All of David Johnston's Comments + Replies

If you're in a situation where you can reasonably extrapolate from past rewards to future reward, you can probably extrapolate previously seen "normal behaviour" to normal behaviour in your situation. Reinforcement learning is limited - you can't always extrapolate past reward - but it's not obvious that imitative regularisation is fundamentally more limited.

(normal does not imply safe, of course)

Their empirical result rhymes with adversarial robustness issues - we can train adversaries to maximise ~arbitrary functions subject to small perturbation from ground truth constraints. Here the maximised function is a faulty reward model and the constraint is KL to a base model instead of distance to a ground truth image.

I wonder if multiscale aggregation could help here too as it does with image adversarial robustness. We want the KL penalty to ensure that the generations should look normal at any "scale", whether we look at them token by token or read a... (read more)

Is your view closer to:

• there's two hard steps (instruction following, value alignment), and of the two instruction following is much more pressing
• instruction following is the only hard step; if you get that, value alignment is almost certain to follow

Mathematical reasoning might be specifically conducive to language invention because our ability to automatically verify reasoning means that we can potentially get lots of training data. The reason I expect the invented language to be “intelligible” is that it is coupled (albeit with some slack) to automatic verification.

There's a regularization problem to solve for 3.9 and 4, and it's not obvious to me that glee will be enough to solve it (3.9 = "unintelligible CoT").

I'm not sure how o1 works in detail, but for example, backtracking (which o1 seems to use) makes heavy use of the pretrained distribution to decide on best next moves. So, at the very least, it's not easy to do away with the native understanding of language. While it's true that there is some amount of data that will enable large divergences from the pretrained distribution - and I could imagine mathematical ... (read more)

For what it's worth, one idea I had as a result of our discussion was this:

• We form lots of beliefs as a result of motivated reasoning
• These beliefs are amenable to revision due to evidence, reason or (maybe) social pressure
• Those beliefs that are largely resilient to these challenges are "moral foundations"

So philosophers like "pain is bad" as a moral foundation because we want to believe it + it is hard to challenge with evidence or reason. Laypeople probably have lots of foundational moral beliefs that don't stand up as well to evidence or reason, bu... (read more)

I can explain why I believe bachelors are unmarried: I learned that this is what the word bachelor means, I learned this because it is what bachelor means, and the fact that there's a word "bachelor" that means "unmarried man" is contingent on some unimportant accidents in the evolution of language. A) it is certainly not the result of an axiomatic game and B) if moral beliefs were also contingent on accidents in the evolution of language (I think most are not), that would have profound implications for metaethics.

Motivated belief can explain non-purely-se... (read more)

Thanks for your continued engagement.

I’m interested in explaining foundational moral beliefs like suffering is bad, not beliefs like “animals do/don’t suffer”, which is about badness only because we accept the foundational assumption that suffering is bad. Is that clear in the updated text?

Now, I don’t think these beliefs come from playing axiomatic games like “define good as that which increases welfare”. There are many lines of evidence for this. First: “define bad as that which increases suffering” is not equally as plausible as “define good as that whi... (read more)

I think precisely defining "good" and "bad" is a bit beside the point - it's a theory about how people come to believe things are good and bad, and we're perfectly capable of having vague beliefs about goodness and badness. That said, the theory is lacking a precise account of what kind of beliefs it is meant to explain.

The LLM section isn't meant as support for the theory, but speculation about what it would say about the status of "experiences" that language models can have. Compared to my pre-existing notions, the theory seems quite willing to accommodate LLMs having good and bad experiences on par with those that people have.

I have a pedantic and a non-pedantic answer to this. Pedantic: you say X is "usually considered good" if it increases welfare. Perhaps you mean to imply that if X is usually considered good then it is good. In this case, I refer you to the rest of the paragraph you quote.

Non-pedantic: yes, it's true that once you accept some fundamental assumptions about goodness and badness you can go about theorising and looking for evidence. I'm suggesting that motivated reasoning is the mechanism that makes those fundamental assumptions believable.

I added a paragraph mentioning this, because I think your reaction is probably common.

Here's a basic model of policy collapse: suppose there exist pathological policies of low prior probability (/high algorithmic complexity) such that they play the training game when it is strategically wise to do so, and when they get a good opportunity they defect in order to pursue some unknown aim.

Because they play the training game, a wide variety of training objectives will collapse to one of these policies if the system in training starts exploring policies of sufficiently high algorithmic complexity. So, according to this crude model, there's a comp... (read more)

Algorithmic complexity is precisely analogous to difficulty-of-learning-to-predict, so saying "it's not about learning to predict, it's about algorithmic complexity" doesn't make sense. One read of the original is: learning to respect common sense moral side constraints is tricky^[1], but AI systems will learn how to do it in the end. I'd be happy to call this read correct, and is consistent with the observation that today's AI systems do respect common sense moral side constraints given straightforward requests, and that it took a few years to figure out h... (read more)

When do you think is the right time to work in these issues? Monitoring, trust displacement and fine grained permission management all look liable to raise issues that weren’t anticipated and haven’t already been solved, because they’re not the way things have been done historically. My gut sense is that GPT4 performance is much lower when you’re asking it to do novel things. Maybe it’s also possible to make substantial gains with engineering and experimentation, but you’ll need a certain level of performance in order to experiment.

Some wild guesses: maybe... (read more)

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... (read more)

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 ... (read more)

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 medi... (read more)

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 $X$. Let's say an "$X$-wrench" is an event that defeats systems of lower ability but not systems of higher ability (i.e. prevents them from solving $X$).

A system that achieves $X$ with $1-ϵ$ probability must defeat all $X$-wrenches but those with a probability of at most $ϵ$.... (read more)

Two observations:

1. 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.

2. 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 wo

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 $x$. If p2 gets loss $y<x$, 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 wo... (read more)

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... (read more)

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... (read more)

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 usu... (read more)

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 compe... (read more)

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

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

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 le... (read more)

I was just trying to clarify the limits of autoregressive vs other learning methods. Autoregressive learning is at an apparent disadvantage if $P\left(X_t|X_{t-1}\right)$ 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 $X_t$ from $X_{t-2}$, but it’s still at a disadvantage if, for example, that’s relatively high entropy compared to $X_{t-1}$ from $X_t$. That’s it, I’m satisfied.