WorldTUI Fakes a Terminal With a 360M Model That Never Runs a Command
WorldTUI is a small, cheap test of the world-model idea outside games: whether a network can hold a consistent screen state it never actually maintains, and where that consistency breaks.
Reporting from 1 source: GIGAZINE.
Software developer Simon Edwardson released WorldTUI, a browser terminal where an AI model predicts the screen instead of executing commands. It runs a Qwen2.5-Coder-0.5B model retrained on about 336,000 key input and screen change records, trimmed to roughly 360 million parameters and quantized to 4 bits at about 208 MB so it fits a WebGPU browser. Edwardson describes it as an experiment, not a practical terminal.
Edwardson built it after Genie 3, Google DeepMind's world-generating AI, drew his interest to systems that predict environmental change from operations. A 3D game was too expensive to train, so he chose a character grid instead. Line-by-line prediction made updates fast but risked contradictions, so he added a pass that verifies and corrects generated lines. On a GeForce RTX 4090, the 95th percentile time for a large screen update fell from 6.7 seconds to 0.87 seconds.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.