Editorial illustration for Anthropic's 'Logit Lens' Reveals How Claude Puzzles Over Words
Anthropic's Logit Lens Reveals How Claude Thinks
Anthropic says it has found a way to watch Claude think before it speaks, literally, at the level of individual words. Researchers at the company built a tool called the Jacobian lens, or J-lens, and pointed it at Claude Opus 4.6, the flagship model the company released in February. What emerged was a hidden layer inside the network they're calling the J-space, a kind of staging area where words related to whatever the model is about to output start clustering before any text gets generated.
The findings, laid out in a paper posted on Anthropic's website this week, suggest a gap between what a model claims it's doing and what it's actually doing internally. That gap matters for anyone trying to figure out whether an AI system's stated reasoning can be trusted. Anthropic has partnered with Neuronpedia, an open-source platform for exploring the innards of language models, to release a demo so outsiders can poke at the J-space themselves rather than take the company's word for it.
The work lands amid a broader push to crack open these systems, and it's drawing attention from researchers outside Anthropic who build similar tools.
The AI firm Anthropic has developed a technique that has given it the clearest glimpse yet at what’s really going on inside large language models as they answer questions or carry out tasks.
Why this matters
For anyone building on top of Claude or trying to reason about what these models actually do, this is the closest thing to an X-ray we've gotten. Anthropic isn't just claiming interpretability progress in the abstract, it built a specific tool, the J-lens, and pointed it at a specific model, Claude Opus 4.6, and found a specific structure, the J-space, that sits between raw computation and the words that come out. That's a different order of evidence than the usual "trust us, we're being careful" messaging from AI labs.
For researchers, the practical value is in the method: if you can watch a model "puzzle" over concepts mid-computation rather than just judging it by outputs, you get a much sharper tool for catching deception, bias, or reasoning failures before they hit a user. For founders and developers shipping products on Claude, the honest takeaway is that the model's internals are stranger and less predictable than the clean API responses suggest. Worth watching whether Anthropic opens up J-lens access, or keeps it in-house as a safety edge.
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives OpenAI Releases GPT-Live Voice Models That Delegate Reasoning to GPT-5.5
From Archives Bezos-backed startup raises USD 320 million to build AGI from gaming data
From Archives Anthropic Launches Claude Cowork AI Agent for Mobile and Web
From Archives Anthropic Moves Its Claude Agent to Phones as AI Rivals Target Mobile
Common Questions Answered
What is the Jacobian lens (J-lens) that Anthropic developed?
The Jacobian lens is a tool created by Anthropic researchers that allows them to observe the internal thinking process of Claude before it generates output. By pointing the J-lens at Claude Opus 4.6, researchers were able to identify a hidden layer in the network called J-space, where words related to the model's upcoming response begin clustering before any text is actually generated.
How does the J-space function as a staging area in Claude's processing?
The J-space operates as an intermediate layer between the raw computational processes and the final text output of Claude Opus 4.6. In this hidden space, words and concepts related to what the model is about to generate start grouping together before the model produces any actual text, providing researchers with insight into the model's decision-making process.
Why is Anthropic's J-lens technique significant for AI interpretability?
The J-lens represents a major breakthrough in understanding how large language models actually work, providing the clearest glimpse yet into the internal mechanisms of AI systems like Claude. Rather than making abstract claims about interpretability, Anthropic has built a specific, tangible tool that reveals concrete structures within the model, offering evidence-based insights into model behavior that goes beyond theoretical explanations.
What model did Anthropic use to test the J-lens discovery?
Anthropic tested the J-lens on Claude Opus 4.6, which is the company's flagship model released in February. This specific model served as the testing ground where researchers discovered the J-space hidden layer and demonstrated how the tool could reveal the model's internal thinking processes.
Further Reading
- Anthropic's J-Lens Reads Claude's Silent Thoughts Before It Speaks - AlphaSignal
- Anthropic says it can read Claude's 'thoughts,' as detailed in new research paper - Tom's Hardware
- A global workspace in language models - Anthropic
- Tracing the thoughts of a large language model - Anthropic
- On the Biology of a Large Language Model - Transformer Circuits
