Anthropic researchers have developed a tool called the Jacobian lens (J-lens) that uncovered a hidden area inside Claude Opus 4.6, which they named J-space. This space contains individual words related to the words and phrases the model is most likely to produce in its next response, functioning as a window into what the model is "thinking" before it speaks.
The company found that what an LLM is actually doing can often differ from what it says it is doing. Anthropic claims that monitoring words that appear in the J-space provides a new method to understand and control its models. They published their findings in a paper this week and partnered with Neuronpedia, an open-source platform, to create a hands-on demo for public exploration.
The J-lens builds on previous mechanistic interpretability research, adapting an existing tool called a logit lens. While logit lenses examine what words an LLM is likely to produce next, J-lens goes deeper into the middle layers of the model where the complex processing occurs, revealing a previously unseen level of internal model activity that McGrath called "very good and interesting work."