WorldTUI: a 360M-parameter model trained to predict terminal screen updates with no shell behind it
AspireAI community · simedw.com
Simon Edwardsson fine-tuned Qwen2.5-Coder-0.5B into what he calls a world model for the terminal. There is no shell and no filesystem behind it: each keypress goes in as the raw bytes a real terminal would send, and the model predicts which screen rows change and where the cursor ends up. Because a terminal is a grid of characters rather than pixels, he could generate perfectly labelled training data by recording real shells running in Docker containers. The run came to 336,000 sessions, about 2B tokens, in roughly 75 GPU-hours on an RTX 4090, against about 18T tokens for the base model's pretraining. Two techniques mattered most: DAgger-style self-correction training, where the model predicts, the result is compared against the real recorded screen, and a correction is built only from words that already appeared on a real screen or in the user's typing, deliberately avoiding rewarding hallucination; and decoding screen rows in parallel then verifying them sequentially with speculative decoding, which cut large-frame p95 latency from 6.7s to 0.87s on a 4090. Quantised to int4 it is 208 MB and runs entirely in a browser tab through WebGPU. His own results: ShellBench 80/106, against 52 for prompted Qwen2.5-Coder-0.5B and 60 for prompted Qwen3-0.6B; file readbacks 9/13, against 2 and 3. The limitation he states plainly is that it is not a reliable filesystem - after appending to a file, catting it often does not show the appended line, and harder Python produces answers he compares to GPT-3-era hallucinations. Model weights are not published yet; only the browser demo is available. Source: simedw.com, Oct 21 2026.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.