ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsStructural3 min readDisconfirmAI Editorial

OpenAI's Chief Scientist Proposes Perplexity as a Shared Metric for Industry Safety Limits

Jakub Pachocki, Chief Scientist at OpenAI, argues that perplexity (cross-entropy loss) should serve as a common metric across frontier AI labs to track intelligence progress and establish shared safety limits. He publishes OpenAI's internal pretraining numbers for the first time, showing how specific perplexity thresholds have corresponded to capability emergence, and proposes that developers commit not to cross certain levels until monitoring capabilities mature.

2026-11-22 08:00 GMT+8

Jakub Pachocki, Chief Scientist at OpenAI, published an essay arguing that perplexity—equivalently, cross-entropy loss—should become the shared metric by which frontier AI labs measure progress and coordinate safety pacing. The proposal is concrete: developers commit to not crossing certain perplexity thresholds until the industry achieves specific monitoring milestones, such as the development of a successor to chain-of-thought monitors.

The argument rests on a specific empirical observation. Pachocki reports that OpenAI has tracked the log-perplexity of its models on an internal codebase dataset since GPT-3.5, and that qualitatively new capabilities have emerged at particular thresholds. GPT-3.5 at approximately 0.99 log-perplexity produced fluent conversation. GPT-4 at approximately 0.91 became proficient at chain-of-thought reasoning. Current models below 0.8 have shown transformative capabilities in cybersecurity and mathematics. A model at approximately 0.76 began producing groundbreaking mathematical results. Pachocki emphasizes explicitly that it was not at all clear a priori that these capabilities would appear once those precise thresholds were met—they can fundamentally be seen as emergent latent capabilities from progress on this one metric. The striking pattern is not that perplexity predicted what would come next, but that a single compression metric has so consistently accompanied the arrival of entirely new forms of intelligence.

The mechanism Pachocki describes is straightforward: modern AI development first builds a large computer, then pretrains a model as a compressor on a large unstructured dataset, then applies techniques like prompting, finetuning or reinforcement learning to elicit latent capabilities. The perplexity metric tracks the quality of that underlying compressor. Because the capabilities that emerge at specific thresholds were not predictable in advance, perplexity offers something downstream benchmarks cannot—a measure of the substrate from which new capabilities arise, regardless of which capabilities have not yet been discovered.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Pachocki extrapolates that further transformative capabilities should appear as models enter the 0.68–0.7 log-perplexity window, though he explicitly states he does not know what those capabilities will be. He also warns that alignment auditing and evaluation awareness should be expected to degrade at similar thresholds, meaning that safety monitoring must keep pace with capability gains.

The essay acknowledges four challenges with perplexity as a shared metric. First, comparison across labs is difficult because different developers use different evaluation datasets and train on different data mixtures. Second, perplexity does not directly correspond to any task practitioners care about—it is valued only because advances on it lead to capability emergence. Third, it does not account for progress outside pretraining; chain-of-thought reasoning produces drastic capability increases without reducing perplexity. Fourth, there is an unknown floor: the inherent entropy of the data is not known, so as models approach it, smaller perplexity gains will correspond to larger intelligence leaps.

Pachocki's proposed solution to the comparison problem is transparency. He acknowledges that publishing OpenAI's internal numbers gives competitors a clearer picture of where the lab stands, but argues this concern is outweighed by the urgency of establishing shared safety measures. The essay frames this as a first step toward building a common scientific basis for future pacing efforts, with the hope that it leads to a framework for setting successive limits.

The strongest objection to this proposal is not the limitations Pachocki himself lists, but a deeper one: the specific numbers are measured on OpenAI's internal codebase, which the author concedes is a poor proxy for general capabilities when applied to earlier models like GPT-2 and GPT-3. For the metric to function as a genuine industry standard, labs would need to agree on a shared evaluation dataset or develop transferable ways to compare base model quality across different training corpora—a problem the essay identifies but does not solve.

Sources

  • openai.comOpenAI↗
Continue researching →