# Evolve better algorithms with a code-writing agent

DeepMind said in May 2025 that AlphaEvolve recovers on average 0.7% of Google's worldwide compute; it also improved on a 1969 matrix-multiplication result.

Organisation: Google DeepMind
Status: Deployed (internal) and research
Place: United Kingdom, United States
Categories: Science & discovery, Work & productivity
Built with: Gemini
Try it with: Claude Code, Codex, Cursor, Grok Bot, OpenClaw
Time to try: 1 to 3 hours

## Why it matters

Small efficiency gains at data-centre scale save energy and money. An agent that searches for better code, checked by automatic tests, finds gains people missed.

## How it works

Gemini models propose changes to code, automated evaluators score them and the best variants are evolved further. The 0.7% comes from a scheduling heuristic it found, which DeepMind says had been in production for over a year. It also sped up a Gemini kernel by 23%, cutting Gemini training time by 1%, and the 1969 result it improved on is Strassen's algorithm: it found a way to multiply 4x4 complex-valued matrices with 48 scalar multiplications.

## Who it's for

Engineers and researchers with a measurable performance problem.

## The prompt

```
You are working in this repository. Goal: make [function or file] faster without changing its output.

1. Write or locate a benchmark and a correctness test for it. Run both and record the baseline.
2. Propose 5 different optimisation ideas. Implement each one on its own branch or in its own copy.
3. Run the tests and the benchmark for each (at least 5 runs; report the median). Discard any variant that fails a test.
4. Take the 2 fastest, combine or refine them and measure again. Repeat for 3 rounds.
5. Report a table of variant, time and pass/fail, and the final diff.

Do not change public interfaces. Do not commit or push without asking me.
```

## Caveats

Results are DeepMind's own reports.

## Sources

The results are DeepMind's own reports.

1. Google DeepMind: 0.7% compute recovered; 23% kernel speed-up; 1% training time; 48 multiplications. AlphaEvolve: a Gemini-powered agent for designing algorithms (https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/)
2. MIT Technology Review: Independent report on AlphaEvolve and the 0.7% figure. Google DeepMind's new AI agent cracks real-world problems (https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/)

Fact-checked: 16 Oct 2026

Page: https://escapevelocitysociety.com/picks/deepmind-alphaevolve

See Selected Projects: https://escapevelocitysociety.com/picks.md
