HomeThe New IntelligenceAlphaEvolve Writes Better Algorithms Than Humans, and Then Improves Itself

AlphaEvolve Writes Better Algorithms Than Humans, and Then Improves Itself

AlphaEvolve, DeepMind's evolutionary coding agent, surpassed human solutions in 20% of open math problems and broke a 56-year matrix multiplication record.

Youtube video
Share
The New Intelligence · Explore this series
May 19, 2025
Key Takeaways
  • AlphaEvolve beat a 56-year-old matrix multiplication record.
  • It surpassed human solutions in 20% of open math problems.
  • The system improved its own training code and hardware designs.

For 56 years, nobody could beat Volker Strassen's matrix multiplication algorithm. Mathematicians refined it. Computer scientists extended it. Entire careers orbited around it.

Then Matej Balog, Alexander Novikov, and their colleagues at Google DeepMind built a system that cracked it in days.

The system is called AlphaEvolve. It may represent the clearest demonstration yet of what happens when AI stops answering questions and starts solving problems that humans cannot.

Fifty Open Problems, Zero Human Guidance

AlphaEvolve works on a deceptively simple principle. Mark a piece of code you want improved, define how to measure whether a new version is better, and let the system evolve solutions through iterative refinement.

What is evolutionary algorithm design?

An evolutionary algorithm mimics natural selection: generate many candidate solutions, test them against a fitness measure, keep the best performers, and repeat. AlphaEvolve supercharges this process by using large language models to propose code mutations, rather than relying on random changes.

The DeepMind team tested it on 50 open problems spanning mathematics, geometry, and combinatorics. In 75% of cases, it rediscovered the best solutions humans had found.

In 20% of cases, it produced something better.

Key figure

20%

The share of open mathematical problems where AlphaEvolve surpassed the best known human solutions

These are problems that professional mathematicians have worked on for years, sometimes decades. AlphaEvolve did not merely match their answers. It found paths they had missed.

The Strassen Record Falls

The matrix multiplication result stands apart. Strassen published his algorithm in 1969, showing that multiplying two matrices required fewer operations than anyone assumed. His method became a cornerstone of computer science.

AlphaEvolve found a procedure for multiplying 4x4 complex-valued matrices using 48 scalar multiplications, one fewer than Strassen's 49. The system was not built to tackle this specific problem. It is a general-purpose discovery engine that happened to crack one of the field's most stubborn benchmarks.

The implications run deeper than a single record. Matrix multiplication sits at the heart of modern AI training. A faster method speeds up the very systems that power AlphaEvolve itself.

The Machine That Tunes Its Own Engine

AlphaEvolve pairs two versions of Google's Gemini model. Gemini Flash generates a broad spread of candidate solutions quickly. Gemini Pro evaluates the most promising ones with greater depth.

The combination balances exploration with refinement, much as evolution favours both mutation and selection.

I think one day, maybe we can cure all disease with the help of AI. Maybe within the next decade. I don't see why not.

Demis Hassabis, DeepMind CEO and Nobel laureate in chemistry

The practical results caught engineers' attention. AlphaEvolve optimized a data centre scheduling algorithm that now recovers 0.7% of Google's worldwide compute resources, on average. It achieved a 23% speedup in a matrix multiplication kernel used in Gemini's own training pipeline.

Key figure

0.7%

The share of Google's global computing resources recovered by AlphaEvolve's scheduling optimization, now running in production for over a year

It also proposed a circuit design for an upcoming Tensor Processing Unit, Google's custom AI chip. The feedback loop is already real: AlphaEvolve improves the hardware it runs on and the training code that makes it smarter.

A Careful Optimism

Demis Hassabis has spoken about using AI to cure all disease within the next decade. Coming from anyone else, this would sound like bluster. Coming from a Nobel laureate whose AlphaFold system predicted protein structures for nearly every known protein, it carries different weight.

Related reading

Google's 'Willow' Quantum Chip Just Crossed an Impossible Barrier

Google's Willow chip solved a physics problem in 2 hours that would take the world's fastest supercomputer over 3 years.

AlphaEvolve has clear limits, though. It excels at problems with automated evaluation criteria. Whether it can extend to domains where success is harder to measure remains an open question.

DeepMind has announced an Early Access Program for selected academic researchers. The open-source community has already responded with OpenEvolve, an independent implementation of the core architecture.

The 56-year record did not fall because a smarter human came along. It fell because someone built a system that could try billions of possibilities, guided by pattern recognition and the pressure of artificial selection.

The interesting question now is which record falls next.

Sources

Share
Related Articles
AI Consciousness Is Unlikely, Says Neuroscientist Anil Seth

Neuroscientist Anil Seth argues AI consciousness is unlikely without biology. His TED talk lands amid a widening debate over conscious AI, not intuition.

AI In Science Connects the Dots, But Only In Fields That Are Fragmented

An analysis of 80 million papers shows AI boosts originality where knowledge is scattered and connections are weak, but contributes little novelty in structured science.

"Keep Humanity Safe From AI," Urges Pope Leo XIV

Pope Leo XIV's first encyclical reaches the same verdict on AI as the labs building it, then parts ways over the meaning of human limits.

AI Solves Erdős Math Problem: What's Next for AI in Mathematics?

An AI solved an 80-year-old Erdős math problem by walking a path mathematicians had collectively avoided.