Arguments about machine learning, research practice, and systems designed to remain legible.
Why the apparent noise in a trained network can be read as an encoding of program structure.