Anthropic: The Claude model posits an existence akin to "conscious human-like cognition."

BlockBeats News, July 7th, Anthropic released its latest research report, which found that the Claude model spontaneously exhibits an internal "global workspace" called J-space (Jacobian space). This is a dedicated neural activation pattern set for silent contemplation in the model, allowing the model to think about concepts without explicitly expressing them, akin to human reportable "conscious" thoughts.

The research identified neural activation patterns in J-space using Jacobian techniques, enabling the reading of concepts the model has not vocalized, and the modification of these patterns. Experiments showed that disabling J-space would weaken multi-step reasoning ability but would not affect basic tasks and factual memory.

Click on the original article link below to join the BlockBeats · Feishu AI News Channel, which provides 24/7 monitoring of global AI hot topics and news.