One look isn’t enough. Use 8.

Video transcript

One attention look gives one set of percentages. Is that enough for a sentence like the animal was tired?

Not really. One look might link tired to animal, a meaning relation.

But another relation matters too, like grammar, or the nearby word. A single look must blend them all.

So run several looks side by side? Like eight heads, each with its own Q, K and V?

Yes. Each head learns its own view. Before we join them, do the eight heads all have to track the same relation?

Right, training lets each head specialise. Then we glue the eight results side by side.

And a final mix turns the glued result into one output per word. Eight views, one answer.