New "Pacing AI" research project argues slowdowns are already common but haphazard and need systematic study
gleech · x · 2026-09-17
AI safety researchers led by Raymond Douglas, with co-authors including David Duvenaud, Jan Kulveit and Stephen Casper, launched a project making the case for a dedicated research area on "pacing AI" — deliberately slowing AI development.
Key arguments:
- Pacing is already common (delayed releases for safety testing, development pauses after shocks, export controls) but haphazard, unilateral, and blunt.
- Current approaches will fail predictably: isolated unilateral actions are insufficient, while big poorly executed interventions could backfire; good solutions need deliberate design and attention to incentives.
- Rapid, unpredictable progress means pre-specified proposals aren't enough — key decisions depend on private information, must be made quickly, and respond to unforeseen surprises.
- Precedents are being set now, and their shortcomings offer lessons for higher-stakes cases.
The work began before the July "pacing letter" and the authors are struck at how quickly the world has moved in that direction. They joke it is "peer-reviewed — that is, we asked some AIs."
More from AGI Musings
- Gary Marcus: 'Rogue agents' is AI's excuse for irresponsibly built software — GaryMarcus · 2026-09-18
- Roman Yampolskiy returns to DOAC to update his AI safety warning that reached 20M+ viewers — MaxUnfried · 2026-09-18
- Ex-OpenAI's Yacine: enterprises are the next thinking machines — learn AI or get eaten — yacineMTB · 2026-09-18
- yacineMTB Predicts AI Tech Is Too Powerful for a Few to Control — yacineMTB · 2026-09-18
- Cell Focus: Biology Needs World Models That Predict Interventions, Not Just Descriptions — rishabh16_ · 2026-09-18
- AI Agents Find Drug Targets With Single-Cell Evidence Are 48% More Likely to Reach Market — james_y_zou · 2026-09-18