There is a terminal that looks completely ordinary, one that accepts your input of ls, cd, mkdir, and can even open Vim to edit files, run Python, or launch tmux with multiple windows — everything is indistinguishable from a real environment. Yet the truth behind it is quite absurd: there is no Linux, no file system, and not a single command has ever actually been executed; everything that pops up on the screen is made up on the spot by AI. This project, called WorldTUI, does one core thing above all: training a model to learn to "act."
Developer Simon Edwardsson used Qwen2.5-Coder-0.5B as a base and trained a small model with roughly 360 million parameters and a size of 716MB, then compressed it down to 208MB via int4 quantization, so it can easily run locally. Its operating logic is pure: the AI executes no programs at all; it simply predicts, based on the commands you type, what the terminal should display next. Ask it to create a file, and it pretends it did; ask it to read a file, and it strains to recall what it just wrote in there; open Vim, and it switches into the role of Vim and carries on; it can even put on a convincing act as the Python interpreter and tmux multi-window. In a word, as long as it acts convincingly enough, it's hard to spot the flaws.
Even more remarkable, this "drama queen" has memory. Content you wrote into a file with Vim can still be seen when you reopen it a while later, because it stores the information in its context. Of course, once you start a new terminal session, the memory resets to zero — after all, no real files ever landed on disk from start to finish.
This acting skill didn't come out of nowhere—it was earnestly trained. The developer first set up a real Linux terminal, running Bash, Vim, Python, and tmux inside Docker containers, randomly generating usernames, directories, files, and operation tasks; then used scripts to simulate human operation, typing commands character by character, deliberately mixing in 1.2% typing errors, while recording the real screen changes after every keypress. With this method, he accumulated 336,000 operation sessions, about 2 billion tokens of data, used to fine-tune that 0.5B small model, teaching it to infer what should be displayed next based on keypresses and screen history. Whenever the model's prediction was wrong, the real terminal's output was used to construct correction samples, teaching it to patch up its own lies and reduce the bloop rate across consecutive operations. Finally, by predicting different lines of the terminal in parallel, shrinking the vocabulary, and applying 4-bit quantization, the model was squeezed into about 208MB; the whole training took 75 hours, running on an NVIDIA RTX4090.
Of course, the model will eventually give itself away. Complex commands are often executed in a chaotic mess, file contents may suddenly change, and computation results are not necessarily reliable — after all, it never actually ran those programs from start to finish. For ordinary users, this thing has little practical value, but it is an excellent reference: you can use the same approach, having a model observe the operation of real software and then training it to imitate the software's behavior. The developer says he will soon release the model weights for download; the current online experience requires downloading a model of about 208MB, and a desktop browser is recommended. Just don't forget: this thing is fake from start to finish, so don't get too deep into the role.