Google Deepmind's Alphaevolve Designs New Algorithms Powered by Gemini

Google DeepMind has unveiled AlphaEvolve, a Gemini-powered coding agent capable of designing entirely new algorithms. This tool aims to tackle complex problems from the ground up, potentially outperforming traditional methods. DeepMind plans an early access program for academics to explore and provide feedback. AlphaEvolve combines the creativity of large language models (LLMs) with automated evaluators. It targets fundamental problems in math and computer science, unlike typical code assistants. DeepMind reports using AlphaEvolve internally to enhance data center efficiency, chip design, and AI training. AlphaEvolve has limitations: it requires a built-in evaluation method and only describes solutions as algorithms. Internal testing showed AlphaEvolve improved data center scheduling, recovering approximately 0.7% of Google's worldwide compute resources. It also found a new lower bound in an 11-dimensional geometric challenge. DeepMind envisions AlphaEvolve assisting in material science, drug discovery, and sustainability. The company emphasizes the system's evolving potential as LLMs improve, especially in coding. AlphaEvolve displays the progression from discovering algorithms for specific domains to developing more complex algorithms for a wide range of real-world challenges.