AlphaEvolve: DeepMind's AI Breaks a 56-Year Matrix Multiplication Record
In May 2025, DeepMind unveiled the Gemini-powered coding agent AlphaEvolve, which found a 4x4 matrix multiplication method beating Strassen's algorithm.
Google DeepMind unveiled AlphaEvolve in May 2025 — an evolutionary coding agent powered by Gemini for algorithm design. It found an algorithm multiplying 4x4 complex-valued matrices in 48 scalar multiplications, breaking the 56-year record held since Strassen's algorithm (49) in 1969.
AlphaEvolve does more than solve one problem: inside Google it optimizes data-center scheduling, chip design and model training, showing AI's capacity to discover new knowledge autonomously.
From AlphaFold to AlphaEvolve, DeepMind is pushing AI from 'fitting the known' to 'discovering the unknown' — when AI improves a half-century-old math record, the paradigm of scientific discovery is changing.
This article aggregates official announcements and public reporting; original sources are linked below.
Source:DeepMind 官方博客