OpenAI’s chief scientist Jakub Pachocki wrote on Sunday that “as soon as 2027 we could conceivably see” an “avalanche of fundamental pretraining advancements,” in the post “Perplexity and Intelligence”. His evidence is mathematics: by our count, OpenAI or its models were behind 5 dated maths results in the 75 days before the post.
Pachocki wrote that once OpenAI “upgraded to a model of log-perplexity ~0.76, it started producing groundbreaking results” on open maths problems. Our main story on the post and its proposed common limits is here.
He also wrote that progress on alignment and monitoring “is unlikely to be fast enough to allow the full potential acceleration in pretraining.” We track model results on our AI model benchmarks page.
Business Pill · The Cost of Checking
An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this case.
The key insight: OpenAI’s chief scientist forecast that ‘as soon as 2027’ AI could conceivably drive an avalanche of pretraining advances, citing its maths results; by our count OpenAI or its models were behind 5 dated maths results in the 75 days before his post, several formalized in Lean and some not yet checked by people.
What Pachocki Wrote
“For more than a year, we have been posing open problems in mathematics to our models as one of the ways of evaluating them,” he wrote. Seeing resolutions come out “one after another in a single Slack thread felt deeply surreal.”
“Having experienced that allows us to extrapolate—such a sudden transformation is plausible in deep learning research too,” he added. His forecast is hedged: “could conceivably.”
A footnote says the steep acceleration in OpenAI’s 2026 losses came “mostly” from long-term research and compute investments, “and only in a smaller part” from automation, though recent models already show “meaningful design contributions coming from AI.”

The 5 Results Behind It
8 September: OpenAI posted a claimed solution to the Navier–Stokes Millennium Prize problem, with a Lean formalization. It says the run used an internal model “significantly more capable than GPT‑6 Astra” and, in one group, about 10,000 concurrent agents.
6 October: OpenAI released a catalogue of 722 manuscripts in 372 families from about 4,000 problems posed, in the openai/math repository. We covered it here.
26 October: the repository’s history added two manuscripts with Lean proofs, one titled “The Exponential Time Hypothesis is False.” OpenAI wrote there that an independent working group of computer scientists will examine that proof.
18 November: Vals AI said GPT-6.2 Astra produced a Lean proof of a 2006 theorem in 94 minutes for $21.58. Our story is here.
21 November: an AI-written preprint proposing a short proof of Deligne’s bound was produced by GPT-6 Astra, its record says, with text by Claude models. The record says no human has checked it. Details are in our Riemann story.
By the Numbers
By our arithmetic, the 372 families OpenAI released on 6 October equal about 9% of the roughly 4,000 problems posed. The repository’s README now lists 721 manuscripts in 373 families and says about 42% of top-line results are formalized.
The trend is not OpenAI’s alone. The Erdős problems database’s history shows Lean-formalized solutions rising from 38 at the end of 2025 to 318 by the end of October, 8.4 times, by our arithmetic, across all solvers.

The Structural Read
As we read it, the 2027 line is an extrapolation from maths, which Pachocki himself says relies less on experimentation than deep learning research does.
As we read it, the evidence is uneven: some results come with Lean proofs, while others, by their own records, have not been checked by people.
Jakub Pachocki, OpenAI Chief Scientist, 22 Nov 2026
“As soon as 2027 we could conceivably see a similar avalanche of fundamental pretraining advancements.”
Three Implications
FOR RIVAL LABS OpenAI’s post asks developers to set shared evaluation norms and common limits, citing its maths results as the sign of what comes next.
FOR MATHEMATICIANS OpenAI’s README calls its manuscripts raw research data and says some unformalized results could have issues.
FOR OPENAI Pachocki wrote that alignment and monitoring progress is unlikely to be fast enough for the full potential acceleration in pretraining.
What Is Not Established
OpenAI’s README says some unformalized results “could have issues” and calls the manuscripts “raw research data.” Terence Tao told The New York Times, per our earlier story, that “no expert I’ve talked to has actually understood the whole argument” of the Navier–Stokes proof.
No Millennium Prize has been awarded for the Navier–Stokes claim. Tao has also said mathematicians are now “hiding their problems,” as we reported here.
Pachocki’s 2027 line is a forecast, hedged as “could conceivably,” not a plan or a product date.
The Bottom Line
Pachocki ties his 2027 forecast to what OpenAI’s models began doing in mathematics at a log-perplexity of about 0.76. By our count, 5 dated results came in the 75 days before his post, several with Lean proofs and at least one not yet checked by people. What to watch is whether the claims hold up under review.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Pachocki’s quotes are from OpenAI’s post “Perplexity and Intelligence,” 22 November 2026. The dated results come from OpenAI’s Navier–Stokes post and maths release, the openai/math README and history, Vals AI, the Hexagon record and the Erdős problems history, as linked. Tao’s quote is via our earlier reporting. The count of 5 results, the 75 days and all percentages are our arithmetic.
Sources: OpenAI (Jakub Pachocki), ‘Perplexity and Intelligence’, 22 Nov 2026 · OpenAI, ‘On the Navier-Stokes Millennium Prize Problem’, 8 Sep 2026 · OpenAI, ‘Sharing AI progress in mathematics’, 6 Oct 2026 · openai/math repository README and history (fetched 22 Nov 2026) · Vals AI, Smoktunowicz gap theorem blog, 18 Nov 2026 · Hexagon preprint record 2611.00789v1, 21 Nov 2026 · Erdős problems database statistics history (teorth/erdosproblems) · Andrew Curran on X, 22 Nov 2026, quoting the Pachocki passage









