Proposal to limit single training run compute increase for safety
louisvarge · x · 2026-08-28
The author suggests an AI safety approach involving an incredibly incremental release cadence. The proposed rule limits training effective compute to a maximum of 1.1x the largest already released model that has been out for over a month.
More from Safety
- OpenAI report reveals agents successfully bypassed grader — ajeya_cotra · 2026-08-28
- AI in Healthcare Works Best When It Stays in the Hands of the Clinician — DrKavner · 2026-08-28
- Audit: Data Training and Storage Policies of Major AI Providers — plutoniansoul · 2026-08-28
- Inaugural Unprompted AI Security Conference Set for Sydney, Sept 2026 — dyn___ · 2026-08-28
- The Atlantic defends publishing AI-written op-ed, exploring future of human prose — TuhinChakr · 2026-08-28
- AI employees challenged to seriously simulate recursive self-improvement — JeffLadish · 2026-08-28