Deep sparse autoencoders yield interpretable features too — AI Alignment Forum