WorldTUI

WorldTUI comes from Simon Edwardsson’s blog and documents an experiment in learned world models: instead of writing an application in code, he fine-tuned a 360M-parameter language model to behave like a terminal. There is no shell and no filesystem behind the screen; the model simply predicts how the display changes in response to input.

The core features are unusual for a terminal. Every keypress is fed to the model, which predicts the next screen state, and it can create files, run vim and tmux, play snake, and usually remember what was typed a few commands earlier. The author highlights two technical wins: teaching the model to repair its own mistakes and decoding screen rows in parallel.

The demo is intended for desktop browsers with WebGPU and a physical keyboard, with Chrome or Edge recommended; it may run on modern phones but very slowly. It suits AI researchers, developers, and enthusiasts interested in world models and learned interfaces rather than anyone needing a real shell.

To try it, open the browser demo (about 208 MB) and type a misspelling such as pythoon to confirm it is not a real shell. Treat it as an experiment in model-predicted interfaces, not a productivity terminal, and expect imperfect recall and rendering.