Google DeepMind improved the matrix multiplication exponent, known as omega, from 2.371339 to 2.371177 using AlphaEvolve, an evolutionary coding agent built on the Gemini architecture. This tiny but significant improvement represents the most substantial progress on this theoretical bound in decades. The research team included Google DeepMind researchers alongside academics Josh Alman and Virginia Vassilevska Williams. AlphaEvolve had previously improved 4×4 complex-valued matrix multiplication by reducing scalar multiplications from 49 to 48, marking the first improvement over Strassen’s 1969 algorithm. In practical terms, DeepMind reported a 23% speedup on a Gemini training kernel and a 1% reduction in overall training time.
Source: Read the original article
