Someone trained a tiny AI to pretend it is a terminal

Simon Edwardsson built a terminal with nothing behind it. His WorldTUI is a 360-million-parameter language model that takes each keypress and predicts how the screen should change, with no shell and no filesystem underneath, he writes on his blog (October 21). It still runs vim and tmux, plays snake, and usually remembers what you wrote a few commands earlier.

The trick itself is older than you’d think, so the size is the impressive part. In December 2022, Jonas Degrave wrote about prompting ChatGPT into behaving like a whole virtual machine. Google researchers later did the same kind of thing for Doom with GameNGen, which simulates the game at over 20 frames per second on a single TPU. Edwardsson trained his version in about 75 GPU-hours on one RTX 4090, and it ships as a 208 MB download that runs in a browser tab.

Memory is where it pulls ahead of the general models he compared it to. On his file test, WorldTUI read back 9 of 13 files right, against 2 and 3 for two prompted Qwen models. He calls his snake “probably the most compute-intensive snake in the world,” and I can’t argue with him.

He’s honest about where it falls apart. Anything harder than simple Python, he says, produces answers that remind him of GPT-3-era hallucinations, and nested directories can drop you into a different path. Hacker News gave the post 463 points, and one reader asked the question I’d have asked: does it recognize a row of asterisks as a password?

If you want to play with it, I’d use Chrome or Edge on a desktop with a real keyboard, which is where Edwardsson says it works best. Type pythoon with the typo and it launches Python anyway, which is how you know nothing real is listening.

Is it useful? His own answer is “Not really.” That’s why I’d try it.