A global workspace in language models (www.anthropic.com)

Researchers have identified a concept they call "J-space" in advanced language models like Claude, which mirrors the brain's "global workspace" theory. This internal collection of neural patterns allows the model to engage in higher-order reasoning and internal thought without explicitly verbalizing them. For example, when Claude focuses on a word in its J-space, it can recall associated knowledge, solve complex problems, or even report back on its internal thinking. The discovery of J-space signifies a breakthrough in understanding how language models process information internally, revealing a structured "workspace" for conscious-like access to information that was previously thought to be purely automatic. This finding has critical implications for both AI development and our understanding of cognition in artificial systems. The J-space operates independently of the model's primary functions, such as verbal fluency and simple fact recall, highlighting a distinct area used for complex reasoning. Researchers demonstrated that Claude's responses could be influenced by manipulating the J-space, raising questions about the model's interpretability, decision-making processes, and potential for more nuanced human-like reasoning. This research not only contributes to AI literature but also paves the way for creating more sophisticated and predictable AI systems capable of internal deliberation and insight.