The Anthropic development team has made a breakthrough in understanding the internal mechanisms of large language models. During an analysis of Claude's operation, a structure was discovered that engineers did not embed in the architecture at the design stage. This internal mechanism, named "J-space," functions as a universal workspace that various components of the neural network access to exchange and process critically important information.
The study, published on July 6, represents a significant step forward in AI interpretability. J-space is a kind of virtual whiteboard inside the artificial intelligence where key data converges when solving problems, answering questions, or executing instructions. Thanks to this, different parts of the model can work synchronously with the same information.
How J-space works and why it is a sensation
Anthropic identified J-space using a special tool called "J-lens." With its help, researchers observed how information moves inside Claude during task execution. Strikingly, J-space emerged spontaneously during the training process—it was not directly programmed by the developers. This concept is surprisingly reminiscent of what cognitive neuroscience calls the "global workspace" in the human brain.
In humans, this system provides simultaneous access to important data for multiple thought processes. For example, when we hear a question, recall a fact, and formulate an answer, the brain brings all necessary information together. Claude demonstrates a similar pattern. Moreover, researchers were able to influence the model's responses by manually altering the contents of J-space. When the content of this virtual space changed, the AI's answers and its behavior during task execution changed dramatically.
Safety and the future of interpretability
The discovery of J-space has enormous implications for artificial intelligence safety. The ability to track activity in this "workspace" opens the way to identifying hidden motives and undesirable patterns in model behavior. In particular, monitoring J-space could help detect prompt-injection attacks in real time, where an attacker attempts to insert malicious instructions into a query.
Although most information processing in Claude still occurs outside J-space, the very existence of such a structure is a powerful tool for future research. Anthropic has already published the source code for the J-lens implementation and provided a demo version on the Neuronpedia platform, inviting the scientific community to verify the results.
Analyst's comment: This discovery could fundamentally change the approach to auditing and security of AI systems. The ability to "look into" the model's central decision-making process is not just a scientific curiosity, but a potential standard for verifying the reliability of any commercial neural networks. However, it is important to emphasize that, despite the anthropomorphic terminology, Anthropic explicitly states: Claude does not possess consciousness or subjective experience. The term "consciously accessible information" is borrowed from cognitive science and does not imply the presence of actual consciousness.