tl;dr: Two tiny neural networks run on a Raspberry Pi Pico 2 W: a looped int8 transformer "world model" (26k–131k params) that generates game frames and rewards from an action, and an MLP actor that plays inside the dream. Five games (Breakout, Pong, Flappy Bird, Doom-lite, Pac-Man) were implemented, with the world model achieving 83–100% exact frame prediction and fitting within the Pico's 4MB flash and 512KB SRAM. The project also includes on-device reinforcement learning via Stream AC(λ), WebAssembly builds, and reusable PyTorch→no_std Rust export tooling.