EDIT: This comment fails on a lot of points, as discussed in this apology subcomment. I encourage people interested by the thread to mostly read the apology subcomment and the list of comments linked there, which provide maximum value with minimum drama IMO.
Disclaimer: this is a rant. In the best possible world, I could write from a calmer place, but I’m pretty sure that the taboo on criticizing MIRI and EY too hard on the AF can only be pushed through when I’m annoyed enough. That being said, I’m writing down thoughts that I had for quite some time, so don’t just discard this as a gut reaction to the post.
(Later added disclaimer: it's a good idea to add "I feel like..." before the judgment in this comment, so that you keep in mind that I'm talking about my impressions and frustrations, rarely stating obvious facts (despite the language making it look so))
Tl;dr:
- I'm annoyed by EY (and maybe MIRI's?) dismissal of every other alignment work, and how seriously it seems to be taken here, given their track record of choosing research agendas with very indirect impact on alignment, and of taking a lot of time to let go of these flawed agendas in the face of mounting evidence.
- I'm annoyed t
This is an apology for the tone and the framing of the above comment (and my following answers), which have both been needlessly aggressive, status-focused and uncharitable. Underneath are still issues that matter a lot to me, but others have discussed them better (I'll provide a list of linked comments at the end of this one).
Thanks to Richard Ngo for convincing me that I actually needed to write such an apology, which was probably the needed push for me to stop weaseling around it.
So what did I do wrong? The list is pretty damning:
- I took something about the original post that I didn't understand — EY's "And then there is, so far as I can tell, a vast desert full of work that seems to me to be mostly fake or pointless or predictable." — and because it didn't make sense to me, and because that fitted with my stereotypes for MIRI and EY's dismissiveness of a lot of work in alignment, I turned to an explanation of this as an attack on alignment researchers, saying they were consciously faking it when they knew they should do better. Whereas I feel know that what EY meant is far closer to alignment research at the moment is trying to try to align AI as best as we can, instead of just
Similarly, the fact that they kept at it over and over with all the big improvement of DL instead of trying to adapt to prosaic Alignment sounds like evidence that they might be over attached to a specific framing, which they had trouble to discard.
I'm... confused by this framing? Specifically, this bit (as well as other bits like these)
I have to explain again and again to stressed-out newcomers that you definitely don’t need to master model theory or decision theory to do alignment, and try to steer them towards problems and questions that look like they’re actually moving the ball instead of following the lead of the “figure of authority”.
Some of the brightest and first thinkers on alignment have decided to follow their own nerd-sniping and call everyone else fakers, and when they realized they were not actually making progress, they didn’t switch to something else as much as declare everyone was still full of it
...Also, I don’t know how much is related to mental health and pessimism and depression (which I completely understand can color one’s view of the world), but I would love to see the core MIRI team and EY actually try solving alignment with neural nets and prosaic
(Later added disclaimer: it's a good idea to add "I feel like..." before the judgment in this comment, so that you keep in mind that I'm talking about my impressions and frustrations, rarely stating obvious facts (despite the language making it look so))
Okay, so you're completely right that a lot of my points are logically downstream of the debate on whether Prosaic Alignment is Impossible or not. But I feel like you don't get how one sided this debate is, and how misrepresented it is here (and generally on the AF)
Like nobody except EY and a bunch of core MIRI people actually believes that prosaic alignment is impossible. I mean that every other researcher that I know think Prosaic Alignment is possible, even if potentially very hard. That includes MIRI people like Evan Hubinger too. And note that some of these other alignment researchers actually work with Neural Nets and keep up to speed on the implementation details and subtleties, which in my book means their voice should count more.
But that's just a majority argument. The real problem is that nobody has ever given a good argument on why this is impossible. I mean the analogous situation is that a car is driving right at you, a...
I think one core issue here is that there are actually two debates going on. One is "how hard is the alignment problem?"; another is "how powerful are prosaic alignment techniques?" Broadly speaking, I'd characterise most of the disagreement as being on the first question. But you're treating it like it's mostly on the second question - like EY and everyone else are studying the same thing (cancer, in your metaphor) and just disagree about how to treat it.
My attempt to portray EY's perspective is more like: he's concerned with the problem of ageing, and a whole bunch of people have come along, said they agree with him, and started proposing ways to cure cancer using prosaic radiotherapy techniques. Now he's trying to say: no, your work is not addressing the core problem of ageing, which is going to kill us unless we make a big theoretical breakthrough.
Regardless of that, calling the debate "one sided" seems way too strong, especially given how many selection effects are involved. I mean, you could also call the debate about whether alignment is even a problem "one sided" - 95% of all ML researchers don't think it's a problem, or think it's something we'll solve easily. But for fair...
95% of all ML researchers don't think it's a problem, or think it's something we'll solve easily
The 2016 survey of people in AI asked people about the alignment problem as described by Stuart Russell, and 39% said it was an important problem and 33% that it's a harder problem than most other problem in the field.
Thanks for the detailed comment!
I think one core issue here is that there are actually two debates going on. One is "how hard is the alignment problem?"; another is "how powerful are prosaic alignment techniques?" Broadly speaking, I'd characterise most of the disagreement as being on the first question. But you're treating it like it's mostly on the second question - like EY and everyone else are studying the same thing (cancer, in your metaphor) and just disagree about how to treat it.
That's an interesting separation of the problem, because I really feel there is more disagreement on the second question than on the first.
My attempt to portray EY's perspective is more like: he's concerned with the problem of ageing, and a whole bunch of people have come along, said they agree with him, and started proposing ways to cure cancer using prosaic radiotherapy techniques. Now he's trying to say: no, your work is not addressing the core problem of ageing, which is going to kill us unless we make a big theoretical breakthrough.
Funnily, aren't the people currently working on ageing using quite prosaic techniques? I completely agree that one need to go for the big problems, especially ones t...
I really feel there is more disagreement on the second question than on the first
What is this feeling based on? One way we could measure this is by asking people about how much AI xrisk there is conditional on there being no more research explicitly aimed at aligning AGIs. I expect that different people would give very different predictions.
People like Paul and Evan and more are actually going for the core problems IMO, just anchoring a lot of their thinking in current ML technologies.
Everyone agrees that Paul is trying to solve foundational problems. And it seems strange to criticise Eliezer's position by citing the work of MIRI employees.
It's just harder when he writes things like everyone working on alignment is faking it and not giving much details.
As Rob pointed out above, this straightforwardly mischaracterises what Eliezer said.
I worry that "Prosaic Alignment Is Doomed" seems a bit... off as the most appropriate crux. At least for me. It seems hard for someone to justifiably know that this is true with enough confidence to not even try anymore. To have essayed or otherwise precluded all promising paths of inquiry, to not even engage with the rest of the field, to not even try to argue other researchers out of their mistaken beliefs, because it's all Hopeless.
Consider the following analogy: Someone who wants to gain muscle, but has thought a lot about nutrition and their genetic makeup and concluded that Direct Exercise Gains Are Doomed, and they should expend their energy elsewhere.
OK, maybe. But how about try going to the gym for a month anyways and see what happens?
The point isn't "EY hasn't spent a month of work thinking about prosaic alignment." The point is that AFAICT, by MIRI/EY's own values, valuable-seeming plans are being left to rot on the cutting room floor. Like, "core MIRI staff meet for an hour each month and attack corrigibility/deceptive cognition/etc with all they've got. They pay someone to transcribe the session and post the fruits / negative results / reasoning to AF, without individually committing to following up with comments."
(I am excited by Rob Bensinger's comment that this post is the start of more communication from MIRI)
+1 for this whole conversation, including Adam pushing back re prosaic alignment / trying to articulate disagreements! I agree that this is an important thing to talk about more.
I like the 'give more concrete feedback on specific research directions' idea, especially if it helps clarify generators for Eliezer's pessimism. If Eliezer is pessimistic about a bunch of different research approaches simultaneously, and you're simultaneously optimistic about all those approaches, then there must be some more basic disagreement(s) behind that.
From my perspective, the OP discussion is the opening salvo in 'MIRI does a lot more model-sharing and discussion'. It's more like a preface than like a conclusion, and the next topic we plan to focus on is why Eliezer-cluster people think alignment is hard, how we're thinking about AGI, etc. In the meantime, I'm strongly in favor of arguing about this a bunch in the comments, sharing thoughts and reflections on your own models, etc. -- going straight for the meaty central disagreements now, not waiting to hash this out later.
Someone privately contacted me to express confusion, because they thought my '+1' means that I think adamShimi's initial comment was unusually great. That's not the case. The reasons I commented positively are:
- I think this overall exchange went well -- it raised good points that might have otherwise been neglected, and everyone quickly reached agreement about the real crux.
- I want to try to cancel out any impression that criticizing / pushing back on Eliezer-stuff is unwelcome, since Adam expressed worries about a "taboo on criticizing MIRI and EY too hard".
- On a more abstract level, I like seeing people 'blurt out what they're actually thinking' (if done with enough restraint and willingness-to-update to mostly avoid demon threads), even if I disagree with the content of their thought. I think disagreements are often tied up in emotions, or pattern-recognition, or intuitive senses of 'what a person/group/forum is like'. This can make it harder to epistemically converge about tough topics, because there's a temptation to pretend your cruxes are more simple and legible than they really are, and end up talking about non-cruxy things.
Separately, I endorse Ben Pace's question ("Can you make a positive case here for how the work being done on prosaic alignment leads to success?") as the thing to focus on.
Thanks for the kind answer, even if we're probably disagreeing about most points in this thread. I think message like yours really help in making everyone aware that such topics can actually be discussed publicly without big backlash.
I like the 'give more concrete feedback on specific research directions' idea, especially if it helps clarify generators for Eliezer's pessimism. If Eliezer is pessimistic about a bunch of different research approaches simultaneously, and you're simultaneously optimistic about all those approaches, then there must be some more basic disagreement(s) behind that.
That sounds amazing! I definitely want to extract some of the epistemic strategies that EY uses to generate criticisms and break proposals. :)
From my perspective, the OP discussion is the opening salvo in 'MIRI does a lot more model-sharing and discussion'. It's more like a preface than like a conclusion, and the next topic we plan to focus on is why Eliezer-cluster people think alignment is hard, how we're thinking about AGI, etc. In the meantime, I'm strongly in favor of arguing about this a bunch in the comments, sharing thoughts and reflections on your own models, etc. -- going straight for the meaty central disagreements now, not waiting to hash this out later.
Excited about that!
Couple things:
First, there is a lot of work in the "alignment community" that involves (for example) decision theory or open-source-game-theory or acausal trade, and I haven't found any of it helpful for what I personally think about (which I'd like to think is "directly attacking the heart of the problem", but others may judge for themselves when my upcoming post series comes out!).
I guess I see this subset of work as consistent with the hypothesis "some people have been nerd-sniped!". But it's also consistent with "some people have reasonable beliefs and I don't share them, or maybe I haven't bothered to understand them". So I'm a bit loath to go around criticizing them, without putting more work into it. But still, this is a semi-endorsement of one of the things you're saying.
Second, my understanding of MIRI (as an outsider, based purely on my vague recollection of their newsletters etc., and someone can correct me) is that (1) they have a group working on "better understand agent foundations", and this group contains Abram and Scott, and they publish pretty much everything they're doing, (2) they have a group working on undisclosed research projects, which are NOT "better under...
Adam, can you make a positive case here for how the work being done on prosaic alignment leads to success? You didn't make one, and without it I don't understand where you're coming from. I'm not asking you to tell me a story that you have 100% probability on, just what is the success story you're acting under, such that EY's stances seem to you to be mostly distracting people from the real work.
(Later added disclaimer: it's a good idea to add "I feel like..." before the judgment in this comment, so that you keep in mind that I'm talking about my impressions and frustrations, rarely stating obvious facts (despite the language making it look so))
Thanks for trying to understand my point and asking me for more details. I appreciate it.
Yet I feel weird when trying to answer, because my gut reaction to your comment is that you're asking the wrong question? Also, the compression of my view to "EY's stances seem to you to be mostly distracting people from the real work" sounds more lossy than I'm comfortable with. So let me try to clarify and focus on these feelings and impressions, then I'll answer more about which success stories or directions excite me.
My current problem with EY's stances is twofold:
- First, in posts like this one, he literally writes that everything done under the label of alignment is faking it and not even attacking the problem, except like 3 people who even if they're trying have it all wrong. I think this is completely wrong, and that's even more annoying because I find that most people working on alignment are trying far harder harder to justify why they e
Thanks for naming specific work you think is really good! I think it's pretty important here to focus on the object-level. Even if you think the goodness of these particular research directions isn't cruxy (because there's a huge list of other things you find promising, and your view is mainly about the list as a whole rather than about any particular items on it), I still think it's super important for us to focus on object-level examples, since this will probably help draw out what the generators for the disagreement are.
John Wentworth’s Natural Abstraction Hypothesis, which is about checking his formalism-backed intuition that NNs actually learn similar abstractions that humans do. The success story is pretty obvious, in that if John is right, alignment should be far easier.
Eliezer liked this post enough that he asked me to signal-boost it in the MIRI Newsletter back in April.
And Paul Christiano and Stuart Armstrong are two of the people Eliezer named as doing very-unusually good work. We continue to pay Stuart to support his research, though he's mainly supported by FHI.
And Evan works at MIRI, which provides some Bayesian evidence about how much we tend to like his stuff. :)
So ...
Also, I feel like I want to emphasize that, like... it's OK to believe that the field you're working in is in a bad state? The social pressure against saying that kind of thing (or even thinking it to yourself) is part of why a lot of scientific fields are unhealthy, IMO. I'm in favor of you not taking for granted that Eliezer's right, and pushing back insofar as your models disagree with his. But I want to advocate against:
- Saying false things about what the other person is saying. A lot of what you've said about Eliezer and MIRI is just obviously false (e.g., we have contempt for "experimental work" and think you can't make progress by "Actually working with AIs and Thinking about real AIs").
- Shrinking the window of 'socially acceptable things to say about the field as a whole' (as opposed to unsolicited harsh put-downs of a particular researcher's work, where I see more value in being cautious).
I want to advocate 'smack-talking the field is fine, if that's your honest view; and pushing back is fine, if you disagree with the view'. I want to see more pushing back on the object level (insofar as people disagree), and less 'how dare you say that, do you think you're the king of align...
Enjoy your rest! :)
That sounded necessary to me to air out this frustration, and I think this was a good idea given the upvotes of my original post and the couple of people who messaged me to tell me that they were also annoyed.
If you'd just aired out your frustration, framing claims about others in NVC-like 'I feel like...' terms (insofar as you suspect you wouldn't reflectively endorse them), and then a bunch of people messaged you in private to say "thank you! you captured my feelings really well", then that would seem clearly great to me.
I'm a bit worried that what instead happened is that you made a bunch of clearly-false claims about other people and gave a bunch of invalid arguments, mixed in with the feelings-stuff; and you used the content warning at the top of the message to avoid having to distinguish which parts of your long, detailed comment are endorsed or not (rather than also flagging this within the comment); and then you also ran with this in a bunch of follow-up comments that were similarly not-endorsed but didn't even have the top-of-comment disclaimer. So that I could imagine some people who also aren't independently familiar with all the background facts, coul...
I'm a bit worried that what instead happened is that you made a bunch of clearly-false claims about other people and gave a bunch of invalid arguments, mixed in with the feelings-stuff; and you used the content warning at the top of the message to avoid having to distinguish which parts of your long, detailed comment are endorsed or not (rather than also flagging this within the comment); and then you also ran with this in a bunch of follow-up comments that were similarly not-endorsed but didn't even have the top-of-comment disclaimer. So that I could imagine some people who also aren't independently familiar with all the background facts, could come away with a lot of wrong beliefs about the people you're criticizing.
That sounds a bit unfair, in the sense that it makes it look like I just invented stuff I didn't believe and ran with it. When what actually happen was that I wrote about my frustrations, but made the mistake of stating them as obvious facts instead of impressions.
Of course, I imagine you feel that my portrayal of EY and MIRI was also unfair, sorry about that.
(I added a note to the three most ranty comments on this thread saying that people should mentally add "I feel like..." to judgments in them.)
Thanks for adding the note! :)
I'm confused. When I say 'that's just my impression', I mean something like 'that's an inside-view belief that I endorse but haven't carefully vetted'. (See, e.g., Impression Track Records, referring to Naming Beliefs.)
Example: you said that MIRI has "contempt with experimental work and not doing only decision theory and logic".
My prior guess would have been that you don't actually, for-real believe that -- that it's not your 'impression' in the above sense, more like 'unendorsed venting/hyperbole that has a more complicated relation to something you really believe'.
If you do (or did) think that's actually true, then our models of MIRI are much more different than I thought! Alternatively, if you agree this is not true, then that's all I meant in the previous comment. (Sorry if I was unclear about that.)
(Responding to entire comment thread) Rob, I don't think you're modeling what MIRI looks like from the outside very well.
- There's a lot of public stuff from MIRI on a cluster that has as central elements decision theory and logic (logical induction, Vingean reflection, FDT, reflective oracles, Cartesian Frames, Finite Factored Sets...)
- There was once an agenda (AAMLS) that involved thinking about machine learning systems, but it was deprioritized, and the people working on it left MIRI.
- There was a non-public agenda that involved Haskell programmers. That's about all I know about it. For all I know they were doing something similar to the modal logic work I've seen in the past.
- Eliezer frequently talks about how everyone doing ML work is pursuing dead ends, with potentially the exception of Chris Olah. Chris's work is not central to the cluster I would call "experimentalist".
- There has been one positive comment on the KL-divergence result in summarizing from human feedback. That wasn't the main point of that paper and was an extremely predictable result.
- There has also been one positive comment on Redwood Research, which was founded by people who have close ties to MIRI. The current step
So, the point of my comments was to draw a contrast between having a low opinion of "experimental work and not doing only decision theory and logic", and having a low opinion of "mainstream ML alignment work, and of nearly all work outside the HRAD-ish cluster of decision theory, logic, etc." I didn't intend to say that the latter is obviously-wrong; my goal was just to point out how different those two claims are, and say that the difference actually matters, and that this kind of hyperbole (especially when it never gets acknowledged later as 'oh yeah, that's not true and wasn't what I was thinking') is not great for discussion.
I think it's true that 'MIRI is super not into most ML alignment work', and I think it used to be true that MIRI put almost all of its research effort into HRAD-ish work, and regardless, this all seems like a completely understandable cached impression to have of current-MIRI. If I wrote stuff that makes it sound like I don't think those views are common, reasonable, etc., then I apologize for that and disavow the thing I said.
But this is orthogonal to what I thought I was talking about, so I'm confused about what seems to me like a topic switch. Maybe the ...
So, the point of my comments was to draw a contrast between having a low opinion of "experimental work and not doing only decision theory and logic", and having a low opinion of "mainstream ML alignment work, and of nearly all work outside the HRAD-ish cluster of decision theory, logic, etc." I didn't intend to say that the latter is obviously-wrong; my goal was just to point out how different those two claims are, and say that the difference actually matters, and that this kind of hyperbole (especially when it never gets acknowledged later as 'oh yeah, that's not true and wasn't what I meant') is not great for discussion.
It occurs to me that part of the problem may be precisely that Adam et al. don't think there's a large difference between these two claims (that actually matters). For example, when I query my (rough, coarse-grained) model of [your typical prosaic alignment optimist], the model in question responds to your statement with something along these lines:
...If you remove "mainstream ML alignment work, and nearly all work outside of the HRAD-ish cluster of decision theory, logic, etc." from "experimental work", what's left? Perhaps there are one or two (non-mainstream, bare
Not sure if this helps, and haven't read the thread carefully, but my sense is your framing might be eliding distinctions that are actually there, in a way that makes it harder to get to the bottom of your disagreement with Adam. Some predictions I'd have are that:
* For almost any experimental result, a typical MIRI person (and you, and Eliezer) would think it was less informative about AI alignment than I would.
* For almost all experimental results you would think they were so much less informative as to not be worthwhile.
* There's a small subset of experimental results that we would think are comparably informative, and also a some that you would find much more informative than I would.
(I'd be willing to take bets on these or pick candidate experiments to clarify this.)
In addition, a consequence of these beliefs is that compared to me you think we should be spending way more time sitting around thinking about stuff, and way less time doing experiments, than I do.
I would agree with you that "MIRI hates all experimental work" / etc. is not a faithful representation of this state of affairs, but I think there is nevertheless an important disagreement MIRI has with typical ML people, and that the disagreement is primarily about what we can learn from experiments.
I would agree with you that "MIRI hates all experimental work" / etc. is not a faithful representation of this state of affairs, but I think there is nevertheless an important disagreement MIRI has with typical ML people, and that the disagreement is primarily about what we can learn from experiments.
Ooh, that's really interesting. Thinking about it, I think my sense of what's going on is (and I'd be interested to hear how this differs from your sense):
- Compared to the average alignment researcher, MIRI tends to put more weight on reasoning like 'sufficiently capable and general AI is likely to have property X as a strong default, because approximately-X-ish properties don't seem particularly difficult to implement (e.g., they aren't computationally intractable), and we can see from first principles that agents will be systematically less able to get what they want when they lack property X'. My sense is that MIRI puts more weight on arguments like this for reasons like:
- We're more impressed with the track record of inside-view reasoning in science.
- I suspect this is partly because the average alignment researcher is impressed with how unusually-poorly inside-view reasoning has done i
- We're more impressed with the track record of inside-view reasoning in science.
I suspect a third important reason is that MIRI thinks alignment is mostly about achieving a certain kind of interpretability/understandability/etc. in the first AGI systems. Most ML experiments either aren't about interpretability and 'cracking open the hood', or they're not approaching the problem in a way that MIRI's excited by.
E.g., if you think alignment research is mostly about testing outer reward function to see what first-order behavior they produce in non-AGI systems, rather than about 'looking in the learned model's brain' to spot mesa-optimization and analyze what that optimization is ultimately 'trying to do' (or whatever), then you probably won't produce stuff that MIRI's excited about regardless of how experimental vs. theoretical your work is.
(In which case, maybe this is not actually a crux for the usefulness of most alignment experiments, and is instead a crux for the usefulness of most alignment research in general.)
Thanks. For time/brevity, I'll just say which things I agree / disagree with:
> sufficiently capable and general AI is likely to have property X as a strong default [...]
I generally agree with this, although for certain important values of X (such as "fooling humans for instrumental reasons") I'm probably more optimistic than you that there will be a robust effort to get not-X, including by many traditional ML people. I'm also probably more optimistic (but not certain) that those efforts will succeed.
[inside view, modest epistemology]: I don't have a strong take on either of these. My main take on inside views is that they are great for generating interesting and valuable hypotheses, but usually wrong on the particulars.
> less weight on reasoning like 'X was true about AI in 1990, in 2000, in 2010, and in 2020; therefore X is likely to be true about AGI when it's developed
I agree, see my post On the Risks of Emergent Behavior in Foundation Models. In the past I think I put too much weight on this type of reasoning, and also think most people in ML put too much weight on it.
> MIRI thinks AGI is better thought of as 'a weird specific sort of AI', rather than as 'like ex...
I'm not (retroactively in imaginary prehindsight) excited by this problem because neither of the 2 possible answers (3 possible if you count "the same") had any clear-to-my-model relevance to alignment, or even AGI. AGI will have better OOD generalization on capabilities than current tech, basically by the definition of AGI; and then we've got less-clear-to-OpenPhil forces which cause the alignment to generalize more poorly than the capabilities did, which is the Big Problem. Bigger models generalizing better or worse doesn't say anything obvious to any piece of my model of the Big Problem. Though if larger models start generalizing more poorly, then it takes longer to stupidly-brute-scale to AGI, which I suppose affects timelines some, but that just takes timelines from unpredictable to unpredictable sooo.
If we qualify an experiment as interesting when it can tell anyone about anything, then there's an infinite variety of experiments "interesting" in this sense and I could generate an unlimited number of them. But I do restrict my interest to experiments which can not only tell me something I don't know, but tell me something relevant that I don't know. &nb...
are you interested in Redwood's research because it might plausibly generate alignment issues and problems that are analogous to the real problem within the safer regime and technology we have now?
It potentially sheds light on small subpieces of things that are particular aspects that contribute to the Real Problem, like "What actually went into the nonviolence predicate instead of just nonviolence?" Much of the Real Meta-Problem is that you do not get things analogous to the full Real Problem until you are just about ready to die.
One tiny note: I was among the people on AAMLS; I did leave MIRI the next year; and my reasons for so doing are not in any way an indictment of MIRI. (I was having some me-problems.)
I still endorse MIRI as, in some sense, being the adults in the AI Safety room, which has... disconcerting effects on my own level of optimism.
... I find that most people working on alignment are trying far harder harder to justify why they expect their work to matter than EY and the old-school MIRI team ever did.
You've had a few comments along these lines in this thread, and I think this is where you're most severely failing to see the situation from Yudkowsky's point of view.
From Yudkowsky's view, explaining and justifying MIRI's work (and the processes he uses to reach such judgements more generally) was the main point of the sequences. He has written more on the topic than anyone else in the world, by a wide margin. He basically spent several years full-time just trying to get everyone up to speed, because the inductive gap was very very wide.
When I put on my Yudkowsky hat and look at both the OP and your comments through that lens... I imagine if I were Yudkowsky I'd feel pretty exasperated at this point. Like, he's written a massive volume on the topic, and now ten years later a large chunk of people haven't even bothered to read it. (In particular, I know (because it's come up in conversation) that at least a few of the people who talk about prosaic alignment a lot haven't read the sequences, and I suspect that a d...
I don't mean to say that there's critique of prosaic alignment specifically in the sequences. Rather, a lot of the generators of the Yudkowsky-esque worldview are in there. (That is how the sequences work: it's not about arguing specific ideas around alignment, it's about explaining enough of the background frames and generators that the argument becomes unnecessary. "Raise the sanity waterline" and all that.)
For instance, just the other day I ran across this:
...Of this I learn the lesson: You cannot manipulate confusion. You cannot make clever plans to work around the holes in your understanding. You can't even make "best guesses" about things which fundamentally confuse you, and relate them to other confusing things. Well, you can, but you won't get it right, until your confusion dissolves. Confusion exists in the mind, not in the reality, and trying to treat it like something you can pick up and move around, will only result in unintentional comedy.
Similarly, you cannot come up with clever reasons why the gaps in your model don't matter. You cannot draw a border around the mystery, put on neat handles that let you use the Mysterious Thing without
From testimonials by a bunch of more ML people and how any discussion of alignment needs to clarify that you don’t share MIRI’s contempt with experimental work and not doing only decision theory and logic
If you were in the situation described by The Rocket Alignment Problem, you could think "working with rockets right now isn't useful, we need to focus on our conceptual confusions about more basic things" without feeling inherently contemptuous of experimentalism -- it's a tool in the toolbox (which may or may not be appropriate to the task at hand), not a low- or high-status activity on a status hierarchy.
Separately, I think MIRI has always been pretty eager to run experiments in software when they saw an opportunity to test important questions that way. It's also been 4.5 years now since we announced that we were shifting a lot of resources away from Agent Foundations and into new stuff, and 3 years since we wrote a very long (though still oblique) post about that research, talking about its heavy focus on running software experiments. Though we also made sure to say:
...In a sense, you can think of our new research as tackling the same sort of problem that we’ve always been attackin
Thank you for the links Adam. To clarify, the kind of argument I'm really looking for is something like the following three (hypothetical) examples.
- Mesa-optimization is the primary threat model of unaligned AGI systems. Over the next few decades there will be a lot of companies building ML systems that create mesa-optimizers. I think it is within 5 years of current progress that we will understand how ML systems create mesa-optimizers and how to stop it.Therefore I think the current field is adequate for the problem (80%).
- When I look at the research we're outputting, it seems to me to me that we are producing research at a speed and flexibility faster than any comparably sized academic department globally, or the ML industry, and so I am much more hopeful that we're able to solve our difficult problem before the industry builds an unaligned AGI. I give it a 25% probability, which I suspect is much higher than Eliezer's.
- I basically agree the alignment problem is hard and unlikely to be solved, but I don't think we have any alternative than the current sorts of work being done, which is a combo of (a) agent foundations work (b) designing theoretical training algorithms (like Paul is)
Thanks for the examples, that helps a lot.
I'm glad that I posted my inflammatory comment, if only because exchanging with you and Rob made me actually consider the question of "what is our story to success", instead of just "are we making progress/creating valuable knowledge". And the way you two have been casting it is way less aversive to me that the way EY tends to frame it. This is definitely something I want to think more about. :)
I want to leave this paragraph as social acknowledgment that you mentioned upthread that you're tired and taking a break, and I want to give you a bunch of social space to not return to this thread for however long you need to take! Slow comments are often the best.
Appreciated. ;)
I'm sympathetic to most of your points.
highly veiled contempt for anyone not doing that
I have sympathy for the "this feels somewhat contemptuous" reading, but I want to push back a bit on the "EY contemptuously calling nearly everyone fakers" angle, because I think "[thinly] veiled contempt" is an uncharitable reading. He could be simply exasperated about the state of affairs, or wishing people would change their research directions but respect them as altruists for Trying At All, or who knows what? I'd rather not overwrite his intentions with our reactions (although it is mostly the job of the writer to ensure their writing communicates the right information [although part of the point of the website discussion was to speak frankly and bluntly]).
If superintelligence is approximately multimodal GPT-17 plus reinforcement learning, then understanding how GPT-3-scale algorithms function is exceptionally important to understanding super-intelligence.
Also, if superintelligence doesn’t happen then prosaic alignment is the only kind of alignment.
This is already reflected in the upvotes, but just to say it explicitly: I think the replies to this comment from Rob and dxu in particular have been exceptionally charitable and productive; kudos to them. This seems like a very good case study in responding to a provocative framing with a concentration of positive discussion norms that leads to productive engagement.
if EY and other MIRI people who are very dubious of most alignment research could give more feedback on that and enter the dialogue, maybe by commenting more on the AF. My problem is not so much with them disagreeing with most of the work, it’s about the disagreement stopping to “that’s not going to work” and not having dialogue and back and forth.
Just in case anyone hasn't already seen these, EY wrote Challenges to Christiano’s capability amplification proposal and this comment (that I already linked to in a different comment on this page) (also has a reply thread), both in 2018. Also The Rocket Alignment Problem.
Context for anyone who's not aware:
Nerd sniping is a slang term that describes a particularly interesting problem that is presented to a nerd, often a physicist, tech geek or mathematician. The nerd stops all activity to devote attention to solving the problem, often at his or her own peril
I'm annoyed by EY (and maybe MIRI's?) dismissal of every other alignment work, and how seriously it seems to be taken here, given their track record of choosing research agendas with very indirect impact on alignment
For what it's worth, my sense is that EY's track record is best in 1) identifying problems and 2) understanding the structure of the alignment problem.
And, like, I think it is possible that you end up in situations where the people who understand the situation best end up the most pessimistic about it. If you're trying to build a bridge to the moon, in fact it's not going to work, and any determination applied there is going to get wasted. I think I see how a "try to understand things and cut to the heart of them" notices when it's in situations like that, and I don't see how "move the ball forward from where it is now" notices when it's in situations like that.
Oh, I was imagining something like "well, our current metals aren't strong enough, what if we developed stronger ones?", and then focusing on metallurgy. And this is making forward progress--you can build a taller tower out of steel than out of iron--but it's missing more fundamental issues like "you're not going to be able to drive on a bridge that's perpendicular to gravity, and the direction of gravity will change over the course of the trip" or "the moon moves relative to the earth, such that your bridge won't be able to be one object", which will sink the project even if you can find a supremely strong metal.
For example, let's consider Anca Dragan's research direction that I'm going to summarize as "getting present-day robots to understand what humans around them want and are doing so that they can achieve their goals / cooperate more effectively." (In mildly adversarial situations like driving, you don't want to make a cooperatebot, but rather something that follows established norms / prevents 'cutting' and so on, but when you have a human-robot team you do care mostly about effective cooperation.)
My guess is this 1) will make the world a better place in the short run under ...
I share the impression that the agent foundations research agenda seemed not that important. But that point doesn't feel sufficient to argue that Eliezer's pessimism about the current state of alignment research is just a face-saving strategy his brain tricked him into adopting. (I'm not saying you claimed that it is sufficient; probably a lot of other data points are factoring into your judgment.) MIRI have deprioritized agent foundations research for quite a while now. I also just think it's extremely common for people to have periods where they work on research that eventually turns out to be not that important; the interesting thing is to see what happens when that becomes more apparent. I immediately trust people more if I see that they are capable of pivoting and owning up to past mistakes, and I could imagine that MIRI deserves a passing grade on this, even though I also have to say that I don't know how exactly they nowadays think about prioritization in 2017 and earlier.
I really like Vaniver's comment further below:
For what it's worth, my sense is that EY's track record is best in 1) identifying problems and 2) understanding the structure of the alignment problem.
...And, like
My impression from talking with people (but not having direct confirmation from the people who left) was far more that OpenAI was focusing the conceptual safety team on ML work and the other safety team on making sure GPT-3 was not racist, which was not the type of work they were really excited about. But I might also be totally wrong about this.
Interesting! This is quite different from the second-hand accounts I heard. (I assume we're touching different parts of the elephant.)
I am confused about the emphasis on secrecy.
Certainly, if you're working on a substantial breakthrough in AI capability, there are reasons to keep it secret. But why would you work on that in the first place? One answer I can imagine is: "Current AI building blocks (deep learning) are too opaque and hard to prove things about, so we need to develop alternative building blocks. These alternative building blocks, if such are found, might produce a capability breakthrough." But, currently we don't have a solution even modulo the building blocks. Because, even if we allow computationally infeasible, or what I've been calling "weakly feasible" algorithms, such as Bayesian inference, we still don't have a complete solution. Therefore, it seems reasonable to focus on "alignment modulo building blocks" i.e. alignment in the unbounded / weakly feasible regime, and I don't see any reason to be secretive about it. On the contrary, we want to involve as many people as possible to get more progress and we want the AI community to know about whatever we come up with, to increase the probability they will use it.
Maybe the objection is, this path won't lead us to success quickly enough. But then, what alternative path is better? And how would secrecy enable it?
If there's no difference between capability work and alignment work, then how is it possible to influence anything at all? If capability and alignment go hand in hand, then either transformative capability corresponds to sufficient alignment (in which case there is no technical problem) or it doesn't (in which case we're doomed).
The only world in which secrecy makes sense, AFAICT, is if you're going to solve alignment and capability all by yourself. I am extremely skeptical of this approach.
I was halfway through a PhD on software testing and verification before joining Anthropic (opinions my own, etc), and I'm less convinced than Eliezer about theorem-proving for AGI safety.
There are so many independently fatal objections that I don't know how to structure this or convince someone who thinks it would work. I am therefore offering a $1,000 prize for solving a far easier problem:
Take an EfficientNet model with >= 99% accuracy on MNIST digit classification. What is the largest possible change in the probability assigned to some class between two images, which differ only in the least significant bit of a single pixel? Prove your answer before 2023.
Your proof must not include executing the model, nor equivalent computations (e.g. concolic execution). You may train a custom model and/or directly set model weights, so long as it uses a standard EfficientNet architecture and reaches 99% accuracy. Bonus points for proving more of the sensitivity curve.
I will also take bets that nobody will accomplish this by 2025, nor any loosely equivalent proof for a GPT-3 class model by 2040. This is a very bold claim, but I believe that rigorously proving even trivial global ...
You're attacking a strawman of what kind of theorems we want to prove. Obviously we are not going to prove theorems that contain specific datasets as part of the statement. What we're going to do is build a theory founded on certain assumptions about the real-world (such as locality / decoupling of dynamics on different scales / some kind of chaos / certain bounds on computational complexity / existence of simple fundamental laws etc) and humans (e.g. that they are approximately rational agents, for some definition thereof). Such a theory can produce many insights about what factors influence e.g. the possibility of adversarial attacks that you mention, most of which will be qualitative and some of which can be made quantitative by combining with empirical research (such as the research OpenAI does on scaling laws).
And, ofc the theory is only as good as its assumptions. Ofc if there are attack vectors your model doesn't account for, your system can be attacked. Having a theory is not a magical wand that immediately solves everything. But, it does put you in a much, much better position than working off pure guesswork and intuition.
Another angle is that, once we can at least state the theorem we might try to make the AI itself prove it. This can still fail: maybe the theorem is practically unprovable, or maybe we can't safely train the AI to prove theorems. But it does give us some additional leverage.
First, an apology: I didn't mean this to be read as an attack or a strawman, nor applicable to any use of theorem-proving, and I'm sorry I wasn't clearer. I agree that formal specification is a valuable tool and research direction, a substantial advancement over informal arguments, and only as good as the assumptions. I also think that hybrid formal/empirical analysis could be very valuable.
Trying to state a crux, I believe that any plan which involves proving corrigibility properties about MuZero (etc) is doomed, and that safety proofs about simpler approximations cannot provide reliable guarantees about the behaviour of large models with complex emergent behaviour. This is in large part because formalising realistic assumptions (e.g. biased humans) is very difficult, and somewhat because proving anything about very large models is wildly beyond the state of the art and even verified systems have (fewer) bugs.
Being able to state theorems about AGI seems absolutely necessary for success; but I don't think it's close to sufficient.
I want to push back against the idea that ANNs are "vectors of floating points" and therefore it's impossible to prove things about them. Many algorithms involve continuous variables and we can prove things about them. Support vector machines are also learning algorithms that are "vectors of floating points" and we have a pretty good theory of how they work. In fact, there already is a sizable body of theoretical results about ANNs, even if it still falls significantly short of what we need.
The biggest problem is not necessarily in the "floating points". The problem is that we still don't have satisfactory models of what an "agent" is and what it means for an agent to be "aligned". But, we do have some leads. And once we solve this part, there's no reason of principle why it cannot be combined with some (hitherto unknown) theory of generalization bounds for ANNs.
A confusion: it seems that Eliezer views research that is predictable as basically-useless. I think I don't understand what "predictable" means here. In what sense is expected utility quantilization not predictable?
Maybe the point is that coming up with the concept is all that matters, and the experiments that people usually do don't matter because after coming up with the concept the experiments are predictable? I'm much more sympathetic to that, but then I'm confused why "predictable" implies "useless"; many prosaic alignment papers have as their main contribution a new algorithm, which seems like a similar type of thing as quantilization.
An example that springs to my mind is Abram wrote a blog post in 2018 mentioning the "easy problem of wireheading". He described both the problem and its solution in like one sentence, and then immediately moved on to the harder problems.
Later on, DeepMind did an experiment that (in my assessment) mostly just endorsed what Abram said as being correct.
For the record, I don't think that particular DeepMind experiment was zero value, for various reasons. But at the same time, I think that Abram wins hands-down on the metric of "progress towards AI alignment per researcher-hour", and this is true at both the production and consumption end (I can read Abram's one sentence much much faster than I can skim the DeepMind paper).
If we had a plausible-to-me plan that gets us to safe & beneficial AGI, I would be really enthusiastic about going back and checking all the assumptions with experiments. That's how you shore up the foundations, flesh out the details, start developing working code and practical expertise, etc. etc. But I don't think we have such a plan right now.
Also, there are times when it's totally unclear a priori what an algorithm will do just by thinking about it, and then ...
An example: when I first heard the Ought experiments described, I was pretty highly confident how they'd turn out - people would mostly fail to coordinate on any problem without an already-very-obvious factorization. (See here for the kinds of evidence informing that high confidence, though applied to a slightly different question. See here and here for the more general reasoning/world models which underlie that prediction.) From what I've heard of the experiments, it seems that that is indeed basically what happened; therefore the experiments provided approximately-zero new information to my model. They were "useless" in that sense.
(I actually think those experiments were worth running just on the small chance that they'd find something very high value, or more likely that the people running them would have some high-value insight, but I'd still say "probably useless" was a reasonable description beforehand.)
I don't know if Eliezer would agree with this particular example, but I think this is the sort of thing he's gesturing at.
I actually think this particular view is worth fleshing out, since it seems to come up over and over again in discussions of what AI alignment work is valuable (versus not).
For example, it does seem to me that >80% of the work in actually writing a published paper (at least amongst papers at CHAI) (EDIT: no longer believe this on reflection, see Rohin’s comment below) involves doing work with results that are predictable to the author after the concept (for example, actually getting your algorithm to run, writing code for experiments, running said experiments, writing up the results into a paper, etc.)
This just doesn't match my experience at all. Looking through my past AI papers, I only see two papers where I could predict the results of the experiments on the first algorithm I tried at the beginning of the project. The first one (benefits of assistance) was explicitly meant to be a "communication" paper rather than a "research" paper (at the time of project initiation, rather than in hindsight). The second one (Overcooked) was writing up results that were meant to be the baselines against which the actual unpredictable research (e.g. this) was going to be measured against; it just turned out that that was already sufficiently interesting to the broader community.
(Funny story about the Overcooked paper; we wrote the paper + did the user study in ~two weeks iirc, because it was only two weeks before the deadline that we considered that the "baseline" results might already be interesting enough to warrant a conference paper. It's now my most-cited AI paper.)
(I'm also not actually sure that I would have predicted the Overcooked results when writing down the first algorithm; the conceptual story felt strong but there are several other papers where the conceptual story felt strong b...
Aaaaaaaaaaaaahhhhhhhhhhhhhhhhh!!!!!!!!!!!!
(...I'll be at the office, thinking about how to make enough progress fast enough.)
Follow-up
One of Eliezer's claims here is
It is very, very clear that at present rates of progress, adding that level of alignment capability as grown over the next N years, to the AGI capability that arrives after N years, results in everybody dying very quickly.
This is a claim I basically agree with.
I don't think the situation is entirely hopeless, but I don't think any of the current plans (or the current alignment field) are on track to save us.
I'm sure Eliezer has written about this previously, but why doesn't he think corrigibility is a natural stance?
It does seem like existing approaches to corrigibility (IE, the utility balancing approaches in MIRI/Stuart Armstrong's work and the "agent has incomplete information" approaches outlined in Dylan Hadfield-Menell or Alex Turner's work) are incredibly fragile. I do agree that current approaches involving utility balancing/assigning utility to branches never executed are probably way too finicky to get working. I also agree that all the existing approaches involving the agent modelling itself as having incomplete information rely on well-calibrated priors and also all succumb to the problem of fully updated deference.
However, I think it's not at all obvious to me that corrigibility doesn't have a "small central core". It does seem to me like the "you are incomplete, you will never be complete" angle captures a lot of what we mean by corrigibility.
It's possible the belief is empirical - that is, people have tried all the obvious ways to patch/fix this angle, and they've all failed, so the problem is hard (at least relative to the researchers we have working on it)...
I'm on record as early as 2008 as saying that I expected superintelligences to crack protein folding, some people disputed that and were all like "But how do you know that's solvable?" and then AlphaFold 2 came along and cracked the protein folding problem they'd been skeptical about, far below the level of superintelligence.
This is tangential nitpicking (I agree that protein folding is solvable), but I don't think AlphaFold 2 entirely cracked it. AFAIU, AF2 relies on multiple sequence alignment as part of its input: the sequences of homologous proteins from different species. This is a standard method to simplify the problem, because by observing that certain parts of the sequence tend to vary together between homologues, you can guess that they correspond to chain fragments that are adjacent in the folded configuration.
Ofc even so it is very impressive and has plenty of applications. But, if you want to invent your own proteins from scratch, this is not good enough.
Well, if viewing it on that level, AlphaFold 2 didn't crack the full problem because it doesn't let you put in a chemical function and get out a protein which performs that function while subject to other constraints of a surrounding wet system, which is the protein folding problem you have to solve to get wet nanotech out the other end, which is why we don't already have general wet nanotech today.
I think this is my second-favorite post in the MIRI dialogues (for my overall review see here).
I think this post was valuable to me in a much more object-level way. I think this post was the first post that actually just went really concrete on the current landscape of efforts int he domain of AI Notkilleveryonism and talked concretely about what seems feasible for different actors to achieve, and what isn't, in a way that parsed for me, and didn't feel either like something obviously political, or delusional.
I didn't find the part about differ...
I do think that if you get an AGI significantly past human intelligence in all respects, it would obviously tend to FOOM. I mean, I suspect that Eliezer fooms if you give an Eliezer the ability to backup, branch, and edit himself.
What improvements would you make to your brain that you would anticipate yielding greater intelligence? I can think of a few possible strategies:
- Just adding a bunch of neurons everywhere. Make my brain bigger.
- Study how very smart brains look, and try to make my brain look more like theirs.
For an AI, the first strategy is equivalen...
I'm torn because I mostly agree with Eliezer that things don't look good, and most technical approaches don't seem very promising.
But the attitude of unmitigated doomyness seems counter-productive.
And there's obviously things worth doing and working on and people getting on with it.
It seems like Eliezer is implicitly focused on finding an "ultimate solution" to alignment that we can be highly confident solves the problem regardless of how things play out. But this is not where the expected utility is. The expected utility is mostly in buying ti...
Yudkowsky mentions this briefly in the middle of the dialogue:
I don't know however if I should be explaining at this point why "manipulate humans" is convergent, why "conceal that you are manipulating humans" is convergent, why you have to train in safe regimes in order to get safety in dangerous regimes (because if you try to "train" at a sufficiently unsafe level, the output of the unaligned system deceives you into labeling it incorrectly and/or kills you before you can label the outputs), or why attempts to teach corrigibility in safe regimes are unlikely to generalize well to higher levels of intelligence and unsafe regimes (qualitatively new thought processes, things being way out of training distribution, and, the hardest part to explain, corrigibility being "anti-natural" in a certain sense that makes it incredibly hard to, eg, exhibit any coherent planning behavior ("consistent utility function") which corresponds to being willing to let somebody else shut you off, without incentivizing you to actively manipulate them to shut you off).
Basically, there are reasons to expect that alignment techniques that work in smaller safe regime fail in larger, unsafe regimes. For ...
What probability do you assign to the proposition "Prosaic alignment will fail"?
- Purely based on your inside view model
- After updating on everyone else's views
Same question for:
"More than 50% of the prosaic alignment work done by the top 7 researchers is nearly useless"
If you have an untrustworthy general superintelligence generating English strings meant to be "reasoning/arguments/proofs/explanations" about eg a nanosystem design, then I would not only expect the superintelligence to be able to fool humans in the sense of arguing for things that were not true in a way that fooled the humans, I'd expect the superintelligence to be able to covertly directly hack the humans in ways that I wouldn't understand even after having been told what happened.
I can't deny that this is possible. Lying is obviously possible; lyi...
Eliezer Yudkowsky
Alpha Zero scales with more computing power, I think AlphaFold 2 scales with more computing power, Mu Zero scales with more computing power. Precisely because GPT-3 doesn't scale, I'd expect an AGI to look more like Mu Zero and particularly with respect to the fact that it has some way of scaling.
I think this view dovetails quite strongly with the view expressed in this comment by maximkazhenkov:
Progress in model-based RL is far more relevant to getting us closer to AGI than other fields like NLP or image recognition or neuroscience or ML hardware. I worry that once the research community shifts its focus towards RL, the AGI timeline will collapse - not necessarily because there are no more critical insights left to be discovered, but because it's fundamentally the right path to work on and whatever obstacles remain will buckle quickly once we throw enough warm bodies at them. I think - and this is highly controversial - that the focus on NLP and Vision Transformer has served as a distraction for a couple of years and actually delayed progress towards AGI.
...If curiosity-driven exploration gets thrown into the mix and Starcraft/Dota gets solved (for real th
It seems to me that the surprising simplicity of current-generation ML algorithms is a big part of the problem.
As a thought experiment: suppose you had a human brain, with the sort of debug access you'd have with a neural net; ie, you could see all the connections, edge weights, and firings, and had a decent multiple of the compute the brain has. Could you extract something like a verbal inner monologue, a text stream that was strongly predictive of that human's plans? I don't think it would be trivial, but my guess is that you could. It wouldn't hold up against a meditator optimizing against you, but it would be a solid starting point.
Could you do the same thing to GPT-3? No; you can't get language out of it that predicts its plans, because it doesn't have plans. Could you do the same thing to AlphaZero? No, you can't get language out of it that predicts its plans, because it doesn't use language.
This analogy makes me think neural net transparency might not be as doomed as the early results would suggest; they aren't finding human-legible low-dimensional representations of things because those representations aren't present (GPT-3) or have nothing human-legible to match up to (Alp...
The following is a partially redacted and lightly edited transcript of a chat conversation about AGI between Eliezer Yudkowsky and a set of invitees in early September 2021. By default, all other participants are anonymized as "Anonymous".
I think this Nate Soares quote (excerpted from Nate's response to a report by Joe Carlsmith) is a useful context-setting preface regarding timelines, which weren't discussed as much in the transcript:
Further preface by Eliezer:
In some sections here, I sound gloomy about the probability that coordination between AGI groups succeeds in saving the world. Andrew Critch reminds me to point out that gloominess like this can be a self-fulfilling prophecy - if people think successful coordination is impossible, they won’t try to coordinate. I therefore remark in retrospective advance that it seems to me like at least some of the top AGI people, say at Deepmind and Anthropic, are the sorts who I think would rather coordinate than destroy the world; my gloominess is about what happens when the technology has propagated further than that. But even then, anybody who would rather coordinate and not destroy the world shouldn’t rule out hooking up with Demis, or whoever else is in front if that person also seems to prefer not to completely destroy the world. (Don’t be too picky here.) Even if the technology proliferates and the world ends a year later when other non-coordinating parties jump in, it’s still better to take the route where the world ends one year later instead of immediately. Maybe the horse will sing.
Eliezer Yudkowsky
Hi and welcome. Points to keep in mind:
- I'm doing this because I would like to learn whichever actual thoughts this target group may have, and perhaps respond to those; that's part of the point of anonymity. If you speak an anonymous thought, please have that be your actual thought that you are thinking yourself, not something where you're thinking "well, somebody else might think that..." or "I wonder what Eliezer's response would be to..."
- Eliezer's responses are uncloaked by default. Everyone else's responses are anonymous (not pseudonymous) and neither I nor MIRI will know which potential invitee sent them.
- Please do not reshare or pass on the link you used to get here.
- I do intend that parts of this conversation may be saved and published at MIRI's discretion, though not with any mention of who the anonymous speakers could possibly have been.
Eliezer Yudkowsky
(Thank you to Ben Weinstein-Raun for building chathamroom.com, and for quickly adding some features to it at my request.)
Eliezer Yudkowsky
It is now 2PM; this room is now open for questions.
Anonymous
How long will it be open for?
Eliezer Yudkowsky
In principle, I could always stop by a couple of days later and answer any unanswered questions, but my basic theory had been "until I got tired".
Anonymous
At a high level one thing I want to ask about is research directions and prioritization. For example, if you were dictator for what researchers here (or within our influence) were working on, how would you reallocate them?
Eliezer Yudkowsky
The first reply that came to mind is "I don't know." I consider the present gameboard to look incredibly grim, and I don't actually see a way out through hard work alone. We can hope there's a miracle that violates some aspect of my background model, and we can try to prepare for that unknown miracle; preparing for an unknown miracle probably looks like "Trying to die with more dignity on the mainline" (because if you can die with more dignity on the mainline, you are better positioned to take advantage of a miracle if it occurs).
Anonymous
I'm curious if the grim outlook is currently mainly due to technical difficulties or social/coordination difficulties. (Both avenues might have solutions, but maybe one seems more recalcitrant than the other?)
Eliezer Yudkowsky
Technical difficulties. Even if the social situation were vastly improved, on my read of things, everybody still dies because there is nothing that a handful of socially coordinated projects can do, or even a handful of major governments who aren't willing to start nuclear wars over things, to prevent somebody else from building AGI and killing everyone 3 months or 2 years later. There's no obvious winnable position into which to play the board.
Anonymous
just to clarify, that sounds like a large scale coordination difficulty to me (i.e., we - as all of humanity - can't coordinate to not build that AGI).
Eliezer Yudkowsky
I wasn't really considering the counterfactual where humanity had a collective telepathic hivemind? I mean, I've written fiction about a world coordinated enough that they managed to shut down all progress in their computing industry and only manufacture powerful computers in a single worldwide hidden base, but Earth was never going to go down that route. Relative to remotely plausible levels of future coordination, we have a technical problem.
Anonymous
Curious about why building an AGI aligned to its users' interests isn't a thing a handful of coordinated projects could do that would effectively prevent the catastrophe. The two obvious options are: it's too hard to build it vs it wouldn't stop the other group anyway. For "it wouldn't stop them", two lines of reply are nobody actually wants an unaligned AGI (they just don't foresee the consequences and are pursuing the benefits from automated intelligence, so can be defused by providing the latter) (maybe not entirely true: omnicidal maniacs), and an aligned AGI could help in stopping them. Is your take more on the "too hard to build" side?
Eliezer Yudkowsky
Because it's too technically hard to align some cognitive process that is powerful enough, and operating in a sufficiently dangerous domain, to stop the next group from building an unaligned AGI in 3 months or 2 years. Like, they can't coordinate to build an AGI that builds a nanosystem because it is too technically hard to align their AGI technology in the 2 years before the world ends.
Anonymous
Summarizing the threat model here (correct if wrong): The nearest competitor for building an AGI is at most N (<2) years behind, and building an aligned AGI, even when starting with the ability to build an unaligned AGI, takes longer than N years. So at some point some competitor who doesn't care about safety builds the unaligned AGI. How does "nobody actually wants an unaligned AGI" fail here? It takes >N years to get everyone to realise that they have that preference and that it's incompatible with their actions?
Eliezer Yudkowsky
Many of the current actors seem like they'd be really gung-ho to build an "unaligned" AGI because they think it'd be super neat, or they think it'd be super profitable, and they don't expect it to destroy the world. So if this happens in anything like the current world - and I neither expect vast improvements, nor have very long timelines - then we'd see Deepmind get it first; and, if the code was not immediately stolen and rerun with higher bounds on the for loops, by China or France or whoever, somebody else would get it in another year; if that somebody else was Anthropic, I could maybe see them also not amping up their AGI; but then in 2 years it starts to go to Facebook AI Research and home hobbyists and intelligence agencies stealing copies of the code from other intelligence agencies and I don't see how the world fails to end past that point.
Anonymous
What does trying to die with more dignity on the mainline look like? There's a real question of prioritisation here between solving the alignment problem (and various approaches within that), and preventing or slowing down the next competitor. I'd personally love more direction on where to focus my efforts (obviously you can only say things generic to the group).
Eliezer Yudkowsky
I don't know how to effectively prevent or slow down the "next competitor" for more than a couple of years even in plausible-best-case scenarios. Maybe some of the natsec people can be grownups in the room and explain why "stealing AGI code and running it" is as bad as "full nuclear launch" to their foreign counterparts in a realistic way. Maybe more current AGI groups can be persuaded to go closed; or, if more than one has an AGI, to coordinate with each other and not rush into an arms race. I'm not sure I believe these things can be done in real life, but it seems understandable to me how I'd go about trying - though, please do talk with me a lot more before trying anything like this, because it's easy for me to see how attempts could backfire, it's not clear to me that we should be inviting more attention from natsec folks at all. None of that saves us without technical alignment progress. But what are other people supposed to do about researching alignment when I'm not sure what to try there myself?
Anonymous
thanks! on researching alignment, you might have better meta ideas (how to do research generally) even if you're also stuck on object level. and you might know/foresee dead ends that others don't.
Eliezer Yudkowsky
I definitely foresee a whole lot of dead ends that others don't, yes.
Anonymous
Does pushing for a lot of public fear about this kind of research, that makes all projects hard, seem hopeless?
Eliezer Yudkowsky
What does it buy us? 3 months of delay at the cost of a tremendous amount of goodwill? 2 years of delay? What's that delay for, if we all die at the end? Even if we then got a technical miracle, would it end up impossible to run a project that could make use of an alignment miracle, because everybody was afraid of that project? Wouldn't that fear tend to be channeled into "ah, yes, it must be a government project, they're the good guys" and then the government is much more hopeless and much harder to improve upon than Deepmind?
Anonymous
I imagine lack of public support for genetic manipulation of humans has slowed that research by more than three months
Anonymous
'would it end up impossible to run a project that could make use of an alignment miracle, because everybody was afraid of that project?'
...like, maybe, but not with near 100% chance?
Eliezer Yudkowsky
I don't want to sound like I'm dismissing the whole strategy, but it sounds a lot like the kind of thing that backfires because you did not get exactly the public reaction you wanted, and the public reaction you actually got was bad; and it doesn't sound like that whole strategy actually has a visualized victorious endgame, which makes it hard to work out what the exact strategy should be; it seems more like the kind of thing that falls under the syllogism "something must be done, this is something, therefore this must be done" than like a plan that ends with humane life victorious.
Regarding genetic manipulation of humans, I think the public started out very unfavorable to that, had a reaction that was not at all exact or channeled, does not allow for any 'good' forms of human genetic manipulation regardless of circumstances, driving the science into other countries - it is not a case in point of the intelligentsia being able to successfully cunningly manipulate the fear of the masses to some supposed good end, to put it mildly, so I'd be worried about deriving that generalization from it. The reaction may more be that the fear of the public is a big powerful uncontrollable thing that doesn't move in the smart direction - maybe the public fear of AI gets channeled by opportunistic government officials into "and that's why We must have Our AGI first so it will be Good and we can Win". That seems to me much more like a thing that would happen in real life than "and then we managed to manipulate public panic down exactly the direction we wanted to fit into our clever master scheme", especially when we don't actually have the clever master scheme it fits into.
Eliezer Yudkowsky
I have a few stupid ideas I could try to investigate in ML, but that would require the ability to run significant-sized closed ML projects full of trustworthy people, which is a capability that doesn't seem to presently exist. Plausibly, this capability would be required in any world that got some positive model violation ("miracle") to take advantage of, so I would want to build that capability today. I am not sure how to go about doing that either.
Anonymous
if there's a chance this group can do something to gain this capability I'd be interested in checking it out. I'd want to know more about what "closed"and "trustworthy" mean for this (and "significant-size" I guess too). E.g., which ones does Anthropic fail?
Eliezer Yudkowsky
What I'd like to exist is a setup where I can work with people that I or somebody else has vetted as seeming okay-trustworthy, on ML projects that aren't going to be published. Anthropic looks like it's a package deal. If Anthropic were set up to let me work with 5 particular people at Anthropic on a project boxed away from the rest of the organization, that would potentially be a step towards trying such things. It's also not clear to me that Anthropic has either the time to work with me, or the interest in doing things in AI that aren't "stack more layers" or close kin to that.
Anonymous
That setup doesn't sound impossible to me -- at DeepMind or OpenAI or a new org specifically set up for it (or could be MIRI) -- the bottlenecks are access to trustworthy ML-knowledgeable people (but finding 5 in our social network doesn't seem impossible?) and access to compute (can be solved with more money - not too hard?). I don't think DM and OpenAI are publishing everything - the "not going to be published" part doesn't seem like a big barrier to me. Is infosec a major bottleneck (i.e., who's potentially stealing the code/data)?
Anonymous
Do you think Redwood Research could be a place for this?
Eliezer Yudkowsky
Maybe! I haven't ruled RR out yet. But they also haven't yet done (to my own knowledge) anything demonstrating the same kind of AI-development capabilities as even GPT-3, let alone AlphaFold 2.
Eliezer Yudkowsky
I would potentially be super interested in working with Deepminders if Deepmind set up some internal partition for "Okay, accomplished Deepmind researchers who'd rather not destroy the world are allowed to form subpartitions of this partition and have their work not be published outside the subpartition let alone Deepmind in general, though maybe you have to report on it to Demis only or something." I'd be more skeptical/worried about working with OpenAI-minus-Anthropic because the notion of "open AI" continues to sound to me like "what is the worst possible strategy for making the game board as unplayable as possible while demonizing everybody who tries a strategy that could possibly lead to the survival of humane intelligence", and now a lot of the people who knew about that part have left OpenAI for elsewhere. But, sure, if they changed their name to "ClosedAI" and fired everyone who believed in the original OpenAI mission, I would update about that.
Eliezer Yudkowsky
Context that is potentially missing here and should be included: I wish that Deepmind had more internal closed research, and internally siloed research, as part of a larger wish I have about the AI field, independently of what projects I'd want to work on myself.
The present situation can be seen as one in which a common resource, the remaining timeline until AGI shows up, is incentivized to be burned by AI researchers because they have to come up with neat publications and publish them (which burns the remaining timeline) in order to earn status and higher salaries. The more they publish along the spectrum that goes {quiet internal result -> announced and demonstrated result -> paper describing how to get the announced result -> code for the result -> model for the result}, the more timeline gets burned, and the greater the internal and external prestige accruing to the researcher.
It's futile to wish for everybody to act uniformly against their incentives. But I think it would be a step forward if the relative incentive to burn the commons could be reduced; or to put it another way, the more researchers have the option to not burn the timeline commons, without them getting fired or passed up for promotion, the more that unusually intelligent researchers might perhaps decide not to do that. So I wish in general that AI research groups in general, but also Deepmind in particular, would have affordances for researchers who go looking for interesting things to not publish any resulting discoveries, at all, and still be able to earn internal points for them. I wish they had the option to do that. I wish people were allowed to not destroy the world - and still get high salaries and promotion opportunities and the ability to get corporate and ops support for playing with interesting toys; if destroying the world is prerequisite for having nice things, nearly everyone is going to contribute to destroying the world, because, like, they're not going to just not have nice things, that is not human nature for almost all humans.
When I visualize how the end of the world plays out, I think it involves an AGI system which has the ability to be cranked up by adding more computing resources to it; and I think there is an extended period where the system is not aligned enough that you can crank it up that far, without everyone dying. And it seems extremely likely that if factions on the level of, say, Facebook AI Research, start being able to deploy systems like that, then death is very automatic. If the Chinese, Russian, and French intelligence services all manage to steal a copy of the code, and China and Russia sensibly decide not to run it, and France gives it to three French corporations which I hear the French intelligence service sometimes does, then again, everybody dies. If the builders are sufficiently worried about that scenario that they push too fast too early, in fear of an arms race developing very soon if they wait, again, everybody dies.
At present we're very much waiting on a miracle for alignment to be possible at all, even if the AGI-builder successfully prevents proliferation and has 2 years in which to work. But if we get that miracle at all, it's not going to be an instant miracle. There’ll be some minimum time-expense to do whatever work is required. So any time I visualize anybody trying to even start a successful trajectory of this kind, they need to be able to get a lot of work done, without the intermediate steps of AGI work being published, or demoed at all, let alone having models released. Because if you wait until the last months when it is really really obvious that the system is going to scale to AGI, in order to start closing things, almost all the prerequisites will already be out there. Then it will only take 3 more months of work for somebody else to build AGI, and then somebody else, and then somebody else; and even if the first 3 factions manage not to crank up the dial to lethal levels, the 4th party will go for it; and the world ends by default on full automatic.
If ideas are theoretically internal to "just the company", but the company has 150 people who all know, plus everybody with the "sysadmin" title having access to the code and models, then I imagine - perhaps I am mistaken - that those ideas would (a) inevitably leak outside due to some of those 150 people having cheerful conversations over a beer with outsiders present, and (b) be copied outright by people of questionable allegiances once all hell started to visibly break loose. As with anywhere that handles really sensitive data, the concept of "need to know" has to be a thing, or else everyone (and not just in that company) ends up knowing.
So, even if I got run over by a truck tomorrow, I would still very much wish that in the world that survived me, Deepmind would have lots of penalty-free affordance internally for people to not publish things, and to work in internal partitions that didn't spread their ideas to all the rest of Deepmind. Like, actual social and corporate support for that, not just a theoretical option you'd have to burn lots of social capital and weirdness points to opt into, and then get passed up for promotion forever after.
Anonymous
What's RR?
Anonymous
It's a new alignment org, run by Nate Thomas and ~co-run by Buck Shlegeris and Bill Zito, with maybe 4-6 other technical folks so far. My take: the premise is to create an org with ML expertise and general just-do-it competence that's trying to do all the alignment experiments that something like Paul+Ajeya+Eliezer all think are obviously valuable and wish someone would do. They expect to have a website etc in a few days; the org is a couple months old in its current form.
Anonymous
How likely really is hard takeoff? Clearly, we are touching the edges of AGI with GPT and the like. But I'm not feeling this will that easily be leveraged into very quick recursive self improvement.
Eliezer Yudkowsky
Compared to the position I was arguing in the Foom Debate with Robin, reality has proved way to the further Eliezer side of Eliezer along the Eliezer-Robin spectrum. It's been very unpleasantly surprising to me how little architectural complexity is required to start producing generalizing systems, and how fast those systems scale using More Compute. The flip side of this is that I can imagine a system being scaled up to interesting human+ levels, without "recursive self-improvement" or other of the old tricks that I thought would be necessary, and argued to Robin would make fast capability gain possible. You could have fast capability gain well before anything like a FOOM started. Which in turn makes it more plausible to me that we could hang out at interesting not-superintelligent levels of AGI capability for a while before a FOOM started. It's not clear that this helps anything, but it does seem more plausible.
Anonymous
I agree reality has not been hugging the Robin kind of scenario this far.
Anonymous
Going past human level doesn't necessarily mean going "foom".
Eliezer Yudkowsky
I do think that if you get an AGI significantly past human intelligence in all respects, it would obviously tend to FOOM. I mean, I suspect that Eliezer fooms if you give an Eliezer the ability to backup, branch, and edit himself.
Anonymous
It doesn't seem to me that an AGI significantly past human intelligence necessarily tends to FOOM.
Eliezer Yudkowsky
I think in principle we could have, for example, an AGI that was just a superintelligent engineer of proteins, and of nanosystems built by nanosystems that were built by proteins, and which was corrigible enough not to want to improve itself further; and this AGI would also be dumber than a human when it came to eg psychological manipulation, because we would have asked it not to think much about that subject. I'm doubtful that you can have an AGI that's significantly above human intelligence in all respects, without it having the capability-if-it-wanted-to of looking over its own code and seeing lots of potential improvements.
Anonymous
Alright, this makes sense to me, but I don't expect an AGI to want to manipulate humans that easily (unless designed to). Maybe a bit.
Eliezer Yudkowsky
Manipulating humans is a convergent instrumental strategy if you've accurately modeled (even at quite low resolution) what humans are and what they do in the larger scheme of things.
Anonymous
Yes, but human manipulation is also the kind of thing you need to guard against with even mildly powerful systems. Strong impulses to manipulate humans, should be vetted out.
Eliezer Yudkowsky
I think that, by default, if you trained a young AGI to expect that 2+2=5 in some special contexts, and then scaled it up without further retraining, a generally superhuman version of that AGI would be very likely to 'realize' in some sense that SS0+SS0=SSSS0 was a consequence of the Peano axioms. There's a natural/convergent/coherent output of deep underlying algorithms that generate competence in some of the original domains; when those algorithms are implicitly scaled up, they seem likely to generalize better than whatever patch on those algorithms said '2 + 2 = 5'.
In the same way, suppose that you take weak domains where the AGI can't fool you, and apply some gradient descent to get the AGI to stop outputting actions of a type that humans can detect and label as 'manipulative'. And then you scale up that AGI to a superhuman domain. I predict that deep algorithms within the AGI will go through consequentialist dances, and model humans, and output human-manipulating actions that can't be detected as manipulative by the humans, in a way that seems likely to bypass whatever earlier patch was imbued by gradient descent, because I doubt that earlier patch will generalize as well as the deep algorithms. Then you don't get to retrain in the superintelligent domain after labeling as bad an output that killed you and doing a gradient descent update on that, because the bad output killed you. (This is an attempted very fast gloss on what makes alignment difficult in the first place.)
Anonymous
[i appreciate this gloss - thanks]
Anonymous
"deep algorithms within it will go through consequentialist dances, and model humans, and output human-manipulating actions that can't be detected as manipulative by the humans"
This is true if it is rewarding to manipulate humans. If the humans are on the outlook for this kind of thing, it doesn't seem that easy to me.
Going through these "consequentialist dances" to me appears to presume that mistakes that should be apparent haven't been solved at simpler levels. It seems highly unlikely to me that you would have a system that appears to follow human requests and human values, and it would suddenly switch at some powerful level. I think there will be signs beforehand. Of course, if the humans are not paying attention, they might miss it. But, say, in the current milieu, I find it plausible that they will pay enough attention.
"because I doubt that earlier patch will generalize as well as the deep algorithms"
That would depend on how "deep" your earlier patch was. Yes, if you're just doing surface patches to apparent problems, this might happen. But it seems to me that useful and intelligent systems will require deep patches (or deep designs from the start) in order to be apparently useful to humans at solving complex problems enough. This is not to say that they would be perfect. But it seems quite plausible to me that they would in most cases prevent the worst outcomes.
Eliezer Yudkowsky
"If you've got a general consequence-modeling-and-searching algorithm, it seeks out ways to manipulate humans, even if there are no past instances of a random-action-generator producing manipulative behaviors that succeeded and got reinforced by gradient descent over the random-action-generator. It invents the strategy de novo by imagining the results, even if there's no instances in memory of a strategy like that having been tried before." Agree or disagree?
Anonymous
Creating strategies de novo would of course be expected of an AGI.
I think, if the AI will "seek out ways to manipulate humans", will depend on what kind of goals the AI has been designed to pursue.
Manipulating humans is definitely an instrumentally useful kind of method for an AI, for a lot of goals. But it's also counter to a lot of the things humans would direct the AI to do -- at least at a "high level". "Manipulation", such as marketing, for lower level goals, can be very congruent with higher level goals. An AI could clearly be good at manipulating humans, while not manipulating its creators or the directives of its creators.
If you are asking me to agree that the AI will generally seek out ways to manipulate the high-level goals, then I will say "no". Because it seems to me that faults of this kind in the AI design is likely to be caught by the designers earlier. (This isn't to say that this kind of fault couldn't happen.) It seems to me that manipulation of high-level goals will be one of the most apparent kind of faults of this kind of system.
Anonymous
RE: "I'm doubtful that you can have an AGI that's significantly above human intelligence in all respects, without it having the capability-if-it-wanted-to of looking over its own code and seeing lots of potential improvements."
It seems plausible (though unlikely) to me that this would be true in practice for the AGI we build -- but also that the potential improvements it sees would be pretty marginal. This is coming from the same intuition that current learning algorithms might already be approximately optimal.
Eliezer Yudkowsky
I expect that when people are trying to stomp out convergent instrumental strategies by training at a safe dumb level of intelligence, this will not be effective at preventing convergent instrumental strategies at smart levels of intelligence; also note that at very smart levels of intelligence, "hide what you are doing" is also a convergent instrumental strategy of that substrategy.
I don't know however if I should be explaining at this point why "manipulate humans" is convergent, why "conceal that you are manipulating humans" is convergent, why you have to train in safe regimes in order to get safety in dangerous regimes (because if you try to "train" at a sufficiently unsafe level, the output of the unaligned system deceives you into labeling it incorrectly and/or kills you before you can label the outputs), or why attempts to teach corrigibility in safe regimes are unlikely to generalize well to higher levels of intelligence and unsafe regimes (qualitatively new thought processes, things being way out of training distribution, and, the hardest part to explain, corrigibility being "anti-natural" in a certain sense that makes it incredibly hard to, eg, exhibit any coherent planning behavior ("consistent utility function") which corresponds to being willing to let somebody else shut you off, without incentivizing you to actively manipulate them to shut you off).
Anonymous
My (unfinished) idea for buying time is to focus on applying AI to well-specified problems, where constraints can come primarily from the action space and additionally from process-level feedback (i.e., human feedback providers understand why actions are good before endorsing them, and reject anything weird even if it seems to work on some outcomes-based metric). This is basically a form of boxing, with application-specific boxes. I know it doesn't scale to superintelligence but I think it can potentially give us time to study and understand proto AGIs before they kill us. I'd be interested to hear devastating critiques of this that imply it isn't even worth fleshing out more and trying to pursue, if they exist.
Anonymous
(I think it's also similar to CAIS in case that's helpful.)
Eliezer Yudkowsky
There's lots of things we can do which don't solve the problem and involve us poking around with AIs having fun, while we wait for a miracle to pop out of nowhere. There's lots of things we can do with AIs which are weak enough to not be able to fool us and to not have cognitive access to any dangerous outputs, like automatically generating pictures of cats. The trouble is that nothing we can do with an AI like that (where "human feedback providers understand why actions are good before endorsing them") is powerful enough to save the world.
Eliezer Yudkowsky
In other words, if you have an aligned AGI that builds complete mature nanosystems for you, that is enough force to save the world; but that AGI needs to have been aligned by some method other than "humans inspect those outputs and vet them and their consequences as safe/aligned", because humans cannot accurately and unfoolably vet the consequences of DNA sequences for proteins, or of long bitstreams sent to protein-built nanofactories.
Anonymous
When you mention nanosystems, how much is this just a hypothetical superpower vs. something you actually expect to be achievable with AGI/superintelligence? If expected to be achievable, why?
Eliezer Yudkowsky
The case for nanosystems being possible, if anything, seems even more slam-dunk than the already extremely slam-dunk case for superintelligence, because we can set lower bounds on the power of nanosystems using far more specific and concrete calculations. See eg the first chapters of Drexler's Nanosystems, which are the first step mandatory reading for anyone who would otherwise doubt that there's plenty of room above biology and that it is possible to have artifacts the size of bacteria with much higher power densities. I have this marked down as "known lower bound" not "speculative high value", and since Nanosystems has been out since 1992 and subjected to attemptedly-skeptical scrutiny, without anything I found remotely persuasive turning up, I do not have a strong expectation that any new counterarguments will materialize.
If, after reading Nanosystems, you still don't think that a superintelligence can get to and past the Nanosystems level, I'm not quite sure what to say to you, since the models of superintelligences are much less concrete than the models of molecular nanotechnology.
I'm on record as early as 2008 as saying that I expected superintelligences to crack protein folding, some people disputed that and were all like "But how do you know that's solvable?" and then AlphaFold 2 came along and cracked the protein folding problem they'd been skeptical about, far below the level of superintelligence.
I can try to explain how I was mysteriously able to forecast this truth at a high level of confidence - not the exact level where it became possible, to be sure, but that superintelligence would be sufficient - despite this skepticism; I suppose I could point to prior hints, like even human brains being able to contribute suggestions to searches for good protein configurations; I could talk about how if evolutionary biology made proteins evolvable then there must be a lot of regularity in the folding space, and that this kind of regularity tends to be exploitable.
But of course, it's also, in a certain sense, very obvious that a superintelligence could crack protein folding, just like it was obvious years before Nanosystems that molecular nanomachines would in fact be possible and have much higher power densities than biology. I could say, "Because proteins are held together by van der Waals forces that are much weaker than covalent bonds," to point to a reason how you could realize that after just reading Engines of Creation and before Nanosystems existed, by way of explaining how one could possibly guess the result of the calculation in advance of building up the whole detailed model. But in reality, precisely because the possibility of molecular nanotechnology was already obvious to any sensible person just from reading Engines of Creation, the sort of person who wasn't convinced by Engines of Creation wasn't convinced by Nanosystems either, because they'd already demonstrated immunity to sensible arguments; an example of the general phenomenon I've elsewhere termed the Law of Continued Failure.
Similarly, the sort of person who was like "But how do you know superintelligences will be able to build nanotech?" in 2008, will probably not be persuaded by the demonstration of AlphaFold 2, because it was already clear to anyone sensible in 2008, and so anyone who can't see sensible points in 2008 probably also can't see them after they become even clearer. There are some people on the margins of sensibility who fall through and change state, but mostly people are not on the exact margins of sanity like that.
Anonymous
"If, after reading Nanosystems, you still don't think that a superintelligence can get to and past the Nanosystems level, I'm not quite sure what to say to you, since the models of superintelligences are much less concrete than the models of molecular nanotechnology."
I'm not sure if this is directed at me or the https://en.wikipedia.org/wiki/Generic_you, but I'm only expressing curiosity on this point, not skepticism :)
Anonymous
some form of "scalable oversight" is the naive extension of the initial boxing thing proposed above that claims to be the required alignment method -- basically, make the humans vetting the outputs smarter by providing them AI support for all well-specified (level-below)-vettable tasks.
Eliezer Yudkowsky
I haven't seen any plausible story, in any particular system design being proposed by the people who use terms about "scalable oversight", about how human-overseeable thoughts or human-inspected underlying systems, compound into very powerful human-non-overseeable outputs that are trustworthy. Fundamentally, the whole problem here is, "You're allowed to look at floating-point numbers and Python code, but how do you get from there to trustworthy nanosystem designs?" So saying "Well, we'll look at some thoughts we can understand, and then from out of a much bigger system will come a trustworthy output" doesn't answer the hard core at the center of the question. Saying that the humans will have AI support doesn't answer it either.
Anonymous
the kind of useful thing humans (assisted-humans) might be able to vet is reasoning/arguments/proofs/explanations. without having to generate neither the trustworthy nanosystem design nor the reasons it is trustworthy, we could still check them.
Eliezer Yudkowsky
If you have an untrustworthy general superintelligence generating English strings meant to be "reasoning/arguments/proofs/explanations" about eg a nanosystem design, then I would not only expect the superintelligence to be able to fool humans in the sense of arguing for things that were not true in a way that fooled the humans, I'd expect the superintelligence to be able to covertly directly hack the humans in ways that I wouldn't understand even after having been told what happened. So you must have some prior belief about the superintelligence being aligned before you dared to look at the arguments. How did you get that prior belief?
Anonymous
I think I'm not starting with a general superintelligence here to get the trustworthy nanodesigns. I'm trying to build the trustworthy nanosystems "the hard way", i.e., if we did it without ever building AIs, and then speed that up using AI for automation of things we know how to vet (including recursively). Is a crux here that you think nanosystem design requires superintelligence?
(tangent: I think this approach works even if you accidentally built a more-general or more-intelligent than necessary foundation model as long as you're only using it in boxes it can't outsmart. The better-specified the tasks you automate are, the easier it is to secure the boxes.)
Eliezer Yudkowsky
I think that China ends the world using code they stole from Deepmind that did things the easy way, and that happens 50 years of natural R&D time before you can do the equivalent of "strapping mechanical aids to a horse instead of building a car from scratch".
I also think that the speedup step in "iterated amplification and distillation" will introduce places where the fast distilled outputs of slow sequences are not true to the original slow sequences, because gradient descent is not perfect and won't be perfect and it's not clear we'll get any paradigm besides gradient descent for doing a step like that.
Anonymous
How do you feel about the safety community as a whole and the growth we've seen over the past few years?
Eliezer Yudkowsky
Very grim. I think that almost everybody is bouncing off the real hard problems at the center and doing work that is predictably not going to be useful at the superintelligent level, nor does it teach me anything I could not have said in advance of the paper being written. People like to do projects that they know will succeed and will result in a publishable paper, and that rules out all real research at step 1 of the social process.
Paul Christiano is trying to have real foundational ideas, and they're all wrong, but he's one of the few people trying to have foundational ideas at all; if we had another 10 of him, something might go right.
Chris Olah is going to get far too little done far too late. We're going to be facing down an unalignable AGI and the current state of transparency is going to be "well look at this interesting visualized pattern in the attention of the key-value matrices in layer 47" when what we need to know is "okay but was the AGI plotting to kill us or not”. But Chris Olah is still trying to do work that is on a pathway to anything important at all, which makes him exceptional in the field.
Stuart Armstrong did some good work on further formalizing the shutdown problem, an example case in point of why corrigibility is hard, which so far as I know is still resisting all attempts at solution.
Various people who work or worked for MIRI came up with some actually-useful notions here and there, like Jessica Taylor's expected utility quantilization.
And then there is, so far as I can tell, a vast desert full of work that seems to me to be mostly fake or pointless or predictable.
It is very, very clear that at present rates of progress, adding that level of alignment capability as grown over the next N years, to the AGI capability that arrives after N years, results in everybody dying very quickly. Throwing more money at this problem does not obviously help because it just produces more low-quality work.
Anonymous
"doing work that is predictably not going to be really useful at the superintelligent level, nor does it teach me anything I could not have said in advance of the paper being written"
I think you're underestimating the value of solving small problems. Big problems are solved by solving many small problems. (I do agree that many academic papers do not represent much progress, however.)
Eliezer Yudkowsky
By default, I suspect you have longer timelines and a smaller estimate of total alignment difficulty, not that I put less value than you on the incremental power of solving small problems over decades. I think we're going to be staring down the gun of a completely inscrutable model that would kill us all if turned up further, with no idea how to read what goes on inside its head, and no way to train it on humanly scrutable and safe and humanly-labelable domains in a way that seems like it would align the superintelligent version, while standing on top of a whole bunch of papers about "small problems" that never got past “small problems”.
Anonymous
"I think we're going to be staring down the gun of a completely inscrutable model that would kill us all if turned up further, with no idea how to read what goes on inside its head, and no way to train it on humanly scrutable and safe and humanly-labelable domains in a way that seems like it would align the superintelligent version"
This scenario seems possible to me, but not very plausible. GPT is not going to "kill us all" if turned up further. No amount of computing power (at least before AGI) would cause it to. I think this is apparent, without knowing exactly what's going on inside GPT. This isn't to say that there aren't AI systems that wouldn't. But what kind of system would? (A GPT combined with sensory capabilities at the level of Tesla's self-driving AI? That still seems too limited.)
Eliezer Yudkowsky
Alpha Zero scales with more computing power, I think AlphaFold 2 scales with more computing power, Mu Zero scales with more computing power. Precisely because GPT-3 doesn't scale, I'd expect an AGI to look more like Mu Zero and particularly with respect to the fact that it has some way of scaling.
Steve Omohundro
Eliezer, thanks for doing this! I just now read through the discussion and found it valuable. I agree with most of your specific points but I seem to be much more optimistic than you about a positive outcome. I'd like to try to understand why that is. I see mathematical proof as the most powerful tool for constraining intelligent systems and I see a pretty clear safe progression using that for the technical side (the social side probably will require additional strategies). Here are some of my intuitions underlying that approach, I wonder if you could identify any that you disagree with. I'm fine with your using my name (Steve Omohundro) in any discussion of these.
1) Nobody powerful wants to create unsafe AI but they do want to take advantage of AI capabilities.
2) None of the concrete well-specified valuable AI capabilities require unsafe behavior
3) Current simple logical systems are capable of formalizing every relevant system involved (eg. MetaMath http://us.metamath.org/index.html currently formalizes roughly an undergraduate math degree and includes everything needed for modeling the laws of physics, computer hardware, computer languages, formal systems, machine learning algorithms, etc.)
4) Mathematical proof is cheap to mechanically check (eg. MetaMath has a 500 line Python verifier which can rapidly check all of its 38K theorems)
5) GPT-F is a fairly early-stage transformer-based theorem prover and can already prove 56% of the MetaMath theorems. Similar systems are likely to soon be able to rapidly prove all simple true theorems (eg. that human mathematicians can prove in a day).
6) We can define provable limits on the behavior of AI systems that we are confident prevent dangerous behavior and yet still enable a wide range of useful behavior.
7) We can build automated checkers for these provable safe-AI limits.
8) We can build (and eventually mandate) powerful AI hardware that first verifies proven safety constraints before executing AI software
9) For example, AI smart compilation of programs can be formalized and doesn't require unsafe operations
10) For example, AI design of proteins to implement desired functions can be formalized and doesn't require unsafe operations
11) For example, AI design of nanosystems to achieve desired functions can be formalized and doesn't require unsafe operations.
12) For example, the behavior of designed nanosystems can be similarly constrained to only proven safe behaviors
13) And so on through the litany of early stage valuable uses for advanced AI.
14) I don't see any fundamental obstructions to any of these. Getting social acceptance and deployment is another issue!
Best, Steve
Eliezer Yudkowsky
Steve, are you visualizing AGI that gets developed 70 years from now under absolutely different paradigms than modern ML? I don't see being able to take anything remotely like, say, Mu Zero, and being able to prove any theorem about it which implies anything like corrigibility or the system not internally trying to harm humans. Anything in which enormous inscrutable floating-point vectors is a key component, seems like something where it would be very hard to prove any theorems about the treatment of those enormous inscrutable vectors that would correspond in the outside world to the AI not killing everybody.
Even if we somehow managed to get structures far more legible than giant vectors of floats, using some AI paradigm very different from the current one, it still seems like huge key pillars of the system would rely on non-fully-formal reasoning; even if the AI has something that you can point to as a utility function and even if that utility function's representation is made out of programmer-meaningful elements instead of giant vectors of floats, we'd still be relying on much shakier reasoning at the point where we claimed that this utility function meant something in an intuitive human-desired sense, say. And if that utility function is learned from a dataset and decoded only afterwards by the operators, that sounds even scarier. And if instead you're learning a giant inscrutable vector of floats from a dataset, gulp.
You seem to be visualizing that we prove a theorem and then get a theorem-like level of assurance that the system is safe. What kind of theorem? What the heck would it say?
I agree that it seems plausible that the good cognitive operations we want do not in principle require performing bad cognitive operations; the trouble, from my perspective, is that generalizing structures that do lots of good cognitive operations will automatically produce bad cognitive operations, especially when we dump more compute into them; "you can't bring the coffee if you're dead".
So it takes a more complicated system and some feat of insight I don't presently possess, to "just" do the good cognitions, instead of doing all the cognitions that result from decompressing the thing that compressed the cognitions in the dataset - even if that original dataset only contained cognitions that looked good to us, even if that dataset actually was just correctly labeled data about safe actions inside a slightly dangerous domain. Humans do a lot of stuff besides maximizing inclusive genetic fitness, optimizing purely on outcomes labeled by a simple loss function doesn’t get you an internal optimizer that pursues only that loss function, etc.
Anonymous
Steve's intuitions sound to me like they're pointing at the "well-specified problems" idea from an earlier thread. Essentially, only use AI in domains where unsafe actions are impossible by construction. Is this too strong a restatement of your intuitions Steve?
Steve Omohundro
Thanks for your perspective! Those sound more like social concerns than technical ones, though. I totally agree that today's AI culture is very "sloppy" and that the currently popular representations, learning algorithms, data sources, etc. aren't oriented around precise formal specification or provably guaranteed constraints. I'd love any thoughts about ways to help shift that culture toward precise and safe approaches! Technically there is no problem getting provable constraints on floating point computations, etc. The work often goes under the label "Interval Computation". It's not even very expensive, typically just a factor of 2 worse than "sloppy" computations. For some reason those approaches have tended to be more popular in Europe than in the US. Here are a couple lists of references: http://www.cs.utep.edu/interval-comp/ https://www.mat.univie.ac.at/~neum/interval.html
I see today's dominant AI approach of mapping everything to large networks ReLU units running on hardware designed for dense matrix multiplication, trained with gradient descent on big noisy data sets as a very temporary state of affairs. I fully agree that it would be uncontrolled and dangerous scaled up in its current form! But it's really terrible in every aspect except that it makes it easy for machine learning practitioners to quickly slap something together which will actually sort of work sometimes. With all the work on AutoML, NAS, and the formal methods advances I'm hoping we leave this "sloppy" paradigm pretty quickly. Today's neural networks are terribly inefficient for inference: most weights are irrelevant for most inputs and yet current methods do computational work on each. I developed many algorithms and data structures to avoid that waste years ago (eg. "bumptrees" https://steveomohundro.com/scientific-contributions/)
They're also pretty terrible for learning since most weights don't need to be updated for most training examples and yet they are. Google and others are using Mixture-of-Experts to avoid some of that cost: https://arxiv.org/abs/1701.06538
Matrix multiply is a pretty inefficient primitive and alternatives are being explored: https://arxiv.org/abs/2106.10860
Today's reinforcement learning is slow and uncontrolled, etc. All this ridiculous computational and learning waste could be eliminated with precise formal approaches which measure and optimize it precisely. I'm hopeful that that improvement in computational and learning performance may drive the shift to better controlled representations.
I see theorem proving as hugely valuable for safety in that we can easily precisely specify many important tasks and get guarantees about the behavior of the system. I'm hopeful that we will also be able to apply them to the full AGI story and encode human values, etc., but I don't think we want to bank on that at this stage. Hence, I proposed the "Safe-AI Scaffolding Strategy" where we never deploy a system without proven constraints on its behavior that give us high confidence of safety. We start extra conservative and disallow behavior that might eventually be determined to be safe. At every stage we maintain very high confidence of safety. Fast, automated theorem checking enables us to build computational and robotic infrastructure which only executes software with such proofs.
And, yes, I'm totally with you on needing to avoid the "basic AI drives"! I think we have to start in a phase where AI systems are not allowed to run rampant as uncontrolled optimizing agents! It's easy to see how to constrain limited programs (eg. theorem provers, program compilers or protein designers) to stay on particular hardware and only communicate externally in precisely constrained ways. It's similarly easy to define constrained robot behaviors (eg. for self-driving cars, etc.) The dicey area is that unconstrained agentic edge. I think we want to stay well away from that until we're very sure we know what we're doing! My optimism stems from the belief that many of the socially important things we need AI for won't require anything near that unconstrained edge. But it's tempered by the need to get the safe infrastructure into place before dangerous AIs are created.
Anonymous
As far as I know, all the work on "verifying floating-point computations" currently is way too low-level -- the specifications that are proved about the computations don't say anything about what the computations mean or are about, beyond the very local execution of some algorithm. Execution of algorithms in the real world can have very far-reaching effects that aren't modelled by their specifications.
Eliezer Yudkowsky
Yeah, what they said. How do you get from proving things about error bounds on matrix multiplications of inscrutable floating-point numbers, to saying anything about what a mind is trying to do, or not trying to do, in the external world?
Steve Omohundro
Ultimately we need to constrain behavior. You might want to ensure your robot butler won't leave the premises. To do that using formal methods, you need to have a semantic representation of the location of the robot, your premise's spatial extent, etc. It's pretty easy to formally represent that kind of physical information (it's just a more careful version of what engineers do anyway). You also have a formal model of the computational hardware and software and the program running the system.
For finite systems, any true property has a proof which can be mechanically checked but the size of that proof might be large and it might be hard to find. So we need to use encodings and properties which mesh well with the safety semantics we care about.
Formal proofs of properties of programs has progressed to where a bunch of cryptographic, compilation, and other systems can be specified and formalized. Why it's taken this long, I have no idea. The creator of any system has an argument as to why its behavior does what they think it will and why it won't do bad or dangerous things. The formalization of those arguments should be one direct short step.
Experience with formalizing mathematician's informal arguments suggest that the formal proofs are maybe 5 times longer than the informal argument. Systems with learning and statistical inference add more challenges but nothing that seems in-principal all that difficult. I'm still not completely sure how to constrain the use of language, however. I see inside of Facebook all sorts of problems due to inability to constrain language systems (eg. they just had a huge issue where a system labeled a video with a racist term). The interface between natural language semantics and formal semantics and how we deal with that for safety is something I've been thinking a lot about recently.
Steve Omohundro
Here's a nice 3 hour long tutorial about "probabilistic circuits" which is a representation of probability distributions, learning, Bayesian inference, etc. which has much better properties than most of the standard representations used in statistics, machine learning, neural nets, etc.: https://www.youtube.com/watch?v=2RAG5-L9R70 It looks especially amenable to interpretability, formal specification, and proofs of properties.
Eliezer Yudkowsky
You're preaching to the choir there, but even if we were working with more strongly typed epistemic representations that had been inferred by some unexpected innovation of machine learning, automatic inference of those representations would lead them to be uncommented and not well-matched with human compressions of reality, nor would they match exactly against reality, which would make it very hard for any theorem about "we are optimizing against this huge uncommented machine-learned epistemic representation, to steer outcomes inside this huge machine-learned goal specification" to guarantee safety in outside reality; especially in the face of how corrigibility is unnatural and runs counter to convergence and indeed coherence; especially if we're trying to train on domains where unaligned cognition is safe, and generalize to regimes in which unaligned cognition is not safe. Even in this case, we are not nearly out of the woods, because what we can prove has a great type-gap with that which we want to ensure is true. You can't handwave the problem of crossing that gap even if it's a solvable problem.
And that whole scenario would require some major total shift in ML paradigms.
Right now the epistemic representations are giant inscrutable vectors of floating-point numbers, and so are all the other subsystems and representations, more or less.
Prove whatever you like about that Tensorflow problem; it will make no difference to whether the AI kills you. The properties that can be proven just aren't related to safety, no matter how many times you prove an error bound on the floating-point multiplications. It wasn't floating-point error that was going to kill you in the first place.