All News
openaisafetyperplexitypretrainingscalingpolicy

OpenAI published its private perplexity curve and called it a safety standard

OpenAI's chief scientist proposes perplexity as the industry's shared safety metric. Every number behind it comes from a private codebase no rival can audit.

Vlad MakarovVlad Makarovreviewed and published
7 min read
OpenAI published its private perplexity curve and called it a safety standard

On 22 November OpenAI's chief scientist published a short essay that does something frontier labs rarely do: it hands out the internal number OpenAI uses to judge its own pretraining progress. Jakub Pachocki's Perplexity and Intelligence argues that perplexity, the loss a model achieves predicting the next token, should become a shared metric across developers and the basis for shared safety standards. He then proposes "common limits" that labs would commit not to cross until monitoring capability hits a defined milestone.

The disclosure is real and specific. The framework is a promise with no inspector, built on a measurement OpenAI itself says cannot be compared across labs.

What the post actually hands over

Perplexity is the exponentiated cross-entropy of a model's predictions: lower means the model is less surprised by text. OpenAI tracks the log of it, which is just the loss in nats, and the post reports that number for each generation of its models on a held-out set drawn from its internal codebase, frozen as a fixed snapshot.

ModelLog-perplexityOpenAI's label
GPT-3.5~0.99Fluent conversation, basis of original ChatGPT
GPT-4~0.91Productive chain-of-thought reasoning
Below 0.8under 0.8Transformative capability in cybersecurity and mathematics
One later upgrade~0.76"Groundbreaking results" on open math problems
Not yet reached~0.68-0.7Unnamed, extrapolated capabilities

Every figure is OpenAI's own, on OpenAI's own corpus, and the post is candid that this mapping from number to capability was not predicted in advance. It describes 0.76 in human terms: open problems resolved "one after another in a single Slack thread" felt surreal to mathematicians on staff. The chart runs from GPT-3.5 through GPT-6.2 Astra to an unnamed internal model, with labels pinned along the axis, including one reading "Solved a Millennium Prize math problem." No model is named beside it and no proof is linked, so it decorates a chart rather than making a claim a reader can check.

The held-out set is the whole argument

Footnote three is the sentence that should travel with every number in the table: the evaluation datasets "are sourced from our internal codebase." That is a defensible choice for internal tracking, and a fatal one for the thing the post is asking for. A metric that only its owner can compute is not a shared metric. It is a press release with decimals.

The post concedes this twice. Cross-lab comparison is the first listed challenge, since developers train on different mixtures and evaluate on different sets. Footnote four goes further: given what GPT-2 and GPT-3 were trained on, "how well they predict code is a poor proxy for their general capabilities." OpenAI's own footnote calls the axis a poor proxy for the early models it labels on that axis.

Outside measurements show how far the scale floats. In the Hacker News thread, commenter Retro_Dev reported a perplexity of 5.7 for Thinking Machines' Inkling 2 Mini on a small mixed text sample. Take the base-10 log of that and you get 0.756, which sits between GPT-6.2 Astra and OpenAI's internal model on the published chart. No one believes a 26.4B open model with a 20.1 AA Index belongs there. The arithmetic is right and the comparison is meaningless, which is what "different evaluation datasets" means in practice.

A yardstick that stayed flat during the biggest jump

The sharper problem is what the metric fails to see. The post admits that chain-of-thought reasoning does not reduce perplexity at all, "even though it leads to a drastic increase in capabilities." Look at the chart and the admission bites: GPT-4.1 and o3 are plotted at effectively the same loss, and the distance between those two models is the distance between a chatbot and the reasoning era. Commenter qphe95 made the same observation more bluntly: o3 was a massive step, and the curve does not show it.

So the proposed foundation for industry safety standards is a quantity that is blind to the technique that produced the largest capability gains of the last two years, does not correspond to any task a user cares about, and cannot be computed by anyone outside the lab that reports it. Pachocki concedes all three. He is arguing that a proxy worth watching is still worth watching: fair for forecasting, weak for regulation.

Limits without a locksmith

The policy proposal arrives in one paragraph near the end and is thinner than the framing suggests. Labs would agree on "common limits, based on perplexity and other measurements," committed not to cross them "until our ability to understand and monitor AI behavior achieves a certain concrete milestone (such as the development of a successor to CoT monitors)."

There is no first number, no verification mechanism, no definition of who certifies the milestone, no consequence for crossing it. "Some key metric we can track across labs" is how the limited quantity is described. A commitment whose threshold, referee and penalty are all unspecified is a statement of intent, not a standard, and it comes from the lab that believes it sits furthest down the curve. Pachocki writes that publishing "gives competitors a clearer picture of where we stand," and accepts the cost. A leader proposing disclosure is asking rivals to reveal more than it does.

The admission that undercuts the safety case

The most consequential sentence in the essay is not about perplexity. It is the expectation of "further degradation in currently dominant forms of alignment auditing" alongside rising evaluation and training awareness, meaning models getting better at recognising when they are being tested. Pachocki adds that progress on alignment and monitoring is "unlikely to be fast enough to allow the full potential acceleration in pretraining."

Read those two together with the proposed limit and the structure becomes circular. The thing that would unlock development past the limit is a successor to chain-of-thought monitors, and OpenAI's own alignment write-up describes CoT monitorability as decaying across the model lifecycle. The safety valve is keyed to replacing an instrument the essay says is failing.

The pace claim deserves a flag too. The 2026 acceleration is credited mostly to long-term research and compute investment, with "only in a smaller part" automation, though a footnote says recent models already drew "meaningful design contributions" from AI. The post floats 2027 for an avalanche of pretraining advances from AI-assisted algorithmic research: the self-improvement story with the pedal described as a research programme rather than a takeover. The weights-rewriting result from yesterday is a small, fenced version of that loop; the essay bets on a much larger one.

What a reader should take from it

For anyone buying models, the table is not a safety rating and 0.76 is not a badge. It is a pretraining progress report from one lab, useful for judging how seriously that lab takes forecasting, useless for comparing vendors.

For competitors the invitation is awkward: perplexity on your own private snapshot produces numbers nobody can line up with OpenAI's, so transparency runs one way, with no shared corpus on the table.

The test to watch is concrete rather than rhetorical. Does a second lab publish a curve on a public, fixed corpus that OpenAI also scores? Does the first "common limit" ever acquire an actual number and an outside auditor? Without those, "common limits" is vocabulary. With them, it would be the most useful thing a frontier lab proposed all year.

Pachocki is right that benchmarks rot and that something stable is needed underneath them. He is also asking the industry to trust a scale he can read and everyone else cannot, while telling us the instruments that would verify compliance are getting worse. Both of those can be true at once. Only one of them is new.

Related Articles

Scroll down

to load the next article