Use Sparse Autoencoders to Discover Unknown Concepts, Not to Act on Known Concepts

Open in new window