A tiny world model running games on a Raspberry Pi Pico 2 W

A researcher created a tiny world model that can run games on a Raspberry Pi Pico 2 W. The model uses two neural networks: one generates the game based on the current frame and action, while the other plays the generated game. The model consists of a world model and agents that learn to play the games. The model is designed to be small enough to run on the microcontroller, and the size depends on the game being played.

The content includes the two networks, the games, the world model, the agents, quantization, running on the chip, flashing, a browser version, reproducing the model, and code layout. The researcher aimed to get the number of parameters small enough for the model to run on the Pico 2 W. The model runs on core 1 of the RP2350, while core 0 handles USB and the display.