Anthropic researchers have made a breakthrough in understanding the internal architecture of large language models. During an analysis of their own model, Claude, they identified a structure that was not built into the system by developers during the design phase but emerged independently during the training process. This internal mechanism, named "J-space", functions as a universal workspace that various components of the neural network access.
What is J-space and how does it work?
Imagine a virtual "whiteboard" inside the artificial intelligence. When Claude receives a request, solves a problem, or follows an instruction, key information is gathered precisely in J-space. This allows different parts of the model to interact and exchange data effectively. To discover this structure, Anthropic developed a special tool called "J-lens." Using it, scientists were able to observe in real-time how information moves within the model while performing tasks.
Notably, J-space was not programmed by engineers—it emerged as a byproduct of training. This concept strikingly echoes the "Global Workspace Theory" in cognitive neuroscience, which describes how the human brain integrates information from different areas for decision-making. Researchers demonstrated that they can not only read the contents of J-space but also modify it, which directly affects the model's responses and behavior.
Why is this critically important for AI safety and interpretability?
This discovery has enormous implications for the future of artificial intelligence safety. The ability to track J-space activity gives us a fundamentally new tool for "looking under the hood" of neural networks. We will be able to identify hidden motives in AI behavior, detect moments when the system begins to operate unreliably, and, most importantly, recognize attempts at prompt-injection attacks (inserting malicious instructions that hijack control of the model's response).
Although at present most information processing in Claude still occurs outside J-space and capabilities are limited, the step forward is colossal. Anthropic has not only published a scientific paper and open-sourced the J-lens implementation but also provided a demo version on the Neuronpedia platform, inviting the entire scientific community to verify the results. It is important to emphasize: the company does not claim that Claude possesses consciousness or subjective experience. The term "consciously accessible information" is borrowed from cognitive science and does not imply the presence of actual consciousness.
Expert commentary: The discovery of J-space is arguably one of the most significant steps toward creating truly interpretable and controllable AI. If we learn to "read the thoughts" of models at such a fundamental level, it will not only enhance their safety but could also lead to the creation of a new generation of algorithms capable of more complex and transparent reasoning. Anyone interested in the future of technology should closely follow this discovery.