“Interpretability research reveals that when an LLM is asked a difficult math problem it cannot solve, its chain-of-thought reasoning is fabricated and its internal circuits show no meaningful computation.”