GenericModel's Shortform

GenericModel

This is a special post for quick takes by GenericModel. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

It seems that the U.S. has decided to allow the sale of H200s to China. Frankly, I'm not that shocked.

What I am more surprised about is the blind-eye that is being turned to their country by Americans, those concerned about AI development and those not. Do you really think your country has been acting like one that seems capable of stewarding us to AGI?

Such that there's a serious argument to be made for the US to accelerate so that China doesn't get there first? Sure, maybe on the margins it's better for the US to get there before China, and if we were talking about the US of thirty years ago (and the China of thirty years ago) then it would be clear.

Alas, that is no longer the case. We are talking about the US of today. A US that forfeited its lead for seemingly no good reason. What can one say? You get what you deserve.

Now advocate for a slowdown. There's so little to be gained from beating China.

EDIT: To be clear, I am not saying all americans are turning a blind eye. But even those acknowledging the gravity of the situation seem to operate from a baseline assumption that of course it's better if we get there before China. You are better on Freedom, but your country is not taking this Seriously. I think China is more likely to take it Seriously, as a third party observer.

4GenericModel11mo

I guess this depends on your belief about the likelihood of alignment by default and the difficulty of alignment. Seems like it (EDIT: by it, I mean alignment by default/alignment is very easy) could be true! I currently think it's pretty hard to say definitively one way or the other, so am just highly unconfident. But how could this possibly be better than taking your time to make sure the first one has the highest odds possible of being aligned? Right? Like if I said "We're going to create an incredibly powerful Empire, this Empire is going to rule the world" And you said "I hope that Empire has good values," Then I say "Wait, actually, you're right, I'll make 100 different Empires so they're all in conflict they cooperate with each other and at least one of them will have good values" I think you should basically look at me incredulously and say that obviously that's ridiculous as an idea.

1StanislavKrym11mo

Didn't the AI-2027 forecast imply that it's the race between the leaders that threatens to misalign the AIs and to destroy mankind? In the forecast itself, the two leading countries are the USA and China, which, alas, has a bit less compute than OpenBrain, the American leading company, alone. This causes the arms race between OpenBrain and the collective having all Chinese compute, DeepCent. As a result, DeepCent stubbornly refuses to implement the measures which would let the company have an aligned AI. Suppose that the probability that the AI created by a racing company is misaligned is p while[1] the probablity P(the AI created by a perfectionist monopolist is misaligned) is q. Then P(Consensus-1 is misaligned|OpenBrain and DeepCent race) is p^2. The probability for a perfectionist to misalign its ASI is q. Then we would have to compare q with p^2, not with p. The authors of AI-2027 assume that p is close to 1 and q is close to zero. 1. ^ It might also depend on the compute budget of the perfectionist. Another complication is that p could also be different for the USA if, say, OpenAI, Anthropic, GDM, xAI decide to have the four AIs codesign the successor while DeepCent, having far less compute selects one model to do the research.

Not sure how you get P(consensus-1 is misaligned | Race) = p^2 (EDIT: this was explained to me and I was just being silly). Maybe there’s an implicit assumption there I’m not seeing.

But I agree that this is what AI-2027 roughly says. Also does seem that racing in itself causes bad outcomes.

However, I feel like the “create a parliament of AIs that are all different and hope one of them is sorta aligned” is a different strategy based on very different assumptions though? (It is a strategy I do not agree with, but think the AI 2027 logic would be unlikely to convince a person who agreed with it, and this would be reasonable given such a person’s priors).