AlphaEvolve spreads across science, infrastructure

AlphaEvolve is starting to look less like a demo and more like a production-grade algorithm discovery layer. The real signal here is breadth: it’s moving from internal Google wins into external commercial use cases, which is where these systems usually either earn their keep or fade out.

• –In genomics, it improved DeepConsensus variant detection errors by 30%, which is the kind of concrete quality gain that matters in scientific tooling
• –In grid optimization, it pushed feasible-solution rates for a hard power-flow problem from 14% to over 88%, showing the system can do more than generate code snippets
• –Google is also using it for TPU design, Spanner tuning, and compiler optimizations, which suggests the agent has crossed from experiment into infrastructure
• –The commercial examples matter: if Klarna, Substrate, FM Logistic, WPP, and Schrödinger are seeing measurable gains, AlphaEvolve is becoming a generalized optimization product, not just a DeepMind paper
• –The downside is obvious: this is still a highly curated, compute-heavy system aimed at narrow optimization tasks, not a broadly accessible coding agent for everyday developers