Anthropic found a hidden space where Claude puzzles over concepts

Anthropic’s new Jacobian lens reveals words associated with Claude’s future responses, offering researchers clues about its calculations, recognition processes, and potentially deceptive behavior. The technique could support AI safety monitoring, but it cannot provide a complete account of a model’s internal activity.

Anthropic开发的新工具“雅可比透镜”可以显示与Claude后续回答相关的词语，让研究人员看到模型计算、识别信息乃至可能采取欺骗行为时的部分内部线索。这项技术或可用于监测人工智能安全，但无法完整呈现模型的内部活动。