AI labs' claimed maths breakthroughs spark disputes with mathematicians over credit and release

Left · 1Over the past year OpenAI, Anthropic and other AI labs have announced breakthroughs on numerous long-standing mathematical problems, in some cases going well beyond what researchers expected current systems could do, including resolving one of the Millennium Prize problems. 1 Left · 1The Verge says results that might normally have been celebrated have instead drawn backlash, and that the labs say they are learning from earlier mistakes. 1 Left · 1Wired reports that OpenAI has told people it plans to put hundreds of results on GitHub on Tuesday 6 October, according to people familiar with the plans; McCallum said no release time has been set. 2

Left · 1McCallum said OpenAI began training a new internal model on 28 August, and that in addition to resolving the Navier–Stokes Millennium Prize problem it has resolved more than 100 long-standing open problems across most areas of mathematics. 2 Left · 1McCallum said the release would draw on recommendations from the Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study; Wired says OpenAI assembled an advisory group in mid-September to help decide how it assesses and communicates results, though mathematicians say little progress has been made. 2 Left · 1Northwestern University mathematician Bryna Kra said that at an August meeting mathematicians asked OpenAI to publish papers rather than blog posts or tweets, so the work could be absorbed and used, and that this input was apparently ignored. 2

Left · 1Wired describes the most striking dispute as arising in September, when OpenAI deployed thousands of agents at a Millennium Prize problem after hearing rumours that others were close; NYU professor Tristan Buckmaster accused the company of front-running work done with Anthropic employee Levent Alpöge, which the pair had not published. 2 Left · 1Buckmaster claims OpenAI researcher Sébastien Bubeck seemed to imply Alpöge should be left off any paper; meeting notes seen by Wired record Bubeck as suggesting Buckmaster would be ruining a career. Bubeck pointed to a previous public statement denying he asked for Alpöge not to be listed as an author. 2 Left · 1NYU visiting professor Nestor Guillen said there is "a perception of mobster behavior" from AI companies among mathematicians, and McCallum replied that OpenAI disagrees with that characterisation. 2

Left · 1Mathematicians told Wired they feel the field has become a playground for OpenAI and Anthropic to show off models ahead of expected IPOs, and that normal processes for releasing and attributing results have been set aside; many object to blog-post releases that are harder to verify and often leave out prior work. 2 Left · 1Wired notes that piecemeal announcements are not limited to the labs, since Alpöge announced disproving an 87-year-old conjecture by tweet; the community has also built tools such as Hexagon, a repository of mainly AI-generated material, and Palomar, a registry of machine-verified mathematics, which Kra says OpenAI has not changed its behaviour to use. 2 Left · 1Kra, who helped write the Leiden declaration signed by more than 4,000 mathematicians, does not see OpenAI's actions as aligned with it. Bubeck told Wired he sees AI as an opportunity to expand what mathematicians can do, and McCallum said OpenAI does not believe the future of mathematics is set. 2