Why the US Is Moving AI Policy Behind Closed Doors

Anton Leicht (0:00)

I still think it shows that, how much more of a securitized conversation this has become, how much more the security agencies we'd expect them to be involved, how much more of this will be classified, how much more of this will be sort of behind closed doors, and only very limited communication between the labs and the US government, but not that much public oversight.

SPEAKER_2 (0:18)

And we're back. We are live once again with Anton Leicht, who's a fellow at the Carnegie Endowment. He is the author of Threading the Needle on Substack. He's written many, many very good articles.

Some of which we're going to talk about today. Anton, welcome back.

Anton Leicht (0:32)

Great to be back.

SPEAKER_2 (0:33)

Second experience. No, this is like your fourth experience.

Anton Leicht (0:35)

Yeah, second experience here. And I think we've called it...

SPEAKER_3 (0:38)

Second my appearance. Yeah. So true.

SPEAKER_2 (0:40)

alongside you.

SPEAKER_3 (0:41)

Yeah.

SPEAKER_2 (0:42)

So I guess we'll start with this weekend. We were at the Curve. What were your impressions?

Anton Leicht (0:47)

I think it was a great time. It was fun to catch up with everyone. And I think very few places as close on the intersection between the DC and SF conversation. I think it's still my favorite conference of the year and I had a great time.

I will say the intersection between DC and SF is a lot less surprising now than it was a year ago or two years ago. I think even last year, there was a lot of the vibes of, oh my God, there's like a guy in a suit at Lighthaven and like sort of there's the suddenly sort of like words are colliding. I think the worlds have pretty much collided at this point.

SPEAKER_2 (1:16)

Yeah, there are definitely more SF people than DC people.

Anton Leicht (1:19)

Yeah, I think just like the worlds have sort of, there's a lot of overlap between those. I think a lot of DC people have now talked to someone who works at a lab within the last 12 months. And I think it's not quite as much of a culture shock to both sides involved anymore. But I think it's still obviously the sort of most important. Was this your first curve or second curve?

SPEAKER_2 (1:38)

Your second curve.

Anton Leicht (1:38)

I started writing about AI stuff shortly after the first. And that was at the last one. The last one was just like eye-opening super cool experience.

SPEAKER_2 (1:46)

Well, you just started writing about AI stuff within the last year?

Anton Leicht (1:49)

No, after that, I started after the first one, which was two years ago.

SPEAKER_3 (1:52)

Still, that's like a generational run.

SPEAKER_2 (1:55)

Yeah. I think one of my big takeaways from it was it felt very monocultury. There was not a lot of fundamental disagreement. There was like, everyone seemed to be on board with pacing the frontier. Absolutely everyone, with very few exceptions, but almost everyone. Everyone seemed to be on board with pretty high chance of existential risk in the near future. Everyone seemed to be on board with very high chance of less than existential risk, moderate to major cyber risk in the near future.

I didn't notice people disagreeing that much on anything, which I imagine was less the case in previous curves.

Anton Leicht (2:30)

Yeah, I think, I mean, in some way, just reality has just caught up to people and really sort of narrowed the band of available, like reasonable predictions in the world in comparison to like a year or two ago, right? I mean, we've seen clear practical evidence of a lot of these cyber risks. Capability progress is just so fast that it's kind of hard to deny that currently we're just at an incredibly sort of breakneck pace. And I think the combination of things are actually substantially sort of going wrong. And capability progress is very, very fast, just makes sort of narrowing on and agreeing on pacing extremely easy because the combination of risks and pace just means that, well, maybe we should go a bit slower. I do think if you go into the specifics of the conversation, what should pacing specifically mean? Should it be about inputs? Should it be about the compute, for example? Should it be about outputs? Like, should it be about the sort of just like the trajectory and the epoch capability in the index or whatever? Or should it be about safety desiderata? Like, should we just pace by saying, well, you only can internally deploy the next generation if it passes this or that bar of safety? I think if you get into actually defining what pacing is supposed to mean, or we're actually supposed to do, I think you do get a lot of disagreement among the people that attended. But I think that very generally everyone's like, well, it would sure be nice if things moved a little bit slower so we could get ahead of the risks. I think that's just reality catching up to people.