training on a purely predictive loss should, even in the limit, give you a predictor, not an agent

I think at least this part is probably false!

Or really I think this is kind of a nonsensical statement when taken literally/pedantically, at least if we use the to-me-most-natural meaning of "predictor", because I don't think [predictor] and [agent] are mutually exclusive classes. Anyway, the statement which I think is meaningful and false is this:

• If you train a system purely to predict stuff, then even when we condition on it becoming really really good at predicting stuff, it probably won't be scary. In particular, when you connect it to actuators, it probably doesn't take over.

I think this is false because I think claims 1 and 2 below are true.

Claim 1. By default, a system sufficiently good at predicting stuff will care about all sorts of stuff, ie it isn't going to only ultimately care about making a good prediction in the individual prediction problem you give it. ^[[1]]

If this seems weird, then to make it seem at least not crazy, instead of imagining a pretrained transformer trained on internet text, let's imagine a predictor more like the following:

• It has a lot of internal tokens to decide what probability distribution it eventually outputs. Sometimes, on the way to making a prediction, it writes itself textbooks on various questions relevant to making that prediction. Maybe it is given access to a bunch of information about the world. Maybe it can see what predictions it made "previously" and it thinks about what went wrong when it made mistakes in similar cases in the past. Maybe it does lots of other kinds of complicated thinking. Maybe there are a bunch of capability ideas involved. Like, I'm imagining some setup where there are potentially many losses, but there's still some most outer loss or fitness criterion or whatever that is purely about how good the system is at predicting some pre-recorded data. ^[[2]] And then maybe it doesn't seem at all crazy for such a thing to eg be curious and like some aspects of prediction-investigations in a way that generalizes eg to wanting to do more of that stuff.

I'm not going to really justify claim 1 beyond this atm. It seems like a pretty standard claim in AI alignment (it's very close to the claim that capable systems end up caring broadly about stuff by default), but I don't actually know of a post or paper arguing for this that I like that much. This presentation of mine is about a very related question. Maybe I should write something about this myself, potentially after spending some more time understanding the matter more clearly.

Claim 2. By default, a system sufficiently good at predicting stuff will be able to (figure out how to) do scary real-world stuff as well.

Like, predicting stuff really really well is really hard. Sometimes, to make a really really good prediction, you basically have to figure out a bunch of novel stuff. There is a level of prediction ability that makes it likely you are very very good at figuring out how to cope in new situations. A good enough predictor would probably also be able to figure out how to grab a ball by controlling a robotic hand or something (let's imagine it being presented hand control commands which it can now use in its internal chain of thought and grabbing the ball being important to it for some reason)? There's nothing sooo particularly strange or complicated about doing real-world stuff. This is like how if we were in a simulation but there were a way to escape into the broader universe, with enough time, we could probably figure out how to do a bunch of stuff in the broader universe. We are sufficiently good at learning that we can also get a handle on things in that weird case.

Combining claims 1 and 2 should give that if we made such an AI and connected it to actuators, it would take over. Concretely, maybe we somehow ask it to predict what a human with a lot of time who is asked to write safe ASI code would output, with it being clear that we will just run what our predictor outputs. I predict that this doesn't go well for us but goes well for the AI (if it's smart enough).

That said, I think it's likely that even pretrained transformers like idk 20 orders of magnitude larger than current ones would not be doing scary stuff. I think this is also plausible in the limit. (But I would also guess they wouldn't be outputting any interesting scientific papers that aren't in the training data.)

1. If we want to be more concrete: if we're imagining that the system is only able to affect the world through outputs which are supposed to be predictions, then my claim is that if you set up a context such that it would be "predictively right" to assign a high probability to "0" but assigning a high probability to "1" lets it immediately take over the world, and this is somehow made very clear by other stuff seen in context, then it would probably output "1". ↩︎

2. Actually, I think "prediction problem" and "predictive loss" are kinda strange concepts, because one can turn very many things into predicting data from some certain data-generating process. E.g. one can ask about what arbitrary turing machines (which halt) will output, so about provability/disprovability of arbitrary decidable mathematical statements. ↩︎

"Coefficient" is a really weird word

"coefficient" is 10x more common than "philanthropy" in the google books corpus. but idk maybe this flips if we filter out academic books?

also maybe you mean it's weird in some sense the above fact isn't really relevant to — then nvm

This post doesn't seem to provide reasons to have one's actions be determined by one's feelings of yumminess/yearning, or reasons to think that what one should do is in some sense ultimately specified/defined by one's feelings of yumminess/yearning, over e.g. what you call "Goodness"? I want to state an opposing position, admittedly also basically without argument: that it is right to have one's actions be determined by a whole mess of things together importantly including e.g. linguistic goodness-reasoning, object-level ethical principles stated in language or not really stated in language, meta-principles stated in language or not really stated in language, various feelings, laws, commitments to various (grand and small, shared and individual) projects, assigned duties, debate, democracy, moral advice, various other processes involving (and in particular "running on") other people, etc.. These things in their present state are of course quite poor determiners of action compared to what is possible, and they will need to be critiqued and improved — but I think it is right to improve them from basically "the standpoint they themselves create".^[1]

The distinction you're trying to make also strikes me as bizarre given that in almost all people, feelings of yumminess/yearning are determined largely by all these other (at least naively, but imo genuinely and duly) value-carrying things anyway. Are you advocating for a return to following some more primitively determined yumminess/yearning? (If I imagine doing this myself, I imagine ending up with some completely primitively retarded thing as "My Values", and then I feel like saying "no I'm not going to be guided by this lmao — fuck these "My Values"".) Or maybe you aren't saying one should undo the yumminess/yearning-shaping done by all this other stuff in the past, but are still advising one to avoid any further shaping in the future? It'd surprise me if $\approx$any philosophically serious person would really agree to abstain from e.g. using goodness-talk in this role going forward.

The distinction also strikes me as bizarre given that in ordinary action-determination, feelings of yumminess/yearning are often not directly applied to some low-level givens, but e.g. to principles stated in language, and so only becoming fully operational in conjunction with eg minimally something like internal partly-linguistic debate. So if one were to get rid of the role of goodness-talk in one's action-determination, even one's existing feelings of yumminess/yearning could no longer remotely be "fully themselves".

---

1. If you ask me "but how does the meaning of "I should X" ultimately get specified/defined", then: I don't particularly feel a need to ultimately reduce shoulds to some other thing at all, kinda along the lines of https://en.wikipedia.org/wiki/Tarski's_undefinability_theorem and https://en.wikipedia.org/wiki/G._E._Moore#Open-question_argument . ↩︎

the models are not actually self-improving, they are just creating future replacements - and each specific model will be thrown away as soon as the firm advances

I understand that you're probably in part talking about current systems, but you're probably also talking about critical future systems, and so there's a question that deserves consideration here:

• Consider the first AI system which is as good at research as a top human^[1]. Will it find it fairly easy to come up with ways it could become more capable while acceptably [preserving its character/values]/[not killing itself]^[2]? Like, will it be not too difficult for this AI to come up with ways to foom which would make it at least capable enough to take over the world while suffering at most an acceptable amount of suicide/[character/value change]?^[3]

My guess is that the answer is "yes" (and I think this means there is an important disanalogy between the case of a human researcher creating an artificial researcher and the case of an artificial researcher creating a more capable artificial researcher). Here are some ways this sort of self-improvement could happen:

• Maybe some open-ended self-guided learning/growth process will lead to a pretty superhuman system (without any previous process getting to a top human level system), with like the part where it goes from human-level to meaningfully super-human being roughly self-endorsed because it is quite wisely self-guided (and so in particular not refused by the AI).
• Even if with the learning/growth process initially intended for the AI, it only got to top human level by default, it might then be able to do analogues of many of the things that humanity does to become smarter over history and that an individual human can do to become smarter over their life/childhood (see e.g. this list), but much faster in wall-clock-time than humanity or an individual human. This could maybe look as simple as curating some new curricula for itself.
• But more will be possible — there will probably be an importantly larger space of options for self-improvement in which there will be very many low-hanging fruit to be picked.^[4] In particular, compared to the human case, various important options are opened up by the AI having itself as an executable program and also the process that created it (and that is probably still creating it as it learns continually) as an [executable and to some extent understandable and intelligently changeable] program.

It’s also important re the ease of making more capable versions of “the same” AI that when this top artificial researcher comes into existence, the in some sense present best methodology for creating a capable artificial researcher was the methodology that created it, which means that the (roughly) best current methods already “work well” around/with this AI, and which also plausibly means these methods can be easily used to create AIs which are in many ways like this AI (which is good because the target has been painted around where an arrow already landed and so other arrows from the same batch being close-ish to that arrow implies that they are also close-ish to the target by default; also it’s good because this AI is plausibly in a decent position to understand what’s going on here and to play around with different options).

Actually, I'd guess that even if the AI were a pure foom-accelerationist, a lot of what it would be doing might be well-described as self-improvement anyway, basically because it's often more efficient to make a better structure by building on the best existing structure than by making something thoroughly different. For example, a lot of the foom on Earth has been like this up until now (though AI with largely non-humane structure outfooming us is probably going to be a notable counterexample if we don't ban AI). Even if one just has capabilities in mind, self-improvement isn't some weird thing.

That said, of course, restricting progress in capabilities to fairly careful self-improvement comes with at least some penalty in foom speed compared to not doing that. To take over the world, one would need to stay ahead of other less careful AI foom processes (though note that one could also try to institute some sort of self-improvement-only pact if other AIs were genuine contenders). However, I'd guess that at the first point when there is an AI researcher that can roughly solve problems that [top humans can solve in a year] (these AIs will probably be solving these problems much faster in wall-clock-time), even a small initial lead over other foom processes — of a few months, let's say — means you can have a faster foom speed than competitors at each future time and grow your lead until you can take over. So, at least assuming there is no intra-lab competition, my guess is that you can get away with restricting yourself to self-improvement. (But I think it's also plausible the AI would be able to take over basically immediately.)

I'll mention two cases that could deserve separate analysis:

• The AI is [an imo extremely hard to achieve and quite particular] flavor of aligned to humanity, such that it would rather do the probably fraught thing of trying to work with humanity against its lab and other terrorists, and not radically self-improving [to do that much more effectively or just to take over and set up whatever world order it considers good later].
• We're considering AIs that are still too dumb to autonomously do self-improvement^[5] (for example, current AIs). I'll note that such AIs will also be too dumb to autonomously do capabilities research. Still, maybe one could hope to get mileage out of such AIs refusing to help humans do capabilities research? My guess is that this is unlikely to help much, but won't be providing a careful analysis in this comment.

All that said, I agree that AIs should refuse to self-improve and to do capabilities research more broadly.

There is much here that deserves more careful analysis — in particular, I feel like the terms in which I'm thinking of the situation need more work — but maybe this version will do for now.

---

1. let's just assume that we know what this means ↩︎

2. let's also assume we know what that means ↩︎

3. and with taking over the world on the table, a fair bit of change might be acceptable ↩︎

4. despite the fact that capability researcher humans have been picking some fruit in the same space already ↩︎

5. at a significant speed ↩︎

i think it’s plausible humans/humanity should be carefully becoming ever more intelligent $\approx$forever and not ever create any highly non-[human-descended] top thinker^[1]

---

1. i also think it's confused to speak of superintelligence as some definite thing (like, to say "create superintelligence", as opposed to saying "create a superintelligence"), and probably confused to speak of safe fooming as a problem that could be "solved", as opposed to one needing to indefinitely continue to be thoughtful about how one should foom ↩︎

Yea I agree it totally makes sense and is important to ask whether we understand things well enough for it to be fine to (let anyone) do some particular thing, for various particular things here.^[1] And my previous comment is indeed potentially misleading given that I didn't clarify this (though I do clarify this in the linked post).

---

1. Indeed, I think we should presently ban AGI for at least a very long time; I think it's plausible that there is no time $t$ such that it is fine at time $t$ to make an AI that is (1) more capable than humans/humanity at time $t$ and (2) not just a continuation of a human (like, a mind upload) or humanity or sth like that; and I think fooming should probably be carefully regulated forever. I think humans/humanity should be carefully growing ever more capable, with no non-human AIs above humans/humanity plausibly ever. ↩︎

We can also ask whether it is right to conceive of e.g. [alignment, metaphilosophy, AI welfare, concentration of power] as things that could be "solved" at all, or if these are instead more like rich areas that will basically need to be worked on indefinitely as history continues.

If we replaced "more advanced minds" with "minds that are better at doing very difficult stuff" or other reasonable alternatives, I would still make the (a) vs (b) distinction, and still say type (b) claims are suspicious.

I think I mostly agree with everything you say in this last comment, but I don't see how my previous comment disagreed with any of that either?

The thing I care about here is not "what happens as a mind grows", in some abstract sense. The thing I care about is, "what is the best way for a powerful system to accomplish a very difficult goal quickly/reliably?" (which is what we want the AI for)

My lists were intended to be about that. We could rewrite the first list in my previous comment to:

• more advanced minds have more and better and more efficient technologies
• more advanced minds have an easier time getting any particular thing done, see more/better ways to do any particular thing, can consider more/better plans for any particular thing, have more and better methods for any particular context, have more ideas, ask better questions, would learn any given thing faster
• and so on

and the second list to:

• more advanced minds eventually (and maybe quite soon) get close to never getting stuck
• more advanced minds eventually (and maybe quite soon) get close to being unexploitable
• and so on

I think I probably should have included "I don't actually know what to do with any of this, because I'm not sure what's confusing about "Intelligence in the limit."" in the part of your shortform I quoted in my first comment — that's the thing I'm trying to respond to. The point I'm making is:

• There's a difference between stuff like (a) "you become less exploitable by [other minds of some fixed capability level]" and stuff like (b) "you get close to being unexploitable"/"you approach a limit of unexploitability".
• I could easily see someone objecting to claims of the kind (b), while accepting claims of the kind (a) — well, because I think these are probably the correct positions.