L lesswrong.com Gradient Descent in Activation Space: A Tale of Two Papers 4 points kofejnik 3 years ago 1 comments