What a transformer actually is

So is a transformer a program? Partly. The word gets used for several different things, so here is each of them.

Somewhere there is a file of code saying: build this many floors, put attention on each one, wire them together in this order. That file is an ordinary program, and running it is how a tower like that comes into existence on a computer.

But the word does not name that particular file. Many companies have written their own version, in different programming languages, sharing no code with each other, and every one of them is a transformer. What they share is the arrangement their programs build.

That arrangement is half of what you need. The program lays out the network: how many floors, how they connect, where every adjustable number sits. Training decides what those numbers should be. Load the trained numbers into the matching network and you have a working model.

Two separable things, then. A transformer is the shape. A trained model is that shape with its numbers filled in. Which is why one transformer can become thousands of different models, all sharing a design, each holding numbers learned from different text.