We show that a two-layer NTK-trained network can exactly execute algorithmic instructions—including addition, multiplication, and SBN—using only logarithmically many examples, via training set orthogonalization and correlation control.