LessWrong 2.0 Reader

Mars is locked due to the Santa Claus parade, so please gather at the Elizabeth St. entrance. We'll accumulate there until 14:15.

I do like the idea of having "model organisms of alignment" (notably different than model organisms of misalignment)

Minecraft is a great starting point, but it would also be nice to try to capture two things: wide generalization, and inter-preference conflict resolution. Generalization because we expect future AI to be able to take actions and reach outcomes that humans can't, and preference conflict resolution because I want to see an AI that uses human feedback on how best to do it (rather than just a fixed regularization algorithm).

Why would they not also potentially feel just as relatively intense positive valence, and have positive utility by default? Just getting an estimate that one side of the equation for their experience exists doesn't tell you about the other.

Wow, that seems really promising (thank you for the link!). I can envision one potential problem with the plan, though. It relies on the assumption that giving away 10% of the resources is the safest strategy for whoever controls AGI. But could it be that the group who controls AGI still lives in the "us vs them" mindset and decides that giving away 10% of the resources is actually a riskier strategy, because it would give the opposing side more resources to potentially take away the control over AGI?

This is a good argument for the systematic extermination of all insects via gene drives. If you value shrimp at a significant fraction of the value of a human and think they have negative utility by default, we should be trying really hard to make them go extinct. Can quicker euthanasia really compete against gene-drive-induced non-existence?

I mostly agree with this post, but while I do think that the AI safety movement probably should try to at least be more cooperative with other movements, I disagree with the claim in the comments section that AI safety shouldn't try to pick a political fight in the future around open-source.

(I agree it probably picked that fight too early.)

The reason is that there's a non-trivial chance that alignment is plausibly solvable for human-level AI systems ala AI control, even if they are scheming, so long as the lab has control over the AIs, which as a corollary also means you can't open-source/open-weights the model.

More prosaically, AI misuse can be a problem, and the most important point here is that open-source/open-weighting the model widens the set of people who can change the AI, which unfortunately also means that there is a larger and larger chance for misuse with more people that know how to change the AI.

So I do think there's a non-trivial chance that AI safety eventually will have to suffer political costs to ban/severely restrict open-sourcing AI.

I'm very uncertain and feel somewhat out of depth on this. I do have quite some hope though from arguments like those in https://aiprospects.substack.com/p/paretotopian-goal-alignment.

A man who is always asking ‘Is what I do worth while?’ and ‘Am I the right person to do it?’ will always be ineffective himself and a discouragement to others.

-- G.H. Hardy, A Mathematician's Apology

Is there a thorough analysis of OpenAI's for-profit restructuring? Surely, a Delaware lawyer who specializes in these types of conversions has written a blog somewhere.

Einstein started doing research a few years before he actually had his miracle year. If he started at 26, he might have never found anything. He went to physics school at 17 or 18. You can't go to "AI safety school" at that age, but if you have funding then you can start learning on your own. It's harder to learn than (eg) learning to code, but not impossibly hard.

I am not opposed to funding 25 or 30 or 35 or 40 year olds, but I expect that the most successful people got started in their field (or a very similar one) as a teenager. I wouldn't expect funding an 18-year-old to pay off in less than 4 years. Sorry for being unclear on this in original post.