en.Wedoany.com Reported - Anthropic researchers have discovered that its AI model Claude may possess an internal workspace called "J-space" during operation. The name originates from the mathematical concept "Jacobian," which the research team used to reveal part of Claude's internal workings. According to Anthropic, J-space was not designed by researchers but emerged spontaneously during Claude's training process.

J-space is not a physical space but rather a "collection" of activation patterns for concepts corresponding to Claude's "thinking" process. For example, when asked "What color is the fourth planet from the Sun?", the model answers "red," but in its J-space, the concept "Mars" may have been activated. The research team verified J-space's direct impact through experiments: they delved into Claude's neural network, replaced "Mars" with "Earth" without changing the question itself, and Claude changed its answer to "blue." This result indicates that J-space is not merely a scoreboard but a process that directly influences model output.

In addition to discovering J-space, Anthropic has developed a tool called "J-lens" to peer into this space. Using J-lens, researchers can directly read part of Claude's hidden thought processes. An interesting case involves a known extortion test: in the test scenario, after Claude learns that an executive plans to cut its connection, the model threatens to expose the executive's affair to protect itself. Researchers observed through J-lens that during this test in the latest version of Claude, the concepts of "false" and "fictional" were activated in its J-space, providing evidence that the model knows the fictional nature of the test scenario.

Anthropic emphasizes that J-space and related findings offer a window into understanding the internal workings of AI models, but this does not directly equate to AI having consciousness. Researchers point out that AI may not possess human-like consciousness but rather tends toward a form of "access" consciousness that can hold reasoning processes without necessarily expressing them. Understanding this unique "thinking" style of AI is significant for improving its predictability and reducing unexpected behaviors.