KEY POINTS
- OpenAI chief scientist Jakub Pachocki says perplexity could be a shared measure of general intelligence.
- He proposes that companies commit not to pass a set threshold until safety understanding reaches a concrete milestone.
- Sam Altman called the post the most important thing OpenAI has published in a while.
OpenAI chief scientist Jakub Pachocki has published a blog post arguing that a metric called perplexity could be used to set limits on the pace of AI development, TechCrunch reported on Nov. 22, 2026.
The idea arrives as the question of slowing AI has become more prominent. According to TechCrunch, a former researcher at OpenAI and Anthropic said the companies were "gambling with our lives," and Anthropic's IPO filing acknowledged that the technology could pose "catastrophic or existential risks to society."
Anthropic CEO Dario Amodei then released a plan to "pace the frontier." TechCrunch said the plan appeared to be endorsed by OpenAI CEO Sam Altman and Elon Musk, the SpaceX CEO. Amodei called on companies to agree to common safety standards and limits on the rate of unchecked AI progress.
Any such speed limit requires an agreed way to measure speed, and that is the gap Pachocki's post tries to fill. Altman described it as "the most important thing OpenAI has published in a while." The metric is not related to the company that shares its name, TechCrunch noted.
Pachocki argues that benchmarks centered on downstream capabilities tend to become outdated quickly as models improve. Perplexity, by contrast, has stayed informative. He defines it as a measure of how well a model predicts the next token on average, where lower values mean better predictions.
The post includes a chart showing how the perplexity of OpenAI's models has dropped since GPT 3.5. Pachocki wrote that GPT-4, at roughly 0.91, could use chain-of-thought reasoning productively, which led to targeted work on that skill and to modern reasoning models. He added that with losses now below 0.8, transformative capabilities have appeared in areas such as cybersecurity and mathematics.
On that basis, he suggests perplexity could form "a shared benchmark for general intelligence" that companies use to set common limits. Under the proposal, firms would commit not to go beyond a given threshold until the ability to understand and monitor AI behavior reaches a certain concrete milestone.
The report does not say whether other AI developers have accepted the proposal or how any threshold would be set.