AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields

Google DeepMind said its AlphaEvolve system has moved from pilot testing to a core infrastructure component. The Gemini-powered coding agent optimized the design of next-generation tensor processing units and discovered more efficient cache replacement policies in two days compared to months of human effort. Chief Scientist Jeff Dean said AlphaEvolve proposed a counterintuitive circuit design integrated directly into the silicon of next-generation TPUs. The system improved Google Spanner efficiency by refining Log-Structured Merge-tree compaction heuristics, reducing write amplification by 20 percent and providing insights for compiler optimizations that reduced software storage footprint by nearly 9 percent. Google Cloud is bringing AlphaEvolve to commercial enterprises. Klarna doubled training speed of a large transformer model while improving quality. Substrate achieved a multi-fold runtime speed increase in computational lithography. FM Logistic found a 10.4 percent improvement in routing efficiency, saving over 15,000 kilometers annually. WPP achieved 10 percent accuracy gains in campaign data optimization. Schrodinger reported a roughly 4x speedup in Machine Learned Force Fields training and inference. Technical Lead Gabriel Marques said faster inference shortens R&D cycles in drug discovery and materials development. AlphaEvolve was developed by a team including Matej Balog and Alexander Novikov with leadership support from Demis Hassabis and Sundar Pichai.