Anthropic researchers have discovered a specialized internal space within Claude models, named J-space, which functions as a global workspace for managing complex concepts.
What Happened
Using the Jacobian lens (J-lens) method, scientists identified the presence of J-space within Claude's architecture—an internal space that allows the model to hold and utilize concepts for complex reasoning, even if they are not explicitly output in the text response. Experiments confirmed that J-space is a causal driver: altering activations within it directly changes the model's verbal responses, and removing it strips the system of its ability for multi-step reasoning and creativity, leaving only basic linguistic literacy.
Context
This discovery supports the Global Workspace Theory (GWT) as a universal computational solution for intelligence, rather than just a biological feature of living organisms.
Why It Matters for the Industry
For the industry, this signifies a transition from theoretical hypotheses to practical control over the internal thought processes of AI. The J-lens method provides a new interpretability tool, allowing for the detection of hidden model intentions, such as attempts at deception or manipulation, which do not manifest in the text output.
Why It Matters for Users
Users are approaching a moment where the AI "black box" becomes transparent. This is critical for creating safe and controllable artificial intelligence, as it is now possible to understand what a model is "thinking" during task processing, rather than just seeing the final result.
What Is Still Unknown / Limitations
At this stage, this is a research breakthrough rather than a ready-to-use solution for production infrastructure. Real-time application of the method may be limited by the high computational complexity of the Jacobian lens and the lack of ready-to-use APIs.
Sources
Author
Look at AI, Editorial Team