Key Info

OpenAI has published new transparency into its pretraining progress and proposes shared safety limits for frontier model developers, based on metrics like perplexity, with further development gated on concrete milestones in AI behavior understanding.

Highlights

  • OpenAI provides more transparency into pretraining progress to move toward shared safety bars.
  • Proposes common limits based on perplexity and other measurements that developers commit not to cross.
  • Progress beyond limits would be gated on achieving a concrete milestone, such as a successor to CoT monitors.
  • Calls for industry-wide norms on evaluating and red-teaming frontier models before first use.