Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations(alignmentforum.org)