Posts
Comments
The AI system builders’ time horizon seems to be a reasonable starting point
Nora and/or Quentin: you talk a lot about inductive biases of neural nets ruling scheming out, but I have a vague sense that scheming ought to happen in some circumstances - perhaps rather contrived, but not so contrived as to be deliberately inducing the ulterior motive. Do you expect this to be impossible? Can you propose a set of conditions you think sufficient to rule out scheming?
What in your view is the fundamental difference between world models and goals such that the former generalise well and the latter generalise poorly?
One can easily construct a model with a free parameter X and training data such that many choices of X will match the training data but results will diverge in situations not represented in the training data (for example, the model is a physical simulation and X tracks the state of some region in the simulation that will affect the learner’s environment later, but hasn’t done so during training). The simplest x_s could easily be wrong. We can even moralise the story: the model regards its job as predicting the output under x_s and if the world happens to operate according to some other x’ then the model doesn’t care. However it’s still going to be ineffective in the future where the value of X matters.
Another comment on timing updates: if you’re making a timing update for zoonosis vs DEFUSE, and you’re considering a long timing window w_z for zoonosis, then your prior for a DEFUSE leak needs to be adjusted for the short window w_d in which this work could conceivably cause a leak, so you end up with something like p(defuse_pandemic)/p(zoo_pandemic)= rr_d w_d/w_z, where rr_d is the riskiness of DEFUSE vs zoonosis per unit time. Then you make the “timing update” p(now |defuse_pandemic)/p(now |zoo_pandemic) = w_z/w_d and you’re just left with rr_d.
Sorry, I edited (was hoping to get in before you read it)
If your theory is: there is a lab leak from WIV while working on defuse derived work then I’ll buy that you can assign a high probability to time & place … but your prior will be waaaaaay below the prior on “lab leak, nonspecific” (which is how I was originally reading your piece).
You really think in 60% of cases where country A lifts a ban on funding gain of function research a pandemic starts in country B within 2 years? Same question for “warning published in Nature”.
If people now don’t have strong views about exactly what they want the world to look like in 1000 years but people in 1000 years do have strong views then I think we should defer to future people to evaluate the “human utility” of future states. You seem to be suggesting that we should take the views of people today, although I might be misunderstanding.
Edit: or maybe you’re saying that the AGI trajectory will be ~random from the point of view of the human trajectory due to a different ontology. Maybe, but different ontology -> different conclusions is less obvious to me than different data -> different conclusions. If there’s almost no mutual information between the different data then the conclusions have to be different, but sometimes you could come to the same conclusions under different ontologies w/data from the same process.
Given this assumption, the human utility function(s) either do or don't significantly depend on human evolutionary history. I'm just going to assume they do for now.
There seems to be a missing possibility here that I take fairly seriously, which is that human values depend on (collective) life history. That is: human values are substantially determined by collective life history, and rather than converging to some attractor this is a path dependent process. Maybe you can even trace the path taken back to evolutionary history, but it’s substantially mediated by life history.
Under this view, the utility of the future wrt human values depends substantially on whether, in the future, people learn to be very sensitive to outcome differences. But “people are sensitive to outcome differences and happy with the outcome” does not seem better to me than “people are insensitive to outcome differences and happy with the outcome” (this is a first impression; I could be persuaded otherwise), even though it’s higher utility, whereas “people are unhappy with the outcome” does seem worse than “people are happy with the outcome”.
Under this view, I don’t think this follows:
there is some dependence of human values on human evolutionary history, so that a default unaligned AGI would not converge to the same values
My reasoning is that a “default AGI” will have its values contingent on a process which overlaps with the collective life history that determines human values. This is a different situation to values directly determined by evolutionary history, where the process that determines human values is temporally distant and therefore perhaps more-or-less random from the point of view of the AGI. So there’s a compelling reason to believe in value differences in the “evolution history directly determines values” case that’s absent in the “life history determines values” case.
Different values are still totally plausible, of course - I’m objecting to the view that we know they’ll be different.
(Maybe you think this is all an example of humans not really having values, but that doesn’t seem right to me).
You're changing the topic to "can you do X without wanting Y?", when the original question was "can you do X without wanting anything at all?".
A system that can, under normal circumstances, explain how to solve a problem doesn’t necessarily solve the problem if it gets in the way of explaining the solution. The notion of wanting that Nate proposes is “solving problems in order to achieve the objective”, and this need not apply to the system that explains solutions. In short: yes.
If we are to understand you as arguing for something trivial, then I think it only has trivial consequences. We must add nontrivial assumptions if we want to offer a substantive argument for risk.
Suppose we have a collection of systems of different ability that can all, under some conditions, solve . Let's say an "-wrench" is an event that defeats systems of lower ability but not systems of higher ability (i.e. prevents them from solving ).
A system that achieves with probability must defeat all -wrenches but those with a probability of at most . If the set of events that are -wrenches but not -wrenches has probability , then the system can defeat all -wrenches but a collection with probability of at most .
That is, if the challenges involved in achieving are almost the same as the challenges involved in achieving , then something good at achieving is almost as good at achieving (granting the somewhat vague assumptions about general capability baked into the definition of wrenches).
However, if is something that people basically approve of and is something people do not approve of, then I do not think the challenges almost overlap. In particular, to do , with high probability you need to defeat a determined opposition, which is not likely to be necessary if you want . That is: no need to kill everyone with nanotech if your doing what you were supposed to.
In order to sustain the argument for risk, we need to assume that the easiest way to defeat -wrenches is to learn a much more general ability to defeat wrenches than necessary and apply it to solving and, furthermore, this ability is sufficient to also defeat -wrenches. This is plausible - we do actually find it helpful to build generally capable systems to solve very difficult problems - but also plausibly false. Even highly capable AI that achieves long-term objectives could end up substantially specialised for those objectives.
As an aside, if the set of -wrenches includes the gradient updates received during training, then an argument that an -solver generalises to a -solver may also imply that deceptive alignment is likely (alternatively, proving that -solvers generalise to -solvers is at least as hard as proving deceptive alignment).
Two observations:
-
If you think that people’s genes would be a lot fitter if people cared about fitness more then surely there’s a good chance that a more efficient version of natural selection would lead to people caring more about fitness.
-
You might, on the other hand, think that the problem is more related to feedbacks. I.e. if you’re the smartest monkey, you can spend your time scheming to have all the babies. If there are many smart monkeys, you have to spend a lot of time worrying about what the other monkeys think of you. If this is how you’re worried misalignment will arise, then I think “how do deep learning models generalise?” is the wrong tree to bark up
C. If people did care about fitness, would Yudkowsky not say “instrumental convergence! Reward hacking!”? I’d even be inclined to grant he had a point.
I can't speak for janus, but my interpretation was that this is due to a capacity budget meaning it can be favourable to lose a bit of accuracy on token n if you gain more on n+m. I agree som examples would be great.
there are strong arguments that control of strongly superhuman AI systems will not be amenable to prosaic alignment
In which section of the linked paper is the strong argument for this conclusion to be found? I had a quick read of it but could not see it - I skipped the long sections of quotes, as the few I read were claims rather than arguments.
I don’t disagree with any of what you say here - I just read Anton as assuming we have a program on that frontier
The mistake here is the assumption that a program that models the world better necessarily has a higher Kolmogorov complexity.
I think Anton assumes that we have the simplest program that predicts the world to a given standard, in which case this is not a mistake. He doesn't explicitly say so, though, so I think we should wait for clarification.
But it's a strange assumption; I don't see why the minimum complexity predictor couldn't carry out what we would interpret as RSI in the process of arriving at its prediction.
I think he’s saying “suppose p1 is the shortest program that gets at most loss . If p2 gets loss , then we must require a longer string than p1 to express p2, and p1 therefore cannot express p2”.
This seems true, but I don’t understand its relevance to recursive self improvement.
I think it means that whatever you get is conservative in cases where it's unsure of whether it's in training, which may translate to being conservative where it's unsure of success in general.
I agree it doesn't rule out an AI that takes a long shot at takeover! But whatever cognition we posit that the AI executes, it has to yield very high training performance. So AIs that think they have a very short window for influence or are less-than-perfect at detecting training environments are ruled out.
An AI that wants something and is too willing to take low-probability shots at takeover (or just wielding influence) would get trained away, no?
What I mean is, however it makes decisions, it has to be compatible with very high training performance.
If I can make my point a bit more carefully: I don’t think this post successfully surfaces the bits of your model that hypothetical Bob doubts. The claim that “historical accidents are a good reference class for existential catastrophe” is the primary claim at issue. If they were a good reference class, very high risk would obviously be justified, in my view.
Given that your post misses this, I don’t think it succeeds as an defence of high P(doom).
I think a defence of high P(doom) that addresses the issue above would be quite valuable.
Also, for what it’s worth, I treat “I’ve gamed this out a lot and it seems likely to me” as very weak evidence except in domains where I have a track record of successful predictions or proving theorems that match my intuitions. Before I have learned to do either of these things, my intuitions are indeed pretty unreliable!
There is a situation in which information markets could be positive sum, though I don't know how practical it is:
I own a majority stake in company X. Someone has proposed an action A that company X take, I currently think this is worse than the status quo, but I think it's plausible that with better information I'd change my mind. I set up an exchange of X-shares-conditional-on-A for USD-conditional-on-A and the analogous exchange conditional on not-A, subsidised by some fraction of my X shares using an automatic market maker. If, by the closing date, X-shares-conditional-on-A trade at a sufficient premium to X-shares-conditional-on-not-A, I do A.
In this situation, my actions lose money vs the counterfactual of doing A and not subsidising the market, but compared to the counterfactual of not subsidising the market and not doing A I gain money because the rest of my stock is now worth more. It's unclear how I do compared to the most realistic counterfactual of "spend $Y researching action A more deeply and act accordingly".
(note that conditional prediction markets also have incentive issues WRT converging to the correct prices, though I'm not sure how important these are in practice)
I don't see how you get default failure without a model. In fact, I don’t see how you get there without the standard model, where an accident means you get a super intelligence with a random goal from an unfriendly prior - but that’s precisely the model that is being contested!
I can kiiinda see default 50-50 as "model free", though I'm not sure if I buy it.
You raise some examples of the generator/critic gap, which I addressed. I’m not sure what I should look for in that paper - I mentioned the miscalibration of GPT4 after RLHF, that’s from the GPT4 tech report, and I don’t believe your linked paper shows anything analogous (ie that RLHFd models are less calibrated than they “should” be). I know that the two papers here investigate different notions of calibration.
“Always say true things” is a much higher standard than “don’t do anything obviously bad”. Hallucination is obviously a violation of the first, and it might be a violation of the second - but I just don’t think it’s obvious!
Sure, but the “lying” probably stems from the fact that to get the thumbs up from RLHF you just have to make up a believable answer (because the process AFAIK didn’t involve actual experts in various fields fact checking every tiny bit). If just a handful of “wrong but believable” examples sneak in the reward modelling phase you get a model that thinks that sometimes lying is what humans want (and without getting too edgy, this is totally true for politically charged questions!)."Lying" could well be the better policy! I am not claiming that GPT is maliciously lying, but in AI safety, malice is never really needed or even considered (ok, maybe deception is malicious by definition).
One thing I'm saying is that we don't have clear evidence to support this claim.
I don’t agree. There is a distinction between lying and being confused - when you lie, you have to know better. Offering a confused answer is in a sense bad, but with lying there’s an obviously better policy (don’t) while it’s not the case that a confused answer is always the result of a suboptimal policy. When you are confused, the right course of action sometimes results in mistakes.
AFAIK there’s no evidence of a gap between what GPT knows and what it says when it’s running in pure generative mode (though this doesn’t say much; one would have to be quite clever to demonstrate it). There is a generator/discriminator and generator/critical gap, but this is because GPT operating as a critic is simply more capable than GPT as a generator. If we compare apples to apples then there’s again no evidence I know of that RLHFd critic-GPT is holding back on things it knows.
So I don’t think hallucination makes it obvious that behavioural safety is not solved.
I do think the fact that RLHFd models are miscalibrated is evidence against RLHF solving behaviour safety, because calibration is obviously good and the base model was capable of it.
I think this is an interesting proposal. It strikes me as something that is most likely to be useful against “scalable deception” (“misinformation”), and given the utility of scalable deception such technologies might be developed anyway. I think you do need to check if this will lead to deception technologies being developed that would not otherwise have been, and if so whether we’re actually better off knowing about them (this is analogous to one of the cases against gain of function research: we might be better if not knowing how to make highly enhanced viruses).
I have a paper (planning to get it on arxiv any day now…) which contains a result: independence of causal mechanisms (which can be related to Occam’s razor & your first point here) + precedent (“things I can do have been done before”) + variety (related to your second point - we’ve observed the phenomena in a meaningfully varied range of circumstances) + conditional independence (which OP used to construct the Bayes net) implies a conditional distribution invariant under action.
That is, speaking very loosely, if you add your considerations to OPs recipe for Bayes nets and the assumption of precedent, you can derive something kinda like interventions.
Maybe it’s similar, but high U is not necessary
Thanks for explaining the way to do exhaustive search - a big network can exhaustively search smaller network configurations. I believe that.
However, a CPU is not Turing complete (what is Turing universal?) - a CPU with an infinite read/write tape is Turing complete. This matters, because Solomonoff induction is a mixture of Turing machines. There are simple functions transformers can’t learn, such as “print the binary representation of the input + 1”; they run out of room. Solomonoff induction is not limited in this way.
Practical transformers are also usually (always?) used with exchangeable sequences, while Solomonoff inductors operate on general sequences. I can imagine ways around this (use a RNN and many epochs with a single sequence) so maybe not a fundamental limit, but still a big difference between neural nets in practice and Solomonoff inductors.
I think there is an additional effect related to "optimization is not conditioning" that stems from the fact that causation is not correlation. Suppose for argument's sake that people evaluate alignment research partly based on where it's come from (which the machine cannot control). Then producing good alignment research by regular standards is not enough to get high ratings. If a system manages to get good ratings anyway, then the actual papers it's producing must be quite different to typical highly rated alignment papers, because they are somehow compensating for the penalty incurred by coming from the wrong source. In such a situation, I think it would not be surprising if the previously observed relationship between ratings and quality did not continue to hold.
This is similar to "causal Goodhart" in Garrabrant's taxonomy, but I don't think it's quite identical. It's ambiguous whether ratings are being "intervened on" in this situation, and actual quality is probably going to be affected somewhat. I could see it as a generalised version of causal Goodhart, where intervening on the proxy is what happens when this effect is particularly extreme.
they can obviously encode a binary circuit equivalent to a CPU
A CPU by itself is not universal. Are you saying memory augmented neural networks are practically close to universality?
as long as you have enough data (or can generate it ) - big overcomplete NNs with SGD can obviously perform a strict improvement over exhaustive search
Sorry, I'm being slow here:
- Solomonoff does exhaustive search for any amount of data; is part of your claim that as data -> infinity, NN + SGD -> Solomonoff?
- How do we actually do this improved exhaustive search? Do we know that SGD gets us to a global minimum in the end?
Neural networks being universal approximators doesn't mean they do as well at distributing uncertainty as Solomonoff, right (I'm not entirely sure about this)? Also, are practical neural nets actually close to being universal?
in the worst case you can recover exhaustive exploration ala solomonoff
Do you mean that this is possible in principle, or that this is a limit of SGD training?
known perhaps experimentally in the sense that the research community has now conducted large-scale extensive (and even often automated) exploration of much of the entire space of higher order corrections to SGD
I read your original claim as "SGD is known to approximate full Bayesian inference, and the gap between SGD and full inference is known to be small". Experimental evidence that SGD performs competitively does not substantiate that claim, in my view.
Do you have a link to a more in-depth defense of this claim?
I’m not convinced the indifference conditions are desirable. Shutdown can be evidence of low utility
I can see why feasibility + individual rationality makes a payoff profile more likely than any profile missing one of these conditions, but I can’t see why I should consider every profile satisfying these conditions as likely enough to be worth worrying about
Why? The biggest problem in my mind is algorithmic progress. If we’re outside (C), then the “critical path to TAI” right now is algorithmic progress
Given that outside C approaches to AGI are likely to be substantially unlike anything we’re familiar with, and that controllable AGI is desirable, don’t you think that there’s a good chance these unknown algorithms have favourable control properties?
I think LLMs have some nice control properties too, not so much arguing against LLMs being better than unknown, just the idea that we should confidently expect control to be hard for unknown algorithms.
One of the contentions of this post is that life has thoroughly explored the space of nanotech possibilities. This hypothesis makes the failures of novel nanotech proposals non independent. That said, I don’t think the post offers enough evidence to be highly confident in this proposition (the author might privately know enough to be more confident, but if so it’s not all in the post).
Separately, I can see myself thinking, when all is said and done, that Yudkowsky and Drexler are less reliable about nanotech than I previously thought (which was a modest level of reliability to begin with), even if there are some possibilities for novel nanotech missed or dismissed by this post. Though I think not everything has been said yet.
I was just trying to clarify the limits of autoregressive vs other learning methods. Autoregressive learning is at an apparent disadvantage if is hard to compute and the reverse is easy and low entropy. It can “make up for this” somewhat if it can do a good job of predicting from , but it’s still at a disadvantage if, for example, that’s relatively high entropy compared to from . That’s it, I’m satisfied.
Are hash characters non uniform? Then I’d agree my point doesn’t stand
It’s the final claim I’m disputing - that the hashed text cannot itself be predicted. There’s still a benefit to going from e.g. to probability of a correct hash. It may not be a meaningful difference in practice, but there’s still a benefit in principle, and in practice it could also just generalise a strategy it learned for cases with low entropy text.
I can see why your algorithm is hard for GPT — unless it predicts the follow up string perfectly, there’s no benefit to hashing correctly — but I don’t see why it’s impossible. What if it perfectly predicts the follow up?
the Boltzmann distribution is the maximum entropy distribution subject to a constraint
for both expectation and energy doesn't lend itself to fast reading. is sometimes standard for expectation
A transformer at temp 0 is also doing an argmax. I’m not sure what the fundamental difference is - maybe that there’s a simple and unchanging evaluation function for direct optimisers?
Alternatively, we could say that the class of approximators all differ substantially in practice from direct optimisation algorithms. I feel like that needs to be substantiated, however. It is, after all, possible to learn a standard direct optimisation algorithm from data. You could construct a silly learner that can implement either the direct optimisation algorithm or something else random, and then have it pick whichever performs better on the data. It might also be possible with less silly learners.
Which is closer to Nate’s position: a) competition leads to highly instrumentally efficient AIs or b) inductive biases lead to highly instrumentally efficient AIs?
A quick guess is that at about 1 in 10 000 chance of AI doom working on it is about as good as ETG to GiveWell top charities
Yes, because I thought the why was obvious. I still do!
If doom has tiny probability, it's better to focus on other issues. While I can't give you a function mapping the doom mechanism to correct actions, different mechanisms of failure often require different techniques to address them - and even if they don't, we want to check that the technique actually addresses them.
I think the question of whether doom is of moderate or tiny probability is action relevant, and also how & why doom is most likely to happen is very action relevant
You could also downweight plans that are too far from any precedent.
Ok, I guess I just read Eliezer as saying something uninteresting with a touch of negative sentiment towards neural nets.
Would you say Yudkowsky's views are a mischaracterisation of neural network proponents, or that he's mistaken about the power of loose analogies?
In contrast, I think we can explain humans' tendency to like ice cream using the standard language of reinforcement learning.
I think you could defend a stronger claim (albeit you'd have to expend some effort): misgeneralisation of this kind is a predictable consequence of the evolution "training paradigm", and would in fact be predicted by machine learning practitioners. I think the fact that the failure is soft (humans don't eat ice cream until they die) might be harder to predict than the fact that the failure occurs.
I do not think that optimizing a network on a given objective function produces goals orientated towards maximizing that objective function. In fact, I think that this almost never happens. For example, I don't think GPTs have any sort of inner desire to predict text really well.
I think this is looking at the question in the wrong way. From a behaviourist viewpoint:
- it considers all of the possible 1-token completions of a piece of text
- then selects the most likely one (or randomises according to its distribution or something similar)
on this account, it "wants to predict text accurately". But Yudkowsky's claim is (roughly):
- it considers all of the possible long run interaction outcomes
- it selects the completion that leads to the lowest predictive loss for the machine's outputs across the entire interaction
and perhaps in this alternative sense it "wants to predict text accurately".
I'd say the first behaviour has high priors and strong evidence, and the second is (apparently?) supported by the fact that both behaviours are compatible with the vague statement "wants to predict text accurately", which I don't think is very compelling.
My response in Why aren't other people as pessimistic as Yudkowsky? includes a discussion of adversarial vulnerability and why I don't think points to any irreconcilable flaws in current alignment techniques.
I think this might be the wrong link. Either that, or I'm confused about how the sentence relates to the podcast video.