US–China cooperation in “AI 2040” plan A

A post examining Kokotajlo and coauthors’ “AI 2040” outlines several assumptions behind its “plan A” for pacing AI progress. The US and China would cooperate, declare essentially all large datacenters and AI compute, and allow inspections sufficient to reassure each side that the other is hiding less than roughly 1% of its compute. The post describes a temporary pause on all new AI training, with existing models still usable and datacenters retrofitted with “inference-only” verification hardware. Nearly all frontier research would be shared, while model weights, some training data and other sensitive information would remain protected. It also describes “Mutually Assured Compute Destruction”: US-owned datacenters in Mongolia and Chinese-owned ones in Canada would make compute vulnerable to destruction if either side defects.