Math 2.0: When AI Published 722 Proofs Overnight

Terence Tao, often described as the unofficial spokesman of mathematics, has now answered the flood: hundreds of AI-generated mathematical breakthroughs that OpenAI published within days. The host, Dr. Samuel Allen Alexander, reads Tao's Math 2.0 memo line by line and argues with it. The question is plain but heavy: when machines prove theorems, what is mathematics for?

In Tao's account of the classical order, or Math 1.0, the proof of a long-standing conjecture sets off a wide chain reaction: the authors are invited to lecture, they meet specialists, workshops are organized, new collaborations form, follow-up problems circulate, and proofs gradually get streamlined into lecture notes. This digestion is what turns a result into common property of the field.

The promise of Math 1.0: digested proof

Tao's objection is sharp: a solved problem cannot be made unsolved again, and even the mere knowledge that a solution exists contaminates the search for alternate routes. In his view, AI-assisted teams now gather open problems on an industrial scale, a logic of proof harvesting that consumes the very ground that makes fields fertile. Where the premium on being first disappears, the machinery that rewards deep thought weakens too.

The debate was ignited in early September by OpenAI's announcement on the Navier-Stokes regularity question. In an interview with IBM, Tao called the development quite concerning: when ninety-year-old knots such as Navier-Stokes are untied before methods and side insights get distilled, the by-products of the process are lost. He argues that AI has moved past benchmarks and begun collecting open research problems one by one.

In early October the picture grew: 722 papers from an unreleased frontier model, 372 families of results across some 20 subfields. According to NewScientist, Francis Johnson, who worked 25 years on Wall's D(2) problem and wrote two books on it, greeted the outcome with astonishment. SiliconAngle reports that a standout item was a quasi-Riemann piece touching the 150-year-old Riemann hypothesis, alongside more than eighty studies on matrix multiplication and theoretical computer science.

The October flood: 722 papers overnight

An Implicator review notes that only about 42 percent of the headline results carried machine-checked Lean certificates; the unreleased model cannot be rerun from outside and no prompts were shared. A day later three manuscripts were withdrawn over a sign error. The episode shows why Lean verification must separate certified headlines from raw claims, a caveat the company itself concedes for unformalized results.

Alexander, the host, finds Tao's picture drawn from the window of the elite. He recounts publishing in a leading journal of mathematical biology last year without receiving a single invited talk. Ordinary researchers, he says, never see this glamour; Andrew Wiles guarding his Fermat work for years proves that secrecy predates AI. And with millions of proofs of Pythagoras around, nobody is barred from seeking one more.

The Hacker News front: combining versus creating

On Hacker News, Love Parade summed up the pattern crisply: models are superb at combining vast knowledge but thin at stepping back to propose an elegant abstraction. Fotron countered with the tight feedback loop thesis: where output is easy to check, models excel, which explains the success of proof assistants paired with AI. Alexander answers from experience: viewers used AI to settle open problems from his own papers, and the solutions look genuinely original to him.

Fubble recalls how labs declared superiority at Go and chess and then walked away: will mathematics be abandoned once milestones run out? Ant-Man bets that hallucinations will fade, outputs will simplify, and humans will leave the loop. The host is exultant: this is the most thrilling golden age since Euclid, and the only ones having a bad day sit in elite chairs.

Bner's line could summarize the debate: the future of mathematics is not solving but communicating what was solved and why it matters. Thurston's psychology remark belongs here too: mathematics is the discipline of ideas that resist moving from mind to mind. As the QuantaMagazine profile shows, LLM-assisted checkers let a problem be split into small chunks, solved bit by bit, and reassembled with confidence.

Beauty, funding, and philosophy

Russell's 1902 praise of cold and austere beauty and Hardy's yardstick that admits no permanent place for ugly mathematics get long airtime in the video. The host turns them around: if pure mathematics is art and social science, taxpayers will be less happy to fund it. The primes-to-cryptography story, he says, carries survivorship bias: a handful of useful results hides millions of forgotten ones. The beauty test may endure, but it cannot serve as a funding case.

A guest post on Tao's WordPress blog sharpens the frame with Rav's maxim: theorems are headlines, proofs are the inside story. The September 8 Navier-Stokes claim, it explains, came from some 10,000 agents exploring in parallel, then formalized in Lean: correctness machine-supported, insight not yet digested. Gowers adds the same note: poor machine write-ups may be a temporary annoyance. Without the inside story , the headline stays thin.

What remains is the homework of Math 2.0: turning the flood of solutions into digested understanding, putting exposition and community-building at the center. In an academy where the first-past-the-post premium erodes and paper mountains keep growing, the AI shock also opens a door to reform. The host's pencil-factory parable fits: faced with a glut of pencils, the answer is not more pencils. Read more, sift, explain; distill meaning, not tokens.