the first one’s predictions for 2022/2023/2024). It’s probably not that extreme, but I would be surprised if there was zero such effect, which would mean these timelines are biased downwards or upwards.
Ok I admit I read over it. I must say though that this makes the whole thing more involved than it sounded at fist, since it would maybe require essentially escalating a conflict with all major military powers and still coming out on top? One possible outcome of this would be that the entire global intellectual public opinion turns against you, meaning you also possibly lose access to a lot of additional humans working with you on further alignment research? I’m not sure if I’m imagining it correctly, but it seems like this plan would either require so many elements that I’m not sure if it isn’t just equivalent to solving the entire alignment problem, or otherwise it isn’t actually enough.
Here is my partial honest reaction, just two points I’m somewhat dissatisfied with (not meant to be exhaustive):
2. “A cognitive system with sufficiently high cognitive powers, given any medium-bandwidth channel of causal influence, will not find it difficult to bootstrap to overpowering capabilities independent of human infrastructure.” I would like there to be an argument for this claim that doesn’t rely on nanotech, and solidly relies on actually existing amounts of compute. E.g. if the argument relies on running intractable detailed simulations of proteins, then it doesn’t count. (I’m not disagreeing with the nanotech example by the way, or saying that it relies on unrealistic amounts of compute, I’d just like to have an argument for this that is very solid and minimally reliant on speculative technology, and actually shows that it is).
6. “We need to align the performance of some large task, a ‘pivotal act’ that prevents other people from building an unaligned AGI that destroys the world.”. You name “burn all GPU’s” as an “overestimate for the rough power level of what you’d have to do”, but it seems to me that it would be too weak of a pivotal act? Assuming there isn’t some extreme change in generally held views, people would consider this an extreme act of terrorism, and shut you down, put you in jail, and then rebuild the GPU’s and go on with what they were planning to do. Moreover, now there is probably an extreme taboo on anything AI safety related. (I’m assuming here that law enforcement finds out that you were the one who did this). Maybe the idea is to burn all GPU’s indefinitely and forever (i.e. leave nanobots that continually check for GPU’s and burn them when they are created), but even this seems either insufficient or undesirable long term depending on what is counted as a GPU. Possibly I’m not getting what you mean, but it just seems completely too weak as an act.
“I have sat down to make toy models ..”
reference?
There is a general phenomenon where:
Person A has mental model X and tries to explain X with explanation Q
Person B doesn’t get model X from Q, thinks a bit, and then writes explanation P, reads P and thinks: P is how it should have been explained all along, and Q didn’t actually contain the insights, but P does.
Person C doesn’t get model X from P, thinks a bit, and then writes explanation R, reads R and thinks: …
It seems to me quite likely that you are person B, thinking they explained something because THEY think their explanation is very good and contains all the insights that the previous ones didn’t. Some of the evidence for this is in fact contained in your very comment:
“1. Pointing out the “reward chisels computation” point. 2. Having some people tell me it’s obvious, or already known, or that they already invented it. 3. Seeing some of the same people continue making similar mistakes (according to me)”
So point 3 basically almost definitively proves that your mental model is not conveyed to those people in your post, does it not? I think a similar thing happened where that mental model was not conveyed to you from RFLO, even though we tried to convey it. (btw not saying the models that RFLO tried to explain are the same as this post, but the basic idea of this post definitely is a part of RFLO).BTW, it could in fact be that person B’s explanation is clearer. (otoh, I think some things are less clear, e.g. you talk about “the” optimization target, which I would say is referring to that of the mesa-optimizer, without clearly assuming there is a mesa-optimizer. We stated the terms mesa- and base-optimizer to clearly make the distinction. There are a bunch of other things that I think are just imprecise, but let’s not get into it).
“Continuing (AFAICT) to correct people on (what I claim to be) mistakes around reward and optimization targets, and (for a while) was ~the only one doing so.”
I have been correcting people for a while on stuff like that (though not on LW, I’m not often on LW), such as that in the generic case we shouldn’t expect wireheading from RL agents unless the option of wireheading is in the training environment, for basically these reasons. I would also have expected people to just get this after reading RFLO, but many didn’t (others did), so your points 1/2/3 also apply to me.
“I do totally buy that you all had good implicit models of the reward-chiseling point”. I don’t think we just “implicitly” modeled it, we very explicitly understood it and it ran throughout our whole thinking about the topic. Again, explaining stuff is hard though, I’m not claiming we conveyed everything well to everyone (clearly you haven’t either).
I would be interested in reading a draft and giving feedback (FYI I’m currently a researcher in the AI safety team at FHI).
It seems to me that the basic conceptual point made in this post is entirely contained in our Risks from Learned Optimization paper. I might just be missing a point. You’ve certainly phrased things differently and made some specific points that we didn’t, but am I just misunderstanding something if I think the basic conceptual claims of this post (which seems to be presented as new) are implied by RFLO? If not, could you state briefly what is different?
(Note I am still surprised sometimes that people still think certain wireheading scenario’s make sense despite them having read RFLO, so it’s plausible to me that we really didn’t communicate everyrhing that’s in my head about this).