Definition
Plain language
A search method that breeds and mutates candidates over many generations, keeping the best.
As stated in the literature
A population-based optimization paradigm inspired by biological evolution, using selection, mutation, and crossover; in LLM contexts, often paired with model-driven proposal and ranking (e.g., AlphaEvolve, FunSearch).
Also called: evolutionary
Why it matters: It works in domains where gradients are unavailable, such as discrete program synthesis or combinatorial design, and pairs naturally with LLMs as mutation operators.
For example, you might start with a hundred candidate functions, keep the top-scoring ten, mutate each into ten variants, and repeat for many generations until a strong solution emerges.
Heard on the show
“A scalar fitness score, like in evolutionary search, can't encode that.”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure