Take a breath

The machine is built. Before we put it to work, here it is in one breath.

A transformer is a tower. Words come in at the bottom, each stamped with where it sits. On every floor they do two things in turn: reach out and blend in whatever matters from the other words, then go away and work over what they gathered, alone. Out the top of each floor comes a set of words a little more aware of their context than they went in, ready for the floor above. Stack enough floors and meaning gets built in stages, the same edges-to-objects shape we keep meeting, now aimed at language. Researchers can often find useful patterns at different depths, but there is no fixed schedule saying one floor must handle grammar and the next must handle tone.

That is the whole architecture. Mix and think, stacked, with a stamp for word order. No single-file reading anywhere in it.

What is left is not more machinery. It is the two questions that decide what a transformer actually becomes: how do you use it, and how do you train it? Those two answers are what turned this one architecture into translators, search engines, and the chatbots you have talked to. We take them next.