Anthropic has developed the Jacobian Lens, a new method to analyze its Claude language model's internal working memory. This tool reveals how the AI processes concepts internally without explicitly mentioning them in its output. Researchers found that Claude's internal memory, called J-Space, can adjust its conclusions when concepts within it are modified. The findings have already led to a new training approach that reduces hallucinations and misleading outputs.
The Jacobian Lens shows that Claude has developed a distinct set of internal neural patterns that play a role different from the rest of its processing. These patterns, referred to as J-Space, are classified under Global Workspace Theory from consciousness research. According to Anthropic, J-Space holds word-like thoughts that the model does not vocalize. It can report on stored content, modify it, and use it for multi-step inferences. The company had previously explored reading out and steering internal states in self-awareness studies.
Anthropic emphasizes that J-Space emerged naturally during training, suggesting that mental working memory may be a general solution for learning systems under certain conditions, not unique to biological brains. The researchers do not conclude that AI experiences consciousness but note that their experiments touch on access consciousness, which requires a system to report on its internal states and process them flexibly. The findings have sparked discussion among neuroscientists, who see the results as a milestone in consciousness research.
Source: thedecoder
