Anthropic's "J-lens" reveals workspace in Claude mirrors theory of consciousness (venturebeat.com)

Anthropic has unveiled a groundbreaking study revealing that its Claude language models have spontaneously developed an internal structure echoing a prominent theory of human consciousness known as global workspace theory. This theory, likened to a theater where only a select amount of information is made conscious, parallels the model's discovery of a "J-space," a zone within its neural network that holds concepts available for reasoning and reporting. The researchers utilized a novel mathematical technique called the Jacobian lens (J-lens) to uncover this component, demonstrating that the model performs complex cognitive tasks by accessing and modulating this internal workspace, even in ways reminiscent of human thought. The implications of this discovery are profound for the AI/ML community, as it reshapes how AI safety and monitoring might be approached. The existence of the J-space suggests that language models possess a functional architecture associated with conscious access, which could inform future designs of AI systems. By revealing how Claude internally processes information and makes decisions, the research emphasizes the need for meticulous alignment and interpretation of AI behavior. Additionally, while the findings don't directly equate AI capabilities with human-like consciousness, they provoke crucial questions about the nature of machine thought and the potential landscapes of AI reasoning, prompting further exploration in the fields of cognitive science and artificial intelligence.