A learning system’s architecture shapes how it uses information. Neural networks learn representations for prediction; transformers reorganize those representations around relationships among tokens. section 13.4.2
Attention lets tokens relate directly rather than relying entirely on information passed through successive steps. Parallel processing expands training capacity, but shifts the main constraints toward data, compute, and model capacity. These architectural changes leave intact the broader framework of uncertainty reduction under constraints. section 13.4.2

