Random-policy hill climbing beats LLMs for Brainfuck: inside a superhuman agent swarm stunt

cephaloform · x · 2026-09-11

@cephaloform announced he's building a "superhuman Brainfuck agent swarm," arguing that completely random policies already achieve single-digit percent pass rates on reasonably sized programs, and hill climbing can be done without any natural language data — a tongue-in-cheek but technically pointed challenge to using LLMs for everything.

Original post →

More from Fun

Fun channel →