OpenAI Chief Scientist Proposes Shared AI Metric

Jakub Pachocki suggests cross-entropy loss per token as a common pretraining measure that labs could use to compare models.

Updated on Nov. 22, 2026 in Artificial Intelligence

OpenAI Chief Scientist Proposes Shared AI Metric

Live Poll

Should AI companies slow down development until safety standards are better understood?

OpenAI Chief Scientist Jakub Pachocki proposed perplexity as a shared pretraining metric for AI labs in a post titled "Perplexity and Intelligence" published November 22, 2026. He framed the metric as cross-entropy loss per token, allowing outside observers to compare models using a single shared number.

Why it matters

Labs currently promote their models with a patchwork of benchmarks, each with its own quirks, making direct comparison difficult. A common metric could make model progress easier for outsiders to assess across the industry.

The proposed measure is cross-entropy loss per token, and OpenAI's internal work linked a log-perplexity of approximately 0.76 to breakthroughs on complex math problems. Whether other labs will adopt the metric, and whether the 0.76 threshold holds across models, remains unclear.

The players

OpenAI

OpenAI is an artificial intelligence research company whose models and products are among the most widely used in the industry.

Jakub Pachocki

Jakub Pachocki is OpenAI's Chief Scientist, a role he has held since May 2024 after succeeding Ilya Sutskever.

Sam Altman

Sam Altman is the chief executive of OpenAI and one of the most prominent figures in the commercial AI industry.

The details

Pachocki also called for limits on further AI scaling until monitoring improves and for voluntary pauses in development until safety standards are upheld. OpenAI CEO Sam Altman publicly backed the transparency initiative, which Pachocki has pursued since succeeding Ilya Sutskever as Chief Scientist.

Timeline

  1. May 2024: Pachocki became OpenAI Chief Scientist, succeeding Ilya Sutskever.

  2. September 6, 2026: Pachocki published the essay "An Alien Mind" calling for industry slowdowns.

  3. November 22, 2026: Pachocki published "Perplexity and Intelligence" proposing a shared pretraining metric.

  4. 2027: Pachocki projected possible major pretraining breakthroughs.

The Tech Race

Pachocki's September 6, 2026 essay "An Alien Mind" called for industry slowdowns, and the November 22 metric proposal moves that argument toward a concrete, shared measurement. The shift matters because competitive labs have so far relied on incompatible benchmarks to signal progress.

If labs adopt a single shared pretraining metric, users and developers could gain a clearer basis for comparing AI models than the mix of benchmarks used today. The proposal remains a suggestion from one lab's chief scientist and has not yet changed how models are evaluated.

The takeaway

A single shared number for model pretraining would let outside observers compare systems directly rather than relying on each lab's chosen benchmarks. Whether that standard takes hold will depend on whether competitors agree to measure progress the same way.

Further reading

Explore more on Artificial Intelligence coverage across the industry.

Source note: This article includes information reported by Crypto Briefing.

Live Poll

Should AI companies slow down development until safety standards are better understood?