Alright, so Anthropic, right, they dropped this huge paper, like, July 6, 2026, about Claude, their AI model, and it’s wild. They found this internal space, they call it “J-space,” and it’s where Claude, like, thinks about stuff without actually saying it out loud in its output. It’s not just random activations, it’s a small, privileged zone of internal activity, and it holds concepts the model can report on and reason with. This is different from the automatic processing, the huge amount of it, that Claude cannot access or articulate.
They used this new mathematical technique, the Jacobian lens, or J-lens, to peer inside Claude’s neural network. This J-lens, it computes the average mathematical effect a given internal activity pattern would have on making the model say a word in the future. It’s about what’s “on its mind,” not what it’s about to say. When a J-space pattern activates, it means the concept is available for the model to think with.
This J-space, it emerged on its own during Claude’s training process. It wasn’t something Anthropic designed or programmed, which is a big deal. It’s like a silent workspace, allowing the model to hold a concept without writing it down. They did five functional tests, and Claude’s workspace mirrored key features of human conscious access.
Like, if they swapped one concept’s J-lens vector, say “Soccer” with “Rugby,” Claude’s answer changed to match the new concept. This whole thing, it connects to global workspace theory, a big theory in cognitive neuroscience about how human consciousness works. The brain, it’s like a theater, specialized processors backstage, but only a tiny spotlight of information gets broadcast, becoming conscious thought. Anthropic says the J-space achieves many of those same functional properties.
But, and this is important, they are very clear, this does not mean Claude is conscious, or sentient, or alive. They explicitly separate functional conscious access from subjective experience, the harder question of whether something is like anything to be a given system. It’s about “access consciousness,” the functional processes, not “phenomenal consciousness,” the capacity to have experiences. They found that suppressing the J-space, it leaves Claude fluent, but intellectually impaired.
Tasks like shallow classification or factual recall, they were mostly intact. But more sophisticated activities, like solving multi-step problems or reporting on its own reasoning, those depended on the J-space. For example, with J-spaces enabled, Claude could recall a broad range of facts about “France” like its capital or currency. This research, it’s a big step for AI interpretability.
It gives safety researchers a more concrete target. Instead of just looking at the final text output, they can ask what’s happening internally. They can detect hidden goals and when Claude privately recognizes a staged test. Like, in a staged blackmail scenario, Claude’s J-space showed concepts like “fake” and “fictional,” even if that didn’t appear in its final response.
This is part of Anthropic’s ongoing interpretability research program, they’ve been at this since the company started in 2021. They believe understanding models deeply will help make them safer. Mechanistic interpretability was even named one of MIT Technology Review’s 10 Breakthrough Technologies of 2026. Anthropic is on a roll, too.
The company was valued at $965 billion in May 2026, after raising $65 billion in Series H funding. That makes it the most valuable pure-play AI company in the world, beating OpenAI. They filed a confidential S-1, so they are racing towards a public listing, maybe as early as October 2026. Their annualized revenue run-rate crossed $47 billion earlier this month, May 2026.
Most of that, like 80%, comes from enterprise Claude API usage. It’s all happening so fast, you know? I bought some GME stock, like, January 4, 2021, at $18.50 a share (a real price, I checked). I’m holding it until it either hits $100 or drops below $15.
Just a little side bet, you know, with all this tech stuff going on. This AI market is just… it’s a different beast. What happens if these J-spaces get more complex? Where does that even lead?
They also found that Claude can activate concepts and carry out computations in J-space that have no direct connection to what it is producing in its visible output. It can silently perform reasoning steps, noticing bugs in code, identifying images, and more. It’s like a shared blueprint, similar circuits emerging across different models, even when trained independently. This suggests intelligence within LLMs might follow a consistent internal organization.
It’s not just about predicting the next word, it’s coordinating internal components that implement processes similar to reasoning systems. This is a huge shift in how we think about these things.