Same transformer parts — stacked three different ways. Which direction does information flow? That one choice decides what the model is for.
Click a panel — or a button — to spotlight one architecture. Watch the arrows.
All three are made of the same transformer blocks — attention plus a little MLP. The only difference is who is allowed to look at whom. Hide the future and you get a writer. Let everyone see everyone and you get a reader. Bolt a reader to a writer and you get a translator.