The theoretical speed limit for multiplying matrices just got a little lower.
Researchers refined the combination loss analysis framework that underlies the laser method, the technique responsible for nearly every advance in the matrix multiplication exponent since 2022. They reformulated the core optimization problem so it could be solved at a larger scale than before, then designed a new machine-learning-based optimization algorithm to tackle it. Finally, they fed the result through AlphaEvolve, Google DeepMind's AI system for discovering and refining algorithms, to squeeze out further gains. The combined approach lowered the upper bound on omega, the exponent describing how fast matrix multiplication can theoretically run, from 2.371339 to 2.371177.
Omega has barely moved in decades. Strassen's 1969 algorithm put it at 2.807, and the record has crept down in tiny increments ever since, most recently by researchers in 2022, 2024 and 2025. Each fractional improvement sharpens the theoretical floor for operations that underpin graph algorithms, scientific computing and deep-learning training, even though none of these exotic bounds get used in real software.
Nobody is about to run this algorithm; the gain lives in the fourth decimal place. But watching a system built to discover algorithms get turned loose on the math of algorithm discovery itself says more about where AI research is headed than the bound does.