Can an AI deadline and a safety limit both hold?
Li Gan models a choice between an AI development deadline and a limit on risk. If the benefit of safety work is unknown, a fixed plan for computing power can set a date but cannot promise the risk limit. A proposed learn-first plan relies on strict assumptions about how well safety work carries over. This is a theory model, not a tested policy.
What the source reports
Li Gan asks whether an AI regulator can promise two things at once: an upper limit on risk and a deadline for building a more capable system. In the paper’s theory model, no one knows how much a unit of safety work will help. A fixed schedule for computing power can hold to a date. It cannot also promise the risk limit. The paper describes a learn-first option, but that option assumes safety work pays off at a steady rate and that its gains carry over in full. These are model conditions, not observed facts. The preprint gives no proven safety guarantee or real policy date.
Original source
- Title
- Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity
- Author
- Li Gan
- Publication
- arXiv
- Date
- Friday, October 9, 2026