What AI timelines are highest impact to act on?
I feel torn, and I think it varies a lot depending on your individual circumstances and opportunities. But overall, I think the arguments for prioritising shorter timelines are a bit stronger.
OscarDđ¸
Karma: 2,125
While I donât work in GHD, I still enjoy reading GHD content on the Forum and on Substack. I agree that interesting questions in GHD are far from solved, but I wonder if a lot of the low-hanging intellectual fruit has been picked (your number 5)? I wasnât around in early GiveWell days but I imagine that would have been an amazing time to be thinking about GHD and coming up with lots of new approaches and ideas. I havenât found GiveWellâs research to be that surprising or interesting lately for instance (vibes-based, I donât engage that closely with them anymore).
I would be keen to hear more from CE charities about what things they are learning and what questions they are facing!
Re your solution #2, I think I probably wouldnât want the Forum team to show âfavouritismâ, but the decline of GHD curated posts is interesting, and maybe that should change.
I think the type of early deal that would be most valuable is where the US and China both agree to produce a joint âconsensusâ ASI aligned to âthe goodâ. In more detail:
The US and China, as you note, are unsure who will win, and would be better off making a deal to preserve some minimum amount of future influence. But I think I am more worried than you about the costs of continued multipolarity into space colonisation. You write âEven having two alternative systems might open up the possibility for comparison, healthy competition, and moral trade.â War, threats, and unhealthy (e.g., burning the cosmic commons) competition also seem like important possibilities here.
Instead, I think having a joint superintelligence that coordinates using our cosmic endowment would be better, with some amount of influence within the âmoral parliamentâ of the ASI for each of the US and China.
Just that would be preferable to dividing up the universe into two camps I thinkâit is easier to do moral trades within one agent acting under moral uncertainty than coordinating between two agents.
A better version, though, could involve the US and China agreeing on some core moral precepts, or just a moral reflection process, and then jointly designing a moral curriculum for the proto-ASI including plenty of Western and Chinese texts, and letting the ASI do as it sees fit. Presumably both sides genuinely believe they are right and that an appropriate moral training process for the AI will lead to liberalism/âSocialism with Chinese characteristics. So this exploits the two sides having different credences (where as you note your proposed deals are possible even if both sides have the same credences). This creates a larger surplus for posisble agreements.
Of course, agreeing to create a joint ASI could also have big nearer term benefits, e.g. avoiding racing and slowing down AI progress and investing more in safety.
This proposal is clearly very far outside the overton window currently, but I donât think this is that much worse on feasibility than your proposed great power resource-sharing deals. It also solves the enforcement challenge as well which is convenient since we might have needed to create such a consensus AI to enforce a different sort of deal.
I am tentatively excited about this proposal, but I expect there isnât much to do to further it until the relevant parties are taking things more seriously.
Iâm fairly sympathetic to that, but it also feels like one needs to draw a line somewhere and where they have currently drawn it seems not unreasonable to me. Though another place to draw the line kind of on the opposite extreme which could also work is just anyone who supports effective giving and is planning to donate/âsalary sacrifice a lot of their money. Maybe the worry is that is too fuzzy and diluting the core 10% message though.
fyi @Luke Moore đ¸
Great article! I sometimes find myself explaining cash benchmarking to people and why some charities still beat cash, and this will be a useful thing to link to going forwards :)
Seems great! Insofar as you feel comfortable saying, why isnât this (fully) funded by cG?
Reasonable if you donât want to publicly go into internecine tensions, but the obvious question seems to be how you see this relating to principles-first EA, which is, on its face, a similar idea.
That is encouraging! Scottâs post linking to various prediction markets for Antrhopicâs implied valuations was also heartening.
Good point re communal values of the forum, seems right.
Ah, maybe I interpreted the original question differently to what you intended. SInce you said it is not about âpost qualityâ I was trying to put that aside and imagine AI-written posts that are better than human-written posts, and I think in that case I would be happy to read them. But I agree that currently I am turned off by AI writing and far prefer people write themselves in most cases. I suppose I was answering the question more in principle, i.e. if an AI-written post was amazing I would be comfortable with it, but currently they are not. So for me it is more a quality issue than fundamentally and AI-written issue (except for the communal/âsentimental aspects, which I agree have value).
How much of a post are you comfortable for AI to write?
Currently, I think AI writing isnât good enough to be better than good human users of the Forum, but I think this will quickly change, and I want to prioritise ideas and impact over who wrote the final words. I expect it will be longer before AIs are at the frontier of doing EA research and cuase-prioritization, so I think posts with only AI ideas will be bad for a longer time to come. But posts with human ideas written up well by an AI I could well imagine being better quality than most Forum writerâs posts within a year or two.
I feel differently if someone is writing something to me personally, if someone writes me a poem or a birthday card or something that has sentimental value, then AI writing reduces that. But the Forum I see as primarily content-value rather than sentimental value.
Nice post! Overall, I am quite sympathetic to this case. One skepticism I have is that the sorts of agents that are scope-sensitive in their ethics (and therefore linear in consensium) are probably also the ones who are fairly altruistic and therefore donât over-weight their own interests extremely, so would fund consensium (or rather hedonium, or their preferred impartial good) regardless of what others do. It feels like you either get that welfare in distant galaxies is a huge deal, or you donât.
Interesting! (And troublingâwell above the lizardman constant.) It would be interesting to do some qualitative follow-up on this, maybe with having these consistently retributivist people chat with an LLM instructed to do qualitative data collection and gently nudge them towards more suffering-averse views to see how deeply held or changeable those beliefs are.
Yep, that all makes sense, and I think this work can still tell us something, just it doesnât update me too much given the lack of compelling theories or much consensus in the scientific/âphilosophical community. This is harsher than what I actually think, but directionally, it has the feel of âcargo cult scienceâ where it has a fancy Bayesian model and lots of numbers and so forth, but if it all built on top of philosophical stances I donât trust then it doesnât move me much. But that said it is still interesting e.g. how wide the range for chickens is.
Most areas of capabilities research receive a 10x speedup from AI automation before most areas of safety research
The biggest factors seem to me to be feedback quality/âgood metrics and AI developer incentives to race
Nice! It strikes me that in figure 1, information is propagating upward, from indicator to feature to stance to overall probability, and so the arrows should also be pointing upward.
I think the view (stance?) I am most sympathetic to is that all our current theories of consciousness arenât much good, so we shouldnât update very far away from our prior, but that picking a prior is quite subjective, and so it is hard to make collective progress on this when different people might just have quite different priors for P(current AI consciousness).
Why does METR not receive cG money?
I have Thoughts about the rest of it, which I am not sure whether I will write up, but for now: I am sad for your Dadâs death and glad you got to prioritise spending some time with him.
I expect there is a fair bit we disagree about, but thanks for your integrity and effort and vision.
Perhaps the main downside is people may overuse the feature and it encourages people to spend time making small comments, whereas the current system nudges people towards leaving fewer more substantive comments and less nit-picky ones? Not sure if this has been an issue on LW, I donât read it as much.
Also, for those of us working in AI governance, âcGâ and âCIGIâ (the Center for International Governance Innovation) sound the same out loud. But in writing, I tend to use cG too.