The article discusses the limitations of deep neural networks when it comes to forming abstractions, which are not accessible to humans, and highlights the possibility of a new form of deep learning that incorporates symbolic manipulation.