Physix Frontier · News Briefing Card (TMTPost · Oct 9, 2026)

OpenAI Internal Model Produces 722 Math Manuscripts in 3 Hours

KEY FACTS

An unreleased OpenAI internal model generated 722 mathematical manuscripts, grouped into 372 result families.
These manuscripts cover classic hard problems including the Kakeya problem, the quasi-Riemann hypothesis, and Hilbert's tenth problem.
The model consumed roughly 3 hours of ChatGPT Pro thinking compute per problem on average, evaluating about 4,000 problems.
The manuscripts claim to prove that the Hausdorff dimension of the four-dimensional Kakeya set is 4, versus the previous human best of 3.059.
The manuscripts hit 10,000 GitHub stars within 24 hours of release, with the model and prompts kept undisclosed.

KEY DATA

722Manuscript count

372Result families

~3 hoursAverage time per problem

3.059Previous best dimension in 4D

PHYSIX OBSERVATION

AI is mass-producing mathematical proofs at an astonishing rate, but verification mechanisms lag far behind. With the model and prompts undisclosed, peers cannot reproduce the work, and the math community's skepticism is justified. This exposes a trust crisis in AI-driven research: output speed far outstrips humanity's capacity to digest it, and without verifiable publication standards, so-called breakthroughs amount to little more than marketing hype.

Source: TMTPost report