AIster · The agentic age · Breakthroughs
Anthropic traces how Claude thinks
An “AI microscope” showed Claude planning rhymes before writing them
Key facts
- Date
- 27 Mar 2025
- Era
- The agentic age
- Interest
- Breakthroughs
- Threads
- Model wars, Safety & governance
- Model studied
- Claude 3.5 Haiku
- Papers
- “Circuit tracing” and “On the biology of a large language model”
- Example
- To find the capital of the state containing Dallas, it first represents “Dallas is in Texas”
- Recognition
- MIT Technology Review, 10 Breakthrough Technologies 2026 (12 Jan 2026)
Anthropic used a technique it called an AI microscope to follow the steps inside Claude 3.5 Haiku. It found the model plans rhymes before writing a line, chains facts together to answer multi-step questions, and shares concepts across languages. MIT Technology Review later named mechanistic interpretability, the study of how a model’s internal parts produce its answers, one of its 10 Breakthrough Technologies of 2026.
Sources
- Anthropic, Tracing the thoughts of a large language model, 2025-03-27
- MIT Technology Review, Mechanistic interpretability: 10 Breakthrough Technologies 2026, 2026-01-12