Terence Tao vs OpenAI: From Workshop Co-Organizer to “I Don’t See a Way to Constructively Engage”
Updated
Live timeline · last updated · 2 entries in the update log
Tao did not turn against AI proofs. He turned against how this batch arrived: no one attached to the results, and years of junior research disrupted in a day.
Terence Tao went from arguing for engagement with AI labs to saying on Oct 16 that he sees no way to engage constructively with OpenAI, and his stated objection is about how the Oct 6 release was done, not about whether AI can prove theorems.
Every timeline node links a public source, read on 2026-10-17. Tao’s page is a living summary drafted with an AI assistant and reviewed by him, as it says itself; quotes from his own posts are taken from it. The New York Times page was reachable only in part, so NYT quotes are limited to lines we could read in full. We add no scenes or quotes beyond what these sources contain. On 2026-10-22 we reread the OpenAI post and Tao’s page. That page says it was last updated on 20 October 2026.
Timeline
-
Tao helps organize an IPAM workshop sponsored by OpenAI. His own page explains the context: the sponsorship came out of emergency fundraising after the 2025 suspension of NSF and NIH funding, and it was worth $60,000 to $70,000. During the workshop OpenAI recorded an hour-long interview with him and used a few snippets in an advertisement, which he says in retrospect he should have pushed back harder on.
-
At the International Congress of Mathematicians he gives the public lecture “Mathematics in the age of AI”. He calls the moment a crisis of values and practices rather than of mathematical arguments, and warns that proof-solving as a target can detach from the goals it used to stand in for.
-
He posts a statement for the record. He identifies with neither a “pro-AI” nor an “anti-AI” position, describes two paths (responsible adoption, or indiscriminate use for short-sighted goals), and writes that the field is “unfortunately rapidly approaching” the second.
-
A declaration titled “A Severe Misalignment of AI in Mathematics” appears with 25 Fields Medallists as its initial signatories, Tao among them. It says the goals of AI companies and of the mathematical community are severely misaligned, because solving problems is a proxy for understanding, not the goal itself.
-
Shipped: OpenAI publishes “Sharing AI progress in mathematics”. The post says the lab is releasing a broad range of new mathematical results from an internal frontier model, publishing them in a GitHub repository, sharing Lean formalizations of many of the proofs, and including 10 summaries of the model’s reasoning. It says the average result used about three hours of ChatGPT Pro thinking.
-
Reported, not in the OpenAI post: Tao’s page, citing The New York Times, describes the release as more than 350 findings across nearly every subfield, including algebra, number theory, theoretical computer science, mathematical logic and topology. The same page says the Times reports that OpenAI had announced a solution to the Navier–Stokes problem. The OpenAI post we read on 2026-10-22 does not name Navier–Stokes and does not give the 350 count.
-
Tao answers on the Proofs and Prompts blog. He writes that none of the humans involved in the proofs are available to take questions, give talks, submit to journals or train students, and that he is “horrified” by person-years of slow research, especially by young mathematicians, being casually disrupted.
-
He gives the Caltech public lecture “Math 2.0”. Its argument is that proof scarcity organized the institutions of Math 1.0, that proof abundance breaks that arrangement, and that the field’s bottleneck is imagination, not technology.
-
The Simons Foundation announces a $15 million effort with Anthropic and OpenAI to help mathematicians evaluate AI-assisted results, learn the tools and shape standards. OpenAI adds an update to its Oct 6 post: $5 million of that is its own, plus model and compute access for Simons-supported researchers.
-
The New York Times publishes its profile. Tao tells the paper “I’m on a war footing”, says he is doing very little math, and says of OpenAI: “At this point, I don’t see a way to constructively engage.”
The New York Times: Why One of the World’s Greatest Mathematicians Is on a ‘War Footing’
A workshop in March, and a bargain that made sense at the time
To understand why Terence Tao’s October sounds the way it does, start in March, because the March version of Tao was not an opponent of AI companies. According to the disclosure on his own page, he helped organize an IPAM workshop that OpenAI sponsored. The money did not appear out of a taste for corporate partnerships. His page says it came out of emergency fundraising after the 2025 suspension of NSF and NIH funding, and that the sponsorship was between $60,000 and $70,000. A research institute that has lost federal money takes the offers that exist.
The same disclosure records a cost. During the workshop OpenAI recorded an hour-long interview with Tao and later used a few snippets in an advertisement. Tao says that in retrospect he should have pushed back harder. It is a small detail, and it is the kind that matters later: the first sign that a sponsor’s interest in a mathematician’s voice and the mathematician’s interest in the sponsor’s money were not the same thing.
His stated reason for engaging at all was strategic, and he put it plainly. According to his page, he long held that engaging the industry was better than being “uniformly hostile”, because incentives work both ways: dealing with the labs could move the balance sheet of the technology “in a better direction”. He also argued that a tenured academic with no AI-company salary can still speak honestly. That is the position to keep in mind for the rest of the story. He was not recruited into a camp. He made a wager that talking would shape the outcome.
Summer: describing the problem before it arrived
Between March and October, Tao spent a growing share of his public writing on a worry that had little to do with whether AI could do mathematics. His page says the answer to that was increasingly yes. At the International Congress of Mathematicians on July 24, he framed the moment as a crisis in the foundations of mathematical values and practices. His argument was that problem solving, historically the most visible goal of research mathematics, had always been a proxy for slower goals: understanding, theory-building, training the next generation. While solving problems was hard, the proxy tracked the goals. Once machines could produce solutions cheaply, he warned, the proxy would stop tracking them, and pushing harder on it would do damage.
He used a name from economics for the mechanism, Goodhart’s law: when a measure becomes a target, it ceases to be a good measure. The same page records a September essay in which he argues that open problems in mathematics have become something like a non-renewable resource, because a problem whose solution is already public is permanently less useful for evaluating anything. In the same stretch he wrote a counterfactual about bounded gaps between primes, imagining how a decades-long collaborative story, with its failed attempts and new methods, might have gone if benchmark-hungry AI companies had been present at the start: a bound that falls quickly, a problem declared saturated, and little left that could be taught.
None of this is about October 6th, and that is the point. Weeks before OpenAI’s release, Tao had written down the failure mode he feared, in detail, with a mechanism. When the release came, his reaction was not that of a man discovering a new danger. It was closer to watching a forecast arrive on schedule. His September 9 statement for the record says as much: he had seen two possible paths from 2023, responsible adoption or indiscriminate use for short-sighted goals, and “we are unfortunately rapidly approaching” the second. Two days later, he was one of 25 Fields Medallists who signed the initial text of “A Severe Misalignment of AI in Mathematics”. It conceded the capability question outright, saying the tools can now solve major outstanding problems, and moved straight to the thing that worried them: that solving problems is only a proxy for conceptual understanding, and that mass-producing true-or-false statements could destroy fertile ground instead of breathing life into new ideas.
October 6: what OpenAI released
On October 6 OpenAI published a post titled “Sharing AI progress in mathematics”. The New York Times, in its profile of Tao ten days later, describes what landed that day as more than 350 findings across nearly every subfield of the profession, including algebra, number theory, theoretical computer science, mathematical logic and topology. It adds that mathematicians, some of whom had been close to solving problems they had worked on for years, found themselves beaten in an instant by AI. (Counts vary elsewhere, from a few hundred to higher figures. We use the Times’ number and attribute it, because we cannot reconcile the others.)
The release came after a headline claim. The Times reports that OpenAI had earlier announced it had solved the Navier–Stokes problem, one of the seven Millennium Prize problems. That is the context in which Tao told the paper “There are no rules any more” and said he was “just numb to what would have been jaw-dropping events” for him a year earlier. A man who spent a career absorbing mathematical surprises was describing a state in which surprise no longer registered. It is an odd thing for a mathematician to say, and it is worth reading as information about the pace, not as rhetoric.
Shipped, reported, and not a date
- Shipped. On 6 October 2026 the OpenAI post says the lab is releasing results from an internal frontier model, in a GitHub repository, with Lean formalizations of many proofs and 10 summaries of the model’s reasoning. It says the average result used about three hours of ChatGPT Pro thinking. An update on the same post, dated 15 October 2026, says OpenAI is contributing $5 million and will provide model and compute access. Read on 2026-10-22.
- Reported, not in that post. Tao’s page, citing The New York Times of 16 October 2026, says OpenAI announced a solution to the Navier–Stokes problem and that the 6 October release contained more than 350 findings. The OpenAI post we read does not name Navier–Stokes and does not give 350. Treat both as reported.
- Tao’s own words, as his page quotes them. The 8 October post on Proofs and Prompts includes “horrified” and the line that none of the humans involved are available to take questions. The lines “I’m on a war footing” and “I don’t see a way to constructively engage” are what the page attributes to the 16 October Times interview.
- Not a date. The OpenAI post says it will share more about workshops in the near future and that it is working to release the model that produced the results. Neither sentence names a day. This page does not add one.
October 8: what Tao actually objected to
Two days later Tao published his reaction, and the reaction is more precise than the headlines made it sound. According to his page, which quotes the post directly, he did not say the proofs were worthless. He said they “appear to introduce clever new ideas that will be fruitful once digested, while also building upon the existing contributions of countless human mathematicians past and present”, and he said he was excited that the community might use such tools to take on ambitious, large-scale projects that earlier generations could not have imagined.
His objection was to the arrangement around the results. “None of the humans involved in these proofs are available to take questions, give talks, attend conferences, submit papers to journals, train students, or otherwise participate in the subsequent development of these results.” In mathematics that sentence carries weight, because a proof is not finished when it is correct. It is finished when someone has explained it, placed it in context, and made it teachable. He wrote that the best opportunity to digest a result is the moment of its discovery, and that such moments are “increasingly wasted when delegated entirely to AI tools”.
Then came the line that explains the word “horrified”. Tao wrote that he was horrified by “the many person-years of ongoing patient and deliberately slow research efforts — particularly by graduate students and postdocs — towards many motivating problems in mathematics being casually disrupted or destroyed by such a release.” His analogy is homely: much as one cannot unhear a movie spoiler or a crossword clue, one cannot explore a problem as profitably once one knows an answer exists. For a graduate student three years into a problem, the spoiler is not abstract. It changes what the thesis can be.
He also described what he would have liked instead. The same post, per his page, sketches a path not taken: labs submitting frontier models to independent researchers for scientific evaluation, coordinating with the research community so that the tools complement human researchers rather than compete with them, using them to foster collaboration and sharing rather than competition and secrecy, and “opening new doors, without closing old ones.” The post ends with a call inward: the community should declare its own standards and values, build its own tools and practices, support its most vulnerable members, and chart its own path. “Let’s get to work.”
Reading the output: “technically readable”
The most concrete part of the story is how Tao says the output reads. Speaking to the Times, he described the Navier–Stokes proof as “technically readable”, but added that “no expert I’ve talked to has actually understood the whole argument.” His page, summarizing the same interview, says the proof was reportedly 165 pages produced by a swarm of some ten thousand agents over 88 hours, and that before the October release he had called the coming results “toxic waste”, and afterwards, “maybe treated waste”.
The Times adds that the writing was better in the latest release, which he called a “mixed bag”. Some papers cited his own work appropriately and appeared to have advanced it in a way he found plausible. That detail cuts against a simple reading of the story as a man angry at being outpaced. He is describing a body of work where some of it is good and is still unhappy, because of what he says is missing: “there is no author that I can have a real conversation with about the results.”
That sentence is the thesis of the whole episode. In his ICM argument the finished pipeline for a result has stages: generation, verification, exposition, publication and canonicalization, with the last being the point at which the result becomes part of what the next generation is taught. AI had made the first stages cheap. The release, in his account, left the later ones unowned.
October 9: the Caltech lecture, and what he says to do instead
The day after his post, Tao gave the Caltech public lecture, announced as “Machine Assisted Proof” but which he now prefers to call “Math 2.0”. The New York Times, describing it the following week, says he argued that despite OpenAI’s releases most problems at the frontier of mathematics remain unsolved, that the difficulty of problems is boundless, and that because AI lets mathematicians collect data at vast scale, they will increasingly work in large groups on projects that once would have been unimaginable.
His page gives the longer version. Math 1.0, he says, ran on proof scarcity: finding a solution and proving it was rare and difficult, and the institutions, traditions and incentives of the field were organized around that. When a breakthrough proof appeared, the field’s response generated invitations, meetings, workshops, new collaborations and newcomers, until the proof was digested into textbooks. Problems solved autonomously, he argues, generate very little of that cycle, because the people who prompted the AI often have no interest in the field once the target is solved and cannot answer questions about the result. His summary of the bottleneck is blunt: it “is not technology, or even institutional culture. It is imagination.”
The advice the Times reports is notably gentle for a man who says he is on a war footing. Problem solving is not a destination but a journey, and lessons learned along the way are often more important than the answer. Rather than solving a single problem alone, mathematicians can use AI to solve ten and look for patterns and new questions. It is acceptance born of necessity, as the Times puts it, rather than a call to fight.
October 15: the money arrives
The day before the Times profile ran, the Simons Foundation announced a new set of initiatives. Under a partnership, the Foundation, Anthropic and OpenAI committed $15 million in combined funding and resources. The Foundation will manage and award the funding, including through an open call for proposals, to support workshops and initiatives that help researchers evaluate recent AI-assisted results, learn and use new tools, share perspectives and shape collective standards for AI use. Anthropic and OpenAI will also provide tools and computing resources to researchers the Foundation supports, in basic science as well as mathematics. The announcement notes that Simons was an early funder of the Lean proof assistant.
OpenAI’s part appears as an update on its own October 6 post, dated October 15: it is contributing $5 million to the initiative, “to help mathematicians examine, understand, and build on results produced by AI, including those shared in this post”, and will provide model and compute access. The same update says the money will pay for workshops, conferences and special programs.
We do not know whether this was a response to the criticism, and neither source says so; the sequence of dates is all we can state. What can be said is that the stated purpose of the money, helping the community understand results after the fact, is the digestion work Tao said on October 8 was best done at the moment of discovery. Whether funding workshops after a release can substitute for having authors available during it is exactly the question his objection raises, and nothing in the announcements answers it.
October 16: “I’m on a war footing”
The Times profile, by Elisabeth Bumiller, opens with Tao grabbing coffee during a two-day math workshop at UCLA, where, it says, talk of AI’s existential crisis for the field was everywhere. Tao, 51, tells the reporter he is doing very little math. “I’m on a war footing.” His page and the paper describe how his time now goes: convening mathematicians, strategizing, lecturing, counseling colleagues, in what he calls community organizing.
The reversal on engagement is stated plainly. A year earlier, he tells the Times, he had good relationships with many people at the company, who “shared a very common vision of how A.I. should be used — you know, to empower scientists.” But the rush to publish, he says, shows what they actually used the field for: they “used math for what they needed, which was to evaluate their models.” And then the sentence that headlines this article: “At this point, I don’t see a way to constructively engage.”
It is worth noticing what he does not say. He does not say the mathematics is false, and he does not say AI should be kept out of the field. The same page that records his October position also records his continuing support for responsible use, formal verification and open tools. His complaint is about a particular way of releasing results, with no author available, no coordination with the people working on the problems, and a company that, in his words, is more interested in evaluating its models than in the field.
What we could not verify, and left out
Summaries of this story, including some written in the last two days, carry details we have not put in the timeline because they are either single-source or come from automatic captions of a video interview: how many of the new papers Tao personally read, how many were withdrawn, and a ranking of one result as the biggest advance on its problem in fifty years. Reported counts of the release also differ between outlets. We kept only what a public source states in a form we could read. We also did not use the “Mozart of mathematics” label that appears in popular coverage; we could not find it in the sources above.
Our take
The useful way to read this story is as a dispute about who owns the last mile of a result, not as a fight between a mathematician and a machine. Tao’s position has been consistent since the summer: generation got cheap, and what stayed expensive, explanation, context, teaching, is where the field’s value lives. A lab can produce hundreds of proofs in a day. It cannot, in the same day, produce hundreds of people who can discuss them.
The October 15 funding is an attempt to pay for some of that digestion after the fact. The announcements do not say it answers Tao’s objection, and this page does not forecast a later release or assign it a date.
Keep reading on Front
Every post, newest first, is on the Front blog index.
- AI mathematicians timeline: the wider October 2026 record of AI and research mathematics.
- Google Argon timeline: the dated record of Google’s Gemini 4 Argon launch.
- Grok 4.8: what xAI has actually said: xAI’s Grok 4.5 to 4.7 releases and what has been said about 4.8.
- Reflection Beam 501B timeline: the dated timeline of the Reflection Beam 501B weights release.
FAQ
What did Terence Tao say about OpenAI’s math release?
Per the New York Times, he said “I’m on a war footing” and, of OpenAI, “At this point, I don’t see a way to constructively engage.” On Oct 8 he wrote that none of the humans involved were available to take questions or give talks and that he was horrified by disrupted research, especially by young mathematicians.
When did OpenAI release the math results?
October 6, 2026, according to OpenAI’s post. That post is the shipped page. The New York Times describes more than 350 findings across nearly every subfield. The 350 count is reported, not a figure in the OpenAI post.
Did Tao work with OpenAI before this?
He helped organize an IPAM workshop sponsored by OpenAI in March 2026, according to his own disclosure. He says the sponsorship was $60,000 to $70,000 and came out of emergency fundraising after the 2025 suspension of NSF and NIH funding.
Did OpenAI respond?
On Oct 15 OpenAI added an update to its post committing $5 million to a Simons Foundation initiative, part of a $15 million effort with Anthropic, to help mathematicians examine and build on AI-produced results. The sources do not say it was a response to Tao.
Update log
- — Reread the OpenAI post and Tao’s page on 2026-10-22. Labeled the 6 October post as shipped, and the Times account of 350 findings plus a Navier–Stokes announcement as reported. Removed a sentence that speculated about a later release.
- — First published.
Sources
- The New York Times: Why One of the World’s Greatest Mathematicians Is on a ‘War Footing’ (Oct 16, 2026)
- Terence Tao on AI: a living summary (updated Oct 16, 2026)
- OpenAI: Sharing AI progress in mathematics (Oct 6, 2026; updated Oct 15)
- Simons Foundation: $15M effort to help mathematicians navigate the AI era (Oct 15, 2026)