Anthropic has developed a novel method called the Jacobian lens that reveals a hidden dimension within its language model Claude, providing a clearer picture of the complex internal computations that precede its answers.
- Jacobian lens uncovers a hidden ‘J-space’ inside Claude
- J-space reveals intermediate concepts related to model outputs
- New insights advance mechanistic interpretability of LLMs
What happened
Anthropic built a diagnostic tool called the Jacobian lens (J-lens) and used it to study Claude Opus 4.6, its large language model released earlier in the year. This tool allowed researchers to identify a previously unnoticed internal region, named J-space, where words related to the model’s intended output appear before the final response is produced. Unlike existing methods that predict only the immediate next word, the J-lens reveals concepts the model considers for near-future tokens as well.
This discovery gives a novel view into the model’s internal processing — akin to seeing what words are on Claude’s 'mind' but not yet spoken aloud. The company has shared these findings publicly via a research paper and partnered with the open-source platform Neuronpedia to enable others to explore the J-lens in action.
Why it matters
Understanding the internal operations of large language models is a key challenge in AI research known as mechanistic interpretability. Anthropic’s J-lens technique offers the clearest insight yet into how models handle intermediate reasoning steps and manage competing possibilities while formulating answers. This transparency can help developers better control and refine model behavior, improving reliability and safety.
The findings reveal that what a model is actually calculating internally can diverge from its explicit output. Sometimes intermediate words or concepts hint at underlying thought processes or problem-solving strategies, such as the numerical steps Claude took while doing math. This deepened understanding could guide future innovations in AI explainability and development.
What to watch next
Observers will be keen to see how Anthropic and other AI researchers build upon the J-lens approach to further demystify the workings of large language models. Expanding these insights across different models and tasks could reveal common patterns or unique strategies in AI cognition and decision-making.
Moreover, the availability of hands-on demos through Neuronpedia offers a valuable resource for the AI community to test and refine interpretability tools. Continued progress here will be vital as LLMs become more integral in high-stakes applications, necessitating robust methods to understand and control their internal logic.