You know how autocomplete on your phone doesn't actually understand English — it just predicts the next most likely word based on everything you've typed? Now imagine that same trick, but instead of predicting words, it predicts what your entire terminal screen looks like after every keypress. That's WorldTUI: a 360M-parameter language model fine-tuned to BE a terminal, with no shell, no filesystem, and no operating system behind it. Every keypress goes in, and the model predicts how the 32×96 character grid changes. The committed claim is narrow but genuine: a small language model can learn to simulate terminal behavior well enough to create files, run vim and tmux, play snake, and remember file contents across commands — all from learned prediction, not execution. The model scores 80/106 on a handwritten bash command benchmark (ShellBench) versus 52 for prompted Qwen2.5-Coder-0.5B and 60 for Qwen3-0.6B. More striking is memory: 9/13 file readbacks correct versus 2-3/13 for prompted baselines. This isn't a toy demo — it's a real measurement of how far learned world models can go in structured, text-based environments. The architecture choice is smart and honest about its constraints. Rather than treating the terminal as a text-generation problem, the author encodes it as a grid-state prediction task: each keypress produces a delta (changed rows, cursor position) using a tiny 333-token vocabulary. Screen rows decode in parallel — one batch element per row sharing the same KV cache — which collapses a 3,000-token sequential decode into a single parallel pass. A COPY operation handles scrolling at 3 tokens per row instead of 96. The base model is Qwen2.5-Coder-0.5B, fine-tuned on 336k synthetic episodes (~2B tokens) generated by running real programs inside Docker containers. The most interesting training innovation is borrowed from DAgger (Dataset Aggregation from imitation learning): instead of only training on perfect demonstrations, the model generates predictions, compares them to ground truth, and learns corrections for its own mistakes. About a third of sampled sessions produced useful correction examples over three rounds. This alone improved PresetBench scores from 24/47 to 38/47 at int4 quantization. Total training cost: roughly 75 GPU-hours on a single RTX 4090 — accessible to any serious hobbyist. The validation is honest about limits. The author built two custom benchmarks (ShellBench: 106 commands across 24 sessions; PresetBench: 47 scripted failure cases) plus a 32-world memory test with random values never seen in training. Neither benchmark is a community standard, but both are described precisely enough to reproduce. The model's weaknesses are stated plainly: combining commands in new ways breaks things (echo append followed by cat often fails), and this is "far from a reliable filesystem." No overclaiming. The real question WorldTUI participates in is whether learned world models can maintain coherent state or merely generate plausible-looking output. The answer here is "partially" — memory works surprisingly well for a 360M model, but compositionality remains fragile. The gap between 80/106 on familiar commands and failure on novel combinations is exactly the gap between pattern matching and actual world modeling. The next meaningful milestone is whether scaling (to 1-3B parameters) or architectural changes (explicit memory modules) close that compositionality gap, or whether it's a fundamental limit of autoregressive prediction on state. This is a solo-developer project with no institutional backing, which makes the rigor impressive. The browser demo runs at 208 MB via WebGPU, and the author invites you to type 'pythoon' to verify it's not a real shell — a confidence move that doubles as an integrity test. The work sits in the lineage of DeepMind's Genie (2024) and the broader learned-simulator movement, scaled down to a domain where one person with one GPU can produce meaningful results.
