Mechanistic Interpretability

Research into how AI 'thinks' internally — enabling better understanding and control of AI behavior. 2026 breakthrough per MIT.

MIT Technology Review selected this as one of the 10 Breakthrough Technologies of 2026. Anthropic used an 'AI microscope' to discover 34 million features in Claude — from 'sarcasm' to 'DNA sequence...