Google's evolutionary AI 'AlphaEvolve' can discover unknown algorithms and new solutions to unsolved mathematical problems, and is already being used internally by Google to improve the efficiency of AI development and chip design.



Google DeepMind, Google's AI research team, announced the coding agent ' AlphaEvolve ' on Wednesday, May 14, 2025. AlphaEvolve is an AI agent that combines Gemini and evolutionary algorithms, and is capable of discovering unknown algorithms and finding new solutions to unsolved mathematical problems.

AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms - Google DeepMind

https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/

DeepMind unveils 'spectacular' general-purpose science AI
https://www.nature.com/articles/d41586-025-01523-z

AlphaEvolve is an AI agent that combines the problem-solving capabilities of the fast and efficient AI model 'Gemini Flash' and the high-performance AI model 'Gemini Pro' with evolutionary algorithms. Human instructions to AlphaEvolve are first converted into prompts for the AI model, and the results of the prompts are then repeatedly 'verified,' 'executed,' and 'scored' to derive optimal answers.



We asked AlphaEvolve to improve our matrix multiplication algorithm, and they generated an algorithm that can perform 4x4 complex matrix multiplication with 48 scalar multiplications. This result exceeds the results of

AlphaTensor , which was announced in 2022. In addition, we had AlphaEvolve solve over 50 unsolved problems in analysis, geometry, combinatorics, and number theory, and we succeeded in 'rediscovering cutting-edge solutions' for about 75% of the problems. Furthermore, for about 20% of the problems, we achieved the result of 'improving the solution method that was previously considered to be the best, and taking a step forward in solving the unsolved problem.'



Google has deployed AlphaEvolve in the fields of data centers, chip design, and AI development. In data centers, it proposed system improvements for Google's cluster manager

Borg , and successfully restored an average of 0.7% of Google's global computing resources. It also proposed design changes to Google's AI processing chipTPU , and the proposals have been incorporated into the TPU.

AlphaEvolve accelerated the matrix multiplication kernel language Pallas , which is also used in the development of Geimini, by 23%, and also reduced Geimini's training time by 1%. In addition, it was possible to optimize low-level GPU instructions, and it was possible to speed up FlashAttention by up to 32.5%.



AlphaEvolve is planned to run an early access program for 'selected academic institutions' before being made available more broadly. Google has a sign-up form for people who want to use AlphaEvolve.

AlphaEvolve: Explore AI-Driven Algorithm Discovery & Optimization
https://forms.gle/WyqAoh1ixdfq6tgN8



In addition, AlphaEvolve's research paper is available at the following link:

AlphaEvolve: A coding agent for scientific and algorithmic discovery
(PDF file) https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf



Related Posts:

in Software,   Science, Posted by log1o_hf