Anthropic reveals Natural Language Autoencoders, training Claude AI to decode internal activations into readable text, boosting AI interpretability in 2026.