News

OpenAI chief scientist proposes shared limits on frontier AI development

Under the proposal, labs would use perplexity and other measures to set a first threshold and wait for a defined advance in model monitoring before crossing it.

D
Nov 22, 2026 · 1 min read

OpenAI chief scientist Jakub Pachocki proposed on November 22 that frontier AI developers agree on a shared development limit. Under his proposal, labs would commit to staying within that limit until they reach a milestone in understanding and monitoring model behavior.

Labs would set the first limit using perplexity and other measures they could track in common. Perplexity measures how well a model predicts the next token, or unit of text, on average. A lower score means better predictions on the dataset tested.

Pachocki gave a successor to current chain-of-thought monitors as one example of a milestone. Chain-of-thought monitors examine the reasoning steps a model generates for signs of unwanted behavior. OpenAI says it uses those traces to monitor some deployed models.

The proposal specifies neither a numerical cutoff nor a test for whether a successor would meet the milestone. Pachocki said he hoped the first limit would lead to a framework for setting further limits.

Pachocki pointed to differences in both the data labs train on and the datasets they use for evaluation. When labs test on different datasets, a lower score may reflect text that was easier to predict, making it hard to compare results directly.

OpenAI published historical pretraining measurements from a fixed set of examples drawn from its internal codebase and held out of training. On that dataset, Pachocki reported log-perplexity, also called cross-entropy loss, of about 0.99 for GPT-3.5 and about 0.91 for GPT-4.

More news