we indentified a new and interesting property of nonlinear activations: better feature separation for similar inputs and better NTK conditioning