AlphaEvolve: Gemini-powered coding agent scaling impact across fields (deepmind.google)

AlphaEvolve, powered by Google DeepMind's Gemini, has transitioned from pilot testing to a vital component of Google's AI infrastructure. It has optimized the design of next-generation TPUs, achieving in two days what traditionally took months of human effort. By proposing a novel circuit design integrated directly into TPU silicon, and improving Google Spanner's efficiency—reducing 'write amplification' by 20%—AlphaEvolve showcases AI's capacity to enhance foundational hardware and software components. Now set to scale across various industries through Google Cloud, AlphaEvolve has already demonstrated significant impact in sectors like finance, logistics, and materials science. For instance, Klarna enhanced its transformer model’s training speed, doubling performance, while FM Logistic optimized routing challenges, achieving a 10.4% improvement in efficiency. Additionally, Schrödinger reported a fourfold speedup in machine learning tasks relevant to drug discovery and materials development. As AlphaEvolve continues to evolve, it emphasizes a future where self-learning algorithms drive breakthroughs across domains, reshaping industries and advancing research capabilities.