Definition
Plain language
A punishing dungeon-crawler video game with permanent death and randomly generated levels, used to test how well an AI agent can survive very long, unpredictable play.
As stated in the literature
A genre of turn-based dungeon-exploration games (e.g., NetHack) featuring procedural generation, permadeath, and deep item interactions; used as long-horizon RL benchmarks with episodes running to tens of thousands of steps.
Why it matters: Its unforgiving, ever-changing design makes it a tough test of whether an AI can plan and adapt over very long stretches rather than just react moment to moment.
For example, in a game like NetHack the dungeon is different every time and one death ends the run, so an agent may need tens of thousands of careful moves to make progress.
Heard on the show
“In NetHack, the roguelike that runs to a hundred thousand steps, the starting agent keeps a file called dungeon map, and it's append-only.”Episode 192 — A 32B Open Model Matched Frontier Systems By Learning to Take Notes