How Language Models Work

Lesson 12

Training

Nobody taught it any rules. It learned by guessing and being wrong.

Here is the entire recipe for training. Hide the next word. Make the model guess it. Measure how wrong the guess was. Nudge every one of its internal settings, its parameters, a tiny bit toward less wrong. Then do it again. And again. A few trillion times.

An ibex folded from bronze paper, long horns swept back over its shoulders.
Guess, measure how wrong the guess was, take one step downhill. That is the whole of training, and this is what a few trillion careful steps produces.

Do thisPress Train it more five times. Watch two things at once: the guesses in step 2 get smarter, and the line in step 3 drops.

1Train it, one round at a time

2The one question we keep asking it

The cat sat on the  

    3How wrong it is, on average

    Engineers call this number the loss. Down and to the right means it is getting less wrong. That is the whole of “learning” here.