Great new work from @ch402 on the differences between representing activations with directions for independently varying features vs a dense yet structureless code. I really appreciate the commitment to writing great exposition!