How did AI help crack a maths problem that stood for 31 years?
Let me tell you a story about a problem that sat unsolved since 1995. It has an awkward name, but the idea is surprisingly friendly. Take a blob of probability in many dimensions. You want to cut it in two. Is the best flat cut always almost as good as the best cut of any weird shape you can dream up, no matter how many dimensions you add? Mathematicians suspected yes. Nobody could prove it.
Why care? Because that cut decides how fast a random walker explores the blob, and that sits underneath a whole area of sampling theory used all over computing.
Act one: a hundred false starts
Before this month, the best known result still crept upward as dimensions grew, very slowly, but it did. A Microsoft Research team set out just to improve it. They worked with AI agents and explored more than a hundred candidate approaches, a lot of them brutal calculations with high-order tensors. A human looked at intermediate results, threw out the dead ends, and passed the promising ones into the next round.
Out of that mess came a new criterion. Then a twist: you can feed the improved bound back into the same argument and get a better one. Do it again, tighter again. That alone shrank the growth to something almost flat, but still not flat enough.
Act two: the final move
The last step was to repeat the whole refinement over and over, and then show that the small errors from each round add up to something bounded, no matter how many dimensions. Done. The paper runs to 140 pages.
So, did the AI solve it, or didn't it?
Both, and that is the interesting bit. In a separate test, the same team let several leading models try the problem on their own, with roughly the same token budget. They got nowhere. What worked was the combination: machines exploring wide and fast, a human choosing what was worth keeping and assembling the final argument. The judgment turned out to be the bottleneck, not the calculation.
Act three: maths checking itself
Within days, another trio of mathematicians posted an independent proof built on the new criterion, just 32 pages. A third group followed with an explicit constant, and then a formal computer-checked version in a proof assistant. Independent proofs are the strongest kind of validation maths has.
What opens up now? Results that quietly assumed this conjecture can be stated outright. Sampling methods may get firmer guarantees. And the recipe, a recursive criterion applied again and again with bounded errors, may fit other long-stuck problems. One caution: most of the details come from the team's own account, so the original paper deserves a read before anyone gets carried away.