Google’s AlphaEvolve AI autonomously creates algorithms that improve computing efficiency and solve longstanding mathematical problems

Google DeepMind has introduced AlphaEvolve, an advanced AI agent that autonomously invents new computer algorithms and deploys them across Google’s vast computing infrastructure. Leveraging Google’s Gemini large language models combined with an evolutionary approach, AlphaEvolve iteratively tests and refines algorithms to enhance performance. It has been operational within Google for over a year, delivering significant improvements such as a 0.7% efficiency gain in Google’s global data center resource management by optimizing ‘stranded resources’ in Borg, Google’s cluster management system. The AI also contributed to hardware design by eliminating unnecessary components in Tensor Processing Units (TPUs), with changes validated for upcoming chip designs. Impressively, AlphaEvolve optimized a matrix multiplication kernel used in training Gemini models, achieving a 23% speedup in that operation and reducing overall training time by 1%, leading to substantial energy and resource savings. Beyond practical applications, AlphaEvolve has solved mathematical problems that have challenged researchers for decades, including breaking Strassen’s 56-year-old record for matrix multiplication algorithms and improving solutions to over 50 open problems in mathematics. Its evolutionary methodology, combining fast and deep Gemini models with automated evaluation, allows it to explore novel solutions beyond its training data. Google DeepMind envisions extending AlphaEvolve’s use to fields like material science and drug discovery and is developing user interfaces and early access programs for academic researchers. This AI represents a major leap in algorithm invention, demonstrating real-world impact at an unprecedented scale.