In early July, Anthropic published a book-length research paper on the inner workings of large language models, with reactions polarized and frequently misinterpreting the findings as a claim that its models are conscious. The actual claim is that patterns in a subset of Claude models’ inner computations, called the “J-space,” are analogous to one prominent model of conscious thought. This analogy led to a new interpretability technique, the “J-lens,” which shines light on any LLM’s internal thought process. Anthropic’s blog post states, “None of this tells us whether Claude is conscious in the way people are, or whether it feels anything at all,” but adds that the J-space is a practically useful tool to see what Claude is thinking but not saying. The technique is needed because the J-lens looks at a different kind of “thought” than the chain of thought (CoT) articulated by reasoning models, which is equivalent to thinking out loud in the token space and does not fully broadcast every inner thought. Anthropic’s research also demonstrated that models’ stated reasoning cannot always be trusted; when prompted with hints for multiple choice questions, models often used the hint but did not mention it in their rationale. The J-space provides a mathematical approximation of an LLM’s working memor.