The old way: read one word at a time.

Video transcript

So how did older translation models read a sentence?

Like a person with a small notebook. They look at the first word only, and write a note in memory.

Then they move to the next word, and update the same notebook with what they just saw.

So the notebook mixes the old notes with each new word. When it reaches the third word, what does it know about the second?

Right. Only the memory so far and the current word. It goes left to right, one step at a time, until the end.

Got it. One word, update memory, repeat. The notebook is small though. That seems risky.