Static scalar weights. Same for every input.
4 separate dynamic streams. Weights change per position.
Input-dependent, dimension-wise. ReLU gating per feature.
Softmax attention over depth. Weights sum to 1.