InterviewsVector

AlphaEvolve: an AI that found faster algorithms — and is already saving Google compute

DeepMind's Gemini-powered evolutionary coding agent found a data-center scheduling heuristic running in production for a year (recovering ~0.7% of Google's global compute) and a 4×4 matrix-multiply that beats a 1969 record.

What happened

On 14 May 2025, Google DeepMind unveiled AlphaEvolve, 'an evolutionary coding agent powered by large language models for general-purpose algorithm discovery and optimization' — it pairs Gemini's idea generation with automated evaluators that verify each candidate, then evolves the best. Concrete results: a scheduling heuristic it found for Google's Borg cluster manager has been 'in production for over a year' and 'continuously recovers, on average, 0.7% of Google's worldwide compute resources'; it found an algorithm to multiply 4×4 complex matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm; and across 50+ open math problems it rediscovered state-of-the-art solutions in ~75% of cases and improved the best known in ~20%.

Why it matters

This is a working template for 'research → production' with AI: wherever a solution can be automatically scored, an LLM can propose candidates and an evolutionary loop can search for better ones — no human labels, just a verifier. The Borg result matters most: it's not a benchmark, it's a durable production win on real infrastructure, which is the proof the pattern generalizes beyond math puzzles.

Staff engineer take

The reusable idea isn't 'DeepMind has a magic optimizer' — it's that if you can write a fast, reliable evaluator for a problem (a scheduler's cost, a kernel's latency, a layout's efficiency), you can put an LLM in an evolutionary loop and let it search the space. The Staff move is to spot the verifiable-optimization problems in your own stack, and to invest in the evaluator — the search is only as good as the score you can compute.

Interview connection

AlphaEvolve improves algorithms by pairing an LLM with an automated evaluator in an evolutionary loop. What property must a problem have for this to work, and where in a real system would you apply it?

Probes verifiability, search/optimization, and where LLM-driven discovery does and doesn't apply.