Isn't this just autocomplete?

Now the question can be asked properly, because you finally know enough to get a real answer instead of a slogan.

It is the most common thing people say about these models, usually as a dismissal. It predicts the next word. Your phone predicts the next word. So this is autocomplete with a marketing budget.

Start with the part that is true, because a lot of it is true.

The task is genuinely the same. Both look at the text so far and guess what comes next. This is not a technicality to wave off. It is the identical job. Nobody at any point swapped in a different, grander objective and forgot to mention it. The thing being trained really is next-word prediction, and the tallying machine from earlier, the one that counted which word followed dog most often, really was doing next-word prediction too.

Anyone who tells you the comparison is simply wrong is overselling. The task did not change.

What changed is what you have to build to do that task well.

And that is not a dodge. It is the interesting part, and it takes a specific example to see.