Briefs · Education and Upskilling
AlphaEvolve Invents, Not Merely Imitates
Verdict: the machine has learned to invent, not merely to imitate.
AlphaEvolve is not a mathematician, but rather a tireless apprentice with infinite patience. Given 67 hard problems in analysis, combinatorics, geometry and number theory, it matched the best-known answers in most cases and improved them in several.
The method is elegantly simple. A language model writes code. An evaluator scores it. Evolution keeps the winners and mutates them again. The search is for the programme that generates the object, not the object itself.
The trophies are real. The 11-dimensional kissing number rose from 592 to 593. Eleven cubes now pack tighter than before. A sofa navigated a three-dimensional corridor with a volume no human had found. Seven cylinders touched in two hours, a task that once consumed months of computing.
Most striking is the finale. On the finite-field Kakeya problem, AlphaEvolve found a construction, Deep Think proved it, and AlphaProof formalised it in Lean. Discovery, proof and verification, in one pipeline.
Mathematicians remain essential. The tool is a partner, not a replacement. But the patient partnership has never been this productive.
Summary in our own words. The original belongs to its publisher.