
Perplexity as a Shared Metric for AI Progress and Safety

OpenAI argues that the central axis of AI progress is pretraining better compressors, and that perplexity—equivalently cross-entropy loss—should become a shared industry metric for comparing models and setting safety limits. The post contrasts this view with the common impression that modern AI is built from many task-specific techniques, noting that most capabilities are elicited from a simpler, less interpretable process: training a predictor on a large corpus. The author proposes tracking perplexity as a way to understand the arc of advances in AI and to create shared safety standards.
To support the case, OpenAI publishes internal pretraining perplexity measurements for the first time. The post maps capability jumps to specific log-perplexity values on its main evaluation dataset: GPT-3.5 at roughly 0.99 enabled fluent conversation and the original ChatGPT; GPT-4 at roughly 0.91 made chain-of-thought reasoning productive; losses below 0.8 brought transformative capabilities in cybersecurity and mathematics; and a model around 0.76 began producing groundbreaking results on open math problems. The author stresses that these capabilities were not predictable in advance and should be seen as emergent latent capabilities driven by progress on a single metric.
The post also warns about a possible step change in AI development speed: AI contributions to algorithmic research on pretraining could trigger an avalanche of advances, plausibly as soon as 2027, similar to what the author observed in mathematics. While AI may help with alignment and monitoring, the author expects that progress will not be enough to keep pace with accelerated pretraining.
OpenAI therefore proposes that frontier developers jointly establish shared norms for evaluating models and red-teaming safety cases before first use, including transparency standards. It further proposes common limits based on perplexity and other measurements, with developers committing not to cross them until a concrete milestone is reached, such as a successor to chain-of-thought monitors. The goal is an industry standard that sets a first limit, specifies what must be achieved before going beyond it, and leads to a framework for successive limits.
The post acknowledges limitations of perplexity as a shared metric: different labs use different datasets and data mixtures, so cross-lab comparison is difficult; perplexity does not directly correspond to an actionable evaluation; it does not capture advances outside pretraining, notably chain-of-thought reasoning; and the inherent entropy floor of the data is unknown, meaning apparent progress will slow near the floor while each small gain becomes more significant. Despite these challenges, the author believes perplexity has stood the test of time and can play an important role in standardizing discussions of AI progress.

