Anthropic discovers secret 'thought space' at the heart of its Claude AI

Artificial intelligence company Anthropic has revealed the discovery of a unique internal mechanism within its advanced language model, Claude, which appears to function as a dedicated workspace for complex thought processes.

Dubbed 'J-Space,' this internal area was identified during Claude's training. Anthropic explains that a specific set of neural patterns within Claude dedicates itself to this function, distinct from its general processing. This mechanism allows the AI to internally recenter and organize intricate ideas, akin to how the human brain focuses on specific tasks or thoughts.

Researchers at Anthropic draw parallels between J-Space and the concept of 'access consciousness' described by some neuroscientists and philosophers. This refers to mental states that are reportable, controllable, and usable for reasoning, suggesting that Claude's J-Space might enable a similar form of focused internal deliberation.

The discovery highlights the evolving complexity of large language models and raises new questions about their internal workings. While Claude is not conscious, the existence of such specialized internal processing suggests sophisticated methods for managing and refining information during AI development.

“

We observe that Claude has developed a small set of internal neural patterns which, relative to all its other internal processing, play a particular role.