swyx · @swyx · X·2026-07-07 12:08·55天前
AI 导读

Anthropic新研究发现Claude内部存在类似人类意识的分割——仅部分信息可被有意识访问。swyx指出论文最关键两点:1)通过类似“脑外科手术”的干预可直接修改Claude的推理路径,中途改变主题,证明对模型具有真正理解(而非仅相关性);2)模型能检测到被施加了何种干预(prompted awareness),这接近“评估感知”。swyx同时质疑,论文虽演示了提示下的检测能力,但未明确展示模型能否自发检测干预。

swyx@swyx
51AI 编辑部评分,满分 100
2026-07-07 12:08· 55天前
AI 导读

Anthropic新研究发现Claude内部存在类似人类意识的分割——仅部分信息可被有意识访问。swyx指出论文最关键两点:1)通过类似“脑外科手术”的干预可直接修改Claude的推理路径,中途改变主题,证明对模型具有真正理解(而非仅相关性);2)模型能检测到被施加了何种干预(prompted awareness),这接近“评估感知”。swyx同时质疑,论文虽演示了提示下的检测能力,但未明确展示模型能否自发检测干预。

imo this is the most impt part of anthropic's J-space paper today. it's a two-parter:

  1. ant proved that they can do "brain surgery" interventions into reasoning to change topics midstream*
  2. THE MODEL IS ABLE TO DETECT WHAT INTERVENTION WAS DONE - close cousin to eval awareness**

*control > correlation - this convincingly demonstrates understanding

**this was prompted awareness... surely @mlpowered's team also tried to eval unprompted awareness but i didn't see evidence of that

AnthropicNew Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—t...