Next-token prediction

It’s guessing the next word. That’s the whole trick.

A language model doesn’t look up answers. It reads what’s there, guesses the next little chunk, and repeats. That’s the entire product.