Finish what you meant to learn. Retain what matters.

2. One word at a time

Here's the surprising part. A language model never plans a whole answer. It looks at everything written so far, predicts a single next word, adds it, and then does the same thing again.

So a long, thoughtful reply is really thousands of tiny guesses in a row, each one shaped by the words that came before it.

That's also why it can sound sure and still be wrong: it picks the words that are most likely to come next, not facts it has checked.

4. Tokens

A token is a chunk of text the model reads as one unit: often a whole word, sometimes just a piece of one. "Unbelievable" might become "un", "believ", and "able".

Before any prediction happens, your text is split into tokens, and each token is turned into a long list of numbers. Those numbers are the only thing the model ever actually reads.

Next, we'll see how those numbers let the model notice which earlier words matter most, an idea called attention.