I have never since 1996 thought that it would be hard to get superintelligences to accurately model reality with respect to problems as simple as "predict what a human will thumbs-up or thumbs-down". The theoretical distinction between producing epistemic rationality (theoretically straightforward) and shaping preference (theoretically hard) is present in my mind at every moment that I am talking about these issues; it is to me a central divide of my ontology.
If you think you've demonstrated by clever textual close reading that Eliezer-2018 or Eliezer-2008 thought that it would be hard to get a superintelligence to understand humans, you have arrived at a contradiction and need to back up and start over.
The argument we are trying to explain has an additional step that you're missing. You think that we are pointing to the hidden complexity of wishes in order to establish in one step that it would therefore be hard to get an AI to output a correct wish shape, because the wishes are complex, so it would be difficult to get an AI to predict them. This is not what we are trying to say. We are trying to say that because wishes have a lot of hidden complexity, the thing you are trying to get into the AI's preferences has a lot of hidden complexity. This makes the nonstraightforward and shaky problem of getting a thing into the AI's preferences, be harder and more dangerous than if we were just trying to get a single information-theoretic bit in there. Getting a shape into the AI's preferences is different from getting it into the AI's predictive model. MIRI is always in every instance talking about the first thing and not the second.
You obviously need to get a thing into the AI at all, in order to get it into the preferences, but getting it into the AI's predictive model is not sufficient. It helps, but only in the same sense that having low-friction smooth ball-bearings would help in building a perpetual motion machine; the low-friction ball-bearings are not the main problem, they are a kind of thing it is much easier to make progress on compared to the main problem. Even if, in fact, the ball-bearings would legitimately be part of the mechanism if you could build one! Making lots of progress on smoother, lower-friction ball-bearings is even so not the sort of thing that should cause you to become much more hopeful about the perpetual motion machine. It is on the wrong side of a theoretical divide between what is straightforward and what is not.
You will probably protest that we phrased our argument badly relative to the sort of thing that you could only possibly be expected to hear, from your perspective. If so this is not surprising, because explaining things is very hard. Especially when everyone in the audience comes in with a different set of preconceptions and a different internal language about this nonstandardized topic. But mostly, explaining this thing is hard and I tried taking lots of different angles on trying to get the idea across.
In modern times, and earlier, it is of course very hard for ML folk to get their AI to make completely accurate predictions about human behavior. They have to work very hard and put a lot of sweat into getting more accurate predictions out! When we try to say that this is on the shallow end of a shallow-deep theoretical divide (corresponding to Hume's Razor) it often sounds to them like their hard work is being devalued and we could not possibly understand how hard it is to get an AI to make good predictions.
Now that GPT-4 is making surprisingly good predictions, they feel they have learned something very surprising and shocking! They cannot possibly hear our words when we say that this is still on the shallow end of a shallow-deep theoretical divide! They think we are refusing to come to grips with this surprising shocking thing and that it surely ought to overturn all of our old theories; which were, yes, phrased and taught in a time before GPT-4 was around, and therefore do not in fact carefully emphasize at every point of teaching how in principle a superintelligence would of course have no trouble predicting human text outputs. We did not expect GPT-4 to happen, in fact, intermediate trajectories are harder to predict than endpoints, so we did not carefully phrase all our explanations in a way that would make them hard to misinterpret after GPT-4 came around.
But if you had asked us back then if a superintelligence would automatically be very good at predicting human text outputs, I guarantee we would have said yes. You could then have asked us in a shocked tone how this could possibly square up with the notion of "the hidden complexity of wishes" and we could have explained that part in advance. Alas, nobody actually predicted GPT-4 so we do not have that advance disclaimer down in that format. But it is not a case where we are just failing to process the collision between two parts of our belief system; it actually remains quite straightforward theoretically. I wish that all of these past conversations were archived to a common place, so that I could search and show you many pieces of text which would talk about this critical divide between prediction and preference (as I would now term it) and how I did in fact expect superintelligences to be able to predict things!
TBC, I definitely agree that there's some basic structural issue here which I don't know how to resolve. I was trying to describe properties I thought the solution needed to have, which ruled out some structural proposals I saw as naive; not saying that I had a good first-principles way to arrive at that solution.
At the superintelligent level there's not a binary difference between those two clusters. You just compute each thing you need to know efficiently.
Lacking time right now for a long reply: The main thrust of my reaction is that this seems like a style of thought which would have concluded in 2008 that it's incredibly unlikely for superintelligences to be able to solve the protein folding problem. People did, in fact, claim that to me in 2008. It furthermore seemed to me in 2008 that protein structure prediction by superintelligence was the hardest or least likely step of the pathway by which a superintelligence ends up with nanotech; and in fact I argued only that it'd be solvable for chosen special cases of proteins rather than biological proteins because the special-case proteins could be chosen to have especially predictable pathways. All those wobbles, all those balanced weak forces and local strange gradients along potential energy surfaces! All those nonequilibrium intermediate states, potentially with fragile counterfactual dependencies on each interim stage of the solution! If you were gonna be a superintelligence skeptic, you might have claimed that even chosen special cases of protein folding would be unsolvable. The kind of argument you are making now, if you thought this style of thought was a good idea, would have led you to proclaim that probably a superintelligence could not solve biological protein folding and that AlphaFold 2 was surely an impossibility and sheer wishful thinking.
If you'd been around then, and said, "Pre-AGI ML systems will be able to solve general biological proteins via a kind of brute statistical force on deep patterns in an existing database of biological proteins, but even superintelligences will not be able to choose special cases of such protein folding pathways to design de novo synthesis pathways for nanotechnological machinery", it would have been a very strange prediction, but you would now have a leg to stand on. But this, I most incredibly doubt you would have said - the style of thinking you're using would have predicted much more strongly, in 2008 when no such thing had been yet observed, that pre-AGI ML could not solve biological protein folding in general, than that superintelligences could not choose a few special-case solvable de novo folding pathways along sharper potential energy gradients and with intermediate states chosen to be especially convergent and predictable.
From a high-level perspective, it is clear that this is just wrong. Part of what human brains are doing is to minimise prediction error with regard to sensory inputs.
I didn't say that GPT's task is harder than any possible perspective on a form of work you could regard a human brain as trying to do; I said that GPT's task is harder than being an actual human; in other words, being an actual human is not enough to solve GPT's task.
Choosing to engage with an unscripted unrehearsed off-the-cuff podcast intended to introduce ideas to a lay audience, continues to be a surprising concept to me. To grapple with the intellectual content of my ideas, consider picking one item from "A List of Lethalities" and engaging with that.
This is kinda long. If I had time to engage with one part of this as a sample of whether it holds up to a counterresponse, what would be the strongest foot you could put forward?
(I also echo the commenter who's confused about why you'd reply to the obviously simplified presentation from an off-the-cuff podcast rather than the more detailed arguments elsewhere.)
The author doesn't seem to realize that there's a difference between representation theorems and coherence theorems.
The Complete Class Theorem says that an agent’s policy of choosing actions conditional on observations is not strictly dominated by some other policy (such that the other policy does better in some set of circumstances and worse in no set of circumstances) if and only if the agent’s policy maximizes expected utility with respect to a probability distribution that assigns positive probability to each possible set of circumstances.This theorem does refer to dominated strategies. However, the Complete Class Theorem starts off by assuming that the agent’s preferences over actions in sets of circumstances satisfy Completeness and Transitivity. If the agent’s preferences are not complete and transitive, the Complete Class Theorem does not apply. So, the Complete Class Theorem does not imply that agents must be representable as maximizing expected utility if they are to avoid pursuing dominated strategies.
The Complete Class Theorem says that an agent’s policy of choosing actions conditional on observations is not strictly dominated by some other policy (such that the other policy does better in some set of circumstances and worse in no set of circumstances) if and only if the agent’s policy maximizes expected utility with respect to a probability distribution that assigns positive probability to each possible set of circumstances.
This theorem does refer to dominated strategies. However, the Complete Class Theorem starts off by assuming that the agent’s preferences over actions in sets of circumstances satisfy Completeness and Transitivity. If the agent’s preferences are not complete and transitive, the Complete Class Theorem does not apply. So, the Complete Class Theorem does not imply that agents must be representable as maximizing expected utility if they are to avoid pursuing dominated strategies.
Cool, I'll complete it for you then.
Transitivity: Suppose you prefer A to B, B to C, and C to A. I'll keep having you pay a penny to trade between them in a cycle. You start with C, end with C, and are three pennies poorer. You'd be richer if you didn't do that.
Completeness: Any time you have no comparability between two goods, I'll swap them in whatever direction is most useful for completing money-pump cycles. Since you've got no preference one way or the other, I don't expect you'll be objecting, right?
Combined with the standard Complete Class Theorem, this now produces the existence of at least one coherence theorem. The post's thesis, "There are no coherence theorems", is therefore falsified by presentation of a counterexample. Have a nice day!
I see several large remaining obstacles. On the one hand, I'd expect vast efforts thrown at them by ML to solve them at some point, which, at this point, could easily be next week. On the other hand, if I naively model Earth as containing locally-smart researchers who can solve obstacles, I would expect those obstacles to have been solved by 2020. So I don't know how long they'll take.
(I endorse the reasoning of not listing out obstacles explicitly; if you're wrong, why talk, if you're right, you're not helping. If you can't save your family, at least don't personally contribute to killing them.)
Expanding on this now that I've a little more time:
Although I haven't had a chance to perform due diligence on various aspects of this work, or the people doing it, or perform a deep dive comparing this work to the current state of the whole field or the most advanced work on LLM exploitation being done elsewhere,
My current sense is that this work indicates promising people doing promising things, in the sense that they aren't just doing surface-level prompt engineering, but are using technical tools to find internal anomalies that correspond to interesting surface-level anomalies, maybe exploitable ones, and are then following up on the internal technical implications of what they find.
This looks to me like (at least the outer ring of) security mindset; they aren't imagining how things will work well, they are figuring out how to break them and make them do much weirder things than their surface-apparent level of abnormality. We need a lot more people around here figuring out things will break. People who produce interesting new kinds of AI breakages should be cherished and cultivated as a priority higher than a fair number of other priorities.
In the narrow regard in which I'm able to assess this work, I rate it as scoring very high on an aspect that should relate to receiving future funding. If anyone else knows of a reason not to fund the researchers who did this, like a low score along some metric I didn't examine, or because this is somehow less impressive as a feat of anomaly-finding than it looks, please contact me including via email or LW direct message; as otherwise I might run around scurrying trying to arrange funding for this if it's not otherwise funded.