👁️ Attention

Scaled dot-product attention, multi-head splits, cross-attention, and masking — derived from one idea instead of memorized as a formula.