The AI That Broke a 56-Year Math Record — And Is Now Saving Millions at Google

Imagine you’re a student in 1969. Your math teacher writes a problem on the board — a tricky puzzle about the fastest way to multiply two large grids of numbers together. The class thinks hard. Nobody can crack it. Then a brilliant young German mathematician named Volker Strassen walks up, picks up the chalk, and scribbles down a solution that changes computing forever.

For 56 years, nobody did better.

Then, last year, an AI named AlphaEvolve did.

The Puzzle Inside Every Computer

Here’s something most people don’t know: every time you open a spreadsheet, stream a video, or send a message, your phone or laptop is doing billions of tiny multiplications — multiplying matrices. These are grids of numbers, and the speed at which a computer can multiply them affects everything: how fast AI models run, how quickly games render, how efficiently data centers operate.

In 1969, Strassen’s algorithm figured out that you could multiply two 4x4 grids using 49 operations instead of the obvious 64. That doesn’t sound like a big deal — but when you’re doing this billions of times per second across millions of servers, every single operation saved translates to real money, real energy, and real speed.

For over half a century, mathematicians tried to shave off even one more operation. Nobody could.

Enter AlphaEvolve

Google DeepMind built AlphaEvolve as what they call an “evolutionary coding agent.” Think of it like this: instead of just answering a question, AlphaEvolve experiments. It writes thousands of different solutions, tests them, keeps the best ones, throws out the bad ones, and keeps evolving — just like nature does with living creatures, but in software.

It’s powered by Gemini, Google’s most advanced AI model. And when Google’s researchers pointed it at the matrix multiplication problem, something remarkable happened.

AlphaEvolve came back with an answer that uses just 48 operations instead of Strassen’s 49.

One operation. Fifty-six years. Solved.

“It’s not just that it found a solution,” said researchers at DeepMind. “It’s that it found a solution we hadn’t considered — from a completely different direction.”

The Part That Really Matters: It’s Already Working

Here’s where the story gets even more interesting. AlphaEvolve isn’t just a research project that sits in a lab. It’s been quietly deployed inside Google’s own infrastructure for over a year now.

The results? It has continuously recovered 0.7% of Google’s worldwide computing resources. That sounds tiny, until you realize Google runs some of the world’s largest data centers. 0.7% of that is an enormous amount of power and cost — reclaimed automatically, every day, without anyone flipping a switch.

It also sped up a key part of Gemini’s architecture — the very AI it’s built on — by 23%. The AI improved the AI. The snake eating its own tail, but productively.

It designed better scheduling algorithms for data centers. It simplified chip designs. It even improved the training process for the next generation of AI models. AlphaEvolve isn’t just solving math problems — it’s solving Google’s problems, quietly and relentlessly.

Why This Feels Different

We’ve heard a lot of AI hype. AI that writes poetry. AI that generates images. AI that summarizes your emails. But AlphaEvolve represents something qualitatively different: an AI that is genuinely discovering things. Not retrieving facts it learned from humans — actually going somewhere no human has gone.

When AlphaEvolve broke Strassen’s record, it wasn’t copying from a textbook. It invented a new page of the textbook.

Terence Tao, widely considered the greatest living mathematician, is now collaborating with Google DeepMind’s team to use AlphaEvolve on some of mathematics’ hardest unsolved problems. If one of humanity’s sharpest minds thinks this tool is worth his time, that tells you something.

What It Means for You

You won’t notice AlphaEvolve. You’ll never see its name in an app. But the next time an AI assistant answers your question a little faster, or your video streams without buffering, or a new drug gets discovered a little sooner — there’s a small chance AlphaEvolve helped make that happen.

This is what AI working in the background looks like. Not a chatbot. Not a photo filter. A quiet, tireless mathematical mind — rewriting the rules of what’s possible, one operation at a time.

This post has been created by Claude AI.

References