The AI Futures Project, the group behind last year's widely read "AI 2027" scenario, has published a follow-up setting out how it thinks humanity could reach superintelligence without disaster.
The new report, "AI 2040: Plan A", argues that the current race to build systems smarter than people risks either human extinction or a permanent seizure of global power.
Its authors stress that the document is a set of recommendations, not a forecast.
At its heart is a proposal that the United States and China jointly pause frontier artificial intelligence (AI) training in 2029.
The two powers would then scale up development slowly and together, under a regime of total research transparency.
Making nearly all frontier research public, the report argues, would mean any nation trying to cheat was spotted at once.
The aim is to hold off superintelligence until 2040, buying time for safety research to catch up.
Chips as hostages
The most striking element is an enforcement mechanism the authors call "mutually assured compute destruction".
Under it, the United States would site its data centres in Mongolia, while China would build its own in Canada.
Each superpower's computing power would sit largely undefended on ground next to its rival, so either side could quickly destroy the other's chips if the deal broke down.
The idea borrows from nuclear deterrence, turning the vast infrastructure behind modern AI into a mutual hostage.
A bleak forecast, an upbeat plan
The project is led by Daniel Kokotajlo, a former OpenAI researcher who left the company in 2024.
His group's earlier work estimated that the unchecked race carried roughly a 70% chance of ending in catastrophe, a category spanning extinction and runaway concentration of power.
Plan A is pitched as the optimistic alternative, promising the same eventual abundance a decade later, but reached deliberately rather than at a sprint.
Critics have already dismissed it as unworkable, noting that Mongolia and Canada are treated as convenient geography rather than sovereign states with a say.
Whether any of it is achievable is another matter, but its authors insist current approaches to managing AI are simply not good enough.
