OpenAI Chief Scientist Proposes Perplexity as Industry Standard for AI Safety Pacing
·2026.11.23 02:06
Key point
Jakub Pachocki argues that cross-entropy loss is the most stable metric for tracking general intelligence and proposes common limits to pause development until monitoring capabilities catch up.
Details
- OpenAI publishes internal pretraining perplexity metrics to establish a shared scientific basis for industry safety standards and pacing.
- The author identifies log-perplexity (cross-entropy loss) as the primary driver of emergent capabilities, noting that losses below 0.8 have already led to transformative advances in cybersecurity and mathematics.
- A potential step change in AI development is expected from AI participation in algorithmic research on pretraining, which could cause an avalanche of fundamental advancements as early as 2027.
- Pachocki proposes common limits based on perplexity that developers should commit to not crossing until specific monitoring milestones, such as successors to Chain-of-Thought monitors, are achieved.
- The proposal acknowledges challenges including difficulty comparing across different evaluation datasets, detachment from actionable task evaluations, and the unknown floor of inherent data entropy.
- Current OpenAI models have reached log-perplexity levels where GPT-3.5 was at ~0.99, GPT-4 at ~0.91, and recent models are below 0.8, with further transformative capabilities expected in the ~0.68-0.7 window.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.