AI & Tech · National
OpenAI's chief scientist suggests a measure called perplexity could anchor shared industry limits on AI development
Jakub Pachocki argues that a measure of how well AI models predict the next token could become a common yardstick, letting companies cap progress until oversight improves, TechCrunch reported.
Listen to this brief · 4 min
Read aloud by an AI voice from the text below. Download MP3

What happened
OpenAI chief scientist Jakub Pachocki has published a blog post about possible limits on AI development, TechCrunch reported on November 22, 2026. The post centers on a measurement called perplexity. Pachocki suggests it could serve as a shared way to track how fast AI systems are advancing.
OpenAI CEO Sam Altman praised the post. According to TechCrunch, Altman described it as "the most important thing OpenAI has published in a while."
Why slowing AI is being discussed
TechCrunch reported that the idea of slowing AI development drew more attention after a former OpenAI and Anthropic researcher said the companies were "gambling with our lives." Anthropic's own IPO filing also suggested the technology could pose "catastrophic or existential risks to society," the outlet noted. Our earlier brief covers how Anthropic's IPO puts its AI safety pledge to the test.
Amodei's plan to pace the frontier
Anthropic CEO Dario Amodei then released a plan to "pace the frontier." TechCrunch said Altman and SpaceX CEO Elon Musk appeared to endorse it. Among other proposals, Amodei asked companies to agree to "common safety standards as well as limits on the rate of unchecked AI progress."
TechCrunch pointed to a practical problem with that approach. Setting a speed limit on AI requires agreement on how to measure speed in the first place. Pachocki's post offers an answer.
How perplexity works as a yardstick
In this context, perplexity is a technical term. It is not the AI company that shares the name, TechCrunch noted. Pachocki's post says it "measures how well a model predicts the next token on average; lower perplexity means better predictions."
Why not use capability tests
Pachocki argues that benchmarks built around AI's "downstream capabilities" have quickly gone out of date as the technology improves. In his view, perplexity has stayed "informative" even as models have advanced. That staying power is why he presents it as a possible common standard.
What OpenAI's own numbers show
The post includes a chart showing that the perplexity of OpenAI's models has fallen since GPT 3.5, according to TechCrunch. Pachocki wrote that GPT-4, at about 0.91, was good at using chain-of-thought reasoning. He said that ability led to the creation of modern reasoning models.
He tied lower scores to new abilities. "Now, as we have reached losses below 0.8, we have seen the emergence of transformative capabilities in fields like cybersecurity and mathematics," Pachocki wrote, as quoted by TechCrunch.
The proposed limit
Pachocki suggests perplexity could form the basis of "a shared benchmark for general intelligence." AI companies could then use that benchmark to set "common limits," TechCrunch reported.
A threshold tied to oversight
Under the idea, companies would agree not to push their models past a set perplexity threshold. The pause would last "until our ability to understand and monitor AI behavior achieves a certain concrete milestone," according to the post. TechCrunch's report does not say what that threshold or milestone would be.
The proposal links the pace of AI progress to how well developers can understand and monitor their systems, rather than to any single product test. For more coverage of AI policy and industry moves, visit our AI and tech section.
What to watch next
- Whether AI companies agree to use perplexity as a shared benchmark for general intelligence, as Pachocki proposes.
- What specific perplexity threshold and what "concrete milestone" for monitoring AI behavior would apply, which TechCrunch's report does not specify.
- Whether Amodei's plan to "pace the frontier," which TechCrunch said Altman and Musk appeared to endorse, leads to common safety standards and limits among companies.
