A working language model · nothing faked · no code from you

You won't be told
how an LLM works.

You'll be handed one with the panel open. Every dial on the next few pages drives a real model — turn it and the thing genuinely changes, or genuinely breaks.

The fastest way to stop being mystified by this technology is to put your hands on a small one and ruin it a few times.

Step in about an hour, or ten minutes at a time

No code from you

Every number here came out of real Python, and you can run it yourself — but nothing asks you to write any. You get the instrument, not the source file.

Nothing faked

Every curve, sample and score is from an actual run. Where a run died, the line stops at the step it died on — no invented tail.

You'll train one

Not an animation of training. A real mini model, real loss, noise turning into language in front of you.

It's one river, and it reads best from the source — each station is built on the one before it. But nothing is locked: the whole run is laid out below, and you can drop in wherever you already know the water.

The whole river

Every station says the one move it wants from you.