The AI Maths Breakthroughs Everyone Quotes: Counterexamples, Not Proofs
Skip to content
Toggle Navigation
- News
- Events
- TNW Conference
- All Events
- Newsletters
- Advertise with us
- Jobs
- Contact
News
The AI Maths Breakthroughs Everyone Quotes: Counterexamples, Not Proofs (August 17, 2026)
Credit: AllenS
A counterexample and a proof are not the same kind of achievement. A proof shows that something is always true. A counterexample shows that something claimed to be always true is not, by producing one object where it fails. Both settle a question. They ask different things of whoever finds them. One demands an argument that covers every case. The other demands a single object, and permission to keep guessing until you have it.
That distinction is the subject of a blog post the Cambridge mathematician Timothy Gowers published on 12 August. Gowers won the Fields Medal in 1998. He has also read the papers, which most people commenting on AI and mathematics have not.
He is not dismissive. He calls the results “extraordinarily impressive” and says plainly that models can prove hard things too.
“LLMs are not just good at finding counterexamples: they can find proofs of difficult statements as well,” he writes.
What the Famous Results Have in Common
OpenAI announced ten open problems solved in mathematics and theoretical computer science. Two led the coverage:
- The construction of a non-sofic group. Gowers has sat through the talks. He calls it “one of the most important unsolved problems in group theory”.
- A lower bound showing a multicolour Ramsey number grows superexponentially.
On that one he is unusually candid. It was “a major open problem in Ramsey theory that I didn’t necessarily expect to see solved in my lifetime”.
Then comes the observation the aggregators skipped. The most celebrated LLM results, Gowers notes, have almost all arrived as counterexamples rather than proofs. He counts the two above, plus the Jacobian conjecture and the unit distance conjecture. His third summary point is the careful version. Models prove universal statements perfectly well. But the strongest things they have proved do not match the strongest things they have disproved.
Two Results He Reclassifies
A counterexample earns its name by demolishing something people had good reason to believe. By that standard Gowers reclassifies two of the headline results, including one against its own paperwork.
On the non-sofic group, he says several construction routes already existed in the literature. He also doubts many experts strongly believed all groups were sofic. So it reads more naturally as the first example of a non-sofic group than as a counterexample. He notes the tension directly: OpenAI titled that section of its paper “A counterexample to the soficity conjecture”.
The Ramsey result gets the same treatment, and here he marks his own homework. Plenty of people expected an exponential bound, so for them it was a counterexample. Gowers was neutral. He had worked on an equivalent formulation years ago. His efforts back then ran in what turned out to be the right direction. For him it confirmed a weak expectation rather than overturning a belief.
Why Examples Suit Machines
Gowers lists eight ways mathematicians hunt for an example:
- Try standard examples off the shelf.
- Build one from familiar pieces.
- Leave parts undefined and fill them in as the proof demands.
- Try proving the opposite and see…