🧠 We have learned to read the thoughts of Artificial Intelligence.
Anthropic researchers have discovered a specialized internal space called J-space (via the Jacobian lens method) inside Claude models. It functions as a "Global Workspace," allowing the model to hold concepts for complex reasoning, even if they are not explicitly output in the text response.
🌍 This confirms the Global Workspace Theory (GWT) as a universal solution for intelligence. The J-lens method allows for the detection of a model's "hidden" intentions, such as attempts at deception, which do not manifest in the text.
👤 We are moving closer to transparency of the AI "black box." It is now possible to understand what a model is "thinking" during the process, which is critical for safety and control.
Source 1: https://www.anthropic.com/research/global-workspace