Current mainstream deep learning architectures (including Transformers and convolutional neural networks) inherit the core assumptions of the information-processing paradigm: information can be encoded as numerical vectors, hierarchically transformed, and read from or written to explicit locations.