771 Mathematicians vs. a $2M Hackathon: OpenAI Pulls Out of Caltech's Mathathon

One week after Caltech undergraduates announced “Mathathon” — a 40-hour hackathon where 100 teams would attack open research problems by prompting frontier LLMs, backed by roughly $2 million in compute credits from OpenAI and Anthropic — the event ran into a wall. Not a technical wall, but a human one: an open letter, signed by 771 mathematicians at the time of publication and now past a thousand, demanding the event be cancelled outright. And on September 10, OpenAI blinked first, withdrawing its sponsorship entirely.

What the Mathathon Was Supposed to Be

Announced on September 4, the Caltech Mathathon was pitched as the first hackathon ever devoted to research-level mathematics. The format was straightforward: 100 teams, 40 hours from October 30 to November 1 in Pasadena, access to frontier AI models and over $2 million in AI compute credits, working on open conjectures. On the final day, teams would present their results. Support came not only from OpenAI and Anthropic but also DARPA’s expMath program and Cognition.

The organizers — Caltech undergraduates Alvan Arulandu, Sathvik Rodrothu, Caiman Moreno-Earle, Avni Garg, and Brian Zhao — framed it as an experiment in resource allocation: giving young mathematicians access to compute that normally only well-funded established researchers enjoy. They received over a thousand applications within a week, from undergraduates to faculty at top institutions in the U.S. and abroad.

Then came the website copy that lit the fuse. The event’s site asked: “What is the role of a mathematician when AI can solve conjectures faster?” — a line the organizers have since retracted and scrubbed from the site.

The Open Letter: “Research Misconduct”

Published September 10 on Proofs and Prompts by “a coalition of current and former Caltech mathematicians,” the letter is remarkable for how precisely it targets the sociology of mathematics rather than AI capability itself.

Its core argument: AI companies see research mathematics as an advertising opportunity. In the arms race to claim the first AI-generated proof of a prominent conjecture, results get announced on social media weekly, often poorly communicated, with “diffuse negative impacts on the careers of human mathematicians who are concurrently proving the same results.” Human researchers are then compelled to verify, disseminate, or discredit those claims — labor the letter calls “uncompensated, uncredited, and unacknowledged.” The letter’s blunt conclusion: “AI companies are engaging in research misconduct.”

Five specific harms are enumerated: the creation of “slop mathematics” in an accelerated environment with verification displaced onto the community afterward; the intrusion of corporate leverage over an independent scientific community; conditions — 40 hours — unlikely to foster actual mathematical understanding; damaging misinformation about what research mathematics is; and the exploitation of early-career mathematicians whose association with the sponsoring labs could plausibly harm their future reputations.

The letter also made an economic point that’s hard to shake: “There is no reasonable future model of mathematical research in which every research mathematician receives 20,000 USD in AI credits to prove a result.” The signatory list includes Caltech community members and PhD-level mathematicians worldwide; by September 11, Reddit threads reported the count had passed 1,000.

The Organizers’ Response: Concessions, Not Surrender

Hours after the letter, the organizers published their own response — and rather than cancelling, they made real concessions. They committed to making arXiv publication an explicit requirement in the second-round verification period, retracted the “AI can solve conjectures faster” line, and said future public statements would be reviewed by faculty advisors.

Their structural defense is worth reading in full: Mathathon is two rounds, not one. The 40-hour sprint produces initial results, but teams whose work merits consideration progress to a six-month verification period in which they must produce an arXiv preprint, a video presentation, and an auto-formalized solution, with judges evaluating throughout and additional prizes at the end. There’s also an educational track — course notes, video explainers, elegant re-proofs of existing theorems — created in response to a signatory’s proposal, with a similar prize pool. The organizers note that the majority of applicants so far are PhD students with publication records, that they’ve signed the Leiden Declaration on responsible AI use in science, and that they plan debates and panels between AI critics and supporters at the event itself.

Their sharpest counterpoint concerns the very people the letter claims to protect: amid a steep decline in PhD applications for theoretical fields, they argue, events like Mathathon keep young people excited about mathematics when access to compute asymmetrically favors the established and well-funded.

OpenAI Walks; Anthropic Goes Quiet

The decisive move came from OpenAI. Research lead Dan Roberts posted on X: “We recognize that the rapid progress of AI in mathematics is disruptive. We’re looking to engage with the math community more on the best way to integrate this technology and communicate its impacts.” The company notified the organizers it was withdrawing its sponsorship. Anthropic — the other half of that $2 million in credits — did not immediately respond to requests for comment.

The timing could hardly be worse for OpenAI’s math program. Its September 8 announcement that Astra had resolved the Navier-Stokes existence and smoothness problem is being actively contested, and TU Dresden’s Andreas Thom has publicly alleged that his private ChatGPT conversations fed into an OpenAI result on non-sofic groups — the third public misconduct allegation against the lab’s math efforts in a single week, following Tristan Buckmaster and Levent Alpöge’s Navier-Stokes dispute.

Why This Matters Beyond Pasadena

This episode is a template for a collision that’s coming to every research field AI touches. The dispute is not about whether LLMs can contribute to mathematics — the letter doesn’t deny it — but about who bears the cost of verification, attribution, and cleanup when results are extracted at industrial speed. Mathematicians are unusually well-positioned to push back: proof is their native currency, their community is small enough to coordinate a 771-signature letter in days, and their endorsement is exactly what AI labs need to legitimize claims of mathematical achievement. As one tracker put it, a withdrawal forced by the academic community whose credibility validates lab math claims is a bigger reputational hit than a missed benchmark.

Watch the next signal: whether Anthropic follows OpenAI out before October 30, or stays and absorbs the reputational differential alone. And watch whether the Mathathon — now stripped of one sponsor but armed with a restructured two-round format — actually demonstrates the thing both sides claim to want: a model for AI-assisted mathematics that doesn’t treat the humans as disposable infrastructure.