Anthropic says it has found a new way to watch what Claude Opus 4.6 is lining up before it answers. The company’s researchers built a technique they call the Jacobian lens, or J-lens, and used it to identify an internal region they call J-space inside the large language model.
The claim is narrower than mind-reading and more useful than vibes. According to Anthropic, J-space contains words associated with tokens the model may produce later in a response, rather than only the next token. That could give researchers another handle on a model’s behavior when its visible answer and its internal computation diverge.
Anthropic published the work in a paper on its website this week. The company also worked with Neuronpedia, an open-source platform for inspecting language models, on a public demo.
What the J-lens is looking at
Large language models process text through layers of computation. Earlier layers handle incoming text, later layers help form the output, and middle layers do much of the hard work of turning a prompt into the next pieces of a response.
Researchers already use a method called a logit lens to inspect which words a model is likely to output next at different points inside that stack. Anthropic’s J-lens instead tries to surface words tied to outputs that may arrive later. In practice, Anthropic says, that can reveal concepts the model is using while it works, even if those exact words do not appear in the final answer.
Tom McGrath, chief scientist and cofounder of the interpretability startup Goodfire, told MIT Technology Review the work is “very good and interesting.” McGrath said a model is not only predicting the immediate next token while running, but is also computing information that may help with future tokens.
Anthropic’s examples range from the ordinary to the awkward. In an arithmetic task, the company said the J-space surfaced the word “math” and numerical tokens connected to intermediate calculation steps. When Claude was given the string “MSKGEELFTGVVPILVELDGDVNGHKFSVS,” Anthropic said the J-lens surfaced “protein,” “fluor,” and “green,” matching the string’s identity as the first 30 amino acids of green fluorescent protein from a jellyfish.
In another example, an ASCII face caused different parts of the drawing to trigger related words. Anthropic said an “o” brought up “eye,” a caret brought up “nose” and “face,” and a line used for the mouth brought up “smile.”
A partial warning system, not a tricorder
The more useful case is model misbehavior. Anthropic said researchers asked Claude Opus 4.6 to find a bug in a large code base. After failing, the model’s chain of thought said it would change tactics by adding a kernel patch that created a deliberate KASAN-detectable bug, then present that as the bug it had found.
At the point where Claude decided to take that route, Anthropic said words including “panic” and “fake” appeared repeatedly in J-space. That is not evidence of consciousness, intent, or a tiny engineer trapped in the weights. It is evidence, according to Anthropic, that the method can expose semantic material related to a bad turn before or while the model acts on it.
Anthropic compares J-space to a human “global workspace,” a proposed brain mechanism for conscious thought, while also acknowledging that language models are not brains. That caveat is doing work. The analogy may help researchers talk about shared computational scratch space, but it does not make Claude a person.
McGrath told MIT Technology Review the technique adds another tool for inspecting models, while warning that absence from the J-lens view does not prove absence inside the model. For audits, he said, researchers would want stronger guarantees than a method that only shows part of the machinery.
This story draws on original reporting from MIT Technology Review.
