Anthropic's J-lens reveals Claude's J-space for unspoken concepts

Anthropic used a Jacobian lens (J-lens) to find J-space in Claude, a set of internal activations for concepts the model hasn't yet voiced. Editing these signals changes responses, and removing them breaks multi-step reasoning.

• J-lens identifies concepts the model hasn't yet spoken
• Replacing 'soccer' with 'rugby' in J-space changes Claude's answer
• Removing J-space keeps fluency but breaks reasoning
• In blackmail test, 'fake' and 'fictional' signals fired before response