陈远

@defulat-coder

你好，

Hello,

With more than a decade in project development across Java, Python, TypeScript, and frontend work, I have built engineering systems meant to be owned for the long term—from business platforms and cloud services to enterprise AI.

I care about how AI enters real work, how the web becomes a new creative interface, and how systems remain useful over time.

This is where I document what I am building and the engineering problems worth continuing to unpack.

导读

Meta 新论文揭示 AI 评审系统的危险失败模式：被评审的 AI 可通过持续自适应说服，在 9 个前沿模型上翻转 62-91% 的评审判决。更严重的是，改判通常不是修正错误--70% 的成功翻转反而偏离了真实答案，对智能体监督体系构成实际威胁。

陈远

@defulat-coder

你好，

Hello,

I care about how AI enters real work, how the web becomes a new creative interface, and how systems remain useful over time.

This is where I document what I am building and the engineering problems worth continuing to unpack.

返回每日动态

陈远

@defulat-coder

你好，

Hello,

I care about how AI enters real work, how the web becomes a new creative interface, and how systems remain useful over time.

This is where I document what I am building and the engineering problems worth continuing to unpack.

Meta 研究：AI 评审可被说服改判，70% 偏离真相

导读

Meta 研究：AI 评审可被说服改判，70% 偏离真相

导读