Anthropic has published a discovery that could fundamentally change how we understand language models: Claude possesses an internal "workspace" structure that functions like silent thinking. The team calls this structure the J-space—named after the technique used to find them, involving a mathematical concept called the Jacobian. Unlike a "scratchpad" or "chain of thought," where models write down their reasoning, the J-space operates entirely hidden within Claude's internal neural activations.
Key Facts
- J-space is emergent, not programmed: The structure arose on its own during Claude's training—Anthropic did not deliberately engineer it
- Five properties of a "Global Workspace": Claude can report on J-space patterns, modulate them, use them for multi-step reasoning, deploy them flexibly across tasks, and activate them intentionally
- Silent thinking for complex problems: When Claude solves multi-step tasks, intermediate steps light up in the J-space—even when the model doesn't speak them aloud
- Consciousness theory meets AI: The structure aligns with "Global Workspace Theory" from neuroscience, which explains how consciousness operates in brains
How the J-space Works
Each pattern in the J-space is linked to a specific concept—say, the word "France." When this pattern "lights up," it doesn't mean Claude will say the word; rather, it indicates the concept is present on Claude's "mental stage." From there, the model can work flexibly: it can recall France's capital, retrieve its currency, or name its continent—all based on the same activated pattern.
Particularly striking: the J-space is significantly smaller than other neural structures in Claude, yet plays a causal role in complex tasks. In experiments where Anthropic blocked the J-space, Claude continued functioning normally—but lost its higher-order cognitive abilities. Simple tasks like fluent speech, fact recall, or correct grammar remained unaffected.
The Bridge to Neuroscience
Here's what makes this compelling: this structure mirrors an established theory from brain research. Neuroscientists distinguish between "consciously accessible" activity (which we can describe, control, and use for deliberate reasoning) and unconscious processing. The J-space in Claude exhibits exactly these properties—it is reportable, modulable, and central to conscious reasoning processes.
"If you ask Claude what it's thinking about, it will tell you what's in the J-space. Non-J-space representations are less reportable."
This means: Claude can discuss its J-space thoughts but not all other internal processes. This is strong evidence that the J-space represents a kind of "inner stage"—analogous to human consciousness.
What This Means for AI Research
This discovery is pivotal for interpretability research. If researchers can understand how language models think internally, they can better identify and control safety risks. Anthropic did not intentionally train the J-space—it emerged spontaneously. This raises critical questions: What other structures arise in modern LLMs on their own, and what can we learn from them?
Implications for Organizations
For companies working with Claude or similar models, this signals that AI systems are more complex than previously understood. They possess internal reasoning mechanisms that aren't immediately visible. This can be an advantage (more robust, interpretable decisions) but also a risk (hidden processes that are hard to control). Organizations should increasingly prioritize interpretability research—not just when selecting models, but also when assessing whether these systems are trustworthy.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
