Talk Around Town: OpenAI Claims 377 New Math Breakthroughs, but Mathematicians Say Hold Your Horses

What Folks Are Saying Down at the Feed Store

Well, slap the mud off your boots and listen up, because OpenAI done showed up to the county fair with a wagon so loaded down it's draggin' axles. The company dumped what it describes as 722 manuscripts onto GitHub, claiming — and Lord, this is a number — 377 fresh mathematical results produced by an internal AI model it hasn't even let the public sniff at yet. OpenAI says these ain't just parlor tricks; the company describes results touching on a wide range of open mathematical problems, including, according to NewsBytesApp, something as gnarly as the Birch-Swinnerton-Dyer leading-term formula.

Word spread faster than gossip at a church potluck, and the reaction from the mathematics community has been about as warm as a January cattle pond. Scientific American — which got actual named mathematicians on the record — reported that OpenAI's reputation in the field for making bold claims while keeping its methods tighter than a rusty jar lid is fueling considerable skepticism. The chatter ain't quiet or polite, neither.

What We Actually Know for Sure

Here's the part where we separate the cornbread from the crumbles. Scientific American confirmed — and this part holds up — that a good chunk of the released proofs have been checked out in Lean, a programming language that acts like a mechanical referee for mathematical logic. A Lean-verified proof is about as logically airtight as a Mason jar full of moonshine, meaning the internal logical steps hold together. But — and this is a but as big as a hay bale — logical correctness don't settle whether a result is genuinely new, genuinely important, or genuinely understandable to a human mathematician who might want to build on it.

Scientific American also confirmed that Terence Tao, a mathematician whose reputation precedes him like thunder before a storm, has publicly characterized the pace at which frontier AI labs are churning out these results as 'insane.' That word came out of his mouth, not ours. Scientific American further confirmed that OpenAI itself has acknowledged — and this is a doozy — that some of its own newly released results are not yet understood by its own mathematicians. The company says, per Scientific American, that it does not intend to slow down, framing math problems as an essential test of whether its AI is genuinely getting smarter.

What Ain't Been Settled by a Long Shot

Now here's where the wagon loses a wheel. MIT mathematician Andrew Sutherland told TokenPost that every single one of OpenAI's claims ought to be treated as unverified until the company releases the model and lets outside researchers reproduce the work. That model, as of this writing, remains locked up tighter than a banker's safe, which means independent reproduction is flat-out impossible right now. Scientific American noted that even with Lean verification in hand, mathematicians may need months — possibly years — to assess the novelty and significance of these results across the specialized subfields involved.

Then there's a separate and uglier dispute simmering on the back burner. SuperPower Daily, citing prior WIRED reporting, noted that researcher Tristan Buckmaster has accused OpenAI of getting ahead of unpublished work he developed alongside Anthropic employee Levent Alpöge — raising thorny questions about attribution and who gets credit for what. That accusation remains contested and unverified; the underlying facts have not been independently confirmed, and a figure named by SuperPower Daily in connection with related allegations has publicly denied them. This one's murkier than creek water after a rainstorm.

The Math Community Is Fixin' to Draw a Line

Northwestern mathematician Bryna Kra didn't just complain into the void — TokenPost confirmed she helped put together something called the Leiden Declaration, which more than 4,000 mathematicians have signed. The gist of it, per TokenPost, is a demand that AI companies meet standards the mathematics community actually sets, rather than just showing up unannounced with a truckload of results and hollering 'you're welcome.' That's a lot of signatures, friend.

The Advisory Group on Mathematics and Artificial Intelligence — AGMAI — went further in its recommendations, TokenPost and NewsBytesApp both confirmed. AGMAI recommended that results like these be deposited in an academic repository that no AI lab controls, and that complete methodological details accompany every single result. AGMAI has also been careful to state publicly that its advisory role does not amount to an endorsement of OpenAI's process or its claimed results. OpenAI's GitHub release, by both outlets' accounts, did not fully satisfy those conditions.

University of Toronto mathematician Daniel Litt offered a somewhat gentler take than his peers, according to Scientific American, saying companies don't have any good reason to sit on mathematical answers and keep them secret. That puts him at odds with colleagues like Kra and Sutherland, who argue that the manner and pace of release matter just as much as the act of disclosing anything at all.

Our Analysis: A Pig Race With No Finish Line in Sight

This is analysis, not reporting, so take it for what it's worth — like advice from a man selling miracle fertilizer at a county fair. The core tension here looks, to this publication, like a collision between two things that genuinely don't fit together real comfortable: the speed at which a well-funded AI company can generate and publish mathematical claims, and the speed at which a global community of human experts can absorb, check, and actually use those claims. OpenAI, by its own account, sees math as a proving ground for AI capability and says it can't pump the brakes. The mathematics community, meanwhile, is pointing out that a proof nobody has time to read is about as useful as a screen door on a submarine.

The Lean verification angle is genuinely interesting and not nothing — a logically verified proof is more than a hand-wavy assertion. But verification of logical structure and verification of scientific value are two entirely different animals, and conflating them is the kind of move that makes mathematicians' eyes twitch. Until OpenAI releases the model, attribution disputes get sorted out, and the field has had a real chance to dig into 377 claimed results across dozens of subfields, this whole affair is going to keep generating more heat than light. The wagons are circled on both sides, and the dust ain't settling anytime soon.