Key point: Anthropic introduces natural language autoencoders that convert Claude’s internal activations into readable text explanations. This technology has already helped identify security issues and improve AI model behavior. The method uses two specialized systems: one explains the activations in language, the other reconstructs them for validation.
Anthropic has developed a revolutionary method to translate the hidden activations of the AI model Claude into understandable language. Natural Language Autoencoders (NLAs) enable researchers for the first time to directly read what is happening inside an AI model during thinking.
Source: www.anthropic.com