Anthropic展示Claude在J-space中默默推理:能识代码bug和图像

By watching the J-space, we can see Claude silently perform reasoning steps in its head—noticing bug...

精选理由

Anthropic新发现:Claude在脑子里悄悄推理,能自己找代码bug和认图,比看手写CoT更直接。

AI 摘要

Anthropic在推文中展示Claude模型通过J-space(内部表示空间)执行推理步骤,无需输出中间文本。Claude能自动发现代码中的错误,并识别图像内容。该发现揭示了模型内部推理过程的可见性,有助于理解Claude的行为机制。研究基于Claude模型的分析,未公布具体版本号。

原文 · Anthropic

By watching the J-space, we can see Claude silently perform reasoning steps in its head—noticing bug...

By watching the J-space, we can see Claude silently perform reasoning steps in its head—noticing bugs in code, identifying images, and more. 💬 5 🔄 54 ❤️ 920 👀 118732 📊 123 ⚡

Anthropic展示Claude在J-space中默默推理:能识代码bug和图像 · AI 热点