Theore proposes that if all possible predicates in Hilda's space are used, there is no need for training data. To achieve this, a challenge was proposed to obtain state-of-the-art MNIST error rates using very few examples per digit.