Google DeepMind today released AlphaEvolve, an artificial intelligence agent with self-evolution capabilities that can independently invent complex computer algorithms and has been widely applied in Google's data centers, chip design, and AI model training, achieving significant results.
AlphaEvolve combines the Gemini large language model with evolutionary optimization methods to automatically test, improve, and enhance the entire codebase, not just a single function. This system has quietly run internally for over a year, improving computing resource scheduling efficiency, accelerating model training, and achieving breakthroughs in mathematical research.
From Servers to Chips: AlphaEvolve Optimizes Google's Underlying Architecture
The scheduling algorithm proposed by AlphaEvolve has already been deployed in Google's global data centers, addressing the "resource stranded" problem and recovering 0.7% of resources. For Google's scale, this means significant cost and energy savings.
It also optimized the key circuit logic of the Tensor Processing Unit (TPU), successfully removing redundant bits, thereby enhancing the upcoming chip design. Meanwhile, AlphaEvolve improved its own AI training kernel, increasing the matrix operation speed of the Gemini model by 23%, and reducing overall training time by 1%.
Figure source note: Image generated by AI, image licensed service provider Midjourney
Breaking a 56-Year Mathematical Problem: Solving the Kissing Number Problem
AlphaEvolve’s contributions to basic scientific research are equally impressive. It rewrote the matrix multiplication algorithm through a newly designed optimizer, surpassing the 1969 Strassen algorithm for the first time on 4×4 complex value matrices, reducing the number of multiplications from 49 to 48, breaking a record that lasted 56 years.
When testing over 50 unsolved mathematical problems, AlphaEvolve matched the existing best solutions in about 75% of cases and proposed better solutions in about 20%. One classic problem is the "kissing number problem": this system found 593 spheres in 11-dimensional space that could simultaneously touch the central sphere, setting a new world record.
AI Inventing AI: How AlphaEvolve Works
Different from traditional AI coding tools, AlphaEvolve does not rely on a single prompt to generate code but uses an evolutionary approach for algorithm invention. It simultaneously calls Gemini Flash and Gemini Pro to propose modification suggestions for the code, which are then screened by the system evaluator to select the optimal solution for the next round of evolution.
DeepMind researcher Alexander Novikov said that this system focuses on "problems with clear evaluation criteria," making automatic optimization more efficient and reliable. That's why AlphaEvolve can span multiple fields, from data center management to mathematical theorem proving, generating highly efficient solutions that are difficult for humans to conceive.
Next Stop: Drug Development, Material Science, and Broader Scientific Collaboration
DeepMind said that the potential of AlphaEvolve goes far beyond Google's internal use. The company is currently collaborating with the "Human + AI" research team to develop user interfaces and plans to provide early access to some academic institutions.
"This is truly a scientific tool that can have a rapid impact in the real world," said researcher Chris Balog. "AlphaEvolve is pushing the boundaries of AI, not only optimizing the systems driving it but also helping us solve long-standing unsolved problems."
As large language models continue to evolve, AlphaEvolve demonstrates how artificial intelligence is constantly evolving toward deeper creativity and scientific discovery.
