The residual stream

A · 10-second intuition Each position carries one 4,096-number vector up through the model, and every sub-layer only adds an update to it.

Not started
Predict before you read

What happens to the residual stream when an attention sub-layer's output is all zeros?