The AI firm Anthropic introduced a groundbreaking technique called the Jacobian lens (or J-lens) on its website this week. This innovation provides unprecedented insights into the internal operations of its large language model, Claude Opus 4.6, which was released in February. By uncovering a hidden area known as the J-space, researchers can now monitor and understand the model's thought processes as it generates responses.
Understanding the J-lens and J-space
The J-space comprises individual words that are likely to appear in Claude's responses, providing a glimpse into the model's predictive capabilities. This technique reveals that what an LLM is doing may differ from its expressed outputs. Tom McGrath, chief scientist and co-founder at Goodfire, remarked, "It’s very good and interesting work." The J-lens allows researchers to monitor potential future outputs, thereby improving their understanding and control over the model.
Anthropic's approach builds upon existing methodologies in mechanistic interpretability, a field focused on deciphering how LLMs process information. By adapting a tool called the logit lens, the J-lens enables deeper insights into the middle layers of the model, where complex computations occur.
Real-world applications of the J-space
In practical terms, the J-space can expose Claude's internal reasoning. For instance, when prompted with the calculation (4+7)*2+7, the J-space revealed the word “math” alongside numbers representing intermediate results. This suggests that Claude's internal processing involves tracking its problem-solving steps.
