WorldTUI: I trained a 360M model to pretend to be a terminal

A developer fine-tuned a 360M-parameter language model to simulate a terminal interface, where each keystroke triggers the model to predict how the screen changes rather than running actual shell commands. Trained on real terminal sessions from Docker containers, the model can perform tasks like creating files, running vim and tmux, and playing snake, though it's designed to demonstrate the concept of learned applications rather than replace real shells. The interactive demo is available in browsers supporting WebGPU and shows that machine learning models can maintain consistent virtual environments by predicting character grid changes instead of generating linear text streams.