Below is the executive summary from our new paper at pacing.tech. The full paper is available on the site and as a PDF. The full author list is Raymond Douglas, Charles Dillon, Nikola Moore, Gavin Leech, Shahar Avin, Mathias Kirk Bonde, Rohit Krishnan, Noah Perez, Nathan Young, Cormac Slade Byrd, Stephen Casper, Jan Kulveit, & David Duvenaud

“Pacing AI” usually refers to how to conclusively handle the most extreme risks in the face of race dynamics. However, even for the goal of handling these highest-stakes cases, it's useful to take a broad view of pacing—one that encompasses all interventions aimed at moderating the pace of AI development, deployment, or diffusion. Thus:

• Haphazard pacing is already common, including: delaying model releases for safety testing, pausing model development in response to shocks, and applying export controls.
• Current approaches will predictably fail. Isolated, unilateral actions addressing only small fractions of the problem are not enough, but poorly executed interventions could easily backfire—good solutions will need to be carefully designed.
• Precedents are being set whether we like it or not. How AI progress is paced now will shape how it is paced in future. We can learn from the shortcomings of existing attempts, and think about what precedents we are now setting for higher-stakes cases.
• Rapid, unpredictable progress requires an existing body of research to navigate well. Pre-specified proposals aren’t enough when key decisions often depend on sensitive information, have to be made quickly, and are responses to surprises.

Given all of this, we think it is time for pacing to be a dedicated research area.

Though there have been many specific proposals, and though many subfields of AI risk largely exist to feed into pacing decisions, pacing as a whole has not yet cohered into a clear field of study. There are massive gaps in our understanding of pacing as a whole. Here we highlight three topics to illustrate the work that needs doing:

• Understanding incentives. The practical effects of an intervention will depend substantially on how the affected parties respond to them, including the actors who enforce the intervention. We recommend prioritizing research on:
  • How different forms of transparency about capabilities can help or hurt coordination;
  • How to prevent mission creep among whoever is empowered to oversee pacing interventions;
  • Ways to align the incentives of different actors around pacing, e.g. by predictably compensating for losses with minimal moral hazard.
• Improving technical and regulatory interventions. All interventions face tradeoffs, but new coordination and oversight mechanisms and prior planning can allow strictly better options. For example:
  • Designing regulatory structures that allow for rapid but limited interventions which can make room for more careful deliberation;
  • Technical pathways to getting high assurance with minimal invasiveness, like LLM-based oversight and cryptographic guarantees;
  • Modeling the consequences of indirect interventions like buyouts, liability, and taxes.
• Investigating the full lifecycle of a pacing intervention. By considering the full sequence from before to after, we can spot gaps. For example:
  • Mapping different ways that interventions can end, and what this implies about different actors’ willingness to participate;
  • Figuring out how to make interventions which are more robust to premature endings or imperfect execution, intentional or otherwise;
  • Dry runs and wargames to stress-test specific interventions.

For each topic, we include recommended reading and starting points in the full text. We invite readers to reach out if they are interested in further work.