Natural Language Autoencoders: Turning Claude's Thoughts Into Text

Natural Language Autoencoders translate Claude's internal activations into readable text, surfacing when the model suspects testing or hides motivations.