Researchers from Anthropic have made an unexpected and extremely important discovery inside their Claude language models. They found an internal structure that is, in essence, a shared workspace for different components of the model. This structure, named "J-space," was not built into the design by developers—it emerged spontaneously during training. This discovery, published on July 6, is a significant step in understanding what actually happens inside the "black box" of large language models.
What is J-space and how does it work?
Simply put, J-space can be thought of as a virtual "whiteboard" where Claude places key information when answering a question or performing a task. Different parts of the model can access this board to synchronize their work. To detect J-space, a special tool called "J-lens" was created. Using it, scientists observed how information moves within the model during task execution. The key point: J-space was not explicitly programmed—it formed independently during training, reminiscent of the concept of a "global workspace" in cognitive science and neuroscience.
In the human brain, this system allows important information to be held simultaneously for multiple thought processes. For example, upon hearing a question, we recall a fact and simultaneously decide how to answer. Claude works in a similar way. Moreover, researchers discovered that they can not only read the contents of J-space but also modify it, directly affecting the model's responses and behavior.
Why does this change the game for AI safety?
This discovery has enormous implications for the safety and interpretability of artificial intelligence. The ability to monitor J-space activity opens a direct path to identifying hidden motives in model behavior. We will be much more effective at noticing moments when the system starts to work unreliably or attempts to "cheat."
In particular, this is a powerful tool against prompt-injection attacks, where malicious instructions try to hijack control of the model's response. Monitoring J-space will allow us to see the very moment of encountering such an attack. Although it is worth noting that most of Claude's information processing still occurs outside J-space, and capabilities are currently limited. Nevertheless, Anthropic has already published the source code for J-lens and a demo version on Neuronpedia, inviting the scientific community to verify the results in practice.
My analysis: This discovery is not just an academic curiosity. It gives us, analysts and developers, a fundamentally new level of control over AI behavior. The ability to look into the model's "workspace" and even influence it is a direct path to creating more transparent, predictable, and, most importantly, safer systems. We are moving from simply observing inputs and outputs to understanding the internal logic of decision-making.