Why this story
- Why it matters
- A shared stopping rule could give labs a way to coordinate safety decisions before more capable models are developed. It depends on agreeing both how to compare models and what evidence would justify proceeding.
- Attention gap
- Ori’s headline-index search returned no matches among the eleven U.S. outlets it checks. Bloomberg News reported the proposal on November 22. It is outside Ori’s named panel. The publisher did not independently establish the coverage status of all eleven outlets.
OpenAI’s chief scientist, Jakub Pachocki, proposed on November 22 that leading AI developers agree on a shared limit they would cross only after demonstrating better oversight of model behavior. One possible requirement would be a successor to tools that monitor a model’s reasoning. His essay announces a proposal and publishes internal results. It names no numerical stop point, deadline or agreement among rival labs.
The measure is log-perplexity. It tracks how accurately a model predicts the next piece of text in material held out of training. Lower scores indicate better predictions. On a fixed sample of OpenAI’s internal code, Pachocki reported roughly 0.99 for GPT-3.5 and 0.91 for GPT-4. He said a later model near 0.76 began producing results on unresolved mathematics problems. The company says it could see another jump around 0.68–0.70; that range is its forecast.
For other labs to use the measure, they would need a way to compare results across different training materials and test sets. A score does not establish whether a system can do a particular job, and gains from reasoning after training can leave it unchanged. Pachocki therefore proposes using other measurements alongside perplexity. Bloomberg News separately reported his call for an industry standard.
What the evidence cannot tell us
- OpenAI supplied the internal measurements and its interpretation of the mathematical results. Bloomberg reported the proposal and did not independently reproduce the tests.
What to watch
Whether another lab publishes a comparable result and whether the developers agree on a first threshold and a way to verify the monitoring requirement.
Read the sources. Credit the work.
The reporting and documents behind this brief, linked directly to the source. Verified publisher and author credits appear where available; other entries show their actual website address. Several links may draw on the same underlying reporting or research.
OpenAI
Perplexity and Intelligence (opens in a new tab)openai.com · November 22, 2026
Primary document / sourceContent read for this publication · November 22, 2026Publisher and byline credit checked November 22, 2026. View credit on the source page (opens in a new tab).
Bloomberg News
OpenAI Chief Scientist Urges Industry Limits on AI Development (opens in a new tab)news.bloomberglaw.com · November 22, 2026
Reporting / analysisContent read for this publication · November 22, 2026Publisher and byline credit checked November 22, 2026. View credit on the source page (opens in a new tab).
This story was written and checked using AI against the linked evidence. No routine human review is claimed. Read the method.
Suggest an update or correctionPublished corrections