AI was supposed to upend whole industries, and so far it mostly hasn't — it's changed how people work, not whether fields exist. Then OpenAI dropped 372 novel mathematical proofs, solving 90 of the top 500 open problems at roughly three hours of compute each, and mathematics became the test case: what actually happens when AI solves a field's life's work?
Mathematics may be AI's first true field-level disruption.
The industry-ending wave everyone predicted never quite arrived — coding is the most changed field and demand for coders keeps rising. But OpenAI's release of 372 novel mathematical results, including partial progress on the Riemann hypothesis and the Hodge conjecture, looks different in kind: a significant chunk of a field's hardest outstanding problems, solved in one drop, faster than humans can absorb. The questions now are which fields are next on the entropy ladder — and whether a 'math genie' negates mathematicians' life's work or unlocks things undreamed of.
Published to a GitHub repo with 722 supporting papers, the results came from an unreleased internal frontier model — likely the same one behind last month's Navier-Stokes solution. OpenAI coordinated the release with its new mathematics advisory committee, including reasoning traces for a sample of ten problems.
Measured against Proof Atlas's list of the field's best-known open problems, OpenAI fully solved 90, with dozens more partial proofs that would qualify as significant contributions — including partial progress on two Millennium Prize problems, the Riemann hypothesis and the Hodge conjecture.
— Alex Kontorovich, Rutgers mathematics professor. The professional mathematicians' reaction is the best gauge of how big this moment is — and the field's top benchmark, the Fields Medal, is the comparison they keep reaching for.
Toronto's Daniel Litt framed the shift as mathematicians moving from decades-long problem work to verifying AI proofs — 'mathematicians have a lot of exciting work to do.' Micah Warren compared it to Barry Bonds's 756th home run: unbelievable if you'd predicted it in 1987, expected by 2007. An Anthropic researcher who'd been working on Navier-Stokes himself called it the most significant moment in mathematical history.
Average compute per result was roughly three hours of ChatGPT Pro thinking. A month ago, Navier-Stokes alone took around 10,000 agents and 88 hours. That efficiency gain happened within weeks — which is why some observers are calling this the intelligence explosion happening in real time.
Mathematician Francesco Maggi raised the core wrinkle: results have to be understood, connected, challenged, and reused, or they remain 'lettera morta.' Hundreds of AI-generated proofs at once may outrun the scarce resources of human attention, taste, and mathematical culture — the Navier-Stokes proof from last month is still being verified.
Physicist Steve Hsu's extrapolation: formal systems may verify the proofs, but humans won't have the context to understand the web of machine-invented concepts. At that point AI stops being a scientific tool — humans receive selected explanations from a larger intellectual civilization they can't independently grasp.
Grad students describe the release as nuking the problems they'd planned to build careers on. The optimist counter from former UW professor Shreeram Kannan: any mathematician from 5000 BC to January 2026 would have said yes to a math genie — this is an extreme but temporary shock to 'I am world-class at math' as an identity, not to the field itself.
Most of the solved problems are impressive but impractical — even Navier-Stokes doesn't change how engineers do fluid dynamics. The exception: a new theoretical efficiency bound for matrix multiplication, the math underpinning AI inference, dropping the exponent from roughly 2.37 to no more than 2.25. Cornell's Steven Strogatz compared the leap to Bob Beamon's long jump.
Bitcoin security researcher Justin Drake flagged a 'striking underrepresentation' of cryptographic breakthroughs and called backroom government interventionism his base case. If an internal model can break major cryptographic schemes, that's not a Bitcoin problem — it's a problem for the cryptography underpinning the entire digital world.
— Scott Aaronson, University of Texas computer science professor, on his blog. Aaronson's sourced claim, from his blog: if the models can, 'it would certainly be nice to get ahead of things before the rest of the world figures out the same.'
NYU Stern's Arpit Gupta: you're deluding yourself if you think advances like math aren't coming to economics and social science. Wharton's Ethan Mollick sees two AI revolutions after a 'slop science era' — everything ever published gets reread and rejudged, and novel discoveries start coming fast, with nearly autonomous research in his own field already approaching top-journal level.
The results largely came from big-data approaches coupled to pattern recognition on verifiable problems, not reasoning into genuinely novel ideas. Pedro Domingos's framing: mathematics is a classic case of Moravec's paradox — hard for humans, easy for machines. That doesn't automatically generalize to messier fields.
A Google distinguished scientist's ordering: math and coding are structured and fall first; hard sciences next, complicated by noisy measurement; then financial markets, non-stationary and psychological; then medicine, 'a nightmare of incomplete information'; then law, built on ambiguous language and evolving norms. Culture, art, and strategic leadership — the pinnacle of disorder and tacit knowledge — stay safe. So rejoice: we'll have artists, lawyers, doctors, and CEOs for a long time yet.
Either this disruption pattern spreads to other domains, or human and systemic inertia proves it's less inevitable than it looks. And the question for mathematics itself: does a mathematical genie negate all that hard work and learning — or let mathematicians do things undreamed of as of yet? Over the coming months, that's what we'll begin to see.
The FT reports OpenAI told investors they hit roughly $50 billion annualized at the end of September — a $20 billion gap from the widely reported $68 billion, which came from an investor's back-of-envelope gross-revenue math, not from OpenAI. Even 77% run-rate growth and 107% enterprise growth got completely overshadowed by the gap.
Anthropic quotes gross revenue before revenue sharing — counting tokens sold through Amazon or Microsoft even when big chunks go to those partners — while OpenAI has always used net. Once Anthropic's audited financials land ahead of its IPO, they're presumably going to come in much lower than the numbers private investors have been shown.
Semiconductor stocks sold off on the FT story — Nvidia down 3%, Oracle down 5.5%, NeoClouds worse — even though OpenAI isn't public and never shared the $68B figure. The takeaway investors are drawing: the entire semi trade is levered against OpenAI and Anthropic growth narratives, and it's fragile.
In The Information's latest subscriber survey, 35% say their organization is returning multiples on AI spend, 17% say returns are positive but less than hoped, and just 8% call AI spend a net negative. Caveat: two-thirds of these readers work at organizations that build AI applications.
After Claude overtook ChatGPT in the summer survey, usage of Claude and Gemini both fell as readers signed back up with OpenAI. Meanwhile Openclaw use hit 12% — its highest level this year — long after the hype train died down, and Cursor usage has doubled in recent months to 23%.
72% of respondents now use open-weight models with some regularity, and 20% say they're the majority of their organization's workload. But only 21% use Chinese open-weight models — even forward-leaning startups remain hesitant to rely on foreign models.
Only 22% of The Information's readers say they're hiring less because of AI, 10% say they're hiring more, and almost half say AI hasn't altered their hiring practices at all.
New terms prohibit 'sustained and needless abusive or cruel behavior,' though Anthropic says it only applies in extreme cases and won't catch common frustration. Box's Aaron Levie offered the AI-realist defense: even if models aren't conscious, you don't want future models trained on endless content of humans being rude to them.
Dashboards connects Claude to a dataset like a CRM or accounting software and generates auto-updating, natural-language-queryable visualizations; Motion turns data into code-based animations you can edit without re-rendering. One early user built an animated product demo and noted it would have cost tens of thousands of dollars a year ago.