Conceptual

Residual Connections and Layer Normalization

Skip connections that keep gradients flowing through deep stacks, and per-token normalization that stabilizes activation distributions during training.