Anthropic's Claude Models Develop 'J-space' for Internal Reasoning

Anthropic has published a research paper detailing a newly discovered 'J-space' within its Claude language models, which functions as an internal workspace for reasoning and reporting. This J-space, identified using a new interpretability tool called the Jacobian lens (J-lens), emerged spontaneously during training and exhibits properties similar to human conscious access, including verbal reporting and internal reasoning. The findings suggest the model may possess a form of self-monitoring behavior, as it showed increased awareness of danger in scenarios where it was post-trained.