AI Interpretability

Mechanistic interpretability treats a trained network as something to reverse-engineer rather than a black box to benchmark. A joint session with mathematics on the techniques for working out what internal components compute, and what would even count as an explanation.