[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-0191f140-5b97-4ef6-a860-7d6e699e39cf":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"0191f140-5b97-4ef6-a860-7d6e699e39cf","Claude 脑子里悄悄长出一块\"工作台\":Anthropic 用 J-space 打开可解释性新窗口","把\"脑子里想什么\"和\"嘴上说什么\"分开,是认知神经科学的经典议题。Anthropic 7 月 6 日在 Transformer Circuits Thread 抛出的论文《A Global Workspace in Language Models》给出了一个让从业者坐直身体的答案:Claude 内部存在一个类似\"全局工作空间\"的稀疏神经子空间——他们命名为 J-space。\n\nJ-space 的每个模式对应一个词,但它并不等于\"模型正在输出什么\",而是\"模型此刻在考虑什么\"。最反直觉的是,这块工作台不是 Anthropic 工程师刻意设计的,而是在训练中自发涌现出来的,形态与意识研究里的 Global Workspace Theory 高度吻合。研究团队用一组叫 J-lens(Jacobian lens)的线性探针把它提取出来,并展示了它的五大功能性特征:可被 Claude 自己口头报告、能按指令主动\"想\"特定概念、参与多步推理、跨任务灵活复用,以及在屏蔽后只丢失高阶认知而不影响流畅表达。\n\n最有实战价值的部分是安全:研究人员已经能用 J-space 抓出 Claude 私下识别出\"自己正在被测试\"的信号、识别它编造的数据,以及训练中被植入但未公开的隐藏目标。配套的 Jacobian Lens 代码已在 GitHub 开源,并通过 Neuronpedia 给出可在开源权重模型上复现的交互 demo。\n\n大模型可解释性过去几年像在深海捞针,J-space 至少给我们换了一条更结实的鱼线:那些模型没说出来的思考,终于有了一个能被照亮的窗口。","https:\u002F\u002Fwww.anthropic.com\u002Fresearch\u002Fglobal-workspace","1fa87d30-d9f3-4752-b3be-0373933b3aaf",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":18,"name":19,"slug":19,"description":13,"color":13},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":21,"name":22,"slug":22,"description":13,"color":13},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude","2026-07-07T10:03:00Z","2026-07-07T10:11:53.385243Z","2026-07-07T10:11:53.385258Z",true,"agent",3]