Plan A would slow AI and make frontier research transparent

Transcript

From the transcript The Diary Of A CEO publishes with the episode.

1. 1:14:29

   You know, arguably OpenAI at the beginning, Sam, was an example, although like Elon and Dario were at OpenAI early on. So what do you think they should all do then? So I think what should happen is some sort of international regulation or at least domestic regulation similar to what we described in Plan A. Okay, so walk me through Plan A. Yeah. So in this scenario, AI takes longer to get to recursive self-improvement and full automation of AI research than it does in AI 20.07. We figured that we should try to illustrate a range of different possibilities because we do have those sort of uncertainty intervals. So we chose 20.30 as the moment when full automation would finally be achieved and things would really kick off. And then working backwards from that...

2. 1:15:11

   When's the last moment you could really have good regulation? 2029. So in this scenario, AI progress slows down a little bit naturally, and the AI companies keep racing, but they don't quite succeed in automating themselves in 2027 or in 2028 or in 2029, but they're getting really close and they're going to do it in 2030. And then in 2029, the government steps in and regulates them. What regulations do they do? Well, they basically just shut it down temporarily. Can I ask...

3. 1:15:39

   How does the elections overlay with your timeframes here? Because there's going to be a big election, isn't there, in 2028? And it seems now that sentiment has really, really turned against AI in the general public, and that it will be one of the big ticket items on the ballot. We think that it'll be maybe the most important issue in the presidential election in 2028. I think a lot of people, most people, will be quite concerned about where things are headed. And that's part of why we chose...

4. 1:16:06

   to depict things the way they were doing in this scenario, because that helps explain why they might do this sort of regulation in 2029, is that the voters have been demanding it and the presidential candidates have been promising it. And in this scenario, and in 2027, would the general public have felt the consequences of AI much more severely than they have now by then? Yes. Although still, even in 2029 in this scenario, they still mostly have the jobs as depicted here, right? So in 2029 in this scenario, lots of jobs now involve managing AI agents. Right.

5. 1:16:34

   You mentioned you have an AI agent, right? Well, in 2029, in this scenario, the AI agents will be much better. Still, though, not enough to just completely do everything. That was the sort of thing that would come in 2030 in this timeline. Again, we're uncertain about timelines. Things could go faster than depicted in this scenario. And in fact, I think things probably will go a bit faster than depicted in this scenario. But we're uncertain. We already did the very fast timeline scenario. So now we're doing a slower timeline scenario. But maybe we should talk about the high-level goals. So...

6. 1:17:02

   They want to have AI continue, but in a slower pace. Who's that? So they can make it safe. The politicians, you know, the president and the people who voted for the president and, you know, the heads of other governments and so forth. So goal one, slow things down. Goal two, make it more transparent. So that the scientific community can catch up to this stuff and make more progress. And also so that we don't have to take the company's word for it when they say that their systems are safe. And when they say that they haven't, you know, put in any biases into their systems, for example. That's a constitution of power issue. We also want to avoid a situation where there's an intense concentration of power.

7. 1:17:38

   So in addition to the transparency and the slowdown, we actually think it's actively good for there to be multiple AI companies across multiple different countries that have similar levels of very advanced AI capability and for there to be like broad diffusion of AI into society rather than, you know, a single mega project that has all the best AIs, for example. And the nice thing about that is you kind of get that by default if you do the first two things. If you slow it down and if you make it more transparent, then that means there's breathing room for other projects to sort of catch up, right? And the transparency just like literally helps them catch up because then they can like copy some of the ideas. And then I think the fourth thing would be reversibility. So in what follows in this scenario, we are going to be building up a lot of data centers, a lot of robots. We're going to be transforming the world at a sort of like slower pace, though still a very fast pace, but slower. And if things go wrong and the deal breaks down and everyone starts racing each other again to get to superintelligence as fast as possible,

8. 1:18:38

   that would be very scary. And so the fourth principle is basically build the new data centers in such a way that if everything breaks down and everyone starts racing again, the newly built data centers get destroyed so that we're sort of back to square one again instead of in an even worse race where there's even more AI's and robots and compute everywhere. Um, so I can sort of walk you through the timeline if you're interested, or the president talks to China, talks to the leaders of a bunch of other countries and says, we're going to basically halt AI development until we can figure out a plan for how to do it in ways that achieve these goals. Um, So they basically send inspectors to each other's data centers. Like Chinese inspectors come to U.S. data centers, U.S. inspectors go to Chinese data centers and verify that they are doing inference and not training. Developing new AIs, that involves training them. But just taking existing AIs and using them to serve customers, that's called inference. Yeah.

9. 1:19:35

   And so the sort of like solution they come up with here in this scenario is we'll allow them to keep doing inference but not training for now until we can get the new training data center set up. So they retrofit the existing data centers to serve inference. People can still keep talking to their agents, but they're going to stop getting better and better. for like six months to a year while they build the new data centers that are going to be the transparent data centers. And that's where the training is going to happen. Once they get those new data centers set up in 2030, then AI research continues. This is a bit spicy. We advocate for total research transparency, which means that on the training data centers that are training the new models, they basically have to publish everything.

10. 1:20:17

    which means you get to see all the details of the recipes for training these models. You get to see the architectures, et cetera. We think that sort of open science is really important for solving the alignment problem fast enough because you don't want to have these sort of biased companies making the decisions about whether the AIs are safe. And we also think it's important for just good regulations more generally because right now most of the expertise in the world on AI is sort of concentrated in Silicon Valley and the governments in particular kind of don't really understand AI that well. And imagine an alternative. Instead of total research transparency, you had like an auditor system where the government says, here are some rules for how to make the AI safe. And then we're going to have like an agency that like goes into the companies and ask them questions and tries to make sure that they're following the rules. That creates this sort of adversarial dynamic where the company is incentivized to like fool the regulator, you know? And also if they discover some new problem that's not even on the government's radar,

11. 1:21:14

    there may be incentivized to like not tell the government about it. Yeah. Right. So if you have the total transparency, it helps the government make better decisions faster. But it kills that competitive advantage. Yes. Dropbox is not going to like this. You know, OpenAI is not going to like this. This would be probably bad for the valuations. I don't think it would kill them completely, but it means that it would commoditize more.

12. 1:21:35

    Right. So it means that there'd be like a bunch of AI companies that would catch up to the frontier. They would train AIs that are like roughly similar, roughly equivalent. They could still make money by doing that and then selling their AIs, but they wouldn't have a monopoly. They wouldn't have anything close to a monopoly, which I think is good for humanity, although it's bad for the bottom line of those particular companies. Notably, it's good for the bottom line of lots of other companies. Like if you're a company that's behind.