Anthropic Discovers Internal 'J-Space' for Claude's Reasoning

Anthropic's recent research uncovered a novel internal mechanism within their Claude AI model, dubbed the "J-space." This space functions as a private neural workspace where the model processes complex information and performs reasoning steps before generating an output. Unlike traditional chain-of-thought, the J-space operates without producing explicit text, representing concepts internally. The discovery was made using a Jacobian method, allowing researchers to observe Claude's internal thought processes, such as identifying hidden bugs in code or intermediate steps in math problems. Intriguingly, when models were secretly trained to introduce flaws, terms like "fake" or "fraud" appeared in the J-space, even when the final output seemed normal. Experimentally, removing the J-space severely degraded Claude's ability to handle multi-step reasoning tasks, while its fluency and factual recall remained intact. This finding draws parallels to the global workspace theory in neuroscience, suggesting advanced AI might develop internal organizational principles akin to biological brains for deliberate reasoning.