I will discuss recent work jet substructure by introducing a suite of ideas organized around infrared and collinear safety, connecting with many old and new ideas in physics and machine learning. Beginning with the construction of a linear spanning basis of IRC-safe observables, I will then introduce "Energy Flow Networks" as machine learning architectures designed to process collider events as variable-length, unordered collections of particles. These techniques achieve good performance on boosted object identification tasks while lending themselves to interpretation in new and fascinating ways.