One neural network predicts the next game frame and reward; another chooses the next action. SheepSeb’s Tinyworldformer fits both on a Raspberry Pi Pico 2 W, with a browser version you can also explore.

The world models learn simplified games from recorded examples. A transformer reuses its weights across several passes, saving memory. These are separate models for each game; training the world models used an RTX 4060, while an actor can learn on the board.

The revealing limitation: an agent can exploit mistakes in its generated world. The project documents extra rewards and impossible shots, then checks generated steps against the original rules.

Code and measurements:
github.com/SheepSeb/tinyworldformer