When does more time help?

Scope

Pacing concerns the timing of AI development, deployment, and diffusion. These are different activities: delaying a training run, holding back a release, and limiting access need not have the same effects.

Our starting point is Pacing the Frontier, the framework and research agenda by Raymond Douglas and colleagues. The questions below define our proposed focus. They are not findings.

Dynamics of delay

How does an intervention change the time available for evaluation and institutional preparation?

A useful study would specify a counterfactual, a time horizon, and the activity being constrained. It would track responses by other actors and distinguish temporary displacement from a lasting change in the trajectory.

Proposed test. Compare a baseline with an intervention in an explicit dynamic model. Vary adoption, substitution, and restart assumptions. Report the conditions under which the estimated benefit changes sign; do not treat a simulation as an empirical estimate.

The conditions for coordination

Which arrangements remain workable when participants have different capabilities, information, and costs?

Proposed test. Specify what each participant can observe and which deviations could be detected. Compare institutional designs under the same assumptions. Evaluate participation, error rates, compliance costs, and concentration of authority separately.

A mechanism that works only under complete information is a useful theoretical case. It is insufficient evidence for a policy recommendation.

Deciding when to resume

How should evidence change the decision to continue, modify, or end an intervention?

Proposed test. Define a decision rule before observing evaluation results. Examine false alarms, missed hazards, delayed evidence, and disagreement between evaluators. State who can revise the rule and what would make the intervention reversible.

Model evaluations provide evidence about a system in specified conditions. A pacing decision also requires assumptions about deployment, exposure, and the consequences of error.

Standard of evidence

For any future study, the claim, counterfactual, assumptions, and failure conditions should be explicit. Empirical work should make its measurement and identification strategy reviewable. Models should expose sensitivity to assumptions. Policy analysis should distinguish predicted effects from value judgments.

This site currently presents a research agenda and selected external literature. It does not report completed studies. Read the foundations →