Before it
Neural translation worked in papers; production systems were still statistical.
The problem
Research models were too slow and too brittle to serve at Google scale.
What it introduced
A deep LSTM encoder-decoder with attention, wordpiece tokenization and the engineering to deploy it.
Why it matters
The first time a neural sequence model replaced a mature statistical system in a product used by millions.
What modern models inherit
WordPiece tokenization, and the proof that neural generation can ship.
What came next
The Transformer, from the same organisation, made this architecture obsolete within a year.