Epoch AI researcher maps the levers for pacing AI: RSI speed limits, compute caps, agent budgets
connoraxiotes · x · 2026-09-14
Epoch AI researcher Josh You published a taxonomy of mechanisms for deliberately slowing frontier AI development.
- Global measures: capping total AI R&D compute, limiting model capabilities via benchmark thresholds, and restricting recursive self-improvement (RSI).
- Targeted measures: FLOP limits on training runs, or slowing discovery/implementation of specific classes of AI innovation.
- Pacing RSI: either a speed limit on recursive development (e.g. internally deployed agents used for AI R&D must be >X months old, or weights refresh only at fixed frequency) or budget caps on internal agents, per-company or per-head.
He also flags the prior question of which dangerous capabilities to pace — cyberoffense, bioweapons research, or AI research itself — suggesting direct inhibition might suffice. He stresses the policy design problems are very hard and he holds no confident answers.
More from AGI Musings
- We Unite or We Fight: The Long-Term Case for International AI Governance — danfaggella · 2026-09-14
- iamtrask: Collectively Controlled AI Is Easier to Regulate With Less Regulatory Capture — iamtrask · 2026-09-14
- iamtrask: Nobody Can Be Trusted With Unilateral Superintelligence Control — Only Collective Control Works — iamtrask · 2026-09-14
- AI safety debate: dangerous AI must self-bootstrap physical affordances, and capability scaling is outpacing skeptics — lu_sichu · 2026-09-14
- CV researchers ask: is computer vision 'mostly done' in the age of large models? — CSProfKGD · 2026-09-14
- AI pause proposal: a mandatory simultaneous lab-wide vacation — herbiebradley · 2026-09-14