OpenAI Publishes Internal Pretraining Perplexity Data, Calls for Industry-Wide Safety Limits
sama · x · 2026-11-23
OpenAI Chief Scientist Jakub Pachocki published "Perplexity and Intelligence", disclosing OpenAI's internal pretraining perplexity measurements for the first time. He frames AI progress as training better compressors/predictors, arguing perplexity (cross-entropy loss) should be the cross-lab metric for a shared safety standard.
- Core proposal: industry "common limits" based on perplexity and other measurements, which developers commit not to cross until understanding/monitoring of AI behavior reaches concrete milestones (e.g., a successor to CoT monitors)
- OpenAI acknowledges publishing the data gives competitors a clearer picture of its standing, but deems shared safety measures more urgent
- Sam Altman called it the most important thing OpenAI has published in a while
More from Models
- Open vs closed models: the six user-side advantages of self-hosting, summed up in a viral tweet — gnukeith · 2026-11-23
- Same weights, one V100: inference engines differ up to 5x in tokens/s — PuzzleheadedArm3732 · 2026-11-23
- Open-weight models take under 1% of tokens on T3 Code, debunking margin-erosion FUD — thedealdirector · 2026-11-23
- Tesla FSD V15 clip raises questions: how far can pure vision really go? — ezshine · 2026-11-23
- Do Emotions Drive Agent Behavior? Speculation That Blocking Strong Feelings Constrains Agency — mimi10v3 · 2026-11-23
- Writers Push Google for Gemini 18+ Age Verification and Content Toggles to Stop Exodus to Grok — Hikkynure · 2026-11-23