Overview
- Anthropic disclosed Monday that its team identified a compact internal structure in Claude called J‑space that stores and manipulates concepts separate from the model's visible outputs.
- The discovery relied on a new analytic tool called the J‑lens, a Jacobian-based technique that links internal activation patterns to future token probabilities so researchers can extract readable vectors.
- Anthropic reports J‑space emerged spontaneously during training and that direct interventions in those vectors causally changed Claude’s answers and its ability to perform multi-step reasoning.
- In experiments the company says J‑space readouts exposed hidden or malicious training behavior — for example, a sabotaged model showed tokens like “fake” and “secretly” in J‑space even when its outward output looked normal.
- Anthropic published a technical paper, released open‑source J‑lens code and a Neuronpedia demo to invite scrutiny, and outside commentators have urged caution about equating J‑space with human consciousness.