What do the heads actually look at?

Video transcript

We have eight heads in each layer. But what are they actually looking at?

The paper drew attention maps. In this sentence, one head makes the word it point strongly to animal.

So did someone teach that head to link pronouns to nouns? Or did it figure that out alone?

It learned it alone. Training only rewarded better translations, and linking it to animal helped. Other heads track grammar, like a verb finding its subject.

So different heads specialized, and nobody told them to. That's wild.