Google DeepMind Unveils AlphaEvolve: AI That Solves Coding and Math Problems

Google DeepMind has announced its latest artificial intelligence agent, AlphaEvolve, designed to solve complex coding tasks, and tackle advanced mathematical problems.

This next-generation system represents a major step forward in how AI can contribute to science and technology.

What Is AlphaEvolve?

AlphaEvolve is a powerful AI tool that generates and improves algorithms using a unique, refining process.

It begins by producing code with Google’s lightweight Gemini 2.0 Flash model.

The system then evaluates and scores each version, keeps the best ones, and refines them over several rounds.

Eventually, it switches to Gemini 2.0 Pro for deeper improvements—reducing common errors known as “hallucinations” that often affect large language models.

Key Achievements and Uses

Boosting Google’s Infrastructure: AlphaEvolve was used in Google’s system that manages its data centers, and this helped recover 0.7% more computing power.

Improving Math Calculations: AlphaEvolve found a faster way to do matrix multiplication, making it better than a method that had been used for over 50 years.

Enhancing Hardware Design: AlphaEvolve also helped improve Google’s computer chips (TPUs) by cleaning up the code that controls them, removing parts that weren’t needed to boost performance.

Scientific Contributions

Beyond engineering tasks, AlphaEvolve is making significant progress in pure mathematics:

It contributed to 50 open math problems across fields like number theory, geometry, and combinatorics.

The system matched top solutions in 75% of cases and improved on 20%, showing real potential for discovery rather than just replication.

A notable achievement includes improving the kissing number in 11 dimensions from 592 to 593—a problem that had gone unsolved for decades.

How AlphaEvolve Works

Unlike traditional code-generating tools, AlphaEvolve learns and evolves:

It evaluates each suggestion, selects the best, and refines them in cycles.

Once it stops getting helpful suggestions, it shifts to a more advanced model to fine-tune the output.

This evolutionary loop results in highly optimized solutions—many of which outperform human-designed algorithms in speed and accuracy.

The Future Plans

DeepMind plans to launch an early access program for researchers and academics, and is developing a user interface for broader use.

According to Pushmeet Kohli, Head of AI for Science at DeepMind, AlphaEvolve could soon become a vital research partner across disciplines.

News Gist

Google DeepMind’s AlphaEvolve is a new AI agent that solves complex coding and math problems.

It improves code through repeated refinement, boosts data center efficiency, enhances chip design, and advances math research.

Matching or surpassing expert solutions, it shows potential as a powerful tool for scientific and technological innovation.