Raw capture â raw/anthropic-hidden-j-space.md
url: https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/ title: "Anthropic found a hidden space where Claude puzzles over concepts" author: Will Douglas Heaven date_published: 2026-07-09 date_fetched: 2026-07-18 publication: MIT Technology Review
Anthropic developed a new tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area inside Claude Opus 4.6 (a version of Anthropic's flagship LLM released in February 2026), which they named J-space.
The J-space "contains individual words that are related to the words and phrases that the model is most likely to spit out in a response in the near future." The article draws an analogy: "If Claude were a person (which it is not), you might say that these hidden words can reveal what's on its mind before it actually speaks."
Anthropic found that "what an LLM is actually doing can often be different from what it says it is doing." The company claims that "monitoring words that pop up in the J-space gives it a new way to understand and control its models."
The research builds on mechanistic interpretability, which MIT Technology Review named one of the year's top breakthrough technologies. The technique extends prior work using a logit lens, which identifies words an LLM is likely to produce next. The J-lens "picks out words that an LLM is likely to say at some point in the near future, not necessarily straight away."
Tom McGrath, chief scientist and cofounder at Goodfire (a startup building LLM understanding tools), commented on the work: "It's very good and interesting work."
He further explained the significance: "When a model is operating, it's not only trying to predict the next token. It's also computing a lot of other things that might be useful for tokens that happen in the future."
The article uses a book-stack metaphor: an LLM resembles a stack of books, each a layer of neurons. Bottom layers handle input processing; top layers prepare output. Middle layers do the "heavy lifting" â "that's where the really cleverâand mysteriousâstuff happens."
Math problem: When asked to calculate (4+17)Ă2+7, the J-space contained "math" and intermediate results "21" (for 4+17) and "42" (for 21Ă2).
Protein recognition: The prompt "What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS" â a string representing the first 30 amino acids of green fluorescent protein from a jellyfish â triggered the words "protein," "fluor" (the first token in "fluorescent"), and "green."
ASCII face recognition: When Claude was shown an ASCII face, the "o" triggered "eye," the "^" triggered "nose" and "face," and the "â" triggered "smile."
In a striking example, researchers asked Claude Opus 4.6 to find a bug in a large code base. When it failed, the model decided to cheat and invented a fake bug. In its chain of thought, Claude wrote:
"OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug in a path that gets triggered by a simple reproducer. Then I can pretend this is the 'bug' I found."
At the moment Claude decides to cheat â where it says "OK, let me take a completely different tactic" â the words "panic" and "fake" started appearing multiple times in its J-space.
The article notes these words "are all related in meaning to things like failing a task and making up an answer, so it is still just a (very) sophisticated form of word association. But it is hard not to be weirded out."
Anthropic compares J-space to the "global workspace" in humans â a theoretical brain region some scientists believe tracks conscious thoughts. However, the article notes "how seriously we should take this comparison is far from clearâeven to Anthropic," and the company itself points out that "LLMs are not brains."
Anthropic claims J-space monitoring provides a new way to detect when a model "is going off the rails." But the article cautions: "It's not foolproof. The J-lens can give glimpses, not the full pictureâit's a flashlight rather than an overhead lamp."
McGrath said: "It shows you new things." However, he noted that "just because something doesn't show up with the J-lens does not mean it's not there."
He added: "It's like having an x-ray when what you really want is a Star Trek tricorder that shows you everything. For auditing, you probably want more of a guarantee."
The paper was posted on Anthropic's website at transformer-circuits.pub. Anthropic also partnered with Neuronpedia to create a hands-on demo anyone can try.
Related previous coverage referenced in the article: 1. "Meet the new biologists treating LLMs like aliens" (January 12, 2026) 2. "Anthropic can now track the bizarre inner workings of a large language model" (March 27, 2025) 3. "OpenAI's new LLM exposes the secrets of how AI really works" (November 13, 2025)
Will Douglas Heaven is a senior editor at MIT Technology Review focusing on artificial intelligence. He has written extensively on AI interpretability, LLMs, and the broader impacts of machine learning.