Google DeepMind has unveiled AlphaEvolve, an AI agent capable of inventing entirely new computer algorithms and immediately deploying them within Google's extensive computing infrastructure. AlphaEvolve integrates Google’s Gemini large language models with an evolutionary method that automatically tests, refines, and improves algorithms. It is already boosting efficiency across Google’s data centers, chip designs, and AI training systems, tackling mathematical problems unresolved for decades. AlphaEvolve, described by DeepMind researcher Matej Balog as a Gemini-powered AI coding agent, can create highly complex algorithms spanning hundreds of lines with advanced logical structures beyond simple functions. Unlike prior work with FunSearch, which evolved single functions, AlphaEvolve evolves entire codebases, marking a major advancement in developing sophisticated algorithms for both scientific and practical computing challenges. The system has been quietly operating within Google for over a year, yielding notable results. One algorithm discovered by AlphaEvolve enhances Borg, Google’s massive cluster management system, recovering an average of 0. 7% of worldwide computing resources by addressing “stranded resources” — machines limited by one resource type but idle in others. Remarkably, the AI produces simple, human-readable code that engineers can easily debug and deploy. Beyond data centers, AlphaEvolve improved Google’s hardware design by simplifying a critical arithmetic circuit for Tensor Processing Units (TPUs). After validation from TPU designers, this improvement will feature in forthcoming chip designs. Furthermore, AlphaEvolve enhanced its own foundational systems by optimizing a matrix multiplication kernel used in training Gemini models, achieving a 23% speedup for that operation and reducing overall training time by 1%. Such efficiency gains translate to significant energy and resource savings on large-scale AI training. In mathematical innovation, AlphaEvolve designed a novel gradient-based optimization procedure yielding multiple new matrix multiplication algorithms, surpassing a record held for 56 years. Specifically, it found an algorithm for multiplying two 4×4 complex-valued matrices using 48 scalar multiplications instead of Strassen’s 49—an achievement that had eluded mathematicians since 1969.

This advancement improved the state of the art for 14 matrix multiplication algorithms. AlphaEvolve’s mathematical prowess extends beyond matrix problems. Tested on over 50 open problems in mathematical analysis, geometry, combinatorics, and number theory, it matched state-of-the-art solutions about 75% of the time and improved on them in roughly 20% of cases. For example, it broke a centuries-old geometric record in the “kissing number problem” by finding a configuration of 593 non-overlapping unit spheres touching a central sphere in 11 dimensions, surpassing the previous record of 592. The core of AlphaEvolve's innovation is its evolutionary approach combined with Gemini language models. It uses Gemini Flash for speed and Gemini Pro for depth to propose and modify code, which is then evaluated automatically. The best-performing algorithms guide subsequent evolution cycles. This process doesn't rely solely on training data but actively explores novel solutions, refining them via automated feedback loops based on clear evaluators of validity and quality. This methodology allows AlphaEvolve to tackle any problem with a measurable evaluation metric—whether optimizing energy use in data centers or improving mathematical proofs. Looking ahead, Google DeepMind envisions applications extending into material sciences, drug discovery, and other complex algorithm-dependent fields. The team is developing a user interface with the People + AI Research group and plans an Early Access Program for select academic researchers, with wider availability under consideration. AlphaEvolve represents a rare scientific tool that concurrently achieves significant real-world impact at massive scale. As large language models continue to advance, AlphaEvolve’s capabilities are expected to grow in tandem. The system exemplifies AI’s evolution: it begins within Google’s digital infrastructure, optimizes the very hardware and software sustaining it, and now addresses longstanding human intellectual challenges across science and technology.