Overview

• AlphaEvolve has been deployed within Google for over a year, improving data center scheduling by 0.7% and enhancing next-generation TPU chip designs.
• The AI surpassed Strassen’s 1969 matrix multiplication record, discovering a faster method for 4×4 matrices using fewer scalar multiplications.
• In tests on over 50 open mathematical problems, AlphaEvolve improved state-of-the-art solutions in 20% of cases and matched them in 75%.
• The system optimized a matrix multiplication kernel, reducing Gemini model training time by 1% through a 23% speedup in key operations.
• DeepMind is developing a user interface and Early Access Program to extend AlphaEvolve’s capabilities to academic researchers.