How to Better Report Sparse Autoencoder Performance — LessWrong