Three Words, Three Rows

Each row has a single dominant weight, so attention is picking the one best-matching input word per step and passing that state through unchanged.

These weights are parameters the model learned during training, so this same table is reused for every sentence it translates.

Attention lets the decoder weight the French tokens directly, bypassing the encoder states entirely.

The blend is rebuilt every step: step 2 leans on input position 3 while step 3 leans on position 2, so attention can follow a reordering that one fixed summary — identical at every step — could never express.