LLMs evolve Core War warriors that beat or tie 96% of 294 human-written ones, says MLST podcast

Machine Learning Street Talk · youtube · 2026-10-10

Machine Learning Street Talk hosts MIT PhD student Akarsh Kumar (Phillip Isola's group, collaborates with Sakana AI, first author of the Fractured Entangled Representation paper with Stanley, Clune and Lehman).

Key argument

Evolution is not a dumb random fallback: selection preserves partial solutions, so mutations only need to be useful 1% of the time to keep progress going; the path a learner takes shapes the structure of what it learns — a regularity-based view of intelligence distinct from the statistical one.

ASAL: foundation models as critics for artificial life

Instead of predicting what a rule does, ASAL runs the simulation and asks a foundation model what happened.
Sweeping all 262,144 Life-like rules, the most interesting worlds cluster on a small island of open-endedness; the episode covers Game of Life, Lenia (with a clip from Bert Chan), neural cellular automata, Boids, Particle Life and the emergence of persistence.

Core War experiment

LLMs asked cold write poor Redcode, but used as the mutation step in an evolutionary loop with a verifier, their warriors beat or tied 96% of 294 human-written ones, trending toward generalists over rounds.

Original post →

LLMs evolve Core War warriors that beat or tie 96% of 294 human-written ones, says MLST podcast

More from AGI Musings