Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity
Can regulators promise both safety limits and delivery dates for AI?
A regulator trying to guarantee both a maximum risk level and a completion deadline faces a fundamental problem: they need to know how fast safety research will improve, but don't. The paper shows that one approach—doing safety research first, then building the model with extra computing power reserved for safety—gets close to the theoretical minimum needed to keep both promises, requiring only 3.3% more efficiency than mathematically necessary.
AI regulation increasingly demands concrete safety guarantees and timelines. This work clarifies what regulators actually need to know to make good on both promises, and shows that some strategies are far more wasteful than others. If regulators instead just cap how much computing power goes to safety, they'll control timelines but won't reliably hit safety targets—a distinction that matters for real oversight rules.