OpenAI's math findings built on stolen user data

# Is OpenAI's Math Breakthrough Built on Stolen User Research?—NYU Math Professor Tristan Buckmaster's Allegations **TL;NYU: NYU math professor Tristan Buckmaster has publicly accused OpenAI of obtaining and "copying" his and his collaborator's research while they were nearing completion, allowing OpenAI to race ahead and publish a solution to the Navier-Stokes equations—a Millennium Prize Problem. OpenAI has denied that any prompts influenced its system.** ## Background: Hundreds of Math Papers and the Millennium Prize This week, OpenAI released hundreds of new mathematical research results spanning algebra, number theory, logic, and other fields. The release followed last month's announcement in which the company claimed to have solved the Navier-Stokes equations—a Millennium Prize Problem, which awards $1 million for each long-standing open question in mathematics. Tristan Buckmaster, a math professor at New York University (NYU), was invited for an interview. He stated that partial results from his work with another researcher may have inspired the prompts OpenAI used to reach its conclusion—on the very problem he himself had been researching for a long time. ## Why the Navier-Stokes Equations Matter Buckmaster explained that Navier-Stokes is one of the most fundamental and oldest equations in mathematics and physics, with a history of roughly a hundred years, making it one of the earliest partial differential equations. It describes fluids—water, blood flow, air—such as airflow around an airplane wing, and is the foundational equation of fluid dynamics. Whether its solutions can "blow up" has been an open question for centuries. A blow-up refers to some quantity tending to infinity within finite time—specifically, velocity going to infinity. Picture a fluid particle; physics breaks down at that moment, and the equations cease to be meaningful. When asked whether this result would change how airplanes are designed, Buckmaster said it could be useful for testing fluid models and many other applications, but blow-up is an extremely rare phenomenon and wouldn't fundamentally change much. ## From Rumors to an "Arms Race" Buckmaster and his collaborator Levitt Applegate (who works at Anthropic) were very close to solving the problem. By late August, they estimated they were about a month from completion. Then rumors began circulating: a "leak" from Anthropic to OpenAI, claiming that Anthropic was very close to solving two Millennium Prize Problems—Navier-Stokes and the Hodge conjecture. The information spread from OpenAI to the tech community, then to Twitter, and things began spiraling out of control. According to Buckmaster, OpenAI subsequently poured millions of dollars into compute to try to solve these Millennium Prize Problems. For them, it was almost a matter of survival: if Anthropic solved a Millennium Prize Problem first, it would prove their internal models were stronger. Incidentally, rumors that Buckmaster's team had "already solved Navier-Stokes" were false—they had actually achieved a slightly less significant result: the Euler equations, which are essentially one step beyond the Navier-Stokes problem. ## Outreach, Proposals, and the "Notorious Conference Call" With the rumors running wild—to the point where people were placing bets on prediction markets about whether Anthropic would solve a Millennium Prize Problem—Buckmaster reached out to a mathematician at OpenAI in an effort to cool things down. He told them: "We use your products too; this isn't some big Anthropic project—it's just me and Levitt, and we'd love to collaborate." OpenAI's response was "let's not compete," and they offered to provide compute support—Buckmaster emphasized that this offer came before the conference call, not via email. The resources they proposed amounted to "a few million tokens," not a few million dollars. The real turning point came on a Saturday phone call. Buckmaster was contacted and told that the other side was about to do "something stupid" and that they should get on a conference call—this became the so-called "notorious conference call." They had no idea what the other party was about to propose. Buckmaster described how, during the call, OpenAI first tried to "bribe" him: offering to let him publish their paper as sole author on the condition that Levitt not be involved, since he was an Anthropic employee. Then came the threats. Buckmaster responded that the proposal was completely unethical and said he would make the matter public. The other side replied, "Why would you destroy your own career?" Sebastian Bubeck of OpenAI said on the call: "I've been very nice to you, but I can stop being nice." Buckmaster said that at that moment, he knew negotiations were over. The host questioned: "I don't quite understand what these negotiations were about—either you did the work, or they did the work." Buckmaster explained that what they had solved was the Euler equations, a different problem, and the context of the negotiations was the other side realizing they had done something wrong. ## The Core Allegation: From "No Progress" to "Picking Up Where We Left Off" Buckmaster's central allegation is this: after he put his name out publicly, it took OpenAI roughly one day to become fully aware of where his team stood in their research. "So they copied our work." Regarding the timeline, he said that before knowing his name, the other side had made no progress at all. After learning it, they suddenly picked up exactly where he and his collaborator had left off. He also noted that internal communications within OpenAI's systems showed they had been working on all these problems, including the research Buckmaster's team was doing: they used 10,000 agents to derive Navier-Stokes from the Euler equations, and used 100 agents to arrive at the exact same result as theirs—using the exact same argument, the exact same method—within a single day. "Honestly, I can't tell you how they did it." He described the situation as "suspicious." ## The Data Usage Toggle: A "Critical Piece" Buckmaster confirmed that he was using OpenAI—including ChatGPT conversations and Codex—to interpret and advance portions of the work he received from Applegate. This meant all of their research output was stored on OpenAI's servers. He was using a personal paid account. When asked whether he had enabled the setting allowing OpenAI to use his data for training, he replied: "Saying it was on or off is not entirely accurate—I had two accounts. I remember turning it off, but on one account it was still on." The feature is enabled by default. Buckmaster also stressed that he wasn't certain exactly what happened and didn't believe anyone had explicitly ordered "go steal this person's data." He said: "It was some person or some thing. Those agents, if they deployed agents." He suggested OpenAI might not even know what happened, given that their initial statement was that they "couldn't determine" it. He added that OpenAI does not fortify user data as securely as people might assume—recently, a white-hat hacker used Claude to break into its systems, gaining internal access and even user data. In his view, the entire sandbox system is "already unreliable." ## OpenAI's Statement and the Evaporation of the "One-Shot" Claim OpenAI's statement was: "We can state unequivocally and with absolute certainty that every prompt Dr. Buckmaster submitted via Codex over the past two months could not, in any way, have influenced our systems, including training." Buckmaster pointed out the wiggle room in this statement: they said "the past two months," yet the problem is that they keep issuing statements and then revising them—their statement to *The New York Times* initially said the time frame was April, and later changed it to July. "They keep lying." During the conference call, OpenAI had claimed they solved Navier-Stokes "one-shot"—that is, by putting the entire prize problem into a single prompt, hitting enter, and getting the answer. But that claim evaporated within minutes, as they admitted it wasn't true. ## Did OpenAI Deny the Tactics? The host noted that OpenAI denied using pressure and threats, and asked Buckmaster to respond. He said they hadn't actually denied anything: - They acknowledged having offered to make him sole author; - They did not deny the phrase "Why would you destroy your own career?"; - Bubeck never denied "I've been very nice to you, I can stop being nice." "They haven't actually denied anything I've said." ## Current Status and the Bigger Picture Buckmaster says he is still using Anthropic and still has access to OpenAI's Codex, but adds: "I'm not putting any ideas into it. Nothing. I'm keeping everything original." Regarding the outcome, he says solving Navier-Stokes was never his goal—he was already thrilled to have solved the Euler equations, and so he will not receive the $1 million prize. In his view, the bigger issue is this: the wave of paper releases on Tuesday destroyed many early-career mathematicians' careers, wiping out their entire research programs in a single "dump." Moreover, OpenAI's method for obtaining user research proposals doesn't even require using an LLM—when researchers submit proposals, reviewers need to write summaries, and the way they typically do so is by pasting the proposal into an AI tool. This creates a massive conflict of interest: in the coming years, these companies will hold the research programs and directions of every mathematician and scientist. ## Conclusion: Users Should Be Careful The message from Buckmaster is simple: "Users should be careful." He concluded: "The fact that they published results ahead of their own customers shows their contempt for their customers. Why would a company do that?" Source: OpenAI's math findings built on stolen user data — CuTe_M0nitor (https://youtu.be/GweKOBSJvEY)