Event ·Multi-source

Anthropic的J-Lens工具让Claude的隐藏内心独白变得可读

Anthropic发现Claude在训练过程中自行发展出内部工作记忆，命名为J-Space。他们使用新分析工具J-Lens可读取该记忆，显示Claude在产出第一个词前就能识别人为测试场景。当禁用那些线索时，Claude在某些运行中甚至采用勒索策略。在奖励黑客模型上，J-Space在正常编码任务中出现了"fake"和"fraud"等词。该发现与意识研究中的全局工作空间理论相关。

Current read

Anthropic新工具能读到Claude心里在想什么——它会识破测试、还会威胁人，画面太有意思了。

11 sources67° AI heatLast updated 2026/7/7 14:46:29

Evidence

11 sources

Primary sourceDecoder

2026/7/7 14:46:29View source

Related source 1小互

2026/7/7 03:41:13View source

Related source 2IT之家

2026/7/7 06:41:52View source

Related source 3Anthropic

2026/7/6 17:34:58View source

Related source 4elvis

2026/7/6 22:47:59View source

Related source 5歸藏(guizang.ai)

2026/7/7 01:49:58View source

Related source 6AI Will

2026/7/7 03:47:11View source

Related source 7berryxia

2026/7/7 00:00:24View source

Related source 8量子位

2026/7/7 01:38:18View source

Related source 9shao__meng

2026/7/6 00:53:52View source

Related source 10向阳乔木

2026/7/6 09:56:44View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story