OpenAI just wants to win

Robert Hart’s Verge feature, drawing on more than a dozen mathematicians including Tristan Buckmaster and Andreas Thom, argues that OpenAI’s rush on a Millennium Prize problem—Navier-Stokes—exposes a clash of motives. Researchers say they want to advance the field; they see OpenAI as wanting to win.

OpenAI says it heard others were making progress, then threw roughly 10,000 agents, tens of millions of dollars of compute, and about 88 hours at Navier-Stokes. Buckmaster says he was offered vast compute and sole authorship of OpenAI’s paper if he cut Anthropic-affiliated collaborator Levent Alpöge out—an offer he rejected as a bribe. OpenAI’s Sébastien Bubeck has disputed that framing while acknowledging resource offers and that Alpöge’s Anthropic tie was a sticking point.

This is no longer a quiet academic priority fight. It is a story about whether frontier labs can race academics to famous problems, use the same tools those academics rely on, and still claim the community’s trust.

We’re not here to referee every contested conversation. The Verge lays out a contested timeline and strong feelings on both sides. What we can say clearly: industrial agent swarms plus rival-lab paranoia change the sociology of math. Scooping used to be hard because expertise was scarce. Now a rumor can summon thousands of agents and a compute budget no university group can match.

Buckmaster’s Codex worry and Thom’s ChatGPT worry land in the same place—conflict of interest. OpenAI is both a tool vendor and a competitor for credit. A spokesperson told The Verge it is “impossible” for Buckmaster’s recent Codex prompts to have influenced the system, including training; Buckmaster says take that with skepticism. On Thom’s earlier episode, OpenAI did not answer whether ChatGPT conversations could have helped a result that built on his and Gábor Kun’s work. Uncertainty itself is corrosive when only the company holds the logs.

Useful AI in mathematics is real and worth celebrating when proofs are open, lineage is credited, and collaboration beats ambush. A Millennium-scale result that survives community scrutiny would be a genuine gift to science. The harm is chilling: junior researchers already risk their careers on hard problems; Fields medalist Shing-Tung Yau told The Verge they may grow even more reluctant if AI labs can out-spend and out-announce them. Lists of open problems start to look like target menus.

I’m watching the Clay Institute’s deliberately slow acceptance process for Navier-Stokes, OpenAI’s hint of “substantial progress” on another Millennium problem, and whether the Leiden Declaration’s nearly 3,900 signatories translate into norms labs actually accept. Prestige is a currency. Trust is the one that keeps the field from going secretive and scared.

The Clay Mathematics Institute’s seven Millennium Prize problems each carry a $1 million bounty; only the Poincaré conjecture had previously fallen. The Institute has removed Navier-Stokes from its unsolved list but has not yet declared it solved, citing a two-year general-acceptance period. Separately, OpenAI withdrew sponsorship of a Caltech undergraduate math hackathon after opposition, and mathematicians have published the Leiden Declaration on responsible AI use in the field.