AI Safety Plan A risks accelerating intelligence explosion if US-China deal breaks

Plan A addresses a recognized existential risk: an uncontrolled, rapidly accelerating intelligence explosion driven by software innovation. The proposal to pause software progress at a critical moment—around 2030—attempts to create a window for safety research and infrastructure development before superintelligence emerges. The mechanism relies on international cooperation to hold that pause steady.

However, the body identifies a structural weakness in this approach. Building massive compute capacity during the software pause creates an asymmetry: the safety infrastructure is physically present and operational, but its effectiveness depends entirely on continued adherence to the software freeze. The analogy used in the body compares this to a city that halts an approaching fire at its gates, then constructs fire-study facilities entirely from dry tinder. The cure—the compute buildup—is materially incompatible with the worst-case scenario it is designed to mitigate. This is not a flaw in the compute itself, but a recognition that the plan's success rests on an assumption (continued cooperation) that, if violated, leaves the world with more dangerous tools than it would have had without the agreement.