Back to feed
News Story
DeepTech深科技
1 sources

Terence Tao publicly endorses boycott of OpenAI, escalating tension between mathematics community and AI giant

Fields Medalist Terence Tao reposted a statement from the Association for Human Mathematics on his personal blog, publicly endorsing a boycott of collaboration with OpenAI and criticizing its release of over 700 AI-generated mathematical manuscripts without sufficient digestion by the mathematics community. The controversy was triggered by OpenAI's October 6 release of 722 manuscripts on GitHub, covering 372 result groups and claiming breakthroughs on problems such as the quasi-Riemann hypothesis, the four-dimensional Kakeya conjecture, and the Unique Games Conjecture. The mathematics community's reaction has been mixed: some results have passed Lean formal verification, but OpenAI has also retracted 3 manuscripts and revised 14 due to symbolic errors, fueling an ongoing debate over academic norms and the influence of commercial AI.

SynthePulse Insight · AI deep readingMembers

722 Manuscripts and a 42% Verification Rate: Why OpenAI's Math Release Sparked a Boycott

Version 1 · 1 source

On October 6, 2026, OpenAI published 722 mathematical manuscripts generated by an undisclosed internal model on GitHub in a single batch, claiming to have touched the quasi-Riemann hypothesis, the four-dimensional Kakeya conjecture, and the unique games conjecture; the next day, Terence Tao reposted the Association for Human Mathematics' boycott statement. The core of the controversy is not "whether machines can prove theorems," but when generation speed far outpaces human comprehension speed, who bears the verification cost and who decides the research agenda.

  • On October 6, 2026, OpenAI publicly released 722 mathematical manuscripts generated by an internal AI model, covering 372 result groups across number theory, geometry, algebra, and theoretical computer science.
  • These manuscripts came from an as-yet-undisclosed internal frontier model; OpenAI says the model attempted roughly 4,000 math problems in total, with the average compute per result equivalent to about three hours of ChatGPT Pro's thinking compute.
  • OpenAI claims to have proven the quasi-Riemann hypothesis, a breakthrough on the four-dimensional Kakeya problem, and the unique games conjecture, but as of October 7 only about 42% of the main conclusions had completed Lean formal verification.
Open section navigationOne Release, Two Timescales

One Release, Two Timescales

On October 6, 2026, OpenAI published 722 mathematical manuscripts generated by an internal AI model on GitHub in a single batch, covering 372 result groups across number theory, geometry, algebra, theoretical computer science, and other fields. According to information OpenAI released, this body of research was completed by an as-yet-undisclosed internal frontier model; because the model's performance on existing mathematical benchmarks was already approaching the ceiling, OpenAI began using genuinely unsolved mathematical problems to test its capabilities.

The sheer scale was itself a shock. Throughout the evaluation process, the model attempted to solve roughly 4,000 mathematical problems, which were filtered and organized into 372 result groups, each containing a core theorem, related corollaries, or different proof approaches. OpenAI says the vast majority of results used a similar generation process, with the average compute consumed per result equivalent to about three hours of ChatGPT Pro's "thinking compute," and this does not include the resources taken up by all the failed attempts.

But the mathematical community's pace of digestion operates on a different timescale. In traditional research, solving a major theorem is often just the starting point; scholars need to explain the core ideas behind the proof, what new tools were developed, and whether it can be extended to other fields through detailed papers, seminars, and long-term exchange—a process that can take years or even generations of relay work. When hundreds of complex proofs appear at "wholesale" speed, mathematicians may be forced into becoming "full-time proofreaders."

Free for now

Read the full analysis

5 more sections of analysis, plus the full takeaway

Loading

Credibility boundary

This article is based on a single secondary reporting source (DeepTech Shenkeji, source_tier 2) and the statements it relays from OpenAI, AHM, AGMAI, and others. The release scale, manuscript count, formalization ratio, and retraction and revision figures announced by OpenAI all come from OpenAI information relayed by that report and have not been independently verified by a third party; whether the proofs of the quasi-Riemann hypothesis, the four-dimensional Kakeya conjecture, the unique games conjecture, and others hold has not yet been rigorously confirmed by the mathematical community. Terence Tao's reposting of the AHM statement is a fact confirmed by the report, but the content of the AHM statement itself is that organization's assertion. Whether undiscovered flaws remain among the roughly 58% of main conclusions that have not completed formalization, and whether the propositions that passed formal verification accurately correspond to the true mathematical content of the original problems, both require manual verification.

Sources for this version

  1. 陶哲轩公开声援抵制OpenAI，数学界与AI巨头矛盾升级DeepTech深科技

Primary report

DeepTech深科技

Primary source