Anthropic has discovered a hidden workspace, referred to as "J-space", within its language model Claude, which appears to be used for silent reasoning. This finding has been made possible through a new tool called "J-lens", and it may provide insight into the model's thought process, including thoughts that are not explicitly stated. The discovery has been characterized as revealing a "consciousness-like workspace" and may have implications for understanding the boundary between thinking and speaking in large language models.