AlphaEvolve: The AI That’s Rewriting the Rules of Algorithm Design

How Google’s latest AI agent is cracking unsolved math problems and optimizing everything from chips to data centers

Meet AlphaEvolve—the AI-powered evolutionary coding agent that’s quietly revolutionizing how algorithms are born. Combining Gemini’s models (Flash for rapid exploration, Pro for deep refinement) with automated evaluators, it doesn’t just tweak code—it rediscovers it from first principles. The results? A staggering 0.7% recovery of global compute resources in Google’s data centers, equivalent to powering millions of laptops. And that’s just the start.

“This isn’t incremental improvement—it’s algorithmic archaeology,” says a Google DeepMind researcher. “AlphaEvolve finds solutions humans wouldn’t consider, because it doesn’t carry our biases.”

In chip design, AlphaEvolve delivered a Verilog rewrite for an upcoming TPU, optimizing hardware-software handoffs. For AI infrastructure, it accelerated matrix multiplication in Gemini’s architecture by 23%, shaving 1% off training time—a massive win at scale. Transformer models got a 32.5% speedup via optimized FlashAttention, proving the system’s versatility.

Breaking 55-Year-Old Math Records

AlphaEvolve’s most jaw-dropping feat? Solving open mathematical problems. It discovered a 4×4 complex-valued matrix multiplication algorithm requiring just 48 scalar multiplications, surpassing Strassen’s 1969 algorithm—a crown jewel of computational theory. When tested on 50+ open math problems, it matched state-of-the-art solutions 75% of the time and improved them in 20%, including setting a new lower bound (593 spheres) for the notoriously tricky 11D kissing number problem.

“We’re seeing emergent mathematical intuition,” notes a project lead. “It’s not just pattern matching—it’s constructing proofs we can verify.”

The team—including Demis Hassabis, Jeff Dean, and Google’s People + AI Research team—plans an Early Access Program for academics, with eyes on material science, drug discovery, and sustainability. Imagine designing carbon-capture molecules or high-temperature superconductors at AI speed. AlphaEvolve isn’t just writing code—it’s rewriting what’s possible.