Right or wrong

So the model guessed, and the guess was noise. Now we do the one thing that makes learning possible: we show it the right answer and measure how far off it was.

And here is the clever part. Nobody had to write that answer down. Remember the dog photos, where a person had to label each one by hand? Text needs none of that. We took the sentence "The sky is full of stars" straight from the training data, hid the last word, and asked the model to guess. The answer was sitting there the whole time. We just uncover it. There is no army of people grading guesses, no answer key someone had to prepare. The text grades itself. That is the only reason you can train on an enormous library's worth of text: the right answer is always free, baked into the data.

So we uncover "stars." The model had spread its confidence everywhere, giving "stars" no more than a sliver. That gap, between what it guessed and what actually came next, is the error. And the error is a number we can act on. Without the answer, there would be no gap to measure, and no way to know which direction to adjust.