We built LLMs but don't actually know how they work
Models are trained, not programmed — no one writes the rules. You show billions of parameters enormous text and they self-organize into something that passes the bar exam but whose internal logic is opaque even to its makers. Interpretability is the field trying to open that black box.
