tl;dr: A developer fine-tuned Qwen2.5-Coder-0.5B (shrunk to 360M params, quantized to int4/208MB) to simulate a terminal by predicting screen deltas from keypresses, with no actual shell or filesystem underneath. Key tricks included parallel per-row decoding with speculative verification for speed, and DAgger-style self-correction training to prevent compounding errors; the result can run vim, tmux, snake, and Python in-browser via WebGPU, and even remembers file contents written in earlier commands. It's not practically useful, but demonstrates that a small model can maintain a surprisingly consistent "world" without encoding any underlying rules.