This allows computation to proceed much faster while still giving the model a large context to make
predictions at each step.