OpenAI opens up pretraining progress metrics, proposes perplexity-based safety bars
eliebakouch · x · 2026-11-23
An OpenAI researcher announced increased transparency into pretraining progress, publishing "Perplexity and Intelligence," which argues for common capability limits based on perplexity and other measurements, with further progress gated on the lab's ability to understand AI behavior. Commenters call it the first public pretraining progress metric from a frontier lab and a key signal for tracking pretraining-driven RSI.
More from AGI Musings
- Musk: superintelligence at exawatt scale will be like a caveman playing with sticks — elonmusk · 2026-11-23
- LeCun: we supposedly have superintelligence but still no level-5 autonomous driving — ylecun · 2026-11-23
- Pedro Domingos mocks AI anxiety: imagine if we actually had AGI — pmddomingos · 2026-11-23
- "Pure math is AGI-complete": viral argument says Riemann-proof AI can master any field — inductionheads · 2026-11-23
- Steep acceleration: medical-superintelligence AI could arrive by 2028 on current trajectory — Neurogence · 2026-11-23
- Math Copes Poorly With AI: The 'Just Pivot' Advice Finally Hits Home — AccBalanced · 2026-11-23