[Interim research report] Taking features out of superposition with sparse autoencoders — LessWrong