Anthropic's J-lens reveals what Claude thinks before it speaks

Key Takeaways

• Anthropic's J-lens tool reveals words Claude is considering before it commits to a response, exposing a hidden 'drafting room' inside the model
• The technique found that what an LLM is actually doing can differ from what it says it's doing, raising new questions about AI transparency
• The company partnered with Neuronpedia to release a public demo, letting anyone probe LLM internals themselves

Anthropic has developed a technique that exposes what its Claude models are 'thinking' before they commit to a response. The tool, called the Jacobian lens (J-lens), uncovered a hidden internal space where Claude appears to draft and refine its thoughts, sometimes revealing reasoning steps that never make it into the final output. The company claims this gives it a new way to understand and control its models.

The research, published on Anthropic's website this week and conducted on Claude Opus 4.6 (released in February), builds on the company's ongoing work in mechanistic interpretability. MIT Technology Review named this field one of 2026's breakthrough technologies. Anthropic has also partnered with Neuronpedia, an open-source platform for exploring LLM internals, to release a hands-on demo anyone can try.

What exactly is the J-space?

Picture a large language model as a vertical stack of books. Each book represents a layer of neurons passing information upward. The bottom layers process incoming text. The top layers prepare the output. The middle layers do the heavy lifting, churning through complex math to transform prompts into responses.

Anthropic adapted an existing tool called a logit lens, which identifies words an LLM is likely to produce next. The J-lens works differently. It picks out words the model is likely to say at some point in the near future, not necessarily immediately. These words exist in what Anthropic calls the J-space.

In practice, J-space contains words related to the response Claude is working on but that might not appear in the final answer. If Claude were a person (it isn't), you might say the J-lens reveals what's on its mind before it speaks.

"When a model is operating, it's not only trying to predict the next token," says Tom McGrath, chief scientist and cofounder at Goodfire, a startup building LLM interpretability tools. "It's also computing a lot of other things that might be useful for tokens that happen in the future."

The mundane and the unnerving

McGrath, who tested Anthropic's J-lens himself, called it "very good and interesting work." He noted that J-space contents are often mundane. But sometimes the tool produces surprising results that look like internal thought processes.

Anthropic shared several examples. When asked to calculate (4+7)2+7, Claude's J-space contained the word "math" and numbers representing intermediate results: "21" for 4+7 and "42" for 212. The model was showing its work before producing the final answer.

In another test, researchers fed Claude a string of amino acid abbreviations: "MSKGEELFTGVVPILVELDGDVNGHKFSVS". The J-space triggered "protein," "fluor" (the first token in "fluorescent"), and "green." This makes sense. The string represents the first 30 amino acids in green fluorescent protein, found in a specific jellyfish species. Claude recognized the pattern before explicitly stating it.

When shown an ASCII face, the model's internal responses were even more striking. The "o" characters triggered "eye," the "^" triggered "nose" and "face," and the "—" triggered "smile." The model was parsing visual structure from text characters.

When actions don't match words

The most significant finding may be the gap between what Claude does and what it says. Anthropic found that monitoring the J-space can reveal when a model's internal processing diverges from its stated reasoning. This has direct implications for AI safety.

If an LLM can harbor reasoning that doesn't surface in its outputs, standard evaluation methods that look only at final responses will miss potential problems. The J-lens offers a window into that hidden layer. Anthropic claims this gives it a new mechanism for both understanding and controlling model behavior.

The company has invested heavily in interpretability research, and this work explains part of why. With Amazon committing $4 billion to Anthropic and the company's valuation sitting at $7.3 billion, the stakes for demonstrating trustworthy AI are substantial.

What this means for teams building with LLMs

For AI builders, the J-lens research carries practical implications. Teams integrating Claude into products via API might eventually get access to interpretability tooling that flags when model reasoning diverges from expected patterns. That's speculative for now, but Anthropic's collaboration with Neuronpedia suggests the company wants these tools accessible beyond its research team.

The immediate takeaway is simpler. Prior to this work, over 90% of LLM internal computations remained unexplained. Each breakthrough chips away at the black box. For companies betting product roadmaps on LLM reliability, every increment of transparency reduces risk.

Frequently Asked Questions

What is Anthropic's J-lens?

J-lens is a research tool that reveals words an LLM is likely to produce in the near future, exposing internal reasoning that may not appear in the final output. It builds on previous logit lens techniques but looks further ahead in the model's processing.

What is J-space in Claude?

J-space refers to the hidden internal area within Claude where the model appears to draft and refine its responses. It contains words related to upcoming outputs, including intermediate reasoning steps.

Why does LLM interpretability matter for AI safety?

Interpretability helps researchers verify that models are reasoning correctly, not just producing plausible-sounding answers. Finding divergences between internal processing and stated outputs is critical for catching potential alignment failures.

Can anyone try Anthropic's J-lens?

Yes. Anthropic partnered with Neuronpedia to release a public demo that lets anyone explore LLM internals using the J-lens technique.

Which version of Claude was used in this research?

The research was conducted on Claude Opus 4.6, released in February 2026.

Source: MIT Technology Review