Don't think LLM, think specific training data. instead of tapping out T H I S, you press down at the T, swipe over to the h, up to the I and left to the S then release.

The NPU takes that squiggle you just made and runs it against a model that's been trained on tens or hundreds of thousands of squiggles to deciper the word you just swiped. It's not trained on books, or the whole of humanity and the internet, it's just trained on the keyboard app you're using and other people who have used it. if they swipe and keep going, the word was right and the squiggle counts, if they went back and corrected the text, that's the answer for the squiggle.

There are all kinds of AI things out there that aren't LLM based.