Step back and look at what this module built.

It started with a wall. Machines read language one word at a time, holding a single fading summary, unable to connect words across any real distance. Then came one idea, attention: let every word reach out, weigh every other word, and blend in what matters, directly, all at once. That idea dissolved the wall. Distance stopped mattering, and the work ran fast on the GPUs the field already had.

Stack that idea into a tower, stamp each word with its position so order survives, and you have the transformer. Choose whether each word may peek ahead and you get either a deep reader or a fluent writer. Pretrain it once on an ocean of text and it absorbs language itself; fine-tune it cheaply and it becomes whatever you need. That is the machine underneath nearly every AI system you have heard of.

So we have, at last, assembled the thing. We know what a model is made of, how it turns words into numbers, and now how it reads them all together. One enormous question remains, the one this chapter kept pointing at: what happens when you make a transformer very, very large?

The answer was strange enough that it surprised the people building them, and it is the whole of the next module.