AI's Internal "Draft": Anthropic Reveals Claude's Unexpected Architecture
An event has occurred in the world of artificial intelligence that is changing our understanding of how large language models work. Researchers at Anthropic, the company behind the development of Claude, have discovered a surprising structure in its architecture that engineers did not originally design. This is the so-called "J-space" — an internal workspace that the model created itself during the learning process.
What is J-space and how does it work?
J-space is nothing other than a virtual "board" where Claude gathers key information when performing any task. Whether it's answering a complex question, writing code, or analyzing text — important data is first collected in this unified space and then distributed to various components of the model. To detect it, the Anthropic team developed a special tool — "J-lens" — which allows real-time observation of information movement within the neural network.
The most astonishing thing is that this structure emerged spontaneously. It was not programmed but evolved during the learning process, resembling what cognitive science calls the "global workspace" in humans. When we hear a question, recall a fact, and formulate an answer, our brain brings all these processes together. Claude does the same, but in digital form.
Why does this change the rules of the game in AI safety?
For the crypto industry and the high-tech sector, this discovery has enormous significance. Access to the model's "consciousness" — albeit in quotation marks — opens new horizons for security. If we can track J-space activity, we gain the ability to identify hidden motives in AI behavior, detect vulnerabilities, and, critically, recognize prompt-injection attacks, where an attacker tries to hijack the response.
Experiments have shown that manually changing the contents of J-space alters both Claude's final response and its behavior. This means we gain a direct lever to control the model's "thought process," rather than just adjusting its output.
However, an important caveat must be made: most of Claude's computations still occur outside J-space. It is merely the visible part of the iceberg. Nevertheless, Anthropic has already open-sourced the J-lens code and posted a demo version on Neuronpedia, inviting the entire scientific community to verify the results in practice.
Expert opinion: The discovery of J-space is not just an academic curiosity. It is a step toward creating "transparent" AI, whose decisions can not only be verified but also predicted. For the DeFi market and automated management systems, this means a potential breakthrough in security: if we can "see the thoughts" of an AI agent, we can prevent catastrophic failures before they occur. Anthropic emphasizes that it is not claiming Claude has consciousness, but the very fact of such a structure emerging raises more questions for us than it provides answers.