In partnership with

World of AI | Edition # 38

Google’s Alpha Evolve: A New Era of AI-Driven Algorithm Discovery

Google DeepMind has unveiled a groundbreaking AI system called Alpha Evolve, a new breed of intelligence that does more than just write code—it evolves it. This innovation marks a shift from traditional code generation tools to autonomous algorithm discovery, with implications that could redefine how humans and machines collaborate on complex problems.

A Machine That Evolves Code

Alpha Evolve blends language models—specifically, Google's Gemini Pro and Gemini Flash—with evolutionary computation. Instead of relying on human-created templates or auto-completing code snippets, Alpha Evolve proposes new algorithmic structures, tests them automatically, and refines them through many generations. The result? Algorithms that are often more efficient than those created by humans.

Already Saving Google Millions

This isn’t just a lab experiment. Alpha Evolve has been integrated into Google’s internal infrastructure, including Borg, their powerful data center scheduler. Just one of its evolved algorithms has led to a 0.7% recovery in global compute resources, translating to millions of dollars saved in energy and performance costs.

Enhancing Gemini and Redesigning Hardware

Alpha Evolve also improved the core training kernel used in Google’s Gemini models. By optimizing a matrix multiplication operation, it achieved a 23% speed boost in that operation, ultimately reducing total training time by 1%. At Google’s scale, that represents a monumental efficiency gain.

In an even more astonishing feat, Alpha Evolve designed hardware-level improvements to Google’s custom Tensor Processing Units (TPUs), rewriting Verilog code—a language used in chip design. These changes passed human verification and are now part of future TPU releases, demonstrating that AI can now operate effectively at both software and hardware levels.

Breaking a 50-Year-Old Mathematical Record

Alpha Evolve recently achieved a historic milestone by breaking a mathematical record held since 1969. It improved upon Strassen’s algorithm for multiplying two 4x4 complex matrices—reducing scalar multiplications from 49 to 48. This breakthrough could ripple through countless applications that rely on matrix operations, including computer graphics and machine learning.

Outperforming Human Solutions

Alpha Evolve was tested on over 50 different math and algorithmic problems, ranging from number theory to geometry. It matched or exceeded known human solutions in 95% of the cases, and in 20% of them, it surpassed human benchmarks. For example, it advanced the ancient “kissing number problem” in 11 dimensions, increasing the record from 592 to 593 spheres touching a central one.

How It Works: An Evolutionary Loop

The system operates in an automated feedback loop. Gemini Flash generates thousands of possible solutions, which are then evaluated instantly using custom metrics like execution time, memory use, and correctness. The top performers are selected as “parents” for the next round, evolving toward optimal performance. If progress stalls, Gemini Pro steps in to provide deeper insights. This natural selection for code allows Alpha Evolve to rapidly iterate through solutions that would take human teams weeks or months.

Designed for Speed and Scalability

With automatic evaluations and high-throughput iteration, Alpha Evolve can explore thousands of code variations in days. It excels at problems with objective metrics, such as improving data center efficiency or solving mathematical equations—ideal for high-impact applications with clear performance goals.

The Road Ahead: Science, Sustainability, and Drug Discovery

Looking forward, DeepMind plans to expand Alpha Evolve's reach into materials science, sustainability, and drug discovery—anywhere that involves complex computation and measurable results. An early access program for academic researchers is in development, along with a user-friendly interface through the People + AI Research initiative.

Limitations and Future Potential

While Alpha Evolve shows immense promise, it’s not without limitations. It currently struggles with subjective problems—those requiring human intuition or interpretive judgment—and doesn’t always offer theoretical explanations for the solutions it finds. Nonetheless, its practical impact is unmistakable.

This system represents a paradigm shift in problem-solving. No longer bound by human trial-and-error, researchers can now collaborate with AI agents that bring a different kind of creativity—relentless, tireless, and evolution-driven.

Find out why 1M+ professionals read Superhuman AI daily.

In 2 years you will be working for AI

Or an AI will be working for you

Here's how you can future-proof yourself:

1. Join the Superhuman AI newsletter – read by 1M+ people at top companies

2. Master AI tools, tutorials, and news in just 3 minutes a day

3. Become 10X more productive using AI

Join 1,000,000+ pros at companies like Google, Meta, and Amazon that are using AI to get ahead.

Sign up and start learning AI