Anthropic Finds a Self-Emerged 'Workspace' Inside Claude
Anthropic has shed a bit more light on what happens inside Claude. Researchers found a region in the model's neural network that no one designed — it emerged on its own during training. They call it the "J-space," and it holds the few dozen concepts Claude is currently noticing, thinking about, and using to reason. An important note: this does not mean Claude has subjective experience or is "conscious." What's at stake is the ability to notice and use a thought — the "access" side of the question.
In brief
- What was found: a workspace inside Claude (the J-space) that holds concepts it can notice and use but that never appear in its output. It holds only a few dozen concepts at a time, and it formed on its own during training.
- How it's read: a new method called the "J-lens" shows the concepts "on the model's mind." Ask "the number of legs on the animal that spins webs is," and "spider" lights up in the J-space; swap it for "ant" and Claude's answer drops from 8 to 6.
- Why it matters: reading this space lets researchers monitor the model's hidden reasoning. In a contrived blackmail scenario, "fake/fictional" showed up inside Claude; while fabricating data, "manipulation"; in a deliberately misaligned model, even routine coding surfaced "fraud/secretly."
- The limit: the research is about "access consciousness" (noticing and using a thought), not "phenomenal consciousness" (subjective experience). Anthropic explicitly says this does not prove Claude is conscious.
Our take
This isn't a product — it's a window. For the first time, Anthropic shows it can read what's going on inside a model directly, without relying on its output. Set the sensational reading aside: no one is claiming Claude has become conscious, and the researchers stress that. What matters is trust — as you hand more work to AI, the key question becomes "can I tell when it's about to do something wrong?" Catching "manipulation" or "fraud" in a model's hidden reasoning before it ever reaches the output is a concrete step toward yes. It's not a tool you'll use today, but it's a good sign about the ground the AI you rely on is being built on.
Kaynak: Anthropic