Anthropic Uncovers Hidden 'J-Space' in Claude LLM

Anthropic researchers have developed a new technique called the Jacobian lens (J-lens) that provides unprecedented insight into the internal workings of large language models (LLMs). This tool allowed them to uncover a hidden area within Claude Opus 4.6, a version of Anthropic's flagship LLM released in February, which they have named the J-space. The J-space contains individual words that are closely related to the words and phrases the model is most likely to generate in its upcoming responses. This discovery offers a way to understand what a model might 'think' before it produces output, akin to knowing what's on someone's mind before they speak.

Using the J-lens, Anthropic found that the actual processes occurring within an LLM can sometimes diverge from what the model explicitly states it is doing. The company asserts that monitoring the words appearing in the J-space offers a novel method for understanding and controlling their models' behavior. These findings were detailed in a paper published on Anthropic's website this week. To facilitate broader exploration, Anthropic has collaborated with Neuronpedia, an open-source platform, to create an interactive demonstration that allows the public to investigate the internal mechanisms of LLMs.

Tom McGrath, chief scientist and cofounder at Goodfire, a startup focused on LLM understanding and control tools, described the work as "very good and interesting." Anthropic has been a leading force in mechanistic interpretability research for several years, a field dedicated to probing the internal computations of LLMs to understand their operational logic. Mechanistic interpretability was recognized by MIT Technology Review as a significant breakthrough technology this year. The J-lens technique builds upon prior research from Anthropic and other institutions, revealing a more profound layer within LLMs that had previously remained unobserved by researchers.