CrucibleBench: judging LLMs by dropping them in a MUD

What's this about?

A $99 proof-of-concept that evaluates LLM agents by making them play a MUD — a 1990s-style text-based multiplayer dungeon — because the game's strict text interface and persistent world make agent behavior measurable.

What exactly is a "MUD"?

A Multi-User Dungeon: a multiplayer world made entirely of text ("You are in a dark room. Exits: north, east."). The ancestor of World of Warcraft, minus graphics. Perfect for LLMs — the whole world is already words — and cheap: the entire eval ran on $99 of compute.

What's the community saying?

1. Warm nostalgia plus grief that Discord swallowed the MUD communities.
2. One commenter runs local models on a live 30-year-old MUD — they build maps and classify events surprisingly well.
3. Style pushback: "We did not choose a MUD because it is charming" got flagged as an LLM-ism — readers now pattern-match AI-written prose in announcements.