
What happened
Developer Simon Edgerson published WorldTUI, which uses a roughly 360M-parameter model to predict the next terminal screen from keystrokes, with no real command execution or file access; it runs free in WebGPU-capable browsers.
Why it matters
It is a low-cost test of whether an AI can mimic consistent behavior without knowing the underlying system — and it already shows the cracks, sometimes missing text appended to a file or giving wrong answers on complex Python.
Summaries like this, every morning.
Edgerson says he was inspired by Genie 3, Google DeepMind's world-generating AI, and became interested in AI that predicts how an environment changes in response to input. Reproducing complex 3D spaces like those in games would demand large amounts of training data and compute, so he chose the simpler grid of characters in a terminal instead. The model is built on Qwen2.5-Coder-0.5B and fine-tuned on about 336,000 records of key input and screen changes captured from a real terminal.
Speed was the main engineering problem. Generating the whole screen each time was slow, so the system updates only the lines that need to change, and it generates several lines in parallel; in a test on a GeForce RTX 4090, the 95th-percentile time for a large screen update fell from 6.7 seconds to 0.87 seconds. Generating lines separately can make the display inconsistent, so a mechanism checks and corrects the output. Trimming unneeded parameters and applying 4-bit quantization brought the model to about 208MB, small enough to run in a browser.
Edgerson says he did not build WorldTUI as a practical terminal but as an experiment to see whether an AI can reproduce consistent behavior without knowing the internal workings. The limits show up in ordinary use: it can handle simple Python calculations, but since Python is not actually running, complex processing can yield wrong results — a reminder that the screen is a prediction, not a computation.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Microsoft said on October 21 that attackers tied to China's government are automating intrusions with ARTEX, a…

1Password unveiled 1Password for Agents, which lets an AI agent request credentials from a user's vault and, o…

A new paper found Raw-RAG—dense retrieval over raw history—ranked first in 6 of 7 benchmark-model settings on…

Nvidia has won a CPU supply order from AI startup Modal for its Vera CPU, moving beyond GPUs into general-purp…

In a Google blog post summarizing the video "The Agent Factory" on Sept 25, 2026, Ryan Lopopolo defined an age…

ModelCompare asked 12 AI models three times each for their most controversial view of humans; all 12 gave the…
