Anthropic used circuit tracing to reveal more internal processes of Claude. It thinks in a conceptual space shared across languages like universal language of thought. It can plan responses several words in advance when writing rhyming poetry. It computes both rough estimations and precise calculations in parallel when solving math problems. It can combine independent facts to arrive at an answer, instead of simply recalling a memorized response. #LLM #AI #ML anthropic.com/research/tracing…
Tracing the thoughts of a large language model
Anthropic's latest interpretability research: a new microscope to understand Claude's internal mechanismswww.anthropic.com