DeepMind's AlphaEvolve Sets New Matrix Multiplication Exponent Record

Google DeepMind, together with academic collaborators and its Gemini-powered coding agent AlphaEvolve, improved the upper bound on the matrix multiplication exponent ω from 2.371339 to below 2.371177, setting a new record. The result was first announced by DeepMind research lead Pushmeet Kohli on August 18. Matrix multiplication underpins modern computing, including AI, making ω a long-standing open problem in complexity theory.

Confirmed

• AlphaEvolve, a Gemini-powered coding agent, combined with modern optimization methods achieved the improved bound of ω < 2.371177 (m1, @pushmeet; m6, @PetarV93).
• The technical approach involved reformulating the optimization problem, improving combinatorial loss analysis, and applying the ML-based AlphaEvolve algorithm (m4, @deepmind).
• The work is a collaboration between DeepMind and academics (m2, @DrSingularity).

Why it matters

• As @PetarV93 noted, although the improvement is incremental, it marks a substantive breakthrough for AI agents in fundamental algorithm discovery, pointing toward greater potential for AI-assisted mathematical research.

2026-08-18 ~ 2026-08-19 · 6 related posts

• [source] DeepMind Improves Matrix Multiplication Exponent with AlphaEvolve — deepmind · 2026-08-18
• [source] DeepMind's AlphaEvolve sets new record for matrix multiplication exponent, ω < 2.371177 — pushmeet · 2026-08-18

4 near-duplicate retellings: PetarV_93 · yoavgo · Dr_Singularity · Dr_Singularity