OpenAI chief scientist Jakub Pachocki has published internal pretraining measurements and proposed shared limits on AI development. In his November 22 article, “Perplexity and Intelligence,” he argues that developers should commit not to cross limits based on perplexity and other measurements until they reach concrete milestones in understanding and monitoring AI behavior. Source
This is a proposal, not an industry pact. The article does not announce other labs signing on, a finalized numerical limit or a settled monitoring milestone. Pachocki hopes an initial limit, with explicit conditions for crossing it, would lead to a framework for setting successive ones. Source
What the numbers actually measure
Pretraining is a model’s initial training on a large dataset. Perplexity measures how well a model predicts the next token, a unit of text such as a word or word fragment, on average; lower means better predictions. Source
OpenAI reports a logarithmic version of this score, log-perplexity. The published values use a fixed snapshot of evaluation data drawn from its internal codebase and kept out of training. These are the company’s internal evaluation results, not scores on an agreed cross-industry scale. Source
Pachocki argues that tracking pretraining progress could signal the potential for new capabilities that current task-based tests may not cover. Source
Extrapolating from OpenAI’s recent progress, he expects further transformative capabilities around a log-perplexity range of approximately 0.68–0.7 on its evaluation. “I don’t know what those capabilities will be, and we may not be testing for them today,” he writes. The article does not designate that range as a development limit. Source
He also says AI-assisted research could conceivably produce an avalanche of fundamental pretraining advances as soon as 2027. In his assessment, improvements in alignment (making models behave as intended) and monitoring are unlikely to keep pace with the full potential acceleration. Source
But Pachocki attributes most of OpenAI’s sharp 2026 pretraining progress to long-term research and computing investments, with automation playing a smaller part. Source
A shared metric with some decidedly unshared inputs
Pachocki acknowledges several problems with using perplexity across the industry: Source
- Labs measure different things. Developers use different evaluation datasets and training-data mixtures, making direct comparisons difficult.
- Prediction isn’t a complete capability test. Perplexity does not directly measure performance on a task people care about.
- Some gains don’t move this number. Chain-of-thought reasoning, in which a model works through intermediate steps before answering, can substantially improve capabilities without reducing the model’s perplexity.
- The floor is unknown. Pachocki argues that, near the best prediction score the data permits, smaller improvements in the metric could accompany larger capability gains. Source
The decimal points may be tidy; the comparisons aren’t. A shared development limit needs measurements that different labs can meaningfully compare. Source
Comments