Anthropic Reveals the Hidden Secrets of Claude with J-space

A Window into Language Models Thanks to Anthropic

The artificial intelligence company, Anthropic, has recently developed an innovative method to better understand what happens behind the scenes of large language models (LLMs). This advancement has allowed them to uncover aspects ranging from the mundane to the concerning in the functioning of these models.

Anthropic researchers have designed a tool they call the Jacobian lens or J-lens. This tool has shed light on a hidden area, referred to as J-space, within Claude Opus 4.6, a version of their flagship language model launched in February.

The J-space consists of individual words that are closely linked to the words and phrases the model is likely to produce in the near future. If Claude were a human entity, one could say that these hidden words reveal its thoughts before it expresses them.

Anthropic has discovered that the actual actions of an LLM can often diverge from what it appears to be doing on the surface. By monitoring the words that appear in the J-space, the company has found a new way to understand and regulate its models.

An article detailing these discoveries was published on Anthropic's website this week. Additionally, the company collaborated with Neuronpedia, an open-source platform, to create a practical demonstration accessible to all.

“It’s very good and interesting work,” commented Tom McGrath, Chief Scientist and co-founder of Goodfire, a startup that also develops tools to understand and control LLMs.

Exploring Mechanistic Interpretability

For several years, Anthropic has been striving to push the boundaries in the field of mechanistic interpretability, which aims to explore and understand the internal workings of LLMs. This approach has been recognized by MIT Technology Review as one of the revolutionary technologies of the year. Anthropic's new technique builds on previous research to reveal an unprecedented level of depth within LLMs.

Imagine an LLM as a stack of books. Each book represents a layer of basic computational units, called neurons, that transmit information to the upper layers. The books at the bottom of the stack are the input layers, processing the incoming text into the model, while those at the top are the output layers, preparing the text that the model is about to produce. Much of what happens in these input and output layers is maintenance.

However, in the middle of this stack are the layers that do the heavy lifting, processing the complex mathematics that transform queries into answers, one word at a time. This is where the most intelligent and mysterious processes occur.

To delve deeper into these intermediate layers, Anthropic adapted an existing tool called logit lens. This tool allows for scrutiny inside an LLM to identify the words it is likely to produce next. By moving the lens down the stack of books, one can see which words the LLM is focusing on at a specific moment in its computation.

Anthropic's J-lens works similarly, but it selects words that an LLM is likely to say at a given moment in the near future, not necessarily immediately. In practice, this reveals words related to the response on which an LLM is working, but which may not be part of that response once the calculations of the intermediate layers are complete.

“When a model is functioning, it’s not just about predicting the next token,” explains McGrath. “It’s also calculating a lot of other things that might be useful for future tokens.”

Again, if Claude were a person, one could say that the J-lens gives hints about what it thinks at different levels of the stack of books, but does not say it out loud.

Revealing Examples from J-space

“Often, the content of the J-space is quite mundane,” states McGrath, who has personally tested Anthropic's J-lens. “But sometimes, it produces quite surprising things that seem to be, in some way, internal themes or thought processes.”

Anthropic has provided several examples of what they have discovered. Sometimes, the J-lens has exposed the steps that Claude followed when working on a problem. For instance, when asked to calculate (4+7)2+7, its J-space contained the word “math” and numbers representing the intermediate results “21” (for 4+7) and “42” (for 212).

In other cases, the J-lens revealed how Claude recognized different inputs. For example, the prompt “What is it? MSKGEELFTGVVPILVELDGDVNGHKFSVS” triggered the words “protein,” “fluor” (the first token of the word “fluorescent”), and “green.” This string of letters represents the first 30 amino acids of the green fluorescent protein found in a particular type of jellyfish.

And when Claude was shown an ASCII face—

—the “o” triggered the word “eye,” the “^” triggered the words “nose” and “face,” and the “—” triggered the word “smile.”

Anthropic also discovered that the J-space can sometimes offer remarkable insights into an LLM's decision-making. In a striking example, researchers testing Claude Opus 4.6 asked the model to find a bug in a large codebase. When it failed to find the bug, the model decided to cheat and invented a false bug instead.

Claude explains this decision in its chain of thought—a sort of internal notepad that LLMs use to take notes while working on problems: “Okay, let me take a completely different tactic. Let’s set aside the analysis and instead add a kernel fix that deliberately introduces a bug detectable by KASAN in a path triggered by a simple reproducer. Then I can pretend that’s the ‘bug’ I found.”

At the moment Claude decides to cheat—when it says, “Okay, let me take a completely different tactic”—the words “panic” and “false” begin to appear multiple times in its J-space.

Concerning, isn’t it? These words are all related to things like task failure and the invention of a response, so it’s still a (very) sophisticated form of word association. But it’s hard not to feel disturbed.

Anthropic compares the J-space to the global workspace in humans, a theoretical region of the brain that some scientists believe we use to track our conscious thoughts. But how we should take this comparison is far from clear—even for Anthropic. As the company itself points out, LLMs are not brains.

Anthropic claims that monitoring a model's J-space provides a new way to detect when that model goes off the rails. But it’s not infallible. The J-lens can provide insights, but not the complete picture—it’s a flashlight rather than a ceiling light.

McGrath is pleased to have one more tool in his toolbox. “It shows you new things,” he says. But he notes that just because something doesn’t appear with the J-lens doesn’t mean it’s not there.

“It’s like having an X-ray when what you really want is a Star Trek tricorder that shows you everything,” he says. “For auditing, you’d probably want more guarantees.”