DeepMind's new agent designs code that beats human mathematicians

Google DeepMind just took the AI self-improvement loop from science fiction to production.

Meet AlphaEvolve, a Gemini-powered coding agent built to discover and optimize complex algorithms from scratch across math and modern computing.

Instead of just predicting code completions, AlphaEvolve pairs Gemini Flash for rapid idea generation with Gemini Pro for deep reasoning. It spits out code, puts it through an automated evaluator, scores the result, and iterates using an evolutionary framework.

Shaving 0.7% off Google's data centers

This isn't an academic demo. Google has quietly been running AlphaEvolve-generated code in production for over a year.

Its simplest win might be its most lucrative: it wrote a clean heuristic for Borg, Google's data center cluster manager, that recovered an average of 0.7% of Google's global compute resources. At Google's scale, that translates to massive power and hardware savings.

It didn't stop at software. AlphaEvolve rewrote Verilog code to strip redundant bits from arithmetic circuits used in matrix multiplication. That improvement is already baked into Google's upcoming TPU.

When applied to AI training itself, it accelerated a vital matrix multiplication kernel inside Gemini by 23%, cutting overall Gemini training time by 1%. It also squeezed a 32.5% speedup out of FlashAttention kernels on GPUs, turning weeks of painful human optimization into days of automated compute.

Beating Strassen and 300-year-old math

DeepMind also unleashed AlphaEvolve on pure mathematics, and the results are wild.

It discovered a new algorithm for 4x4 complex matrix multiplication using just 48 scalar multiplications, finally beating Strassen's iconic 1969 record. That's a massive leap beyond DeepMind's previous AlphaTensor, which only beat binary arithmetic on 4x4 matrices.

Out of 50 open problems in geometry, number theory, and analysis, AlphaEvolve matched state-of-the-art human solutions in 75% of cases. In 20% of cases, it improved them.

That includes the 300-year-old kissing number problem, where it found a 593-sphere configuration in 11 dimensions to set a brand-new mathematical lower bound.

Why it matters

We are looking at the foundational pieces of recursive AI self-improvement.

AlphaEvolve uses Gemini to optimize the kernels that train Gemini, the data center schedulers that host Gemini, and the TPUs that will run the next generation of Gemini. Once that loop closes, engineering velocity starts moving at machine speed rather than human hiring speed.

What makes DeepMind's approach clever is pairing LLM hallucination-prone creativity with hard, uncompromising automated evaluation. If code doesn't verify, it dies. That makes the solutions practical, deployable, and surprisingly readable for human engineers.

The catch? AlphaEvolve only works where success can be mathematically quantified and automatically verified. It won't solve messy, subjective product problems anytime soon. But for hardware, compilers, and core infrastructure, the bar for human optimization just got moved permanently higher.