Project · Kernels and compilers

AlphaEvolve

An evolutionary coding agent that pairs Gemini models with automated evaluators to propose, test and keep improvements to code. Google applied it to math problems and to its own computing stack: data-center scheduling, Gemini training kernels, compiler-generated attention kernels and TPU circuit design.

The loop

Gemini models inside AlphaEvolve propose code changes and automated evaluators keep the versions that score best. Applied to Google's AI stack, it found a better way to split a matrix multiplication kernel used to train Gemini and proposed a Verilog change that was integrated into an upcoming TPU, so the agent lowers the cost of the hardware and software that train the next Gemini models, which in turn power the next AlphaEvolve.

The loop

AlphaEvolve

Project · Google DeepMind

Gemini models propose code changes
Evaluators keep best scoring versions
Finds better Gemini training kernels
Lowers cost to train next Gemini models

Gemini models propose code changes
Evaluators keep best scoring versions
Finds better Gemini training kernels
Lowers cost to train next Gemini models

↻ The improved system does the next round, and the loop turns again.

Why it is a road to recursion

It is a documented case of a model family measurably shortening its own training run, and the same agent works on kernels, compilers and chip circuits, so each Gemini generation can make the next one cheaper to build.

Evidence

Google reports that AlphaEvolve sped up a matrix multiplication kernel in Gemini's architecture by 23%, cutting Gemini's training time by 1%, reached up to a 32.5% speedup on a FlashAttention kernel implementation, and found a scheduling heuristic that has recovered on average 0.7% of Google's worldwide compute for over a year.

evolutionary-searchgeminigpu-kernelstpudata-center