The Month AI Conquered Math: The Full Story
The Algorithmic Bridge Alberto Romero ● Covered by 8 sources
AI models spent July 2026 tearing through decades-old math conjectures like Twitter threads. Mathematicians are split between calling it a miracle and mourning their field.
July 2026 will go down as the month pure mathematics got ambushed. It started quietly in May, when an unnamed OpenAI research model disproved a 1946 Erdős conjecture on the unit-distance problem, finding an infinite family of counterexamples that beat the best known bound. Mathematicians called it stunning, unprecedented, the first time a major open problem in a subfield had fallen to an autonomous system rather than a human with a whiteboard.
Then things sped up fast. On July 10, OpenAI's public GPT-5.6 Sol Ultra model was told to spend at least eight hours grinding on the Cycle Double Cover Conjecture, a puzzle that had sat unsolved for fifty years. It came back with a proof in under an hour. Ten days later, Anthropic's Fable model disproved the 87-year-old Jacobian conjecture in the middle of a World Cup final, and, separately, an OpenAI model apparently found the exact same counterexample the same day. Terence Tao himself pored over the result and concluded it almost certainly wasn't brute-forced, which is a polite way of saying the machine got creative. More scalps followed within two weeks: the Dinitz-Garg-Goemans conjecture, gone after thirty years; the Maxwell conjecture, gone after a century and a half.
OpenAI capped the run on August 1 with a blog post listing ten resolved or significantly advanced open problems, all credited to an internal model called Astra. The eyebrow-raising detail wasn't the math — it was the price tag. OpenAI estimated the total compute cost at roughly two thousand dollars, about what a grad student makes in a month, for work that would represent career-defining output for a tenured professor.
And yet the researchers closest to it keep insisting math isn't solved. OpenAI's Noam Brown pointed out that Astra isn't inventing new branches of mathematics or posing fresh conjectures, and that attempts on the Millennium Prize problems have so far come up empty. There's also a harder, less flattering truth buried in all this: someone still has to understand what the machine found. Tao spent thousands of words unpacking Fable's Jacobian counterexample before anyone else could make sense of it. Without that translation work, a disproof is just symbols nobody can use — which means the bottleneck on progress hasn't disappeared, it's just moved. It now sits squarely on the humans left to interpret what the machines hand them.
My take
None of this should be read as math getting solved so much as the marketing cycle around math getting solved — a steady drip of tweets timed for maximum shock value, which is exactly the playbook AI labs already run on benchmarks. The real story isn't that a model found a counterexample nobody expected; it's that verification and meaning-making are now the scarce resource, and there aren't enough Terence Taos to go around. Anyone cheering this as proof that human intellect is obsolete is skipping the part where a machine's discovery is worthless until a person figures out what it means.
Read more about this at: The Algorithmic Bridge
