Anthropic Uncovers 'JSpace': A Hidden Global Workspace for Thought in Claude's Neural Network

Anthropic has published a new paper, “A Global Workspace and Language Models,” detailing the discovery of a bizarre, spontaneously emerged neural region within its Claude model, dubbed the ‘JSpace.’ This internal ‘mental whiteboard’ appears to be where Claude quietly formulates and reasons with thoughts before generating external outputs, sparking philosophical debate given its eerie resemblance to human global workspace theories of consciousness.

Researchers utilized a novel tool, the Jacobian lens (J-lens), a grid of partial derivatives, to view and manipulate tokens within the JSpace. Experiments demonstrated its critical role in reasoning: swapping internal concepts (e.g., ‘spider’ for ‘ant’ in a reasoning task) directly altered Claude’s logical conclusions without changing the prompt or output. Deleting the JSpace entirely resulted in Claude retaining fluent, confident English but losing its ability to reason. Interestingly, fundamental skills like grammar, fluency, and basic fact recall operated automatically, external to the JSpace. While some interpret this as a step towards Artificial General Intelligence, Anthropic explicitly states the findings do not confirm consciousness, though the emergence of such an internal scratch pad through training alone is considered a significant, unengineered development in AI architecture.