I agree with this statement denotatively, and my own interests/work have generally been "driven by open-ended curiosity and a drive to uncover deep truths", but isn't this kind of motivation also what got humanity into its current mess? In other words, wasn't the main driver of AI progress this kind of curiosity (until perhaps the recent few years when it has been driven more by commercial/monetary/power incentives)?
Interestingly, I was just having a conversation with Critch about this. My contention was that, in the first few decades of the field, AI researchers were actually trying to understand cognition. The rise of deep learning (and especially the kind of deep learning driven by massive scaling) can be seen as the field putting that quest on hold in order to optimize for more legible metrics.
I don't think you should find this a fully satisfactory answer, because it's easy to "retrodict" ways that my theory was correct. But that's true of all explanations of what makes the world good at a very abstract level, including your own answer of metaphilosophical competence. (Also, we can perhaps cash my claim out in predictions, like: was a significant barrier to more researchers working on deep learning the criticism that it didn't actually provide good explanations of or insight into cognition? Without having looked it up, I suspect so.)
consistently good strategy requires a high amount of consequentialist reasoning
I don't think that's true. However I do think it requires deep curiosity about what good strategy is and how it works. It's not a coincidence that my own research on a theory of coalitional agency was in significant part inspired by strategic failures of EA and AI safety (with this post being one of the earliest building blocks I laid down). I also suspect that the full theory of coalitional agency will in fact explain how to do metaphilosophy correct, because doing good metaphilosophy is ultimately a cognitive process and can therefore be characterized by a sufficiently good theory of cognition.
Again, I don't expect you to fully believe me. But what I most want to read from you right now is an in-depth account of which the things in the world have gone or are going most right, and the ways in which you think metaphilosophical competence or consequentialist reasoning contributed to them. Without that, it's hard to trust metaphilosophy or even know what it is (though I think you've given a sketch of this in a previous reply to me at some point).
I should also try to write up the same thing, but about how virtues contributed to good things. And maybe also science, insofar as I'm trying to defend doing more science (of cognition and intelligence) in order to help fix risks caused by previous scientific progress.
In trying to reply to this comment I identified four "waves" of AI safety, and lists of the central people in each wave. Since this is socially complicated I'll only share the full list of the first wave here, and please note that this is all based on fuzzy intuitions gained via gossip and other unreliable sources.
The first wave I’ll call the “founders”; I think of them as the people who set up the early institutions and memeplexes of AI safety before around 2015. My list:
- Eliezer Yudkowsky
- Michael Vassar
- Anna Salamon
- Carl Schulman
- Scott Alexander
- Holden Karnofsky
- Nick Bostrom
- Robin Hanson
- Wei Dai
- Shane Legg
- Geoff Anders
The second wave I’ll call the “old guard”; those were the people who joined or supported the founders before around 2015. A few central examples include Paul Christiano, Chris Olah, Andrew Critch and Oliver Habryka.
Around 2014/2015 AI safety became significantly more professionalized and growth-oriented. Bostrom published Superintelligence, the Puerto Rico conference happened, OpenAI was founded, DeepMind started a safety team (though I don't recall exactly when), and EA started seriously pushing people towards AI safety. I’ll call the people who entered the field from then until around 2020 "safety scalers" (though I'm open to better names). A few central examples include Miles Brundage, Beth Barnes, John Wentworth, Rohin Shah, Dan Hendrycks and myself.
And then there’s the “newcomers” who joined in the last 5-ish years. I have a worse mental map of these people, but some who I respect are Leo Gao, Sahil, Marius Hobbhahn and Jesse Hoogland.
In this comment I expressed concern that my generation (by which I mean the "safety scalers") have kinda given up on solving alignment. But another higher-level concern is: are people from these last two waves the kinds of people who would have been capable of founding AI safety in the first place? And if not, where are those people now? Of course there's some difference in the skills required for founding a field vs pushing the field forward, but to a surprising extent I keep finding that the people who I have the most insightful conversations with are the ones who were around from the very beginning. E.g. I think Vassar is the single person doing the best thinking about the lessons we can learn about failures of AI safety over the last decade (though he's hard to interface with), and Yudkowsky is still the single person who's most able to push the Overton window towards taking alignment seriously (even though in principle many other people could have written (less doomy versions of) his Time op-ed or his recent book), Scott is still the single best blogger in the space, and so on.
Relatedly, when I talk to someone who's exceptionally thoughtful about politics (and particularly the psychological aspects of politics), a disturbingly large proportion of the time it turns out that worked at (or were somehow associated with) Leverage. This is really weird to me. Maybe I just have Leverage-aligned tastes/networks, but even so, it's a very striking effect. (Also, how come there's no young Moldbug?)
Assuming that I'm gesturing at something real, what are some possible explanations?
- There was a unique historical period during which blogging culture was coming online, during which a bunch of ideas and people could come together. This is hard for anyone to replicate now, and so they can't "level up" in the same way.
- This is just what it's like to be "inside" a paradigm in general. Founding it seems like a really impressive achievement that nobody can match by pushing it forward incrementally; and the founders seem brilliant because they can operate the paradigm better than anyone else. Eventually the issues with this paradigm will pile up enough that someone else can found a new paradigm.
- The "takeover" of AI safety by EA changed the kinds of people who were attracted to it. The kinds of people who could found a movement have gone elsewhere (but where?)
- The world produces fewer of the kinds of people who are capable of founding movements like this now (but why?)
This is all only a rough gesture at the phenomenon, and you should be wary that I'm just being pessimistic rather than identifying something important. Also it's a hard topic to talk about clearly because it's loaded with a bunch of social baggage. But I do feel pretty confused and want to figure this stuff out.
1
1Yepp, makes sense, and it's a good reminder for me to be careful about how I use these terms.
One clarification I'd make to your original comment though is that I don't endorse "you have to deeply understand intelligence from first principles else everyone dies". My position is closer to "you have to be trying to do something principled in order for your contribution to be robustly positive". Relatedly, agent foundations and mech-interp are approximately the only two parts of AI safety that seem robustly good to me—with a bunch of other stuff like RLHF, or evals, or (almost all) governance work, I feel pretty confused about whether they're good or bad or basically just wash out even in expectation.
This is still consistent with risk potentially being reduced by what I call engineering-type work, it's just that IMO that involves us "getting lucky" in an important way which I prefer we not rely on. (And trying to get lucky isn't a neutral action—engineering-type work can also easily have harmful effects.)
I agree that there are some ways in which my comment did not meet the standard that I was holding your post to. I think this is defensible because I hold things to higher standards when they're more prominent (e.g. posts versus shortforms or comments), and also because I hold things to higher standards when they're making stronger headline claims. In my case, my headline claim was "I feel confused". If I had instead made the headline claim "Mikhail is untrustworthy", then I think it would have been very reasonable for you to be angry at this.
I think that my criticism contains some moves that I wish your criticism had more of. In particular, I set a standard for what I wanted from your criticism:
I think of good critiques as trying to identify standards of behavior that should be met, and comparing people or organizations to those standards, rather than just throwing accusations at them.
and provide a central example of you not meeting this standard:
"Anthropic is untrustworthy" is an extremely low-resolution claim
I also primarily focused on drawing conclusions about the post itself (e.g. "My overall sense is that people should think of the post roughly the way they think of a compilation of links") and relegate the psychologizing to the end. I accept that you would have preferred that I skip it entirely, but it's a part of "figuring out what's up with Mikhail", which is an epistemic move that I endorse people doing after they've laid out a disagreement (but not as a primary approach to that disagreement).
Some examples of statements where it's pretty hard for me to know how much the statements straightforwardly follow from the evidence you have, vs being things that you've inferred because they seem plausible to you:
- Jack Clark would have known this.
- Anthropic's somewhat less problematic behavior is fully explained by having to maintain a good image internally.
- Anthropic is now basically just as focused on commercializing its products.
- Anthropic's talent is a core pitch to investors: they've claimed they can do what OpenAI can for 10x cheaper.
- It seems likely that the policy positions that Anthropic took early on were related to these incentives.
- Anthropic's mission is not really compatible with the idea of pausing, even if evidence suggests it's a good idea to.
If we zoom in on #3, for instance: there's a sense in which it's superficially plausible because both OpenAI and Anthropic have products. But maybe Anthropic and OpenAI differ greatly on, say, the ratio of headcount, or the ratio of executives' time, or the amount of compute, or the internal prestige allocated to commercialization vs other things (like alignment research). If so, then it's not really accurate to say that they're just as focused on commercialization. But I don't know if knowledge of these kinds of considerations informed your claim, or if you're only making the superficially plausible version of the claim.
To be clear, in general I don't expect people to apply this level of care for most LW posts. But when it comes to accusations of untrustworthiness (and similar kinds of accountability mechanisms) I think it's really valuable to be able to create common knowledge of the specific details of misbehavior. Hence I would have much preferred this post to focus on a smaller set of claims that you can solidly substantiate, and then only secondarily try to discuss what inferences we should draw from those. Whereas I think that the kinds of criticism you make here mostly create a miasma of distrust between Anthropic and LessWrong, without adding much common knowledge of the form "Anthropic violated clear and desirable standard X" for the set of good-faith AI safety actors.
I also realize that by holding this standard I'm making criticism more costly, because now you have the stress of trying to justify yourself to me. I would have tried harder to mitigate that cost if I hadn't noticed this pattern of not-very-careful criticism from you. I do sympathize with your frustration that people seem to be naively trusting Anthropic and ignoring various examples of shady behavior. However I also think people outside labs really underestimate how many balls lab leaders have up in the air at once, and how easy it is to screw up a few of them even if you're broadly trustworthy. I don't know how to balance these considerations, especially because the community as a whole has historically erred on the side of the former mistake. I'd appreciate people helping me think through this, e.g. by working through models of how applying pressure to bureaucratic organizations goes successfully, in light of the ways that such organizations become untrustworthy (building on Zvi's moral mazes sequence for instance).
1I regret using the word "marginalist", it's a bit too confusing. But I do have a pretty high bar for what counts as "ambitious" in the political domain—it involves not just getting the system to do something, but rather trying to change the system itself. Cummings and Thiel are central examples (Geoff Anders maybe also was aiming in that direction at one point).
I think me using the word "marginalist" was probably a mistake, because it conflates two distinct things that I'm skeptical about:
- People no longer trying to make models more aligned (but e.g. trying to do work that primarily cashes out in political outcomes). This is what I mean by "not even aspiring to be the type of thing that could solve alignment".
- People using engineering-type approaches (rather than science-type approaches) to try to make models more aligned.
The list I gave above was of things that fall into category 1, whereas (almost?) all of the things you named fall into category 2. What I want more of is category 3: science-type approaches. One indicator that something is a science-type approach is that it could potentially help us understand something fundamental about intelligence; another is that, if it works, we'll know in advance (I used to not care about this, but have changed my mind).
I think there are versions of most of the things you named that could be in category 3, but people mostly seem to be doing category-2 versions of them, in significant part because of the sort of EA-style reasoning that I was criticizing from Neel's original post.
When I wrote "pragmatic interpretability feels like another step in that direction" I meant something like: ambitious interpretability was trying to do 3, and pragmatic interpretability seems like it's nominally trying to do 2, and may in practice end up being mostly 1. For example, "Stop models acting differently when tested" could be a part of an engineering-type pipeline for fixing misalignments in models, but could also end up drifting towards "help us get better evidence to convince politicians and lab leaders of things". However, I'm not claiming that pragmatic interpretability is a central example of "not even aspiring to be the type of thing that could solve alignment". Apologies for the bad phrasings.
1Thanks for writing this up. While I don't have much context on what specifically has gone well or badly for your team, I do feel pretty skeptical about the types of arguments you give at several points: in particular focusing on theories of change, having the most impact, comparative advantage, work paying off in 10 years, etc. I expect that this kind of reasoning itself steers people away from making important scientific contributions, which are often driven by open-ended curiosity and a drive to uncover deep truths.
(A provocative version of this claim: for the most important breakthroughs, it's nearly impossible to identify a theory of change for them in advance. Imagine Newton or Darwin trying to predict how understanding mechanics/evolution would change the world. Now imagine them trying to do that before they had even invented the theory! And finally imagine if they only considered plans that they thought would work within 10 years, and the sense of scarcity and tension that would give rise to.)
The rest of my comment isn't directly about this post, but close enough that this seems like a reasonable place to put it. EDIT: to be more clear: the rest of this comment is not primarily about Neel or "pragmatic interpretability", it's about parts of the field that I consider to be significantly less relevant to "solving alignment" than that (though work that's nominally on pragmatic interpretability could also fall into the same failure modes). I clarify my position further in this comment; thanks Rohin for the pushback.
I get the sense that there was a "generation" of AI safety researchers who have ended up with a very marginalist mindset about AI safety. Some examples:
- the evals that Beth Barnes (and maybe Dan Hendrycks?) are focusing on
- the scenarios that Daniel Kokotajlo is focusing on
- the models of misalignment that Evan Hubinger is focusing on
- the forecasting that the OpenPhil worldview investigations team focused on
- scary demos
- safety cases
- policy approaches like SB-1047
In other words, whole swathes of the field are not even aspiring to be the type of thing that could solve misalignment. In the terminology of this excellent post, they are all trying to attack a category I problem not a category II problem. Sometimes it feels like almost the entire field EDIT: most of the field is Goodharting on the subgoal of "write a really persuasive memo to send to politicians". Pragmatic interpretability feels like another step in that direction (EDIT: but still significantly more principled than the things I listed above).
This is all related to something Buck recently wrote: "I spend most of my time thinking about relatively cheap interventions that AI companies could implement to reduce risk assuming a low budget, and about how to cause AI companies to marginally increase that budget". I'm sure Buck has thought a lot about his strategy here, and I'm sure that you've thought a lot about your strategy as laid out in this post, and so on. But a part of me is sitting here thinking: man, everyone sure seems to have given up. (And yes, I know it doesn't feel like giving up from the inside, but from my perspective that's part of the problem.)
Now, a lot of the "old guard" seems to have given up too. But they at least know what they've given up on. There was an ideal of fundamental scientific progress that MIRI and Paul and a few others were striving towards; they knew at least what it would feel like (if not what it would look like) to actually make progress towards understanding intelligence. Eliezer and various others no longer think that's plausible. I disagree. But aside from the object-level disagreement, I really want people to be aware that this is a thing that's at least possible in principle to aim for, lest the next generation of the AI safety community ends up giving up on it before they even know what they've given up on.
(I'll leave for another comment/post the question of what went wrong in my generation. The "types of arguments" I objected to above all seem quite EA-flavored, and so one salient possibility is just that the increasing prominence of EA steered my generation away from the type of mentality in which it's even possible to aim towards scientific breakthroughs. But even if that's one part of the story, I expect it's more complicated than that.)
3Thinking more about the cellular automaton stuff: okay, so Game of Life is Turing complete. But the question is whether we can pin down properties that GoL has that Turing machines don't have.
I have a vague recollection that parallel Turing Machines are a thing, but this paper claims that the actual formalisms are disappointing. One nice thing about Game of Life is that the way that different programs interact internally (via game of life physics) is also how they interact with each other. Whereas any multi-tape Turing Machine (even one with clever rules about how to integrate inputs from multiple tapes) wouldn't have that property.
I feel like I'm not getting beyond the original idea that Game of Life could have adversarial robustness in a way that Turing Machines don't. But it feels like you'd need to demonstrate this with some construction that's actually adversarially robust, which seems difficult.
I expect it's not worth our time to dig too deep into whose position is more common here. But I think that a lot of people on LW have high P(doom) in significant part because they share my intuition that marginalist approaches don't reliably work. I do agree that my combination of "marginalist approaches don't reliably improve things" and "P(doom) is <50%" is a rare one, but I was only making the former point above (and people upvoted it accordingly), so it feels a bit misleading to focus on the rareness of the overall position.
(Interestingly, while the combination I describe above is a rare one, the converse is also rare—Daniel Kokotajlo is the only person who comes to mind who disagrees with me on both of these propositions simultaneously. Note that he doesn't characterize his current work as marginalist, but even aside from that question I think this characterization of him is accurate—e.g. he has talked to me about how changing the CEO of a given AI lab could swing his P(doom) by double digit percentage points.)