@rohanpaul_ai: Full video on CNBC

Full video on CNBC

https://t.co/2wCR3EtIJW

TL;DR: NYU mathematics professor Tristan Buckmaster accused OpenAI on CNBC of potentially using research data from his personal accounts to reproduce, with agents in a single day, the nearly completed results on the Euler equations that he and his collaborator had developed. He also alleged that OpenAI offered him sole authorship as an inducement and threatened his career during negotiations. OpenAI denied that his prompts had influenced its system.

Background: OpenAI’s Mathematical Releases and the Millennium Prizes

OpenAI released hundreds of new mathematical research results within a week, spanning a wide range of fields including algebra, number theory, and logic. This followed its announcement last month that it had solved the Navier–Stokes equations—one of the Millennium Prize Problems, which offer a $1 million reward for solving each longstanding mathematical challenge.

NYU mathematics professor Tristan Buckmaster joined CNBC remotely and said that recent progress in research by him and another researcher may have inspired OpenAI’s prompts, leading to the solution the company had been pursuing for a long time.

What Are the Navier–Stokes Equations, and Why Do They Matter?

Buckmaster said these are among the most fundamental equations in mathematics and physics, as well as among the oldest, with a hundred-year history, belonging to one of the oldest classes of partial differential equations. They describe fluids—water, blood flow, and air—and are also used in aerodynamics and aircraft design. They are the most fundamental equations in fluid mechanics.

For hundreds of years, an open question has been whether their solutions can blow up.

The host asked whether something approaching infinity in a finite amount of time meant “to infinity and beyond.” Buckmaster corrected him: “Just to infinity.”—The velocity approaches infinity. Imagine a fluid particle moving, the laws of physics breaking down, and the equations no longer making sense.

The host asked whether solving the problem would change how aircraft are designed. Buckmaster thought it could be used in various ways to test fluid models, but “this situation is so rare that it won’t fundamentally [change things].”

For Buckmaster, the problem was not something he had studied for hundreds of years, but “the North Star of almost my entire research career.”

Rumors, Competition, and the Timeline

Buckmaster’s collaborator is Levitt Applegate. The two were already close to solving the problem. At the end of August, Buckmaster estimated that they were about a month away from a solution.

At that point, rumors emerged: information leaked from Anthropic to OpenAI claimed that Anthropic was close to solving two Millennium Prize Problems—the Navier–Stokes problem and the Hodge conjecture. The information spread from OpenAI to the tech community and then to Twitter, and “things got a little crazy.” OpenAI then began spending large sums on computing resources in an attempt to solve these Millennium Prize Problems.

Buckmaster’s assessment was that this was almost a matter of life or death for them—“If Anthropic solves a Millennium Prize Problem, that means their internal models are better.”

One piece of background needs clarification: Levitt works at Anthropic. But Buckmaster emphasized that this was not some major Anthropic project: “It’s just me and Levitt,” and he himself was using OpenAI’s products.

As the rumors spread, Buckmaster reached out to a mathematician at OpenAI to try to calm things down. The reason was straightforward: the rumors were false. They had not solved Navier–Stokes; they had solved the somewhat less advanced Euler equations. They realized that Navier–Stokes was a prized project for these tech companies—if a competitor solved it, the consequences would be serious.

OpenAI initially replied by email: “Let’s not compete,” offering computing resources so they could collaborate. Buckmaster clarified that the offer was not millions of dollars, but millions of tokens to complete the project.

That “Infamous Conference Call”

On Saturday, Buckmaster received a message: someone told them that OpenAI was going to “do something stupid” and asked them to join a conference call. They did not know exactly what it was about, but it was enough to persuade them to participate.

The call unfolded as follows:

• Inducement: OpenAI said that as long as Levitt was not involved—because he was an Anthropic employee—Buckmaster could publish their paper as the sole author.
• Threat: When Buckmaster said the proposal was completely unethical and that he intended to make it public, they responded: “Why would you ruin your career?”
• Showdown: OpenAI’s Sebastian Bubeck said: “I’ve been very nice to you. I don’t need to be nice anymore.” Buckmaster said that at that moment, he knew the negotiations were over.

The host challenged this: “I don’t understand what’s being negotiated here. Either you did the work, or they did the work.” Buckmaster replied that the negotiations concerned “their realization that they had done something wrong.”

Asked whether OpenAI denied using these tactics, Buckmaster said: “Actually, they haven’t denied any of the tactics. They admitted them.” They admitted offering him sole authorship and acknowledged saying, “Why would you ruin your career?” Sebastian also never denied making that remark. “They actually haven’t denied anything I’ve said.”

The Core Allegation: Data and Agents

The host pressed him on the central issue: What exactly are you accusing them of doing?

Buckmaster’s answer was that after he publicly disclosed his name, “they may have taken a day to get to exactly where we were in our research. So they were reproducing our work.”

The timing detail was that he publicly disclosed his name on the third day.

How the Data Entered OpenAI

Buckmaster used OpenAI’s products—ChatGPT and Codex—to help interpret some of the work and advance the research results he received from Levitt. This meant that “all of our research results were on OpenAI’s servers.”

As for his accounts, he used paid personal accounts. Asked whether he had enabled permission to “use your data for training,” he answered “yes and no”—he had two accounts at the time, and he remembered disabling training on one while leaving it enabled on the other. That setting was “on by default.”

He further suggested that OpenAI’s barriers around user data might not be as robust as people assume: recently, a white-hat hacker had used Claude to break into its systems, gain internal privileges, and access user data. The entire sandbox system was “not very reliable.” “So I don’t know what happened. I really don’t know.”

On whether someone deliberately looked at his private data, Buckmaster’s view was: “Someone or some system. Those agents.” He did not think it necessarily involved a person ordering someone to steal this individual’s data. “Honestly, I don’t think they even know”—their initial statement was that they could not determine what had happened.

OpenAI’s Statement and the Time Window

The host quoted OpenAI’s statement: it could definitively say that Dr. Buckmaster’s Codex prompts from the past two months could not have influenced the system in any way, including through training. The host pointed out that the key wording was “the past two months”—did they need data from those two months to arrive at the answer?

Buckmaster responded that the problem was that OpenAI kept changing its story: it initially told The New York Times April, then changed it to July. “They keep lying.”

The “One-Shot” Claim and the Scale of the Agents

During that conference call, OpenAI claimed that it had “solved Navier–Stokes in one shot”—meaning that they entered the prize problem as a prompt, pressed Enter, and the solution appeared. “Then that claim disappeared within a few minutes. They admitted it wasn’t true.”

Buckmaster also said that OpenAI’s communications with its staff on an internal messaging platform showed that it had worked on all the other problems, including the work he himself had been doing: it used 10,000 agents for the Euler equations and 100 agents for Navier–Stokes, reaching the same point as his team within a day, using exactly the same arguments and exactly the same methods.

His conclusion was: “They had made no progress before I knew my name. After they knew my name, they suddenly picked up where we had left off.” He acknowledged that he could not say for certain how they had done it, but “I find it suspicious.”

His Current Situation and the Bigger Picture

Buckmaster is still using Anthropic. He also still has his OpenAI account and Codex, but “I don’t put any ideas in there anymore. Anything. I decided to stay original.”

Regarding the $1 million prize, he said he would no longer receive it, but “for me, honestly, solving the Navier–Stokes problem was never my goal. I’m very happy to have solved the Euler problem.”

He believes the bigger picture is that the batch of papers released on Tuesday destroyed the research projects of some young mathematicians early in their careers—their entire research projects were wiped out. A deeper issue concerns where research proposals end up: after researchers submit proposals, reviewers need to write a summary, and the way they write it is by entering the content into a large language model. This means OpenAI “has the research projects of all mathematicians and scientists for the next few years, along with information about how they will conduct their research,” creating an enormous conflict of interest.

The host summarized his message as “user beware.” Buckmaster responded that OpenAI uses its customers’ data to get ahead of them: “That tells you how they view their customers. A computer company…”

Source: @rohanpaul_ai — Full video on CNBC (https://www.youtube.com/watch?v=GweKOBSJvEY)