Anthropic AI: Claude's "J-space" Mimics Human Brains

New research shows AI models may work more like human brains than previously thought. Anthropic has identified a cluster of internal patterns within its Claude AI. They call this the "J-space." It behaves like a "global workspace," a term neuroscientists use for conscious awareness in humans. This J-space appears to process information differently from the rest of the language model. Researchers used a technique called the "Jacobian lens" to find these patterns. They found that the J-space holds a small collection of concepts, often just a few dozen at a time. However, this small collection carries a significant amount of the model's actual thinking. For example, when Claude silently picks a sport, the J-space shows the answer before the model says it. If this pattern is swapped, Claude reports a different sport. This suggests the model is genuinely using this internal channel. This discovery could change our understanding of how advanced AI models think and process information.