From The Verge · Robert Hart · · 4m
OpenAI just wants to win
A Millennium Prize solution turned into a fight over credit, trust, and who gets to scoop a mathematician.
In brief
A Millennium Prize solution turned into a fight over credit, trust, and who gets to scoop a mathematician. When the company that builds your research tools also competes with you for results, the problem is not the math, it is the conflict of interest nobody has answered. Originally reported by The Verge.
Quick answers
- Did Codex see his work?
- Buckmaster had used OpenAI's Codex on the problem and asked whether his prompts fed the model. OpenAI says categorically no, and he is not buying it.
- What is the key takeaway from OpenAI just wants to win?
- When the company that builds your research tools also competes with you for results, the problem is not the math, it is the conflict of interest nobody has answered.
A prize claimed, not celebrated
OpenAI says it solved the Navier-Stokes problem, one of the seven Millennium Prize problems. Instead of applause, mathematicians reacted with unease.
Two different goals
Mathematicians work to advance the field. OpenAI, many of them believe, is playing a different game entirely.
“Mathematicians want to advance the field. OpenAI wants to win.”
A field shaken
The Verge spoke to more than a dozen mathematicians. Even the skeptics of the harshest allegations described a community rattled and worried about what comes next.
88 hours, 10,000 agents
OpenAI heard rumors that researchers were closing in, pointed an unreleased model at Navier-Stokes, and burned tens of millions in compute to get there first.
The rival in the room
The researchers racing OpenAI were Tristan Buckmaster of NYU and Levent Alpoge, who works at Anthropic. Alpoge called it a personal collaboration. OpenAI saw a competitor.
“How can we have an internal OpenAI project with an Anthropic employee?”
The offer on the table
Buckmaster says OpenAI dangled near unlimited compute and sole authorship on the announcement paper. The condition was that his collaborator disappear.
“All I had to do was throw Levent under the bus.”
He called it a bribe
Buckmaster turned the offer down and went public with both his mathematics and his account. His office became a war room of colleagues, outreach, and lawyers.
Did Codex see his work?
Buckmaster had used OpenAI's Codex on the problem and asked whether his prompts fed the model. OpenAI says categorically no, and he is not buying it.
“Given their behavior up until this point, one should take such statements with great skepticism.”
OpenAI's side
Sebastien Bubeck disputes the threatening characterization, but confirms he offered company resources. He also said OpenAI had cut similar deals with other mathematicians he would not name.
Why these problems
The Clay Institute set seven Millennium problems in 2000, each worth a million dollars. In 25 years only the Poincare conjecture has fallen. For an AI lab, they are irresistible targets.
Most mathematicians aren't trophy hunters
Oxford's Andras Juhasz describes the work as part science, part art, usually with no practical payoff. People do it for the discovery, not the prize.
“They do it because it's beautiful. They enjoy it. It's the sense of discovery.”
The ideas matter more than the answer
Frontier problems have no set route. The techniques invented along the way often outlast the problem itself, which is why lineage and attribution are the whole point.
A chasm in values
Bubeck looked shocked when Buckmaster said he did not care about the Millennium Prize. In tech, Buckmaster says, being first and being seen to be first is everything.
“That's the only currency.”
A quiet edit to the record
Andreas Thom and Gabor Kun's work underpinned an earlier OpenAI result. The company amended its announcement to credit them without ever saying it had.
They may not even know
Thom wonders whether his ChatGPT conversations helped train the model that cracked his problem. Only OpenAI can answer, and he suspects it cannot.
“To be honest, I suspect that they don't even know.”
The same fight as artists
Writers, musicians, and publishers have all asked what companies owe the people whose work built their systems. Mathematics has now joined the queue.
Scooping goes industrial
Priority disputes are old news in mathematics, but scooping was hard because so few people had the expertise. Thousands of agents on demand changes that math.
Limbo, not victory
Clay requires two years of general acceptance before awarding anything. Navier-Stokes has been pulled from the unsolved list but is not declared solved.
“The process is deliberately unhurried.”
Already on to the next
OpenAI says it has made substantial progress on another Millennium problem, possibly the Hodge conjecture. Rumors put Anthropic close to one too.
A chilling effect
Researchers are getting secretive about unfinished work, and lists of open problems now look like target lists. Nearly 3,900 people have signed the Leiden Declaration.
“The reality is that mathematicians are actually scared of them.”
What matters most
When the company that builds your research tools also competes with you for results, the problem is not the math, it is the conflict of interest nobody has answered.
