Anthropic Discovers Hidden 'Thinking' Workspace in Claude AI Model

Anthropic, the artificial intelligence company behind the Claude language model, has announced the discovery of a hidden internal workspace within the model that processes concepts without immediately revealing them in generated responses. Termed the 'J-space', this collection of neural patterns represents what the researchers describe as internal thoughts that do not necessarily appear in the model's outputs.

The findings are detailed in a research paper titled 'Verbalizable representations form a global workspace in language models', released by Anthropic on July 6. The company has also released an open-source implementation of the core methods and an interactive demo in partnership with Neuronpedia.

According to Anthropic, the existence of J-space emerged spontaneously in Claude and was not explicitly programmed. The researchers explain: 'More broadly, these findings have changed our understanding of how Claude’s mind works, revealing a privileged mental workspace that can be used for deliberate reasoning, operating amidst a sea of more automatic, inflexible processing.'

The J-space is named after the Jacobian mathematical concept. It consists of internal neural patterns, each linked to a specific word. When a pattern activates, it indicates that the model is 'thinking' about that word. Importantly, J-space differs from 'chain of thought' reasoning because it operates silently within the model's internal activations, without needing to write down the thought. Anthropic claims that Claude can report what it is thinking in J-space when asked, and these activations can be modulated upon request.

Using J-space appears to enhance Claude's performance. In experiments where researchers prevented access to J-space, the model continued normal interaction but lost higher-order cognitive functions. However, the model does not rely on J-space for basic tasks like fluent speech or recalling simple facts.

The experiments were inspired by the global workspace theory in neuroscience, which views the brain as a collection of specialist systems operating in parallel. Researchers used a technique called the Jacobian lens (J-lens) to identify internal activity patterns that make Claude more likely to produce a specific word later. By applying this technique across different layers of the model, they observed how these silent words evolve as the model processes information.

Despite the intriguing findings, Anthropic has denied that this indicates consciousness in Claude. The company emphasizes that these are mathematical representations, not conscious thoughts. However, the research could advance efforts to make large language models more interpretable and controllable.