A block doesn't run attention once — it runs several heads in parallel, each specialized. Pick a head, and click any word to see what it looks at.
Illustrative specializations — real heads learn their own quirks — but this is the idea: each head attends with its own Q/K/V weights. Their outputs are concatenated and projected (by Wₒ) into one vector per word, so a token gets information from several relationships at once.