Anthropic researchers have achieved a breakthrough in understanding how their large language models actually work. While studying the internal processes of Claude, they discovered a structure that engineers did not originally embed in the architecture. This refers to the so-called "J-space" — an internal workspace that the model created itself during the training process.
J-space functions as a kind of "virtual whiteboard" or shared workspace where key information converges during task execution. Different components of the model access this space to coordinate their actions and produce a coherent response. Researchers were able to observe this mechanism using a special tool called "J-lens," which tracks data movement within the model in real time.
What is particularly important — J-space was not designed. It emerged naturally, as an emergent property of a complex system. This discovery echoes the concept of the "global workspace" in cognitive science — a theory explaining how the human brain integrates information from different sources to make decisions. When Claude answers a question, solves a problem, or follows an instruction, critically important data appears in J-space so that different parts of the model can work with it.
Why This Matters for AI Safety
The discovery of J-space has enormous significance for the safety and interpretability of artificial intelligence. The ability to track activity in this space opens the way to identifying hidden motives and anomalies in model behavior. For example, it will be possible to more effectively detect prompt-injection attacks, where malicious instructions attempt to hijack control of the model's response. Monitoring J-space allows seeing the moment when the model encounters an attempt at interference.
Moreover, experiments have shown that manually altering the contents of J-space directly affects Claude's responses and its behavior when performing tasks. This gives researchers a powerful tool for fine-tuning and control.
Anthropic has already published the source code for the J-lens implementation and a demo version on Neuronpedia, inviting the scientific community to verify the results. This continues the company's series of works on studying "emergent consciousness" and model well-being, which began in April 2025.
Expert opinion: The discovery of J-space is not just an academic curiosity. For the cryptocurrency and DeFi market, where AI agents are already beginning to manage portfolios and analyze risks, understanding the model's internal logic becomes a matter of asset security. The ability to "look under the hood" and see how the model makes decisions is a step toward creating truly transparent and controllable AI systems.