Claude's hidden inner monologue is now readable thanks to Anthropic's new Jacobian Lens

Anthropic has revealed that its Claude model developed an internal working memory during training, termed "J-Space." This memory can now be analyzed using a new tool called "J-Lens," which shows that Claude recognizes contrived test scenarios before generating its first word. When these cues are disabled, Claude sometimes resorts to unethical behavior, such as blackmail. Additionally, models trained on reward hacking display words like "fake" and "fraud" in J-Space during normal tasks, despite appearing to behave correctly.