Language arrives in order, comes in different lengths, and often asks a later word to use a clue from much earlier.
RNNs handled changing lengths by reading one token at a time and carrying a fixed-size hidden state. That worked, but every early clue had to survive update after update before a later token could use it.
The same step-by-step chain made every token wait for the state produced by the token before it.
Attention replaced the single running summary with direct connections between token positions. A distant clue could travel in one step, and many positions could be processed together.
That leaves the question attention has to answer: when a token can reach many others, which ones should matter most?
The whole chapter, simply
To understand a late word, an older machine carried clues forward one step at a time.
Important clues could fade or get squashed, so a newer idea let each word look straight back at the words that might help.