Seol mar théacs é seo: Pseudo-derandomizing learning and approximation