OpenAI pushes for shared safety bars based on perplexity, opens up pretraining progress

daniel_mac8 · x · 2026-11-23

OpenAI says it will provide more transparency into pretraining progress for the sake of shared safety bars. In "Perplexity and Intelligence," it argues for common capability limits based on perplexity and other measurements, with progress beyond them gated on the ability to understand AI behavior.

Related event: OpenAI Discloses Pretraining Perplexity Data, Proposes Industry Safety Threshold(4 posts)→

Original post →

More from Models

Models channel →