The Information Machine
Updated todayDay 41in the edition since 7 Oct 202628 sources

AI proofs of Navier-Stokes and Yang-Mills

The gist

Latest: AI Math Releases Disrupt Field; Governance Debate Continues

AI systems claiming solutions to longstanding problems in physics and mathematics have disrupted research norms and career prospects in those fields before the community has agreed on standards for verifying or incorporating such results. The governance gap between what AI produces and what the field can absorb or validate remains open.

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

The full picture

OpenAI released 722 AI-generated mathematical manuscripts in a batch, then retracted three a day later after a sign error was found. The release included claimed proofs of pure Yang-Mills theory existence, a refutation of the Exponential Time Hypothesis, and work implying the full Kakeya conjecture. OpenAI dates its Yang-Mills existence proof to September 9, 2025, its QCD proof to September 18, 2025, with Lean formal verification of the first completed by October 16, 2025. The Advisory Group on Mathematics and Artificial Intelligence has demanded OpenAI release the natural-language and Lean outputs behind these physics proofs. Some results went directly to GitHub while the physics proofs stayed with reviewers led by David Gross, a handling approach mathematicians have objected to.

Fields medalist Hugo Duminil-Copin said sixteen released manuscripts directly addressed his area and described the release as having 'annihilated' his field. Terence Tao, in a New Scientist interview, said the Navier-Stokes field is 'in disarray' and called the automated approach 'strip mining' that achieves an immediate goal 'at the cost of damaging the future development of the field'. In a Caltech lecture titled 'Math 2.0,' Tao argued that open problems serve as 'lighthouses' guiding exploration of surrounding mathematical territory and that reaching them prematurely via automated tools 'can disrupt the exploration of the paths not taken, and sterilize the surrounding field'. He also observed that mathematicians now hide research directions for fear that sharing them will allow others to use a frontier model to rush out results first. Some Northwestern University mathematics students questioned whether to remain in the field at all following the batch release.

An MIT survey of 268 graduate students and postdocs across nearly 60 institutions found 76.3% expect AI to make tenure-track jobs harder to obtain; among those who use AI in research, 54.8% say they dislike doing so. Scott Aaronson reported that students 'don't know how to build a career when AI can solve the problems they'd train on'. Harvard mathematicians are split on whether AI-generated proofs need to be humanly understandable to count as valid.

The Simons Foundation, Anthropic, and OpenAI committed a combined $15 million for workshops and initiatives supporting the mathematics community, with proposals due December 1, 2026. The initiative extends beyond mathematics to basic science and will include shared standards for checking AI-assisted results. OpenAI launched a new science blog, and Nora Ammann joined the OpenAI Foundation as Program Director of Formal Methods. AGMAI, established by the Institute for Advanced Study, operates independently of AI companies and accepts no payment.

How it developed
16 November 2026

OpenAI announces physicist Nathaniel Craig's collaboration with GPT-6 Astra on Standard Model lattice definition

15 November 2026added 16 Nov

OpenAI published its Standard Model lattice formulation work as the inaugural post on a new science blog, describing physicist Nathaniel Craig's collaboration with GPT-6 Astra to develop a candidate nonperturbative lattice definition.

AGMAI's demands for full release of the natural-language and Lean outputs behind the physics proofs remain outstanding.

12 November 2026in the 13 Nov edition

The David Gross-led committee published its validation of OpenAI's Lean proof that pure Yang-Mills theory exists, the first external check to support the claim on physics results OpenAI had withheld from public release and given to a small review group rather than posting to GitHub.

The committee noted that Lean confirms only that a proof matches its internal formalization, not necessarily the real mathematical question, and that an earlier OpenAI proof had Lean verify a weaker claim than the prose described. The mass gap, the harder remaining piece of Yang-Mills, is not resolved.

8 November 2026in the 9 Nov edition

A TCSExam benchmark published November 8 by Tsinghua and Moonshot AI found GPT-6 Astra proved 70.0% of research-level computer science theorems in Lean, far ahead of other agents, but resolved none of 30 open conjectures in theoretical computer science, with the few accepted conjecture submissions adapting proofs OpenAI had already published rather than generating novel solutions.

The result adds a data point to debate about whether AI systems can produce genuinely novel results, following OpenAI's October 26 announcement that internal models produced a Lean-verified ETH disproof, a proof of the Kakeya conjecture, and a Yang-Mills existence proof.

5 November 2026

Nora Ammann joins OpenAI Foundation as Program Director of Formal Methods

4 November 2026

Sam Altman says human control should serve as an ultimate constraint on AI progress

3 November 2026added 4 Nov

Terence Tao said mathematicians who 'used to share very openly all their promising research directions' are now afraid that sharing will let others input those ideas into a frontier model and rush out results before them, calling the shift a departure from longstanding norms of open collaboration in the field.

1 November 2026in the 2 Nov edition

Terence Tao, in a New Scientist interview, called OpenAI's automated approach to major open mathematics problems 'strip mining,' saying the Navier-Stokes field is 'in disarray' even if the proof is eventually validated.

An MIT survey of 268 graduate students and postdocs conducted in September and early October 2026, mostly before OpenAI's mass release, found 76.3% expect AI to make tenure-track positions harder to obtain, with fear of being scooped causing some to share less unfinished work.

First citedRohan Paul on X
28 October 2026

OpenAI published its ETH disproof and Kakeya materials on GitHub, with independent reviewers Impagliazzo, Paturi, and Williams operating under no NDAs, while the Yang-Mills existence proof, produced September 9, 2025 with Lean verification by October 16, 2025, remains withheld with David Gross's review group.

The Advisory Group on Mathematics and Artificial Intelligence publicly demanded OpenAI release those physics outputs, and mathematicians named the arrangement 'shepherding,' objecting to routing significant results through a small group rather than the broader community.

26 October 2026added 27 Oct

OpenAI on October 26 released a second batch of AI-generated results claiming to have disproved the Exponential Time Hypothesis, which had stood roughly 25 years, solved the local smoothing and Kakeya conjectures, and proved part of the Yang-Mills Millennium Prize problem, following its earlier release of 722 AI-generated papers.

An advisory group called for public evidence, disclosed prompts, and independent review before such results are accepted. Resolution reported spending roughly $135,000 formalizing Hironaka's resolution of singularities in Lean using AI agents.

First citedThe Neuron
22 October 2026

Microsoft researchers proved the Kannan-Lovász-Simonovits conjecture, posed in 1995, using AI agents that explored more than 100 candidate research trajectories under human direction, with Song and Zhang credited for the key recursive method and error bound.

The work produced a single 140-page manuscript. Where OpenAI's September 21 release had an autonomous model generate 722 papers simultaneously, here human researchers directed the search and are credited with the core mathematical innovations.

in the 23 Oct edition

Harvard mathematicians divided publicly over whether an AI-generated proof that humans cannot follow counts as valid mathematics, a question raised by OpenAI's release of 722 manuscripts, including a claimed Navier-Stokes proof.

The Simons Foundation formally opened its grant program, with proposals due December 1, 2026, and confirmed the initiative extends to basic science. Anthropic co-founder Jack Clark, at the IMF annual meeting in Bangkok, cited benchmark gains to argue AI crossed into a technological singularity this year.

20 October 2026

Sam Altman told OpenAI staff the company must choose among three AI safety governance endgames within a year

18 October 2026in the 19 Oct edition

Scott Aaronson of UT Austin reported that graduate students are scared and do not know how to build a career when AI can solve the problems they would train on, as concerns about mathematics careers spread to a second named institution following OpenAI's September 21 release of 722 AI-generated manuscripts including a resolution of the Navier-Stokes Millennium Prize problem.

Northwestern professor Antonio Auffinger has said some of his students are asking whether they should stay in mathematics at all.

First citedRohan Paul on X
16 October 2026in the 17 Oct edition

The $15 million Simons Foundation initiative, backed on October 15 by OpenAI and Anthropic, will include shared standards for checking AI-assisted mathematical results.

Hugo Duminil-Copin, whose area covers 16 of the 722 manuscripts OpenAI released in September, reported he cried the night of the release and it woke up his daughter. An October 7 analysis by Zvi Mowshowitz argued that OpenAI's Navier-Stokes resolution effectively reached the milestone of a Nobel Prize-worthy AI discovery, predicted at a prior conference for around 2032.

15 October 2026in the 16 Oct edition

On October 15, OpenAI and Anthropic announced backing for a $15 million Simons Foundation initiative to fund workshops, conferences, and researcher access to models and computing resources, with each contributing $5 million, Anthropic's share specified as including compute and Claude access.

On the same day, Fields medalist Hugo Duminil-Copin said OpenAI's release of 722 AI-generated mathematics manuscripts had "annihilated" his field of research, with 16 of those papers directly targeting his own area.

14 October 2026in the 15 Oct edition

Johns Hopkins professor Matthew Green on October 14 put a 15% probability on AI eroding confidence in public key encryption, following OpenAI's October 6 release of 722 AI-generated mathematics manuscripts.

Green added that most internet security relies on a small set of elliptic curve algorithms, with X25519 alone securing most connections, and that a break in two of them would hit a large share of encrypted traffic, with fallback options less studied and potentially weaker.

12 October 2026in the 13 Oct edition

OpenAI retracted three of the 722 AI-generated mathematics manuscripts released October 6 due to a sign error and issued 14 revisions, leaving researchers sorting out which of their own projects had been preempted by the AI-generated work.

Of the papers, 42% now have Lean-verified proofs holding up, including five of the nine headline results, while 58% remain largely unvetted by the community.

First citedThe Neuron
11 October 2026in the 12 Oct edition

At a Caltech 'Math 2.0' lecture on October 11, Terence Tao argued that blind AI optimization of problem-solving 'is now actively harmful to the long-term health of mathematics,' sharpening his stance on OpenAI's October 6 release of 722 AI-generated manuscripts covering 372 open problems.

Northwestern professor Antonio Auffinger reported that some students are now questioning whether to remain in mathematics at all, saying his priority has shifted to retaining talented people in the field. Cryptographer Matthew Green placed a 15% probability on AI causing a functional loss of confidence in existing public-key encryption algorithms.

First citedRohan Paul on X
10 October 2026

Security researchers warned on October 9 that OpenAI's October 6 mathematics release may endanger public-key cryptography: the release contains no abstracts touching cryptography, yet Matthew Green stated 'I think we might lose public key cryptography' and Vitalik Buterin recommended multiplying key sizes by 10 for structured schemes.

AGMAI stated the release violated its request that labs not test hard problems on private models. NYU professor Tristan Buckmaster said on October 10 he withholds original ideas from AI tools, citing finite PhD and postdoc windows.

9 October 2026

OpenAI withdrew portions of its release of 722 mathematics manuscripts within two days of their October 6 publication.

A paper published October 8 showed AI-assisted Lean verification is structurally unreliable: chatbots can silently correct errors while producing a passing Lean check, and faithful translation into Lean is provably harder than the Halting problem. Terence Tao said at a Caltech summit that the results are 'not actually creating any value', and the Association for Human Mathematics called on mathematicians to boycott OpenAI.

7 October 2026

OpenAI on October 6 published 722 manuscripts covering 372 results from the model that claimed a Navier-Stokes blowup proof on September 8, formalized in Lean on GitHub and developed in consultation with AGMAI, an unpaid advisory body at the Institute for Advanced Study with no decision-making power at AI companies.

The highlighted result is a theoretical matrix multiplication upper bound of roughly n^2.25, improving on AlphaEvolve's August record of roughly n^2.371177, though not applicable to real GPU workloads. OpenAI had separately stated on September 21 that the same model resolved more than 100 long-standing open problems across most branches of mathematics.

added 8 Oct

A reviewer at The Curve assessed that AI achieving a Nobel Prize-equivalent scientific discovery, forecast at a prior conference for around 2032, was already met in 2026, citing a math professor's remark that a human producing OpenAI's October 6 batch of 722 mathematical manuscripts would be an instant Fields Medal, no questions asked, alongside OpenAI's September 8 claimed Navier-Stokes proof.

The Clay Mathematics Institute has not accepted that proof and still lists Navier-Stokes as unsolved.

First in the edition · 7 October 2026
2 October 2026

Meta AI reported six research papers on previously unsolved mathematical problems produced using Muse Spark 1.1 and 1.2

1 October 2026added 2 Oct

OpenAI announced September 8 that a swarm of roughly 10,000 agents running 88 hours produced a Lean-verified proof targeting a finite-time singularity in the 3D Navier-Stokes equations, a Clay Millennium Prize Problem.

Independent verification is expected to take weeks to months and the Clay Institute has not issued a statement. Anthropic published September 25 that physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude to compute N=4 super Yang-Mills scattering at nine loops, surpassing Lance Dixon's eight-loop record from 2023. Dixon of SLAC independently validated the result.

30 September 2026

GPT-6 Astra used to identify new families of plasma equilibria in fusion mathematics

25 September 2026

Anthropic published a report that its physicists used Claude to compute the nine-loop Yang-Mills amplitude, answering von Hippel's challenge; Lance Dixon independently validated the result

22 September 2026

Harvard professor Wood joins the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study

21 September 2026

OpenAI announced the model had resolved more than 100 long-standing open mathematics problems including Navier-Stokes

11 September 2026

Lean 4 formal proof aspect of Navier-Stokes announcement identified as consequential 2026 trend

9 September 2026added 3 Oct

A mathematical objection entered the record: critics say OpenAI's September 8 claimed Navier-Stokes blowup proof uses a smooth external force formulation most mathematicians exclude from the real problem, and the Clay Institute still lists it unsolved.

OpenAI's figures put the run at 4.9 million messages and 300 billion tokens across all Millennium Prize attempts. NYU professor Buckmaster, who had worked the same approach via Codex for about a year, was offered compute resources rather than priority acknowledgment.

8 September 2026

OpenAI announced a claimed Navier-Stokes finite-time singularity proof using about 10,000 agents over 88 hours

17 August 2026

Axiom Math uses AxiomProver to produce a formal Lean 4 verification of the 246 Prime-Gaps Theorem

Sources
Rohan Paul on X
23 more sources

Get the daily email

When stories like this move, I write about them in an email each morning.

One free email each morning, around 8 AM US Eastern.

See the latest edition · Privacy