Technology

OpenAI's Chief Scientist Proposes Perplexity as a Speed Limit for AI Progress

Martin HollowayPublished 3 min readBased on 1 source
Reading level
OpenAI's Chief Scientist Proposes Perplexity as a Speed Limit for AI Progress
Photo by UK Prime Minister / CC BY 2.0

OpenAI chief scientist Jakub Pachocki has proposed using perplexity as the shared measure for setting limits on how fast AI development proceeds. In a blog post published on November 22, 2026, he argued that AI companies could agree on a common metric and commit to restraint once it reaches a defined threshold, according to TechCrunch.

Perplexity is a long-standing metric for scoring how well a language model predicts a sequence of tokens, the chunks of text a model reads and writes. A lower score means the model is less surprised by the next token. Pachocki's proposal goes further than using it as a model-quality score. He proposed it as the agreed measurement for the speed of AI progress, the quantity against which a speed limit could be set. He also framed it as a shared benchmark for general intelligence that AI companies could use to establish common limits.

The restraint he described is conditional. Pachocki proposed that AI companies commit not to go past a certain perplexity threshold until their ability to understand and monitor AI behavior reaches a concrete milestone. The proposal ties the pace of capability growth to progress in interpretability (the study of what happens inside a model) and monitoring, rather than to a fixed calendar or a compute budget.

OpenAI CEO Sam Altman described the post as "the most important thing OpenAI has published in a while."

Separately, Anthropic CEO Dario Amodei published a plan to "pace the frontier." The plan calls for common safety standards and limits on the rate of unchecked AI progress. Both proposals treat the rate of progress, not only its direction or its safety properties in isolation, as something that can be governed.

The broader context here is the measurement problem. A limit is only as useful as the number it is written against, and perplexity has known sensitivities. The value depends on the evaluation dataset, the tokenization scheme and the context length used to compute it, so two labs reporting the same nominal figure may not be measuring the same thing. Any shared benchmark would need a standardized evaluation protocol before a threshold could be enforced or verified by a third party. Perplexity is also a next-token prediction metric. Whether it tracks what the industry means by general intelligence is contested, and Pachocki's framing asks the industry to accept that equivalence for the purposes of coordination.

In my view, the more interesting feature of the proposal is its trigger condition. Tying a pause point to monitoring capability rather than to a raw capability number puts the emphasis on the ability to see what a model is doing. That makes the proposal dependent on progress in interpretability, evaluation tooling and behavioral auditing, which are themselves still developing. A threshold that cannot be checked is hard to hold a company to, so the milestone is where the proposal will be tested.

The obvious open question is adoption. Both proposals were published by leaders of frontier labs that compete with one another, and each would need counterparts to accept a common yardstick and a common commitment. Whether other developers, standards bodies or governments take up either framework is not yet settled. The proposals do, however, give the field a concrete starting point for a conversation that has mostly been conducted in general terms.

In my view, a shared metric, even an imperfect one, gives labs, auditors and regulators something specific to argue about, and that is a prerequisite for any enforceable coordination.