July 6, 2026
AI caught thinking out loud-ish
A global workspace in language models
AI may have a hidden mental notepad, and commenters are already fighting over it
TLDR: Anthropic says its AI may have a hidden internal workspace for silent reasoning, a bit like keeping ideas in mind before speaking. Commenters swung between excitement about peeking into that process, skepticism about what it really means, and big hopes that it could make AI easier to understand or even smarter.
Anthropic just dropped a brainy claim with big sci-fi energy: its Claude chatbot appears to have a tiny hidden "mental workspace" where ideas can sit before they’re spoken out loud. In plain English, the company says the model seems to have an internal place for concepts it can report on, control, and use for step-by-step thinking—without printing those steps into the chat. That instantly sent the community into full "wait, is this inner monologue or just fancy math?" mode.
And honestly, the comments are the real spectacle. One camp was already racing to product ideas, dreaming of dashboards that show what an AI was "thinking about" during customer support chats, or even flags for sketchy thoughts before they become hallucinations. That’s the practical, slightly spooky crowd: give us the receipts from the robot’s brain. Another set of commenters hit the brakes hard, asking the most grounding question of all: where is this thing actually living inside the model? If this hidden workspace is so important, people want to know whether it can be expanded, controlled, or turned into a performance boost.
Then there was the classic internet flexing: one commenter coolly noted that, without saying so, the researchers seem to be using a certain math-heavy approach; another asked whether this is basically a scaled-up version of older work. Translation: while some readers heard "AI has thoughts," the veterans heard "show your homework." The vibe is a delicious mix of awe, skepticism, and opportunism—and everyone seems to agree on one thing: if this hidden notepad is real, it could become a huge deal fast.
Key Points
- •The article presents a new paper claiming that language models such as Claude develop a small set of internal neural patterns called J-space.
- •J-space was identified using a Jacobian-based method and is described as a silent internal representation system distinct from written scratchpads or chain-of-thought text.
- •The article says Claude can report and intentionally modulate J-space representations, while non-J-space representations are less reportable and harder to control.
- •In multi-step tasks, intermediate reasoning steps reportedly appear in J-space and causally contribute to Claude's performance.
- •The authors compare J-space to neuroscience's global workspace theory and report that disabling J-space preserved normal interaction but impaired higher-order cognitive functions.