![\"FRANCE-TECHNOLOGY-AI\"](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/07\/1783873991_737_0x0.jpg\")
<\/p>\n<\/p>\n
Anthropic logo<\/p>\n
AFP via Getty Images<\/p>\n
All of this can be a little hard to follow.<\/p>\n
It seems that, earlier this month, Anthropic released a post about something called the \u201cJ-Space\u201d that shows us more about how an LLM thinks. It also posits something called a \u201cJ-lens\u201d that can help engineers or human users to \u201csee\u201d what\u2019s in this internal workspace.<\/p>\n
The J-space, theorists explain, is a part of the LLM where hidden signals light up, where the LLM shows what is thinking about without generating any explicit expression. It would be kind of like being able to read a human\u2019s mind, or, perhaps more relevantly, to read a human\u2019s body language or something unspoken.<\/p>\n
The result is that we are learning more about the interactions between humans and AI entities \u2013 but there\u2019s a lot of math and jargon and abstraction involved.<\/p>\n
Investigating the J-Space<\/p>\n
A simple way to describe this is that an LLM might, for example, get certain inputs and provide an answer, while thinking about something else. It might see a picture of a black cat, and return the word \u201cblack\u201d based on tokens, while displaying patterns linked to the word \u201ccat\u201d in the J-space. That\u2019s just one random instance, for the less superstitious among you.<\/p>\n
So why is it called the J-Space?<\/p>\n
I think this helps to explain what\u2019s going on: the J-Space is so-called after something called the \u201cJacobian matrix,\u201d where a function maps how each of many inputs relates to the steering of the LLM\u2019s cognitive process. You might have a certain value for input #1, cat, and another for input #2, dog, etc. The map shows part of the LLM\u2019s discriminatory cognitive process, how features make impressions and guide the LLM toward thinking about a particular subject or concept.<\/p>\n
Here\u2019s how Anthropic explains it in the July 6 announcement<\/a>:<\/p>\n \u201cEach J-space pattern is linked to a particular word. But when one of these patterns lights up, it doesn\u2019t mean the model is saying that word\u2014just that the word is on its mind. If you’ve heard of language models having a \u2018scratchpad\u2019 or \u2018chain of thought\u2019\u2014text they write to themselves while reasoning\u2014the J-space is something different. It operates silently, in the model\u2019s internal neural activations, allowing the model to think about a concept without writing it down. Notably, the J-space wasn\u2019t designed or programmed by us, but instead emerged on its own during Claude\u2019s training process.\u201d (italics theirs).<\/p>\n Using the Lens<\/p>\n So what is the J-Lens? Here, you make use of a view of that Jacobian matrix, to see the hidden patterns that represent what Claude (or some other model) is thinking without telling you.<\/p>\n \u201cApply the lens at each layer, and you get a ranked list of vocabulary words,\u201d writes Yash Takker at explain.ai<\/a>: \u201cERROR on buggy code, fake on prompt injections, intermediate math steps in order \u2014 often absent from chain-of-thought and output. Those patterns collectively form J-space, Claude’s internal global workspace. The lens is the instrument; J-space is the phenomenon.\u201d<\/p>\n Evaluating the Intelligence<\/p>\n I want to keep this part fairly narrow, and let you, the reader, do additional research based on your interest level. I want to share two things, one of which follows from the other, to look at what this research is saying about LLM thought processes.<\/p>\n