Interpretability

Anthropic Wants Claude to Explain Its Hidden Activations in Plain Language

Anthropic introduced Natural Language Autoencoders, a research direction for translating Claude activations into readable explanations.