Skip to content
Theme
OnDevice Tools

AIAI Models & Research

AI development limits proposed by OpenAI chief scientist

OpenAI chief scientist Jakub Pachocki has suggested using perplexity as a shared measure for AI progress, which could help companies agree on limits. The idea follows growing calls for safety standards and controls on unchecked AI development.

Illustration of an AI model chart showing falling perplexity and a proposed safety limit

Key takeaways

  • OpenAI chief scientist Jakub Pachocki proposed perplexity as a shared measure for AI progress.
  • TechCrunch says lower perplexity indicates better next-token prediction.
  • Pachocki links loss figures below 0.8 to advances in cybersecurity and mathematics.
  • The proposal follows warnings and calls for safety standards and limits.
In this article
  1. The new proposal
  2. How perplexity would work
  3. Safety concerns behind the idea
  4. A possible threshold for AI progress
  5. Why it matters

The new proposal

According to TechCrunch, OpenAI chief scientist Jakub Pachocki published a blog post on 22 November 2026. It discusses possible AI development limits. The post centres on perplexity, a technical measure used to judge how well a model predicts text. Pachocki suggests that this measure could become a shared benchmark for AI progress.

TechCrunch says Sam Altman, OpenAI's chief executive, described the post as highly important. The article presents the idea as a proposal, not as a confirmed company rule. It also notes that researchers and executives are debating how to control AI advances.

The topic has become especially heated after public warnings about powerful AI systems. The post does not provide a final rule, but it outlines how one could be designed.

How perplexity would work

Illustration: Perplexity logo, AI, crypto, OpenAI logo and the labels AI Safety, Anthropic, Dario Amodei

Perplexity estimates how accurately a model predicts the next token. A token is a small unit of text. A lower score indicates better predictions. TechCrunch notes that this technical measure should not be confused with the AI company that has the same name.

Pachocki argues that many benchmarks tied to later capabilities become outdated quickly. He says perplexity has stayed useful because it reflects a core prediction task. His post includes a chart showing how perplexity has fallen across OpenAI models since GPT 3.5.

He says GPT-4, at about 0.91, could use step-by-step reasoning effectively. This kind of reasoning helps models work through problems in stages. He says that encouraged later work on reasoning models. He also links loss figures below 0.8 to major advances in cybersecurity and mathematics.

Safety concerns behind the idea

The proposal appears during a wider debate over AI safety. TechCrunch says a former researcher at OpenAI and Anthropic warned that the companies were risking lives. A filing from Anthropic for its IPO also suggested that AI could create catastrophic or existential dangers.

Dario Amodei, chief executive of Anthropic, then published a plan to slow AI progress. He called for shared safety standards and limits on unchecked progress. TechCrunch says Sam Altman, OpenAI's chief executive, and Elon Musk, chief executive of SpaceX, appeared to support it.

These statements show that the debate is no longer limited to researchers. It now involves company leaders and public discussion about advanced AI systems. The warnings have made the question of speed limits more urgent for the industry.

A possible threshold for AI progress

In Pachocki's view, perplexity could provide a shared yardstick for general intelligence. Companies could use it to agree on common limits. They would commit not to pass a chosen threshold. The pause would last until they can understand and monitor AI behaviour at a defined milestone.

TechCrunch says this would turn a broad safety goal into a measurable test. The post does not name a specific threshold or timeline. It also does not say which companies would take part. The threshold would be a chosen perplexity level, and the milestone would relate to oversight.

The idea is therefore a starting point for discussion. It suggests that safety limits could be tied to technical progress rather than only to broad principles.

Why it matters

The proposal matters because it tries to connect AI safety with a concrete measure. If companies accepted a shared metric, they could compare progress and agree on when to pause. That could make debates about limits more practical. It also addresses a basic problem: groups may disagree about the speed of AI progress without a shared measure.

It also matters because the measure would be linked to monitoring and understanding. Pachocki's condition is that limits would hold until researchers reach a clear milestone in AI oversight. For readers, the key point is that the discussion is moving from warnings toward possible technical standards. The proposal could influence how companies describe progress and justify decisions to slow down.

Sources

  1. OpenAI’s chief scientist has thoughts on how to limit AI developmentTechCrunch

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.