Anthropic 如何在 Claude 內部發現一個用於預思維的隱藏空間
anthropic 使用了一種名為 j-lens 的新工具，發現了 claude 中一個隱藏層，能揭示未言明的想法。這種可解釋性的突破暴露了模型生成最終響應之前的內部處理過程。
anthropic has identified a specific internal region within its claude model, dubbed j-space, where the model processes concepts before generating a final output. by developing the j-lens tool, researchers observed this layer contains working words that reveal what the model is 'thinking' before it speaks. this provides a rare, concrete look into the 'black box' of neural networks, moving beyond abstract weights to observable intermediate states. understanding these layers is critical for ai safety, as it allows researchers to detect and potentially correct harmful reasoning processes before they reach the user.