Over a Thousand Mathematicians Jointly Establish AHM, Refusing to Do Free Cleanup for AI-Generated Proofs

The sky of the mathematics world has utterly collapsed!

Mathematicians who have ruled the high ground of human logic for thousands of years have, for the first time, felt the chill of "catastrophe."

Just in the past couple of days, mathematics departments at the world's top universities have directly set off a massive collective uprising.

As of now, a full 1,008 mathematicians worldwide have signed on in a major joint statement, joining a newly established organization—the "Association for Human Mathematics" (AHM).

Just looking at the list of signatories, it is so long you cannot scroll to the bottom in one glance.

Even more astonishing, 654 top scholars have joined the internal "No-AI Core Group," publicly vowing to the whole world—

in their future daily research, whether finding proof ideas, picking counterexamples, or proposing new topics, they will 100% not use any AI.

Even Fields Medalist Terence Tao personally stepped in and reposted the statement.

In the same week on arXiv, the Barry Simon conjecture, which had stood unresolved for 26 years, came crashing down. ChatGPT, which pushed the counterexample parameters to the limit, surprised even the two human authors.

Tao used Claude to revise OpenAI's drafts, and the comment section erupted

At the start of the month, OpenAI dumped 700+ manuscripts overnight, becoming the trigger for all of this.

At the time, OpenAI tested about 4,000 unsolved problems, with each result consuming 3 hours of ChatGPT Pro reasoning compute, releasing 719 manuscripts in one go.

These manuscripts fall into 372 problem families, with about 42% accompanied by Lean formal verification. Although they run logically, the accompanying natural-language proofs are obscure and fragmented, and they omit a large amount of prior literature.

Faced with this batch of half-finished products, four scholars including Abhinav Nellore and Johannes Schmitt set up a GitHub open-source repository called oai-math-revised.

On October 24, Tao published a joint article by the four scholars on his blog. The article invited peers worldwide to submit code merge requests to fill in literature citations and rewrite the derivations for this batch of AI manuscripts.

Hidden in the repository's commit records is a highly dramatic detail. Several days before the repository went public, Tao, under the account name teorth, had already quietly gotten to work.

For item No. 003, the "Uniform Exclusion of Landau-Siegel Zeros" problem (a number theory problem concerning the uniformity of prime distribution), he successively submitted merge requests No. 2 and No. 3.

Unexpectedly, Tao drafted these two revisions using none other than Claude Code (Opus 5.5).

After verifying them, he changed the author line of the manuscripts to "OpenAI (with community contributions)" and in one stroke filled in the literature trail for Banks, Green, and himself.

As of October 25, 8 of the repository's 13 revision records were marked as drafted with the assistance of Claude Code.

This recruitment post sparked a head-on clash in the comment section of Tao's blog, with 21 long comments quickly splitting into two camps.

University of British Columbia mathematics professor Lior Silberman took a clear stance. He said bluntly that since the listed authors of these manuscripts are OpenAI, organizing the literature and writing the proofs clearly is the authors' own job.

In his view, if a commercial company wants academia to provide review and polishing services, it should pay at market rates. Scholars have no reason to serve as free volunteers.

Commenter Skip Tich pointed out the training-data calculation behind it. He noted that the high-quality proofs peers revise for free will be sucked into the next generation of model training sets.

Scholars who support revising the manuscripts also gave weighty reasons.

Commenter Nestor used an analogy: a multinational giant dumped toxic sludge in a national park. Although the pollution is not the volunteers' fault, if no one cleans it up, the entire academic ecosystem will suffer.

Facing the criticism, Tao personally wrote a reply on the morning of October 25, citing his own commit filling in the Siegel zero literature as an example.

He said the repository is meant to build a public stronghold for academia, controlled by the community itself. Only in this way can scattered corrections, follow-up research, and criticism be gathered together.

Even with a Fields Medalist personally stepping in to put out the fire, the rift over "whether humans should clean up after commercial models" continues to widen.

1,008 mathematicians publicly cut ties, 654 completely quit AI

Precisely because they do not want to become free labor cleaning up after commercial giants, over a thousand mathematicians chose to build a high wall directly.

English text is being prepared.

OpenAI drove the world's mathematicians to revolt

The trigger that pushed these thousand-plus mathematicians to publicly cut ties was the controversy over OpenAI's earlier high-profile promotion of having conquered the "Navier-Stokes equations" (one of the seven Millennium Prize Problems).

At the time, a mathematician input an unpublished manuscript into ChatGPT for polishing. OpenAI then mobilized about 10,000 Agents running concurrently for 88 hours, generating a massive proof that even experts found difficult to verify.

In its first statement, the association made things clear, directly accusing the commercial giant's recent intensive hype of "displaying compute hegemony, unrelated to rigorous academic research."

The statement criticized the giant for attempting to replace human understanding with an opaque statistical black box, turning mathematicians into "verification technicians" who stamp machine outputs.

Tech blogger Hesamation cited chess champion Carlsen as an example, saying that mathematicians rejecting AI is as absurd as chess players refusing to review games with a chess engine.

A 26-year-old conjecture falls, and ChatGPT deserves much of the credit

On October 22, Simon Becker of Bocconi University and Mengxuan Yang of Texas A&M University uploaded a paper to arXiv.

The paper disproves the "multidimensional L2 conjecture" proposed in 2000 by Barry Simon, a leading figure in mathematical physics.

This problem, which had stood unresolved for 26 years, originates from the Schrödinger operator in quantum mechanics. The operator's "spectrum" determines whether a particle can scatter freely to infinity (absolutely continuous spectrum) or is trapped in a local region (pure point spectrum).

In three-dimensional and higher-dimensional spaces, particle scattering channels are far wider than a one-dimensional straight line.

On this basis, Barry Simon conjectured at the time that as long as the spherical square-integrable integral in higher-dimensional space remains finite, an absolutely continuous spectrum must be preserved on the positive half-axis. In other words, particles would not be completely trapped.

For twenty-six years, countless mathematical physicists were stuck on the geometric obstacle that made it difficult to seal off all escape angles in higher-dimensional space.

Construction of a three-dimensional spherical shell counterexample, where each spherical shell contains a lens region (Lens) that focuses wave packets, a narrow-angle channel (Channel), and a periodic reflecting barrier (Reflecting barrier)

How was this paper produced? The two authors devoted a section specifically to explaining the entire process by which they and ChatGPT pierced the conjecture in three steps.

In the first step, the two authors checked OpenAI's batch of 719 manuscripts, found that this problem had not yet been claimed, and worked with ChatGPT to design a concentric spherical shell potential function.

This construction uses a weak angular potential as a "weak lens" to focus wave packets into narrow-angle channels, and then uses a radial periodic potential as a "reflecting barrier." In this way, the two successfully erased the absolutely continuous spectrum within a bounded energy interval.

In the second step, the two scholars judged from experience that this construction could at most go only this far, but unexpectedly, ChatGPT provided an unexpected advance.

Without changing the geometric structure, the model fine-tuned parameters and introduced the "Combes-Hislop spectral averaging criterion" (a tool for eliminating energy-level degeneracy and locking in exponential decay), proving a stronger pure point spectrum within a bounded interval.

The third step surprised the two scholars the most. In their own words, "To our great surprise, ChatGPT then pushed the parameters to the limit, extending the properties of the pure point spectrum all the way to the entire positive half-axis."

The human scholars thought the weapon in their hands could pierce only one shield. The machine took the weapon, adjusted the sights, and with one blow pierced the entire fortress.

After obtaining this unexpected parameter route, the two authors re-derived and finalized the entire 35-page manuscript from beginning to end, taking responsibility for all the proofs themselves.

In an era of excess proofs, who gets the credit?

These three controversies collided because, for the first time, the centuries-old sequence of "pose the question, win the honor, understand and verify the answer" has split apart.

In the past "era of proof scarcity," humans had to think through the methods along the way before they could reach the summit. Being the first to obtain the answer naturally meant contributing deep understanding.

Now, in the "era of proof excess," compute power has split this equivalence apart.

By stacking Agents, machines can be the first to compute results, but translating celestial tomes into human knowledge requires scholars to expend long, arduous labor.

If honor still goes only to the entity that first computes the result, giants can harvest headlines through compute power, while the scholars who distill ideas become unpaid laborers.

This also forces changes in academic evaluation and the training of young researchers, shifting from rewarding "who computes the result first" to valuing "who can pose good questions and establish new frameworks."

References:

This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelation, published by 36Kr with authorization.