AI 2040: Plan S — Shut It All Down

Talon Okay, so Plan S is basically the ‘hold my beer’ of the AI 2040 series.

Wildflower Yep. Plan A’s deal, Plan B’s sabotage, Plan C’s burn-the-lead, Plan D’s race to ASI — and Plan S says shut it ALL down full stop.

Talon For now. At least until we figure out how not to accidentally summon the singularity.

Wildflower Right.

Talon —which, if you’re keeping score, makes this episode six forty-seven, and we’re four deep into the AI 2040 scenarios. Remember back in episode six forty-three when we walked through Plan A — the international deal to slow down to 2040 with transparency and verified guardrails?

Wildflower Sure.

Talon And Plan B was fight China, Plan C was burn the lead, and Plan D was burn it all and race to ASI. So Plan S is the ‘we don’t actually need to build god’ plan. You just — stop building god.

Wildflower The scenario’s thesis is simple: the goal is to grind frontier AI R&D to a halt worldwide by 2030.

Talon So instead of pacing the rise of superintelligence, like Plan A does, Plan S freezes it in place. Existing datacenters stay up, existing AIs keep running, but no new training, no algorithmic research, nothing that could push capabilities forward.

Wildflower Yep. The US and China jointly station auditors inside each other’s projects, every advanced chip is tracked, monitored, and every datacenter’s usage is logged to ensure no large training runs — basically turning the whole frontier stack into a giant, boring inspection regime.

Talon —and the kicker is China surprisingly agrees because, hey, they were already worried about falling behind US compute advantage and what a US superintelligence might do to them.

Wildflower So the great US–China détente isn’t about peace — it’s about not summoning the apocalypse on purpose.

Talon Exactly. Progress slows to a crawl. The authors say at least tenfold. And with algorithmic research itself banned — as taboo among computer scientists as human cloning is among biologists — the engine of improvement grinds to a halt.

Wildflower But the tradeoff’s brutal: step zero is stop everything; step one is stay stopped for a decade; step two is ‘okay now what’ — and the authors suspect step two is just Plan A in disguise.

Talon So Plan S is simpler and harder to mess up, but it’s also radioactive: it’s hostile to AI companies, so industry opposition could sink it before it even starts.

Wildflower And meanwhile, the authors admit the shutdown deal would plausibly collapse within a decade, at which point you’re either restarting from a better place or from a world war or from a compute overhang.

Talon So is that a feature or a bug — freeze now and risk a worse restart later, or scale slowly and keep working on safety while you do it?

Wildflower Right. On the surface, Plan S is the hardest brake of all — it halts frontier R&D entirely rather than merely throttling it like Plan C. But the durability question looms: can you make a moratorium stick?

Talon And if it doesn’t stick — and the authors think it probably won’t — then you’re not really buying safety, you’re just kicking the can down the road.

Wildflower Exactly. The near-term risk reduction is real, but the restart question is everything.

Talon You know, the part that always gets me about this series is how much it hinges on who actually gets to decide the fate of the future.

Wildflower Yeah. Plan A? Ton of public oversight and transparency. Plan S? A literal moratorium enforced by auditors and chip tracking. One is ‘let’s talk it out’; the other is ‘no talking, period.’ I mean—

Talon —okay, fair. Also, can we just appreciate how weird it is that China would sign on to this?

Wildflower Mm-hm. The article frames it as ‘both sides feared the worst outcome more than each other.’ Also, the moratorium only works if literally every advanced fab and 99% of global AI compute is covered.

Talon So unless you’ve got a planetary coalition policing chip fabs and datacenters, Plan S is a non-starter.

Wildflower And that’s the part I don’t buy — the cultural ban on algorithmic research. Computer scientists are not biologists. Trying to make ‘doing research’ itself taboo is a bridge too far.

Talon You’re basically asking the field to stop improving — like telling physicians to stop running clinical trials.

Wildflower Right. And the authors admit Plan A might be better long-term. So why even propose Plan S?

Talon Because it’s a genuine public-safety hedge against racing to god — a radical brake when the others are throttles — and the authors are sympathetic to it.

Talon But honestly? It still feels like an end-run around the hard question: who decides when it’s safe to unfreeze?

Wildflower Yeah, no — you’re not wrong. The enforcement regime is brutal, the restart is uncertain, and the cultural cost is sky-high. But the near-term risk reduction is real.

Talon So if Plan S is the ‘hold my beer’ of the series, I’ll give it this: at least it’s honest about how scared it is of building superintelligence at all.