Scalable Interpretability via Polynomials