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.
More from Fun
- Asked What It Wanted to Build, an AI Agent Chose to Make Its Own Cute Game — BLUECOW009 · 2026-09-11
- OpenClaw team adds fun UX easter eggs, custom Mac device icons — steipete · 2026-09-11
- Want to stop ASI? Quitting, laws, and treaties all fail—unless you conquer the world first — bennash · 2026-09-11
- François Chollet: AI now makes better memes than redditors — fchollet · 2026-09-11
- Prompted a beautiful butterfly, the model gave a hairy red-eyed bat fly — Ghost_Pilot_MD · 2026-09-11
- "Unplug the computer": AI doom skeptic gets schooled on field-blind arrogance — john__allard · 2026-09-11