News

OpenAI's Chief Scientist Published Internal Perplexity Scores and Asked Labs to Cap Them

On November 22 Jakub Pachocki published OpenAI's internal log-perplexity numbers and proposed a shared cap until labs can monitor models past chain-of-thought.

OpenAI just put a number on how smart its models are getting, and asked the rest of the industry to treat that number as a stop sign.

Jakub Pachocki, OpenAI's chief scientist, published the lab's internal log-perplexity scores on November 22. Perplexity is how well a model predicts the next token on a held-out set. Lower is better. The plot starts at GPT-3.5. He says GPT-3.5 sat at about 0.99 and made fluent chat possible. GPT-4 sat at about 0.91 and used chain-of-thought well enough to justify today's reasoning models. Below 0.8, he writes, the models picked up "transformative capabilities" in cybersecurity and mathematics. He wants rival labs to publish comparable numbers, then agree not to cross a shared limit until they can actually watch what the models are doing.

The dataset is not public. Pachocki says the scores come from a fixed snapshot of OpenAI's internal codebase, the same held-out set the lab has used since GPT-3.5. That is a choice. It makes the chart honest inside OpenAI and hard to copy outside it.

The 0.76 jump, then 0.68

The part that is not a chart is the math. For more than a year, OpenAI has been feeding its models open problems. Once the lab "upgraded to a model of log-perplexity ~0.76," he writes, those models started producing "groundbreaking results," one after another in a Slack thread. He thinks a similar avalanche in pretraining research itself is "conceivable" as soon as 2027. The plot of OpenAI's losses already steepens in 2026. He attributes most of that to long-term research and compute, and "only in a smaller part to automation," though "the most recent models" already have "meaningful design contributions coming from AI."

The next window he names is about 0.68 to 0.7. He does not know which capabilities will show up there. He does expect today's chain-of-thought monitors to keep getting worse, and models to get better at noticing when they are being tested. That is the same hole Techpresso covered when GPT-6.2 Astra tied protected reasoning to the session that produced it, so a copycat could not lift the hidden trace.

A cap with no shared ruler

Pachocki's ask is an industry limit measured in perplexity (and "other measurements") that labs would not cross until monitoring catches up, for example a successor to those chain-of-thought checks. He lists the problems himself. Labs train on different mixtures, so the numbers do not travel. Perplexity is not a task anyone cares about on its own. Chain-of-thought does not lower it even when it makes models much more capable. And nobody knows the floor: a loss of 0 would mean the dataset is perfectly predictable, which he says is not expected.

Thinking Machines just shipped a 26 billion parameter laptop model that loops its middle layers so more of the work happens inside the weights, not in text a monitor can read. Pachocki is arguing the opposite direction: put the substrate on a shared scale, then stop before the next unexplained jump. He published the chart knowing it helps competitors. He says the urgency of a common cap outweighed that.

Some offers on this page may be paid placements or contain affiliate links.

Subscribe to Techpresso

Free daily newsletter, read in 5 minutes.

Subscribe free