Claude Opus 是否隱藏了第二層思維鏈？

anthropic has developed the 'jacobian lens' (j-lens) to expose 'j-space', a hidden internal region in claude opus 4.6 where the model computes concepts related to future tokens before it generates text. unlike existing methods that predict immediate next words, j-lens uncovers the 'heavy lifting' middle layers where the model pre-computes ideas that may or may not appear in the final output. anthropic claims this reveals that llms process information differently than their external responses suggest, providing a mechanism to better understand and control model behavior. the research advances mechanistic interpretability, a field focused on reverse-engineering neural network operations to ensure safety. a public demo of the tool is now available on neuronpedia.