Anthropic Introduces Natural Language Autoencoders That Convert Claude’s Internal Activations Directly into Human-Readable Text Explanations