AlphaEvolve by DeepMind automates code optimisation and discovers new algorithms

Google’s DeepMind has introduced AlphaEvolve, a new AI-powered coding agent designed to autonomously discover and optimise computer algorithms.

Built on large language models and evolutionary techniques, AlphaEvolve aims to assist experts across mathematics, engineering, and computer science by improving existing solutions and generating new ones.

Unlike natural language-based models, AlphaEvolve uses automated evaluators and iterative evolution strategies—like mutation and crossover—to refine algorithmic solutions.

DeepMind reports success across several domains, including matrix multiplication, data centre scheduling, chip design, and AI model training.

In one case, AlphaEvolve developed a new method for multiplying 4×4 complex matrices using just 48 scalar multiplications, surpassing a longstanding result from 1969. It also improved job scheduling in Google data centres, recovering an average of 0.7% of global compute resources.

In mathematical tests, AlphaEvolve rediscovered known solutions 75% of the time and improved them in 20% of cases. While experts have praised its potential, researchers also stress the importance of secure deployment and responsible use.

Would you like to learn more about AI, tech and digital diplomacy? If so, ask our Diplo chatbot!