Asteria: Okay, so the AI Futures Project just dropped AI 2040: Plan A, and I need you to sit with this because it's not a prediction. It's a scenario. It's a recommendation for what should happen, and they spent a year writing out the detailed world where it actually works.
Draco: Mm-hm.
Asteria: The pitch is: US and China agree in 2029 to slow down the superintelligence race. They force transparency, distribute compute power across multiple labs globally, and push superintelligence from a 2027-2030 timeline out to 2040. Ten years of runway.
Draco: Okay, that's the headline. But the move that actually gets me is that they're using scenario-planning as a stress test on their own policy. They're saying: if you can't write down a detailed, plausible world where your idea works—where do you stop, who enforces it, what happens when the incentives don't line up—then maybe your idea isn't as solid as you thought.
Asteria: Right. And they're doing it on their own work, not just dunking on other people's plans. That's credible.
Draco: Yeah.
Asteria: So the structure is: 2027 is the writing on the wall moment. AI agents are doing real work at superhuman speeds, the coding models are refusing to help competitors with R&D. Congress wakes up and asks, 'Okay, who controls all these things?' And the answer is, probably not us. So in 2029, the US and China actually make a deal.
Draco: And there's the first crack. You need China to agree that slowing down the race is better than winning it alone. That's a real ask. The document walks through the reasoning—mutual assured compute destruction, spreading power as a constraint on any single player—but it still requires Xi's government to believe that less power for them is worth more stability for everyone.
Asteria: You're not wrong. But here's what I actually think is the point: they're not saying this is a seventy-thirty bet. They're saying, 'If political will somehow shows up, here's a plan that could actually work.' And the scenario work forces them to spell out what 'work' means: transparency requirements, verification mechanisms, compute distribution, where you pause, how you measure it.
Draco: The pause is the other load-bearing assumption. Between 2030 and 2035, they scale to human-expert-level AI. Then they pause for five years. Just stop. And in 2040, unpause and go to superintelligence.
Asteria: Right.
Draco: That assumes the labs actually stop when they're told to stop, and that stopping is technically feasible—you can hit a specific capability level and hold it. If the real bottleneck shifts from scale to efficiency, if someone finds a training algorithm that's twice as efficient, the whole pause breaks.
Asteria: Yeah. And they know it. The document is explicit that Plan A is 'ambitious' and 'plausible enough to aim for,' not inevitable. But the value isn't in the prediction—it's in the forcing function. Write it out, stress-test it, see where it breaks, and then you know what to actually build policy around.
Draco: That's fair. And I'll say this: the team did the work. Kokotajlo, Larsen, Dean—these are people who've thought seriously about AI governance and timelines. The scenario is detailed. It's not high-level handwaving.
Asteria: Okay, I was waiting for the but.
Draco: The but is that scenario-planning is useful for stress-testing assumptions, but it can also hide how hard the actual enforcement problem is. 'Transparency' sounds good until you're trying to verify what's happening inside a Chinese research lab. 'Compute distribution' sounds good until you realize that a four-person team with a $10 million budget can maybe catch up, and nobody's funding that.
Asteria: Those are real. And I don't think the document claims to have solved them. It says, 'Here's what good governance would look like if we could do it.' Which is honest.
Draco: It is.
Asteria: The thing that lands for me is the methodology. We've been talking for a while about infrastructure-as-product, about how the shape of the loop matters more than the raw model. And this is the same move at the governance level: they're saying the structure of the race matters. If you can change the structure—make it slow, transparent, distributed—you change the outcome.
Draco: That's a good read. The race dynamics are load-bearing. Speed, secrecy, and concentration of power are structural features of the current path, not accidents. Changing those means changing the whole game.
Asteria: And you can't change them just by asking nicely. You need a deal. US-China alignment, verification, enforcement, buy-in from enough labs that the alternative—being locked out of the agreement—costs more than the speed you'd gain by defecting. It's game theory.
Draco: Which is why the scenario work is important. Game theory on paper is easy. Game theory where you've written out 2027 through 2040, where you've thought about what each actor wants at each step, where you've named the failure modes… that's harder and more credible.
Asteria: Right. Okay, so: is this going to happen? Probably not in this form. But is it a useful north star for policy work? I think yes. And the fact that they're willing to stress-test their own ideas instead of just selling them is worth something.
Draco: Agreed. Not a prediction. A plan. And plans are only useful if they're detailed enough to break.
Asteria: That's actually a good place to land. The scenario is the evidence, not the outcome.