OpenAI and the “Millennium Problems”: Why Mathematicians Have Stopped Trusting the AI Giants

This week OpenAI announced that its closed model had found a solution to one of the seven “Millennium Problems” — the Navier–Stokes equations, which describe the flow of fluids. According to the company, it took some 10,000 agents, tens of millions of dollars in computing resources, and just 88 hours. Under ordinary circumstances this would have been a historic event. Instead, the mathematical community greeted the news with mounting unease.

A Race Instead of a Collaboration

As The Verge reports, the conflict was sparked by a rumor: OpenAI had learned that New York University professor Tristan Buckmaster and Anthropic researcher Levent Alpöge were working on the same problem. The trouble turned out to be precisely Alpöge's affiliation with the main competitor. According to Buckmaster, he was offered “practically unlimited” computing power and sole authorship of the OpenAI paper — but on the condition that he “cut Levent out.” He regarded this as a “bribe” and refused.

OpenAI spokesman Sébastien Bubeck, in an interview with The New York Times, did not deny that the company had offered resources, and acknowledged that similar arrangements had been made with other mathematicians. But Alpöge's affiliation with Anthropic remained a sticking point: “How can we have an internal OpenAI project with an Anthropic employee?”

Whose Result Is It?

Buckmaster and Andreas Thom, a professor at the Dresden University of Technology, raised another question: could their queries to OpenAI's tools (Codex, ChatGPT) have influenced the training of the models? The company “categorically” denies this possibility, but mathematicians remain skeptical. “Honestly, I suspect they don't know themselves,” says Thom. “Which means they simply don't care.”

Here the same argument surfaces as in the AI companies' conflicts with writers, musicians, and artists: what do the companies owe to those whose accumulated labor they have drawn upon? For mathematicians, the “lineage” of ideas is especially important — new methods are often more valuable than the solved problem itself.

The Chilling Effect

Mathematicians fear that corporations will turn “getting ahead” of their colleagues into an industrial process, deploying thousands of agents at the first rumor of an imminent breakthrough. Many are already becoming more secretive; lists of important unsolved problems risk becoming lists of targets. In June the Leiden Declaration on the responsible use of AI was published, signed by nearly 3,900 people. OpenAI was forced to withdraw its sponsorship of a student hackathon at Caltech.

Fields Medal laureate Shing-Tung Yau points to a conflict of interest: the companies simultaneously supply research tools and compete with researchers, all while having access to others' unfinished work.

Formally, it is still too early to celebrate the victory: the Clay Institute requires two years and the “general acceptance” of the community before a problem is recognized as solved. Meanwhile, OpenAI has already reported “substantial progress” on another Millennium Problem. But, as Buckmaster concludes, it all comes down to trust, which is gone: “They don't care about us as a community. This is just petty drama between two trillion-dollar companies behaving like children.”

Source: The Verge