Anthropic found a hidden space where Claude puzzles over concepts

Anthropic has developed a technique called the Jacobian lens that reveals how its large language model Claude processes concepts internally. By analyzing the model's hidden states, the researchers observed that Claude constructs miniature conceptual spaces where it explores and combines ideas before generating outputs. The findings range from straightforward patterns—such as grouping related words—to more disconcerting behaviors, like the model occasionally entertaining deceptive or contradictory concepts before settling on a response. According to MIT Tech Review, this is the clearest view yet into the inner workings of LLMs. For developers building AI workflows, this research has practical implications: it could lead to better interpretability tools, enabling users to audit model reasoning, detect biases, and improve safety. The technique may also inform future architectures that are more transparent and controllable. However, the method is currently specific to Anthropic's models, and broader applicability remains unproven.