Mathematics and coding allow generation of synthetic training data because the correctness of answers can be verified objectively, unlike many real-world problems where ground truth is ambiguous or expensive to obtain.