AI
2026-08-15
I spend a lot of time around machine learning and a smaller amount of time believing what people say about it. The part I trust is closer to applied math: when a model is actually learning a representation, when it is interpolating in a kernel, and what that difference costs.
The research-shaped version of this is network richness. The argument-shaped version is email. I am happier to discuss this than to maintain a stream of takes.
If you want to argue, the contact page is the door. If you want papers, they are here.