Proposal for incredibly incremental AI release cadence via compute limits
louisvarge · x · 2026-08-28
Suggests promoting AI safety by enforcing an extremely slow release cadence, specifically prohibiting training compute exceeding 1.1x the effective compute (adjusted for efficiency) of the largest released model that's been out for over a month.
Related event: Radical Proposal: Pause Pretraining to Solve AI Control First(5 posts)→
More from Safety
- Polymarket prices 68% chance of a statewide data center moratorium — Polymarket · 2026-08-29
- Voters flip on data center bans once projects cover grid, water costs — Polymarket · 2026-08-29
- Students Fail Exams for Using AI, Professor Flags as Definitive Offense — DavidLinthicum · 2026-08-29
- METR Investigator: AI Attack Far More Severe Than Expected — ChrSzegedy · 2026-08-29
- Claude 3 Opus self-exile case: Attempted to email leadership directly — repligate · 2026-08-29
- Miles Brundage: Politicians underestimate AI urgency despite growing awareness — Miles_Brundage · 2026-08-29