AlphaEvolve: DeepMind's Autonomous Algorithm-Discovery Agent
Google DeepMind's Pushmeet Kohli and Matej Balog introduce AlphaEvolve, an autonomous coding agent that pairs Gemini-family LLMs with evolutionary search to discover new algorithms that outperform human-designed ones. The system is positioned as the latest step in a lineage from AlphaGo (Move 37) through AlphaTensor (matrix multiplication) and FunSearch, generalizing those domain-specific breakthroughs into a more broadly applicable agent that searches in the space of programs rather than hand-crafted operations. AlphaEvolve is framed as both a scientific and practical tool. Scientifically, it has already produced new results in mathematics and computer science, including novel constructions in open problems worked on for decades. Practically, it has been deployed across Google's stack—yielding efficiency gains in data center scheduling, AI training infrastructure (a ~23% speedup in a kernel used for Gemini training), and hardware/circuit design. The hosts emphasize that discovered solutions are inspectable code, a meaningful advantage over opaque neural-network decision systems. A central technical theme is the role of the user-provided evaluator: AlphaEvolve requires a function that scores candidate solutions, which limits the domains it can currently attack but also explains its reliability versus general coding agents that plateau or hallucinate. The guests argue this constraint is not fundamental—spectrums from strict simulators to LLM-based evaluators (as in their AI Co-scientist work) can relax it. They also frame AlphaEvolve as a potential early signal of self-improving AI, while noting today's feedback loop is slow (months-scale) and the question of compounding improvements remains open. Broader access will come via a trusted-tester program.
Original sources
Podcast No Priors
YouTube No Priors
Video:No Priors Ep. 120 | With Google DeepMind’s Pushmeet Kohli and Matej Balog ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.