SAN FRANCISCO, May 18 (WSH) — Google DeepMind has unveiled AlphaEvolve, a groundbreaking AI system that has solved a 56-year-old open problem in mathematics, marking a major leap in automated algorithm discovery.
Unlike traditional coding agents, AlphaEvolve evolves programs through large-scale sampling and automatic evaluation. Its most notable achievement to date: a new algorithm that performs 4×4 complex matrix multiplication using just 48 multiplications—an improvement over a record unbroken since 1968.
Trained to optimize problems that can be programmatically expressed and functionally evaluated, AlphaEvolve has been tested on over 50 mathematical challenges. It successfully rediscovered known solutions in ~75% of cases and outperformed human-derived algorithms in about 20%, including constructing a new lower bound for the 11-dimensional kissing number problem.
Beyond math, AlphaEvolve has been deployed internally at Google to optimize chip architectures, large language model training, and scheduling strategies. Experts say this signals a future where algorithm design becomes automated, scalable, and no longer dependent solely on human ingenuity.
DeepMind researchers believe this is just the beginning: “What AlphaEvolve has achieved,” said lead scientist Matej Balog, “is not just solving problems—it’s evolving a new paradigm for scientific discovery.”
