SHELFCRITTER AI Digest

MODEL

AlphaEvolve

modeltopic-noteagentic-discovery

Overview

AlphaEvolve is DeepMind‘s agentic algorithm-discovery system, pairing frontier reasoning models with verifier-and-evolutionary-search loops to find novel optimisations across mathematics, hardware, and infrastructure code. The system continues the AlphaGo / AlphaTensor line — using model-driven search to attack problems where the verification function is cheap but the search space is enormous.

Timeline

  • 2026-05-08-AI-Digest — DeepMind posts an impact retrospective dated May 7 claiming concrete wins across genomics, the Willow quantum chip stack, an Erdős combinatorics problem, and a 0.7% Borg scheduler efficiency gain inside Google’s own infrastructure. The Borg result is the practitioner-relevant one — at Google’s compute footprint, 0.7% scheduler efficiency is an enormous absolute saving and is hard to fake on aggregate metrics. The broader wins carry the usual lab-self-evaluation caveats (framed wins, not externally replicated), but publication of specific verifiable optimisation deltas pushes the AlphaEvolve story past pure capability-demo territory.
  • 2026-05-11-AI-Digest — DeepMind publishes a one-year-on update reporting a 10× lower error rate on the Willow quantum processor via AlphaEvolve-discovered circuit optimizations, and characterizes the system as graduating from pilot to core Google infrastructure. Secondary-coverage corroborated; direct fetch of the DeepMind blog was egress-blocked in this environment.

Key Developments

  1. 0.7% Borg Scheduler Win — Concrete Deployed Saving: First quantitative external claim of an AlphaEvolve-discovered optimisation deployed at hyperscaler scale. The result is meaningful as evidence that agentic-discovery systems can produce production-relevant code rather than benchmark-only artefacts.

  2. Cross-Domain Coverage: Reported wins span genomics, quantum chip stacks (Willow), pure-math combinatorics (Erdős), and systems infrastructure — the breadth pattern suggests the system’s bottleneck is framing-and-verification more than domain-specific tuning.

See also: DeepMind, Google, MOC - AI Infrastructure.