Claude's Internal Architecture: Anthropic Discovers a Spontaneously Emergent "Workspace" J-space
Researchers from Anthropic have made a breakthrough in understanding the internal workings of their Claude language models. During their analysis, they discovered a structure that developers did not originally embed in the architecture — the so-called "J-space." This mechanism functions as a common workspace accessed by various components of the model.
The discovery, dated July 6, marks a significant step forward in studying the "black box" of large language models. J-space is an internal environment where Claude collects and transmits critically important data across the network. The simplest way to envision it is as a virtual whiteboard inside the artificial intelligence.
How J-space Works
When Claude answers a question, solves a problem, or follows an instruction, key information appears in J-space so that different parts of the model can interact with it. To detect this structure, Anthropic specialists developed a special tool called "J-lens." Using it, researchers observed how information moves within the model during task execution.
Notably, J-space emerged spontaneously during the training process — the designers did not directly embed it. This concept largely echoes what neuroscience calls the "global workspace." In humans, this system provides simultaneous access to important information for multiple thought processes. For example, when we hear a question, recall a relevant fact, and decide how to respond, the brain consolidates all necessary data in one place.
Experiments showed that Claude can describe the contents of J-space upon request and even modify it if asked. Moreover, when researchers manually altered J-space, Claude's answers and task behavior changed accordingly.
Implications for AI Safety and Interpretability
This discovery has enormous implications for artificial intelligence safety. If scientists can track J-space activity, they will be able to identify hidden motives in AI model behavior and more effectively detect moments when the system begins to operate unreliably.
This is especially important for detecting prompt-injection attacks — inserting malicious instructions into a query to hijack the model's response. Monitoring J-space allows seeing the moment when the model encounters such attempts. Even partial access to this "conscious" processing layer opens important prospects for further research.
It is worth noting that capabilities are still limited: most of Claude's information processing still occurs outside J-space. Nevertheless, Anthropic has already published the source code for the J-lens implementation and released a demo version on Neuronpedia, inviting the research community to verify the results in practice. This continues the company's series of works on studying model well-being and step by step advances our understanding of what actually happens inside AI systems.
Expert opinion: The discovery of J-space is not just an academic curiosity but a potential key to creating more transparent and controllable AI systems. For the crypto industry, where smart contracts and algorithmic solutions require maximum predictability, such interpretability mechanisms could become the foundation for a new generation of trusted AI agents.