What happened
TokenPost reports that Terence Tao reviewed roughly 10 manuscripts from an OpenAI collection of 719 papers grouped into 372 result families. Tao criticized the release process and said one reported advance on zero-free regions for Dirichlet L-functions does not prove the Riemann Hypothesis. The report does not independently verify the manuscripts, mathematical claims or OpenAI’s stated process.
TokenPost reports that Tao criticized OpenAI’s release as discourteous because researchers already working on some of the problems were allegedly not notified. According to the report, he reviewed about 10 papers: some extended existing research, one or two contained clever new ideas, and many were difficult for mathematicians to understand.
The report says the collection contains 719 manuscripts organized into 372 result families. It attributes to the release materials the claim that an unreleased internal model was given about 4,000 open problems and that most results used a standard process averaging three hours of ChatGPT Pro thinking per result. TokenPost also reports that about 42% of top-line results have Lean formalizations, while cautioning that some unformalized results may contain problems.
TokenPost reports that one result claims a zero-free region for Dirichlet L-functions, including the Riemann zeta function, when the real part exceeds 7/8. Tao reportedly described it as the largest advance on that problem in 50 years, while emphasizing that it does not prove the Riemann Hypothesis and that a further idea would be needed to move beyond 3/4.
Source details: tokenpost.com ↗
Why it matters
The report raises a practical question about how AI-generated research should enter scholarly fields: whether speed and scale can coexist with verification, attribution and fair opportunities for researchers. If the reported collection is accurate, AI systems may increasingly compete with students and postdoctoral researchers on problems that traditionally helped them build expertise and publication records. The broader concern is not simply whether AI can produce novel mathematics, but whether researchers can understand, check and responsibly build on its outputs.
Tao’s criticism highlights a governance issue specific to AI-assisted research: releasing many purported results at once can shift the burden of checking from the system’s developer to the broader research community. Formalization in Lean may help with some verification, but the report does not establish that every headline result has a machine-checked proof or that formalization guarantees the underlying question was addressed correctly.
The source also describes a possible labor and education effect. Moderately difficult mathematical problems can give students and postdoctoral researchers a path to develop skills and establish a record. If researchers fear that undisclosed AI systems may solve their questions first, they could become less willing to share problems or pursue exploratory work. These consequences remain concerns reported through Tao’s interview, not independently measured outcomes.
The reported 7/8 result would be significant if independently validated, but its importance should not be confused with a solution to the Riemann Hypothesis. TokenPost provides no independent assessment, links to proofs or evidence that the claimed advance has been accepted by mathematicians.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Watch for independent mathematical assessments, public access to the manuscripts and formal proofs, clarification from OpenAI about the model and release process, and evidence of whether researchers were notified before work was published. The source does not establish access conditions, pricing, the identity of the internal model or the final status of the reported results.
The most important next step is independent scrutiny of the manuscripts and any associated Lean code. The source does not say whether the full collection, underlying prompts, model outputs, revision history or formalizations are publicly available.
OpenAI’s response could clarify whether the manuscripts were formally released, how affected researchers were contacted, what review was performed and what access or usage restrictions apply. TokenPost does not document pricing, general availability or a public interface for the reported system.
Researchers and institutions may also debate disclosure norms for AI-generated work, including attribution, priority claims, notification of people already working on a problem and standards for human understanding before publication.