Anthropic's research team has achieved a breakthrough in the field of artificial intelligence interpretability. During an analysis of the internal processes of the Claude model, a structure was discovered that the developers did not originally embed in the architecture. This concerns the so-called "J-space" — an internal mechanism that functions as a global workspace for various components of the neural network.
What is J-space and how does it work?
Essentially, J-space is a virtual "board" where Claude pulls together key information while processing a request. When the model solves a problem or answers a question, important data appears in this space, becoming accessible to all parts of the system. Anthropic identified this structure using its own tool, J-lens, which allows real-time observation of data movement within the model.
Notably, J-space emerged spontaneously during the training process. The designers did not program it directly. This discovery from July 6th represents a significant step forward in understanding what exactly happens "under the hood" of large language models.
Analogy with the human brain and implications for safety
The concept of J-space strikingly echoes the neurobiological theory of the "global workspace" in the human brain. In humans, this system provides simultaneous access to important information for multiple thought processes. For example, when we hear a question, recall a fact, and formulate an answer, the brain brings all the data together in one point.
Experiments have shown that Claude can not only describe the contents of J-space upon request but also modify it if asked to do so. Moreover, manual intervention in J-space directly influenced the model's responses and behavior. This opens up enormous prospects for AI safety. Monitoring J-space activity will allow for the identification of hidden motives in the model's behavior, detection of prompt-injection attacks, and tracking of moments when the system begins to operate unreliably.
Prospects and limitations
Anthropic has already published the source code for the J-lens implementation and provided a demo version to the Neuronpedia platform, inviting the research 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 opinion: The discovery of J-space is not merely an academic curiosity. It is the first practical tool that could fundamentally change the approach to auditing and securing AI systems. If we can "read the model's thoughts" in real time, this would give us unprecedented control over its behavior and allow us to prevent failures and attacks before they occur. However, for now, most of the information processing in Claude still takes place outside of J-space, so it is too early to talk about complete transparency of the "black box."