LessWrong Team
I have signed no contracts or agreements whose existence I cannot mention.
LessWrong Team
I have signed no contracts or agreements whose existence I cannot mention.
Yes, true, fixed, thanks!
Dog: “Oh ho ho, I’ve played imaginary fetch before, don’t you worry.”
My regular policy is to not frontpage newsletters, however I frontpaged this one as it’s the first in the series and I think it’s neat for more people to know this is a series Zvi intends to write.
Curated! I think it’s generally great when people explain what they’re doing and why in way legibile to those not working on it. Great because it let’s others potentially get involved, build on it, expose flaws or omissions, etc. This one seems particularly clear and well written. While I haven’t read all of the research, nor am I particularly qualified to comment on it, I like the idea of a principled/systematic approach behind, in comparison to a lot of work that isn’t coming on a deeper, bigger, framework.
(While I’m here though, I’ll add a link to Dmitry Vaintrob’s comment that Jacob Hilton described as “best critique of ARC’s research agenda that I have read since we started working on heuristic explanations”. Eliciting such feedback is the kind of good thing that comes out of up writing agendas – it’s possible or likely Dmitry was already tracking the work and already had these critiques, but a post like this seems like a good way to propagate them and have a public back and forth.)
Roughly speaking, if the scalability of an algorithm depends on unknown empirical contingencies (such as how advanced AI systems generalize), then we try to make worst-case assumptions instead of attempting to extrapolate from today’s systems.
I like this attitude. The human standard, I think often in alignment work too, is to argue why one’s plan will work and find stories for that, and adopting the methodology of the opposite, especially given the unknowns, is much needed in alignment work.
Overall, this is neat. Kudos to Jacob (and rest of the team) for taking the time to put this all together. Doesn’t seem all that quick to write, and I think it’d be easy to think they ought to not take time out off from further object-level research to write it. Thanks!
Thanks! Fixed
Curated. I really like that even though LessWrong is 1.5 decades old now and has Bayesianism assumed as background paradigm while people discuss everything else, nonetheless we can have good exploration of our fundamental epistemological beliefs.
The descriptions of unsolved problems, or at least incompleteness of Bayesianism strikes me as technically correct. Like others, I’m not convinced of Richard’s favored approach, but it’s interesting. In practice, I don’t think these problems undermine the use of Bayesianism in typical LessWrong thought. For example, I never thought of credences being applied to “propositions” rigorously, and more like “hypotheses” or possibilities for how things are that could be framed as models already too. Context-dependent terms like “large” or quantities without explicit tolerances like “500ft” are the kind of things that you you taboo or reduce if necessary either for your own reasoning or a bet
That said, I think the claims about mistakes and downstream consequences of the way people do Bayesianism are interesting. I’m reading a claim here I don’t recall seeing. Although we already knew that bounded reasons aren’t logically omniscient, Richard is adding a claim (if I’m understanding correctly) that this means that no matter how much strong evidence we technically have, we shouldn’t have really high confidence in any domain that requires heavy of processing that evidence, because we’re not that good at processing. I do think that leaves us with a question of judging when there’s enough evidence to be conclusive without complicated processing or not.
Something I might like a bit more factored out is the rigorous gold-standard epistemological framework and the manner in which we apply our epistemology day to day.
I fear this curation notice would be better if I’d read all the cited sources on critical rationalism, Knightian uncertainty, etc., and I’ve added them to my reading list. All in all, kudos for putting some attention on the fundamentals.
Welcome! Sounds like you’re on the one hand at start of a significant journey but also you’ve come a long distance already. I hope you find much helpful stuff on LessWrong.
I hadn’t heard of Daniel Schmachtenberger, but I’m glad to have learend of him and his works. Thanks.
The actual reason why we lied in the second message was “we were in a rush and forgot.”
My recollection is we sent the same message to the majority group because:
Treating it different would require special-casing it and that would have taken more effort.
If selectors of different virtues had received a different messages, we wouldn’t be able to have a properly compared their behavior.
[At least in my mind], this was a game/test and when playing games you lie to people in the context of the game to make things work. Alternatively, it’s like how scientific experimenters mislead subjects for the sake of the study.
Added!
Added!
Money helps. I could probably buy a lot of dignity points for a billion dollars. With a trillion variance definitely goes up because you could try crazy stuff and could backfire. (I mean true for a billion too). But EV of such a world is better.
I don’t think there’s anything that’s as simple as writing a check though.
US Congress gives money to specific things. I do not have a specific plan for a trillion dollars.
I’d bet against Terrance Tao being some kind of amazing breakthrough researcher who changes the playing field.
Your access should be activated within 5-10 minutes. Look for the button in the bottom right of the screen.
Not an original observation but yeah, separate from whether it’s desirable, I think we need to be planning for it.
Just thinking through simple stuff for myself, very rough, posting in the spirit of quick takes
At present, we are making progress on the Technical Alignment Problem[2] and like probably could solve it within 50 years.
Humanity is on track to build ~lethal superpowerful AI in more like 5-15 years.
Working on technical alignment (direct or meta) only matters if we can speed up overall progress by 10x (or some lesser factor if AI capabilities is delayed from its current trajectory). Improvements of 2x are not likely to get us to an adequate technical solution in time.
Working on slowing things down is only helpful if it results in delays of decades.
Shorter delays are good in so far as they give you time to buy further delays.
There is technical research that is useful for persuading people to slow down (and maybe also solving alignment, maybe not). This includes anything that demonstrates scary capabilities or harmful proclivities, e.g. a bunch of mech interp stuff, all the evals stuff.
AI is in fact super powerful and people who perceive there being value to be had aren’t entirely wrong[3]. This results in a very strong motivation to pursue AI and resist efforts to be stopped
These motivations apply to both businesses and governments.
People are also developing stances on AI along ideological, political, and tribal lines, e.g. being anti-regulation. This generates strong motivations for AI topics even separate from immediate power/value to be gained.
Efforts to agentically slow down the development of AI capabilities are going to be matched by agentic efforts to resist those efforts and push in the opposite direction.
Efforts to convince people that we ought to slow down will be matched by people arguing that we must speed up.
Efforts to regulate will be matched by efforts to block regulation. There will be efforts to repeal or circumvent any passed regulation.
If there are chip controls or whatever, there will be efforts to get around that. If there are international agreements, there will be efforts to clandestinely hide.
If there are successful limitations on compute, people will compensate and focus on algorithmic progress.
Many people are going to be extremely resistant to being swayed on topics of AI, no matter what evidence is coming in. Much rationalization will be furnished to justify proceeding no matter the warning signs.
By and large, our civilization has a pretty low standard of reasoning.
People who want to speed up AI will use falsehoods and bad logic to muddy the waters, and many people won’t be able to see through it[4]. No matter the evals or other warning signs, there will be people arguing it can be fixed without too much trouble and we must proceed.
In other words, there’s going to be an epistemic war and the other side is going to fight dirty[5], I think even a lot of clear evidence will have a hard time against people’s motivations/incentives and bad arguments.
When there are two strongly motivated sides, seems likely we end up in a compromise state, e.g. regulation passes but it’s not the regulation originally designed that even in its original form was only maybe actually enough.
It’s unclear to me whether “compromise regulation” will be adequate. Or that any regulation adequate to cost people billions in anticipated profit will conclude with them giving up.
People aren’t thinking or talking enough about nationalization.
I think it’s interesting because I expect that a lot of regulation about what you can and can’t do stops being enforceable once the development is happening in the context of the government performing it.
Thinking through the above, I feel less motivated to work on things that feel like they’ll only speed up technical alignment problem research by amounts < 5x. In contrast, maybe there’s more promise in:
Cyborgism or AI-assisted research that gets up 5x speedups but applies differentially to technical alignment research
Things that convince people that we need to radically slow down
good writing
getting in front of people
technical demonstrations
research that shows the danger
why the whole paradigm isn’t safe
evidence of deception, etc.
Development of good (enforceable) “if-then” policy that will actually result in people stopping in response to various triggers, and not just result in rationalization for why actually it’s okay to continue (ignore signs) or just a bandaid solution
Figuring out how to overcome people’s rationalization
Developing robust policy stuff that’s set up to withstand lots of optimization pressure to overcome it
Things that cut through the bad arguments of people who wish to say there’s no risk and discredit the concerns
Stuff that prevents national arms races / gets into national agreements
Thinking about how to get 30 year slowdowns
By “slowing down”, I mean all activities and goals which are about preventing people from building lethal superpowerful AI, be it via getting them to stop, getting to go slower because they’re being more cautious, limiting what resources they can use, setting up conditions for stopping, etc.
How to build a superpowerful AI that does what we want.
They’re wrong about their ability to safely harness the power, but not if you could harness, you’d have a lot of very valuable stuff.
My understanding is a lot of falsehoods were used to argue against SB1047 by e.g. a16z
Also some people arguing for AI slowdown will fight dirty too, eroding trust in AI slowdown people, because some people think that when the stakes are high you just have to do anything to win, and are bad at consequentialist reasoning.
Quickly written. Probably missed where people are already saying the same thing.
I actually feel like there’s a lot of policy and research effort aimed at slowing down the development of powerful AI–basically all the evals and responsible scaling policy stuff.
A story for why this is the AI safety paradigm we’ve ended up in is because it’s palatable. It’s palatable because it doesn’t actually require that you stop. Certainly, it doesn’t right now. To the extent companies (or governments) are on board, it’s because those companies are at best promising “I’ll stop later when it’s justified”. They’re probably betting that they’ll be able to keep arguing it’s not yet justified. At the least, it doesn’t require a change of course now and they’ll go along with it to placate you.
Even if people anticipate they will trigger evals and maybe have to delay or stop releases, I would bet they’re not imagining they have to delay or stop for all that long (if they’re even thinking it through that much). Just long enough to patch or fix the issue, then get back to training the next iteration. I’m curious how many people imagine that once certain evaluations are triggered, the correct update is that deep learning and transformers are too shaky a foundation. We might then need to stop large AI training runs until we have much more advanced alignment science, and maybe a new paradigm.
I’d wager that if certain evaluations are triggered, there will be people vying for the smallest possible argument to get back to business as usual. Arguments about not letting others get ahead will abound. Claims that it’s better for us to proceed (even though it’s risky) than the Other who is truly reckless. Better us with our values than them with their threatening values.
People genuinely concerned about AI are pursuing these approaches because they seem feasible compared to an outright moratorium. You can get companies and governments to make agreements that are “we’ll stop later” and “you only have to stop while some hypothetical condition is met”. If the bid was “stop now”, it’d be a non-starter.
And so the bet is that people will actually be willing to stop later to a much greater extent than they’re willing to stop now. As I write this, I’m unsure of what probabilities to place on this. If various evals are getting triggered in labs:
What probability is there that the lab listens to this vs ignores the warning sign and it doesn’t even make it out of the lab?
If it gets reported to the government, how strongly does the government insist on stopping? How quickly is it appeased before training is allowed to resume?
If a released model causes harm, how many people skeptical of AI doom concerns does it convince to change their mind and say “oh, actually this shouldn’t be allowed”? How many people, how much harm?
How much do people update that AI in general is unsafe vs that particular AI from that particular company is unsafe, and only they alone should be blocked?
How much do people argue that even though there are signs of risk here, it’d be more dangerous to let other pull ahead?
And if you get people to pause for a while and focus on safety, how long will they agree to a pause for before the shock of the damaged/triggered eval gets normalized and explained away and adequate justifications are assembled to keep going?
There are going to be people who fight tooth and nail, weight and bias, to keep the development going. If we assume that they are roughly equally motivated and agentic as us, who wins? Ultimately we have the harder challenge in that we want to stop others from doing something. I think the default is people get to do things.
I think there’s a chance that various evals and regulations do meaningfully slow things down, but I write this to express the fear that they’re false reassurance–there’s traction only because people who want to build AI are betting this won’t actually require them to stop.
Related:
Curated. I think Raemon’s been doing a lot of work in the last year pushing this stuff, and this post pulls together in one place a lot of good ideas/advice/approach.
I would guess that because of the slow or absent feedback loops, people don’t realize how bad human reasoning and decision-making is when operating outside of the familiar and quick feedback. That’s many domains, but certainly the whole AI situation. Ray is going after the hard stuff here.
And the same time, this stuff ends up feeling like the “eat your vegetables” of reasoning and decision-making. It’s not sexy, or at least it’s not that fun to sit down and e.g. try to brainstorm further plans when you already have one that’s appealing. or backchain from your ostensible goal. I think we’d be in a better place if these skills and practices were normalized, in the sense of there’s a norm that you do these things and if you don’t, then you’re probably screwing up.
Yeah, I think a question is whether I want to say “that kind of wireheading isn’t mypoic” vs “that isn’t wireheading”. Probably fine eitherway if you’re consistent / taboo adequately.
My guess is Ben created the event while on the East Coast and 6pm got timezone converted for West Coast. I’ve fixed it.
Once I’m rambling, I’ll note another thought I’ve been mulling over:
My notion of value is not the same as the value that my mind was optimized to pursue. Meaning that I ought to be wary that typical human thought patterns might not be serving me maximally.
That’s of course on top of the fact that evolution’s design is flawed even by its own goals; humans rationanlize left, right, and center, are awfully myopic, and we’ll likely all die because of it.
Curated. This is a good post and in some ways ambitious as it tries to make two different but related points. One point – that AIs are going to increasingly commit shenanigans – is in the title. The other is a point regarding the recurring patterns of discussion whenever AIs are reported to have committed shenanigans. I reckon those patterns are going to be tough to beat, as strong forces (e.g. strong pre-existing conviction) cause people to take up the stances they do, but if there’s hope for doing better, I think it comes from understanding the patterns.
There’s a good round up of recent results in here that’s valuable on its own, but the post goes further and sets out to do something pretty hard in advocating for the correct interpretation of the results. This is hard because I think the correct interpretation is legitimately subtle and nuanced, with the correct update depending on your starting position (as Zvi explains). It sets out and succeeds.
Lastly, I want to express my gratitude for Zvi’s hyperlinks to lighter material, e.g. “Not great, Bob” and “Stop it!” It’s a heavy world with these topics of AI, and the lightness makes the pill go down easier. Thanks