When does more time help?
Pacing changes when AI capabilities are developed and used. Our research agenda asks when interventions improve outcomes, what they cost, and how they end.
Understanding the pace of AI and the choices that shape it.
Dynamics of delay
How much time does an intervention actually create? Study substitution, catch-up, and the difference between slowing one actor and slowing a frontier.
The conditions for coordination
What makes a commitment to restraint credible? Examine verification, unequal costs, and the incentives to participate or defect.
Deciding when to resume
What evidence should trigger, revise, or lift a restriction? Specify the decision rule, who applies it, and what happens when evidence is inconclusive.
Computing power and AI governance
A foundation for thinking about compute as a policy instrument, including its limitations and the safeguards oversight would require.
Model evaluation for extreme risks
A framework for evaluating dangerous capabilities and alignment, and for using evaluations to inform development and deployment decisions.