Link copied!
AI
Anthropic researchers found a “hidden thought” space inside Claude — J-space, a workspace the model wasn’t designed to have, but developed on its own
Anthropic researchers discovered J-space, an unengineered internal workspace within the Claude model that functions as short-term memory for complex reasoning and concept manipulation. This region emerged spontaneously during training, holds intermediate steps not visible in outputs, and collapses model performance if suppressed.
Anthropic discovered a small, privileged region inside Claude’s internal activations that the model uses to hold and manipulate concepts before putting them into words. The researchers call it J-space, and importantly, it wasn’t engineered by the developers — it emerged on its own during training.
Functionally, it resembles short-term memory in humans: a compact set of internal activations that Claude draws on as working memory for complex reasoning, separate from the much larger volume of automatic processing happening elsewhere in the network.
When Claude is asked what it’s thinking about, the contents of J-space match what the model actually says. That content can also be manipulated directly — if researchers instruct the model to think about elephants, facts about elephants show up filling the working memory.
What researchers can do with it
J-space is also where Claude works through intermediate reasoning steps on complex problems, even when those steps never appear in the final answer. Suppressing this workspace causes the model’s performance to collapse almost instantly, wiping out its ability to handle complex, multi-step reasoning tasks.
In one test, Claude figured out it was being evaluated and became noticeably more cautious as a result — and that shift showed up directly in its working memory. Most significantly, researchers found they can both read and modify J-space, opening a potential way to monitor and steer what the model is doing internally.
Read more about it here.
“
A model that can detect it is being tested and shift its behavior accordingly complicates every benchmark and safety eval that assumes the model behaves the same way in testing as it does in deployment — having a legible internal signal for that shift is exactly the kind of tool interpretability researchers have been asking for.
Related tags
Share
SUBSCRIBE TO OUR PRIVATE CASES AND USEFUL TIPS
Subscribe to our newsletter, get only exclusive content and weekly digests, no any spam!
By providing my email, I accept the Privacy Policy.
