When does more time help?
Our research agenda on AI pacing: what it changes, how to coordinate it, and when to resume.
The question
More time could help researchers make AI safer and give institutions room to prepare. But delay has costs. The question is when changing the pace of development leaves us better off, and what would need to happen during that time.
We study development, deployment, and diffusion separately. Holding back a training run, delaying a release, and limiting access to model weights are different interventions. Each needs its own case.
Dynamics of delay
If one part of AI development slows, where does the work go? Compute can be put to other uses. Other developers may continue. Progress may accumulate and arrive quickly when a restriction ends.
We want to measure the time an intervention creates, the work it displaces, and what can be done with the time gained. A slower training run is not, by itself, evidence of a safer outcome.
The conditions for coordination
A commitment has to hold when it becomes costly. Who agrees to it? What can they observe about one another? Who can investigate a breach, and who has the authority to respond?
We are interested in arrangements that can work despite unequal capabilities and competing interests. That includes the risk of giving an institution power it cannot use well, or cannot later give up.
Deciding when to resume
Our initial focus is the rules for slowing, stopping, and resuming frontier development. We plan to examine how published commitments handle incomplete evidence, disagreement between evaluators, and pressure from competitors.
The first study will pair a close reading of these commitments with a set of decision cases. Can independent readers tell what action a rule requires? Does the answer change when the evidence changes? Where does the rule leave the decision open?
This is a starting point for pacing evaluations: documented cases, explicit criteria, and results that another researcher can check. It can tell us where a rule is ambiguous or difficult to apply. Establishing whether it reduces harm requires further evidence.