Google DeepMind has announced AlphaEvolve, an AI-powered coding agent that evolves algorithms for math and computing applications. The tool combines the creativity of large language models with automated evaluators to discover and optimize complex solutions.

AlphaEvolve uses DeepMind's Gemini models to propose algorithmic solutions. It has been deployed across Google's computing ecosystem, improving data center efficiency by recovering 0.7% of compute resources, accelerating AI model training by 23%, and designing faster matrix multiplication algorithms. It's also found new solutions to open mathematical problems, like advancing the kissing number problem.

The agent leverages an ensemble of Gemini models—Gemini Flash for breadth and Gemini Pro for depth—to generate code proposals. These are evaluated using automated metrics and stored in a database that implements an evolutionary algorithm, determining which programs to use in future prompts.

AlphaEvolve represents a significant advance in AI-powered algorithm discovery. It has the potential to accelerate progress in areas like hardware design, AI training, and mathematical problem-solving by automating complex coding tasks and optimizing performance.