Head of Lightcone Infrastructure. Wrote the forum software that the EA Forum is based on. Often helping the EA Forum with various issues with the forum. If something is broken on the site, it’s a good chance it’s my fault (Sorry!).
Habryka
Yeah, this.
From my perspective “caring about anything but human values” doesn’t make any sense. Of course, even more specifically, “caring about anything but my own values” also doesn’t make sense, but in as much as you are talking to humans, and making arguments about what other humans should do, you have to ground that in their values and so it makes sense to talk about “human values”.
The AIs will not share the pointer to these values, in the same way as every individual does to their own values, and so we should a-priori assume the AI will do worse things after we transfer all the power from the humans to the AIs.
In the absence of meaningful evidence about the nature of AI civilization, what justification is there for assuming that it will have less moral value than human civilization—other than a speciesist bias?
You know these arguments! You have heard them hundreds of times. Humans care about many things. Sometimes we collapse that into caring about experience for simplicity.
AIs will probably not care about the same things, as such, the universe will be worse by our lights if controlled by AI civilizations. We don’t know what exactly those things are, but the only pointer to our values that we have is ourselves, and AIs will not share those pointers.
It’s been confirmed that the donation matching still applies to early employees: https://www.lesswrong.com/posts/HE3Styo9vpk7m8zi4/evhub-s-shortform?commentId=oeXHdxZixbc7wwqna
Your opening line seems to be trying to mimic the tone of mocking someone obnoxiously. Then you follow-up with an exaggerated telling of events. Then another exaggerated comparison.
Weird bug. But it only happens when someone votes and unvotes multiple times, and when you vote again the count resets. So this is unlikely to skew anything by much.
Given that I just got a notification for someone disagree-voting on this:
This is definitely no longer the case in the current EA Funding landscape. It used to be the case, but various changes in the memetic and political landscape have made funding gaps much stickier, and much less anti-inductive (mostly because cost-effectiveness prioritization of the big funders got a lot less comprehensive, so there is low-hanging fruit again).
I’m not making any claims about whether the thresholds above are sensible, or whether it was wise for them to be suggested when they were. I do think it seems clear with hindsight that some of them are unworkably low. But again, advocating that AI development be regulated at a certain level is not the same as predicting with certainty that it would be catastrophic not to. I often feel that taking action to mitigate low probabilities of very severe harm, otherwise known as “erring on the side of caution” somehow becomes a foreign concept in discussions of AI risk.
(On a quick skim, and from what I remember from what the people actually called for, I think basically all of these thresholds were not for banning the technology, but for things like liability regimes, and in some cases I think the thresholds mentioned are completely made up)
You’re welcome, and makes sense. And yeah, I knew there was a period where ARC avoided getting OP funding for COI reasons, so I was extrapolating from that to not having received funding at all, but it does seem like OP had still funded ARC back in 2022.
Thanks! This does seem helpful.
One random question/possible correction:
Is Kelsey an OpenPhil grantee or employee? Future Perfect never listed OpenPhil as one of its funders, so I am a bit surprised. Possibly Kelsey received some other OP grants, but I had a bit of a sense Kelsey and Future Perfect more general cared about having financial independence from OP.
Relatedly, is Eric Neyman an Open Phil grantee or employee? I thought ARC was not being funded by OP either. Again, maybe he is a grantee for other reasons.
(I am somewhat sympathetic to this request, but really, I don’t think posts on the EA Forum should be that narrow in its scope. Clearly modeling important society-wide dynamics is useful to the broader EA mission. To do the most good you need to model societies and how people coordinate and such. Those things to me seem much more useful than the marginal random fact about factory farming or malaria nets)
I don’t think this is true, or at least I think you are misrepresenting the tradeoffs and diversity here. There is some publication bias here because people are more precise in papers, but honestly, scientists are also not more precise than many top LW posts in the discussion section of their papers, especially when covering wider-ranging topics.
Predictive coding papers use language incredibly imprecisely, analytic philosophy often uses words in really confusing and inconsistent ways, economists (especially macroeconomists) throw out various terms in quite imprecise ways.
But also, as soon as you leave the context of official publications, but are instead looking at lectures, or books, or private letters, you will see people use language much less precisely, and those contexts are where a lot of the relevant intellectual work happens. Especially when scientists start talking about the kind of stuff that LW likes to talk about, like intelligence and philosophy of science, there is much less rigor (and also, I recommend people read a human’s guide to words as a general set of arguments for why “precise definitions” are really not viable as a constraint on language)
AI systems modeling their own training process is a pretty big deal for modeling what AIs will end up caring about, and how well you can control them (cf. the latest Anthropic paper)
For most cognitive tasks, there does not seem to be a particularly fundamental threshold at human-level performance (this one is still out in many ways, but we are seeing more evidence for this on an ongoing basis as we reach superhuman performance on many measures)
Developing “contextual awareness” does not require some special grounding insight (i.e. training systems to be general purpose problem solvers naturally causes them to optimize themselves and their environment and become aware of their context, etc.). This was back in 2020, 2021, 2022 one of the recurring disagreements between me and many ML people.
(In general, the salaries which I will work for in EA go up with funding uncertainty, not down, because indeed it means future funding is more likely to dry up, and I have to pay the high costs of a career transition, or self-fund for many years)
You are right! I had mostly paid attention to the bullet points, which didn’t extract the parts of the linked report that addressed my concerns, but you are right that it totally links to the same report that totally does!
Sure, I don’t think it makes a difference whether the chicken grows to a bigger size in total, or grows to a bigger size more quickly, both would establish a prior that you need fewer years of chicken-suffering for the same amount of meat, and as such that this would be good (barring other considerations).
No, those are two totally separate types of considerations? In one you are directly aiming to work against the goals of someone else in a zero-sum fashion, the other one is just a normal prediction about what will actually happen?
You really should have very different norms about how you are dealing with adversarial considerations and how you are dealing with normal causal/environmental considerations. I don’t care about calling them “vanilla” or not, I think we should generally have a high prior against arguments of the form “X is bad, Y is hurting X, therefore Y is good”.
Thank you! This is the kind of analysis I was looking for.
I think the answer to this is “yes”, because your shared genetics and culture create much more robust pointers to your values than we are likely to get with AI.
Additionally, even if that wasn’t true, humans alive at present have obligations inherited from the past and relatedly obligations to the future. We have contracts and inheritance principles and various things that extend our moral circle of concern beyond just the current generation. It is not sufficient to coordinate with just the present humans, we are engaging in at least some moral trade with future generations, and trading away their influence to AI systems is also not something we have the right to do.
(Importantly, I think we have many fewer such obligations to very distant generations, since I don’t think we are generally borrowing or coordinating with humans living in the far future very much).
Look, this sentence just really doesn’t make any sense to me. From the perspective of humanity, which is composed of many humans, of course the fact that AI does not care about the same things as humans creates a strong presumption that a world optimized for those values will be worse than a world optimized for human values. Yes, current humans are also limited to what degree we successfully can delegate the fulfillment of our values to future generations, but we also just share, on-average, a huge fraction of our values with future generations. That is a struggle every generation faces, and you are just advocating for… total defeat being fine for some reason? Yes, it would be terrible if the next generation of humans suddenly did not care about almost anything I cared about, but that is very unlikely to happen, but it is quite likely to happen with AI systems.