Why AI Competition Forces Labs to Race | Dan Robinson

Dan Robinson (0:00)

There is this kind of Prisoners' Dilemma-like component to AI spacing, where you might have every participant in an AI race prefer to, at least on the margin, go slower than they are going. But because of competition from others, they're concerned that they don't actually want to do that. And so we see this in sort of the professed claims of the AI labs, which I think we all think are sincere, that a lot of people at these labs would prefer to slow down, but are concerned that if they do, others won't, and so they're looking for coordination mechanisms there.

SPEAKER_2 (0:28)

All right, we are back. We are live with Dan Robinson, who is General Partner and Head of Research at Paradigm, which is an investment firm that does some pretty sick design and research stuff.

They just built Pace, which is a game about the race dynamics between AI labs pacing the frontier, and we're going to play it on stream. So Dan, welcome.

Dan Robinson (0:49)

Thanks for having me on.

SPEAKER_2 (0:50)

So tell us about this game.

Dan Robinson (0:52)

So the origins of this came from, we put out a previous post that was about recursive self-improvement and a game that was based on an economics paper.

SPEAKER_2 (1:02)

We did that one on stream too. Yes.

Dan Robinson (1:03)

So we saw that. I thought you did very well on it.

One of the ideas we've been having is that some of these concepts, particularly around economics of AI, these somewhat frontier research in concepts and economics, can be best explained through games rather than blog posts or papers. It's really helpful sometimes to get the feeling of actually being, say, a frontier lab facing some of these trade-offs. And so the RSI game was trying to illustrate what is it like to actually try to have a frontier lab go through recursive self-improvement. The pacing game is about what it's like to be an AI lab in the presence of competition. And it's based on another economics paper that came out by Andrew Cohen, Drew Feudenberg, which is about the game theory of AI pacing.

SPEAKER_3 (1:45)

Yeah.

So I'm curious, initially, what do you take away from the game theory of AI pacing? What kind of competition, domestic or international, motivates pacing?

Dan Robinson (1:57)

Yeah. So I think one of the defaults that you come into, and one of the simplest lessons of the game, is there is this kind of prisoners dilemma component to AI pacing, where you might have every participant in an AI race prefer to, at least on the margin, go slower than they are going. But because of competition from others, they're concerned that they don't actually want to do that. And so we see this in sort of the professed claims of the AI labs, which I think we all think are sincere, that a lot of people at these labs would prefer to slow down, but are concerned that if they do, others won't. And so they're looking for coordination mechanisms there. So one of the key lessons of the game is like, you clearly see this dynamic, where I think you would rather that both you and your competitor go slower, but when you're playing you say, okay, like if they're going to go ahead, then think of the best if I'm still following them. Then there are some other sort of like other subtler games where sometimes, under certain conditions in the paper talks about this, you might actually get voluntary coordination, even in the absence of government regulation. And that's when, if the consequences are catastrophic enough, or if safety moves along so slowly, then in fact, it makes more sense for both participants to slow down, even irregardless of what the other person does. There's an equilibrium there. And then there's another interesting set of scenarios where in fact, both speeding ahead and pacing the frontier are equilibria. So if there's a world where just all the labs choose to voluntarily slow down their development to a safe pace, and then that remains in equilibrium, but then there's even another choice that they could all make where they're all speeding ahead and that's in equilibrium. So I think a lot depends really on these just sort of contingent factors about what happens. And I think you see this in the game where these games can go very different directions.

SPEAKER_3 (3:33)

Right.

SPEAKER_2 (3:33)

So should we play? Yeah. Yeah, absolutely. Jinx.

SPEAKER_3 (3:37)

Double jinx.

SPEAKER_4 (3:40)

Yeah.

SPEAKER_2 (3:41)

All right. So I'm playing against you now.

SPEAKER_4 (3:43)

Yep.

SPEAKER_2 (3:44)

So I'm not accelerating at all.

Dan Robinson (3:45)