Google DeepMind's AlphaEvolve becomes generally available, finds 23% speed improvement in Gemini kernel

Google DeepMind/AlphaEvolve

AlphaEvolve, Google DeepMind's algorithm discovery agent, became generally available on the Gemini Enterprise Agent Platform. The system uses evolutionary search with Gemini models to automatically optimize algorithms. Internally, it found a 23% improvement in Gemini's matrix multiplication kernel and reduced Spanner's write amplification by 20% through optimized compaction heuristics.

• Takes a seed algorithm and optimization goals, generates and ranks variants.
• 23% faster matrix multiplication for Gemini architecture.
• Optimized TPU cache replacement policy and arithmetic circuits.
• 20% reduction in write amplification for Google Spanner.
• Available on Gemini Enterprise Agent Platform for external users.