OpenAI Chief Scientist proposes perplexity as a standard for AI safety
November 22, 2026
Jakub Pachocki argues that tracking training better compressors via perplexity or cross-entropy loss is the best way to measure AI progress. He suggests establishing common perplexity limits to inform shared safety standards across model developers.
HOW THIS AFFECTS YOU
●
researcherYou might see more companies publishing loss curves to benchmark their models.
●
policyThis could lead to new industry standards for when model training should be gated based on performance metrics.