Anthropic Maps the 'J-Space' Inside Claude Opus 4.6

Anthropic has developed a new technique to probe the internal operations of its flagship LLM, Claude Opus 4.6, which was released in February. According to reporting from MIT Technology Review, the company created a tool called the Jacobian lens (or J-lens) to uncover a hidden area within the model dubbed the "J-space."

While traditional logit lenses identify the immediate next word a model is likely to produce, the J-lens identifies words and phrases the model is focusing on that may appear later in a response, or may not appear at all. This allows researchers to see what the model is "thinking" about in the middle layers of its computational stack—the area where complex math transforms prompts into responses—before the final output is generated.

Anthropic's findings, shared in a recent paper and a demo hosted by the open-source platform Neuronpedia, show that the J-space can reveal the model's internal logic. For instance, when calculating (4+7)*2+7, the J-space contained the word "math" and intermediate results "21" and "42." In other cases, the J-lens revealed how the model recognized specific inputs: a string of amino acids triggered the words "protein," "fluor," and "green," while parts of an ASCII face triggered "eye," "nose," and "smile."

Tom McGrath, chief scientist and cofounder of the startup Goodfire, noted that while much of the J-space is mundane, it can reveal surprising internal themes. Anthropic further noted that monitoring this space provides a new method for understanding and controlling its models, as the internal processes revealed by the J-lens can differ from what the model explicitly states it is doing.