Beyond Scaling Curves: Internal Dynamics of Neural Networks Through the NTK Lens