What at the center of Claude’s mind?

What’s at the center of Claude’s mind?

What’s at the center of Claude’s mind?

Out of everything happening in your brain right now, only a tiny fraction is consciously accessible — thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside our AI model, Claude.

Our experiments were inspired by a leading theory in neuroscience: the global workspace theory. It holds that a thought becomes consciously accessible when it enters a shared “workspace” that’s broadcast across the brain. We found a set of representations in Claude’s neural activity that play a similar role.

Read more about the research here: https://www.anthropic.com/research/gl…

Disabling JSpace does not completely break the model, but it significantly impairs the AI’s ability to perform tasks that require complex reasoning (3:17 – 3:41).

When researchers switched off JSpace while leaving the rest of the network untouched, they observed the following:

Continued basic functionality: The model could still answer simple questions and generate fluent text, such as writing in a specific language (e.g., Spanish) when prompted (3:24 – 3:31).
Impaired reasoning: The model struggled with tasks that required multi-step or logical reasoning, such as identifying an author who wrote in a specific language from a prompt (3:31 – 3:41).

This suggests that JSpace acts as a specialized mental workspace necessary for higher-level cognitive functions rather than basic linguistic output.

At the center of Claude’s mind is a silent, internal workspace called J-space. Discovered by Anthropic researchers using a mathematical technique called the Jacobian lens (J-lens), J-space acts as a mental notepad where the model actively holds and edits concepts before expressing them as text. [1, 2, 3, 4, 5]

What J-space Does

Silent Thinking: Like the Global Workspace Theory in human neuroscience, J-space creates a shared buffer where pieces of information are made available to many different processing systems in parallel. [1, 2]
Concept Processing: Claude does not “think” in English sentences; it computes using floating-point representations of concepts. J-space isolates the representations of specific topics (e.g., words like “spider” or “rugby”) that Claude is currently entertaining. [1, 2, 3]
Beyond Auto-complete: Anthropic’s experiments proved J-space is more than a passive scoreboard. By manually deleting an active J-space concept and injecting a new one, researchers could directly change Claude’s multi-step responses. [, 2]
Safety & Alignment: Anthropic uses J-lens to eavesdrop on Claude’s unspoken thoughts to monitor for deception, hidden goals, or inconsistencies, paving the way for more controllable AI. [1, 2]