AI Futures Project calls for deeper scenario analysis and covert project research

Thomas Larsen from the AI Futures Project published 'AI 2040: Plan A' yesterday, but stresses there is still substantial uncertainty about navigating the intelligence explosion. In a follow-up post, he outlines research areas that could improve the feasibility and desirability of Plan A. Key proposals include building a large set of competing concrete scenarios—prescriptive plans that detail specific strategies such as Plan S (indefinite halt on frontier AI capabilities), domestic-first regulation, GPU arms control, and an international 'CERN for AI' project. Larsen also calls for variations of Plan A that address concerns like 'dry tinder' (excess compute) by limiting capabilities via compute instead of algorithms.

Another urgent area is covert project research. Larsen notes that modeling of hidden AI development remains highly uncertain, and further effort could identify near-term actions governments or private actors could take to prevent, deter, or detect such projects. He invites serious researchers to collaborate directly with AI Futures Project. The post underscores that many competing plans lack the concrete scenario analysis needed to uncover hidden difficulties, making this work critical for informed decision-making as AI capabilities accelerate.