Terence Tao told a UCLA audience that “mathematicians are hiding their problems”. He spoke at the Science x AI Summit on 16 October, and the SAIR YouTubechannel posted the full talk on 3 November. Researchers who “used to share very openly” their “promising research directions” now fear that “someone with an AI will input their research ideas into” a frontier model and publish ahead of them, “before they can finish their work”, he said.
The remark is 18 days old. It spread on Tuesday through a 27-second clip that @rohanpaul_ai posted on X at 13:46 UTC.
One place to see how fast proofs are arriving is the Erdลs problems database. Its history file shows solutions formalised in the Lean proof language rose from 38 at the end of 2025 to 318 by the end of October, 8.4 times, by our arithmetic. Over the same period, the count of open problems fell from 689 to 642.
The full talk · SAIR
Terence Tao, “Terence Tao: Math 2.0”, Science x AI Summit, UCLA, 16 October 2026, as posted by SAIR on 3 November. The passage on hiding problems starts at about 22:46.
The key insight: As we read it, Tao is describing a shift in what is scarce in research: by the account in his talk, proofs are becoming abundant while good problems are becoming harder to find, and some mathematicians now keep those problems to themselves.
The passage comes about 22 minutes into the 31-minute video, titled “Terence Tao: Math 2.0”, in a chapter the channel labels “Answers without insight and the search for fruitful problems”.
Tao said that “for the first time in history we are running out of problems to solve”. Solving a problem used to suggest two or three new ones, he said, but now “we are running out of really good problems”. He then made the point about hiding problems.
Just before that, he said many recent AI papers in mathematics are hard to follow, and that there was no one at OpenAI he could call to talk through “any of these 300 papers”.
The video description says pure mathematics suits AI unusually well and that the result is “an era of proof abundance”, while “good research problems are becoming harder to find”.
Erdős problems database: open problems, solved problems and solutions formalised in Lean, at the last update of each month from December 2025 to October 2026, from the teorth/erdosproblems history file. The 8.4x rise (318 / 38) is our arithmetic. A Lean-formalised solution has been checked by a machine; the file does not say whether an AI wrote it.Animated: Erdős problems with a solution formalised in Lean, at the last update of each month from December 2025 to October 2026, from the teorth/erdosproblems history file. The 8.4x rise is our arithmetic; Lean-checked does not mean AI-made.
What Was Posted on X
Mathematicians are now hiding their research ideas out of fear of AI ๐
~ Terence Tao
“now, mathematicians are hiding their problems. they used to share very openly all their promising research directions and now they’re afraid that if you do that, someone with an AI will input their research ideas into a frontier model and rush out some people before they can finish their work.”
—- Full video on “SAIR” YouTubechannel, (link in comment)
Rohan Paul (@rohanpaul_ai) posted this 27-second clip of Terence Tao’s 16 October talk at 13:46 UTC on 3 November, with a quoted transcript of the remark. This piece quotes the SAIR upload’s captions of the same passage, embedded above.
What OpenAI Released
On 6 October, OpenAI released a GitHub repository of 372 significant proof claims in mathematics, Thomas Bloom, who runs erdosproblems.com, wrote on 8 October. He counted claims relevant to 25 Erdลs problems; 21 of the 25 proofs had been formalised in Lean, and the PDFs ran to 1,347 pages.
Bloom stressed that he had not checked the proofs and that they should be read as claims. He also noted accusations of plagiarism, and that some of the proofs had appeared online in some form before the release.
What the Erdลs Site Changed
On 6 October, Bloom announced that erdosproblems.com would stop displaying the status of problems and the count or percentage solved. He wrote that “the main way that people publicly interact with the site now is to advertise their AI-generated proofs”, often without explanation, as a way to record an “increasingly meaningless” priority claim.
He also wrote that “many have stopped thinking about Erdลs problems believing that they cannot compete with AI”. He said repositories for AI-generated proofs should exist, “to save others wasting their tokens generating the same proof”, but that he did not want to run one.
The GitHub repository that tracks the same problems moved its solved counts off its front page and froze its progress chart on the same date. It still publishes the data.
What the Numbers Show
The repository’s history file lists 1,221 problems. At its last update in October, 579 were solved and 642 open; 318 solutions had been formalised in Lean, up from 38 at the end of 2025.
Formalised in Lean means a machine has checked the proof. It does not mean an AI found it: the file does not say who or what wrote each proof.
The solved count has stood at 579 since at least 28 September. Bloom’s 8 October post treated the OpenAI release as claims still to be checked.
The Business Lever
The business lever in Tao’s remark is research IP: who captures the value of an idea once a frontier model can turn it into a paper quickly. In open research, a scholar shares problems and directions early and gets credit later, through publication and priority.
Cost matters here. Sam Altman said on the Cheeky Pint podcast, published 29 October, that the Navier-Stokes math cost “Order $10 million per… For Navier-Stokes, I think,” and that “I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done.”
Volume is rising too. arXiv says it received 40,363 submissions in September, a monthly record, and since 1 October it has capped submitters at two papers a month.
As we read the Erdős data, machine-checked solutions on the list rose from 38 to 318 in ten months, by our arithmetic, while open problems fell from 689 to 642; the file does not say how many of those proofs came from AI.
As we read Bloom’s 6 October post, the site’s owner saw the same pattern from the other side: public activity turned into AI-generated proofs posted to record priority, and he removed the solved count from view.
As we read Tao’s remark against Altman’s cost comments, the cost of running a hard problem through a frontier model is a number OpenAI’s chief executive expects to fall; Tao’s concern is about who gets there first once it does.
Terence Tao, Science x AI Summit, UCLA, 16 October 2026 (automatic captions, SAIR upload)
“mathematicians are hiding their problems”
Three Implications
FOR RESEARCHERS Tao says some mathematicians now hold back problems and research directions they used to share openly, for fear of being beaten to publication by someone using AI.
FOR PROBLEM DATABASES erdosproblems.com stopped showing problem statuses and the solved count on 6 October; the GitHub repository still publishes the counts, 579 solved and 642 open at its last October update.
FOR AI LABSOpenAI’s 6 October release held 372 proof claims, 25 of them relevant to Erdős problems, according to Bloom, who had not checked them.
The framework argues: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, Tao’s talk describes proofs moving to the abundant side and good research problems to the scarce side, and his remark describes researchers closing the scarce layer. The framework is about technology companies; it does not discuss mathematics or research norms.
What Is Not Established
In the passage, Tao did not give a number for how many mathematicians are hiding problems, or name any. The remark describes what he says he has heard; we have not measured it.
The quotations here are from YouTube’s automatic captions of the SAIR upload. The X post’s transcription of the clip is a lightly cleaned version of the same passage: it drops false starts and repeated words and adds an article before the words frontier model. Both give “rush out some people”, while the captions of an audience recording of the same talk give “rush out a paper”, so we do not quote that phrase. The audience captions are rougher elsewhere: they read “physicists” for “mathematicians” and drop “promising”, so we quote the SAIR captions only.
It is not established that any of the Lean-formalised Erdลs solutions came from AI, or that the open-problem count fell because of AI.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Terence Tao told a UCLA audience on 16 October that “mathematicians are hiding their problems” for fear of being beaten to publication by someone using a frontier model. On the Erdลs list, Lean-formalised solutions rose 8.4 times in ten months, by our arithmetic, and the site’s owner stopped showing a solved count on 6 October.
A note on sourcing. This piece is based on the SAIR YouTube upload of Terence Tao’s 16 October talk, read through YouTube’s automatic captions and compared with the X post’s transcription of the clip and with an audience recording whose captions differ in places, as described above; Thomas Bloom’s posts on erdosproblems.com of 6 and 8 October; and the teorth/erdosproblems repository on GitHub, read on 3 November. The 8.4x figure is our arithmetic. We did not contact Tao, SAIR or OpenAI.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
