1. A vibrant digital landscape of code, symbolizing AI-driven exploration of complex algorithmic solutions. Google DeepMind’s new AI coding agent, AlphaEvolve, is making waves with a breakthrough in mathematics. AlphaEvolve is a Gemini-powered system that combines a large language model (LLM) with automated evaluators in an evolutionary process

. In a recent announcement, researchers revealed that this AI discovered a faster way to multiply matrices. At the heart of this breakthrough is the Strassen algorithm, a classic method for multiplying matrices introduced in 1969. That algorithm had held the speed record for 56 years

. Astonishingly, AlphaEvolve found a new algorithm to multiply 4×4 complex matrices using only 48 scalar multiplications

– one fewer than the best-known method. Even this small change can significantly speed up large-scale computations.

How AlphaEvolve Works

2. AI tackling advanced math: The collage of hands on a keyboard and mathematical formulas reflects how AlphaEvolve generates and tests new code to solve complex problems. AlphaEvolve works like a digital inventor. It starts with an initial algorithm or idea and then “evolves” it over time. In each generation, the Gemini LLM proposes hundreds of candidate modifications to the code, and automated evaluators test these ideas against clear performance goals. The most successful ideas are kept and recombined, while weaker ones are discarded

. This cycle is reminiscent of natural selection: random “mutations” of code are tried, and the fittest survive. Because it can handle large codebases and complex criteria, AlphaEvolve has been applied to a range of hard problems. For example, DeepMind reports using it to improve task scheduling in Google’s data centers and to optimize the design of AI chips (TPUs)

. In fact, the company says this work helped them save about 0.7% of global compute resources

– a substantial efficiency gain.

Breaking the 56-Year Record

Before AlphaEvolve, all known matrix-multiplication techniques traced back to Volker Strassen’s 1969 method. Strassen’s algorithm uses clever math to multiply matrices with fewer steps than the textbook method, and it stood as the gold standard for decades. Now, AlphaEvolve has changed that by finding a method that outperforms Strassen’s in this case

. DeepMind emphasizes that the discovery is truly novel. Matej Balog, one of the AlphaEvolve researchers, points out that they can prove these new algorithms are correct and new, meaning they weren’t simply memorized from training data

. One outside expert even called DeepMind’s paper “quite spectacular” and said it represents the first time a general AI model has produced genuinely new mathematical solutions

. AlphaEvolve didn’t stop at matrices. It also tackled other famous math puzzles. For example, it found a new solution to the 11-dimensional “kissing number” problem by arranging 593 spheres around a central sphere

. These results suggest the AI isn’t just matching humans’ work; it’s pushing beyond human knowledge.

Real-World Impact and Future Prospects

AlphaEvolve’s accomplishments aren’t limited to pure math. The system has produced better algorithms for real-world applications: improving cloud job scheduling, cutting waste in data centers, and refining the way other AI models are trained

DeepMind’s team has described AlphaEvolve as a “superhuman” coding agent that can explore solutions far beyond what humans have found

This breakthrough hints at a future where AI helps drive innovation across fields. Companies might use similar tools to automatically optimize software, design new materials, or even generate novel scientific ideas. DeepMind is already planning an early access program so outside researchers can use AlphaEvolve themselves

Key accomplishments of AlphaEvolve include:

Matrix multiplication: Discovered an algorithm to multiply 4×4 complex matrices with only 48 multiplications (vs. 49 in Strassen’s method)

Mathematical puzzles: Solved or improved solutions to problems like the 11-dimensional kissing-number problem (finding 593 spheres)

Data-center efficiency: Evolved better scheduling algorithms that reduced Google’s compute waste by about 0.7%

Chip and AI design: Generated more efficient AI chip designs and optimized training methods for large language models like Gemini