On the morning of September 11, 2026, Dor Minzer's phone started filling up with messages from colleagues asking whether he was close to settling one of the most famous open questions in theoretical computer science. The rumours had a name: OpenAI, fresh off a bombshell proof about the behaviour of fluids, had allegedly found a proof of the unique games conjecture and might publish at any moment.
Minzer, a professor at MIT, had not proved the unique games conjecture. But he and his graduate students Yumou Fei and Shuo Wang had just finished a milestone result on a closely related problem — a variant of Subhash Khot's 2002 "2-to-1 games" conjecture — and were months away from writing it up. They dropped everything and finished the draft in three days. On September 14 they posted a 95-page paper with an apology on its first page: "The current version of the manuscript is complete mathematically, but it is not in the shape we wished to share in." From section 6 onward, as Minzer later admitted, there are "literally no connecting words" — just definitions and proofs stacked on top of each other.
The mathematics is the point. Khot proposed the unique games conjecture in 2002 and later defined 2-to-1 games to cover cases the original conjecture could not reach. Minzer, Fei and Wang proved a slightly weaker 4-to-1 version whose consequences are nearly identical. Most notably, it implies a result about a graph-colouring problem that predates Khot's work by decades: even when a graph can be coloured with three colours, no matter how many extra colours you allow, there will always be perverse graphs where finding that colouring stays hard. "You cannot do it even with the entire Crayola box," said Princeton's Mark Braverman.
The rush was vindicated within three weeks. On October 6, OpenAI announced a proof of the unique games conjecture itself, one of 377 results released at once. It was Lean-verified — and, unlike the trio's paper, an AI-generated manuscript that had gone through no human editing and no independent expert review.
Researchers greeted the deluge with the usual mix of alarm and curiosity. "Math by press release is not that healthy for math," Braverman said. Carnegie Mellon's Ryan O'Donnell called Minzer's result "another truly great one". Minzer's own worry is about incentives: "There is a lot of value in failing and knowing why you failed," he said. "Using AI takes all of this out." And for the next generation of mathematicians: "You are human, right? ... You don't know if you're going to get scooped by the trillion-dollar company."
