Overview
- Anthropic reports that Claude developed a distinct activation subspace called J-space that organizes and shares information across the model.
- Researchers used a new Jacobian-based method called the J-lens to identify J-space and map coordinated activation patterns that emerged during training.
- Claude can report the contents of J-space on request, change those contents when asked, and researchers show direct interventions in J-space alter the model’s verbal outputs and task performance.
- Anthropic released the J-lens code and a Neuronpedia demo so external teams can reproduce and probe the finding, stressing community validation is needed to test generality.
- The team frames J-space as a partial, causally actionable window for interpretability and safety work, notes most processing still occurs outside J-space, and stops short of claiming any form of consciousness.