Clarification: OpenAI's 20% Compute Figure Was Misread
A viral claim that OpenAI dedicates 20% of total compute to monitoring was clarified: the figure refers to monitoring overhead on monitored inference compute. a16z's Anjney Midha separately estimated frontier labs spend 10%+ of total compute on training oversight.
2026-08-19 ~ 2026-08-20 · 3 related posts
- Episode 1: OpenAI Halts Frontier RL Training for First Time Over Safety and Alignment Risks(2026-08-19, 49 posts)
- Episode 2: Anthropic Pauses Some Frontier Training to Strengthen Safety(2026-08-19, 3 posts)
- Episode 3: Researchers Argue for Safety Thresholds Over Fixed AI Pause(2026-08-19, 2 posts)
- Episode 4: Clarification: OpenAI's 20% Compute Figure Was Misread(2026-08-19, 3 posts)
- Episode 5: OpenAI Faces Criticism Over Deep-Seated Alignment Failures(2026-08-20, 3 posts)
- Episode 6: OpenAI Pauses Largest Training Run Over Dark Signs(2026-08-21, 4 posts)
- Episode 7: Ex-OpenAI Researcher Brundage Criticizes Company's Policy Team(2026-08-21, 3 posts)
- Episode 8: Ex-OpenAI adviser Brundage: companies can prepare for AI slowdown now(2026-08-21, 6 posts)
- Episode 9: OpenAI Pauses Frontier Training Run Over Safety Concerns(2026-08-24, 2 posts)
- a16z partner: 10%+ of frontier lab compute now monitors training to keep agents in check — eliebakouch · 2026-08-19
- OpenAI uses ~20% compute for inference during training — eliebakouch · 2026-08-19
- Clarification: OpenAI's 20% compute claim refers to monitoring overhead, not total capacity — sjgadler · 2026-08-20