A language model does not arrive at a word all at once. It passes a running state through twelve blocks, and a guess condenses out of noise along the way. This puts you inside that state: a horizon of every word still in play, which you turn to read. Drag up and down and you travel the twelve layers, watching the field rearrange itself as the model makes up its mind. Tap a word to write it and the whole thing re-forms around what comes next.
GPT-2 runs entirely in this tab. The model is fetched once from Hugging Face (548 MB); what you type never leaves your device. build 9