The guy who quit OpenAI over safety concerns now has a plan to keep superintelligent AI from arriving too soon. Daniel Kokotajlo’s AI Futures Project just dropped “AI 2040: Plan A,” a report that argues the world’s best move is to deliberately pump the brakes on artificial intelligence development and push the arrival of superintelligence out to 2040.
The core pitch: get the US and China to agree, by 2029, on a framework for full transparency in AI research and development, backed by verification protocols and what the report calls “mutually assured compute destruction.” Think nuclear deterrence logic, but for data centers.
From doom forecast to deliberate optimism
The AI Futures Project previously published “AI 2027,” a forecast that painted a faster and considerably more alarming trajectory for AI progress. “AI 2040: Plan A” takes the opposite approach, mapping out what could go right if the major players choose cooperation over competition.
The report proposes a phased development plan. Rather than racing toward the most powerful systems as fast as possible, it envisions a gradual scaling of AI capabilities, with expert-level systems arriving around 2035. The final stretch to superintelligence would come five years later, in 2040, with enough safety infrastructure in place to manage it.
