I have a lot of respect for Nick, and I’d be very happy if neural networks were much easier to understand than we thought. But I think the argument here mischaracterises what superposition predicts. Since it’s getting attention, here’s a first pass at some issues.

this is who runs this account