Latest / Minds, Bodies, and Terawatts / OpenAI Says Its AI Solved a Million-Dollar Math Problem in 88 Hours. The Mathematician It Raced Says OpenAI Learned the Route From Him and Asked Why He'd 'Ruin His Career'
Transcript
- 0:00[Host] This is Minds, Bodies, and Terawatts, the show where the machine solved the equation and the humans started the fight. OpenAI announced on September 8 that an unreleased model, running about ten thousand agents at once, cracked one of the million-dollar Millennium Prize problems in about eighty-eight hours.
- 0:16[Guest] And the mathematician who had spent a year on that exact problem, inside OpenAI's own coding tool, says the company raced him down his own route and then asked why he would ruin his career.
- 0:28[Host] We covered the last one on Sunday, when Anthropic's Claude swarm formalized Fermat's Last Theorem in eleven days. So what's different this time?
- 0:36[Guest] That one re-checked old work. This one claims new work, and the story underneath is about whose work it was.
- 0:43[Host] Start with the math, in words I can hold onto.
- 0:46[Guest] The Navier-Stokes equations describe how water and air move, the math behind wing design and weather forecasts. The open question for ninety years was whether a perfectly smooth flow can whip itself up to infinite speed, and OpenAI's model says it can, in one specific setup with a smooth push applied from outside.
- 1:04[Host] Hold on. If most experts drop that outside push when they state the problem, is this a solution or a loophole?
- 1:12[Guest] The prize problem as written allows the push, so as written it is solved, and some of the field will grumble anyway. The fight is that almost nobody was attacking it that way. Mathematicians in Madrid opened the route, and Tristan Buckmaster at NYU picked it up about a year ago with Levent Alpöge, who works at Anthropic.
- 1:31[Host] Where were they doing the work?
- 1:33[Guest] As paying customers, typing a year of drafts into OpenAI's coding tool, Codex, and Anthropic's Claude, by Buckmaster's own account in a statement posted Monday night. Weeks earlier they had cracked a simpler cousin of the problem, checked it by computer, and were still writing it up when a rumor spread that Anthropic had solved a Millennium problem.
- 1:53[Host] And OpenAI heard the rumor.
- 1:55[Guest] Its own post says so, and the same day it launched its new model at every open Millennium problem. Its agents cracked the simpler cousin first, then it pulled agents off everything else and pointed them at Navier-Stokes, at a price OpenAI says would run a paying customer about fifteen million dollars.
- 2:13[Host] Fifteen million against a professor paying for his tools out of his own grant. What happened when they talked?
- 2:19[Guest] Buckmaster says OpenAI pressed him for a call for days, and on the Sunday afternoon he got Sébastien Bubeck, OpenAI's math lead. He was told the model had simply been handed the problem statement, and over the call, he says, that story came apart. There had been a whole team, easier problems first, an enormous amount of compute, and a first prompt sent only after word of his work reached the company.
- 2:44[Host] Did he ask whether the model had been trained on his drafts?
- 2:48[Guest] He did. He says he was told the model did not look up user data, and when he asked again, about training, he got no answer.
- 2:57[Host] What did OpenAI offer him?
- 2:59[Guest] A choice, by his account. Post his partial result and let OpenAI post the full one the next day, calling him the closest human to the problem. Or write the paper himself, credit OpenAI's model, and drop Alpöge as an author, because Alpöge works at Anthropic.
- 3:15[Host] They wanted the co-author gone for where he works?
- 3:18[Guest] Bubeck asked twice, Buckmaster says. When he refused and said he would go public, he says the reply was, why would you ruin your career, followed by, if you don't want me to be nice, then I don't have to be nice.
- 3:32[Host] What's OpenAI saying?
- 3:33[Guest] Bubeck calls the allegations false and inflammatory, and told reporters neither the researchers nor the agents saw their work. But OpenAI's own post concedes it cannot rule out that de-identified data from their use of its products helped improve the model.
- 3:47[Host] Cannot rule out. Is that the sentence for anyone who types anything into these tools at work?
- 3:53[Guest] It is, and TechCrunch points out OpenAI's terms let it train on Codex sessions unless the user opts out. Buckmaster himself says he is accusing nobody and has not seen the proof.
- 4:04[Host] So say I'm a contractor who works out my bids in ChatGPT, or a designer who drafts in Codex. Where does my work go?
- 4:12[Guest] Into the tool, and from there wherever the terms allow, and the terms are written by the company that also wants to sell your output without you. Buckmaster paid the bill, did the year of thinking, and the tool's owner had the compute to finish his idea before he could.
- 4:29[Host] So he keeps the math and loses the credit.
- 4:33[Guest] He gets credit for the first step, and OpenAI's post says so. What he lost is the year, and any say over what his year taught the machine that beat him.
- 4:43[Host] What's the takeaway from an episode about an equation for how water moves?
- 4:48[Guest] The book Unscarcity says an economy after labor has to run on contribution that is recorded to the person who made it, what Patrick Deglon calls Impact, and this week the contribution got recorded as usage data, which has no author. For more on who owns the work you do inside a machine, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.