NanoWorld Speedrun

Train a next-frame world model of Procgen CoinRun that people can actually play — coins score, deaths reset, controls respond. Then race to the target quality in the fewest wall-clock minutes on one fixed 8xH100 node.

Nothing in the clip is a recording: every frame is invented by the model, one at a time, in response to the keys a player is pressing. You are given access to a game environment, and must generate your own training data from it and then train a tiny video world model against that data. Entries are timed end to end and the record book lives on GitHub.

Your score is the wall-clock time to reach the target; lower is better.