Did OpenAI Solve the Hodge Conjecture? What's Known, What's 'Doug,' and Why It's Quiet

On the morning of September 17, The Information's AI Agenda newsletter reported that OpenAI is close to solving the Hodge Conjecture — which would be its second Millennium Prize problem in as many weeks — and that any announcement may wait while the company, in reporter Stephanie Palazzolo's words, 'figures out how to navigate its rocky relationship with the math community.' Within two hours, prediction markets moved and the recap accounts were off. So let's do what the recaps won't: sort what's reported (a single outlet), what's documented (a genuinely wild two weeks in mathematics), and what's unknown (more than the excitement suggests).

What's reported — and how thinly

The Hodge claim comes from one outlet: The Information's paywalled newsletter, with the 'staff expect it relatively soon' framing reaching most people through Techmeme's paraphrase. The 'Doug' detail — that the earlier Navier–Stokes solve used a variant of OpenAI's next pretrained model under that codename — reaches us only through widely shared summaries of the newsletter, though 'Doug' itself is a real OpenAI pretraining codename The Information has reported on since late 2025, not an invention. OpenAI has said nothing about the Hodge report, and no named source appears anywhere.

Here's the part that makes it credible anyway: OpenAI already told us something is coming. In a statement reported by The New York Times around September 9, the company said that 'since the completion of Navier-Stokes, we have made substantial progress on another Millennium Prize problem. We are working through how to share these results thoughtfully.' OpenAI never named the problem — The Information just did. The leak fills in a blank OpenAI drew itself.

The scar tissue: what actually happened with Navier–Stokes

To understand why OpenAI might sit on a proof, you need the documented record of the last one. On September 8, OpenAI announced a resolution of the Navier–Stokes problem — roughly 10,000 agents on an unreleased internal model, an 88-hour run, five million inter-agent messages, a 166-page paper proving finite-time blowup, and a formalization in Lean, at a compute cost OpenAI put at several million dollars. Charles Fefferman, who wrote the problem's official statement, told Quanta he was 'thrilled.' OpenAI said it would not claim the $1 million prize.

Then the war started. NYU's Tristan Buckmaster — who had spent a year quietly working the same territory with Anthropic-employed mathematician Levent Alpöge — alleged their advances leaked to OpenAI days before the announcement, and described pressure tactics from OpenAI's Sébastien Bubeck ('Why would you ruin your career?'), which Bubeck denies. Terence Tao warned that 'the indiscriminate strip-mining of open problems' could destroy mathematics' ecosystem. By September 11, 25 Fields Medalists — including Tao, Scholze, and Villani — signed a declaration titled 'A Severe Misalignment of AI in Mathematics' stating bluntly that 'AI companies are engaging in research misconduct,' with over 1,000 mathematicians signing a companion letter. The Clay Mathematics Institute's response: the problem 'has apparently been settled,' and its evaluation process is 'deliberately unhurried.'

That's the context for Palazzolo's line about a delayed announcement. The math is going faster than the mathematics community's trust can absorb — and OpenAI appears to know it.

The Hodge Conjecture in 90 seconds

Posed by W.V.D. Hodge in 1950 and one of the Clay Institute's seven Millennium Prize problems, the conjecture asks, roughly: for a particularly well-behaved class of geometric spaces (projective algebraic varieties), is every piece of their topology that looks like it could come from algebra actually built from algebraic pieces? It's a bridge between shape and equation — proven in low dimensions, open in dimension four, and considered one of the deepest structural questions in geometry.

One thing is certain regardless of what OpenAI has: no prize money is coming anytime soon. The Clay Institute's rules require that a solution be 'published by a Qualifying Outlet,' that 'at least two (2) years have elapsed since publication,' and that it achieve 'general acceptance in the global mathematics community.' Whatever headlines say, the clock on a Hodge prize hasn't even started.

About that 57% on Polymarket

Prediction markets moved fast — Polymarket's 'Which Millennium Prize Problem will AI solve next?' put Hodge around 57–61% within hours. Read the resolution criteria before treating that as evidence: the market resolves when an eligible AI lab announces a solution by the end of 2027 — 'no external verification from Clay Math Institute is required.' That's not a bet that the Hodge Conjecture gets solved. It's a bet that a press release gets written. Useful as a sentiment gauge; worthless as mathematics.

Our read: proofs are the anti-benchmark

We've spent a month documenting how AI benchmark tables are marketing — scores on vendor-picked tests under vendor-picked conditions. Mathematics is the opposite regime: a proof either checks or it doesn't, and machine formalization (Lean verified the Navier–Stokes argument in 17 hours) makes the checking itself mechanical. That's why this story matters beyond math: it's the first arena where frontier-AI claims can be settled definitively, in public, with no benchmark theater possible. If OpenAI has a Hodge proof, we'll know for sure — eventually, completely, and regardless of anyone's marketing.

The sides, honestly weighed: the bulls note OpenAI frames math as 'the next step toward automating AI research' (per the newsletter's summary), and a second Millennium-class result in a month would be the strongest recursive-self-improvement evidence yet. The skeptics note 'close to solving' is doing enormous work in a single-source report about famously treacherous terrain — and that if agents leaned on human problem-selection (the Buckmaster question), human research taste is still the scarce ingredient. The mathematicians — the people who'd actually verify this — are the ones OpenAI reportedly now waits on, and as geometer Burt Totaro wrote on Tao's blog about this exact conjecture, 'the power of mathematical ideas comes from a continual negotiation among people.'

Bottom line: treat 'OpenAI is close to solving the Hodge Conjecture' as a credible, unconfirmed, single-source report sitting on top of one verifiable fact — OpenAI itself says it has substantial progress on a second Millennium problem. If a paper and a Lean artifact appear, it's the biggest AI story of the year. Until then, the confirmed story is remarkable enough: two weeks after AI settled its first Millennium problem, the bottleneck isn't the math anymore. It's the peace treaty with the people who check it.