你的 AI 审核员与您有相同的盲点

出处: Your AI Reviewer Has the Same Blind Spots You Do

发布: 2026年2月15日

📄 中文摘要

在计划的第126行中,使用了一个正则表达式来捕捉对抗性标记重复,期望的精确度超过95%。经过审查后,发现该模式使用了回溯引用,但由于Rust的正则表达式引擎不支持回溯引用,导致该模式无法编译并在启动时出现错误。随后,将该计划发送给五个AI模型进行独立审查,其中GPT在实际的Rust代码库中花费了五分钟,发现了编译调用,并指出回溯引用不被支持。而Qwen则通过完全不同的视角在两秒内标记了同一模式,显示出未经过测试的潜在问题。

📄 English Summary

Your AI Reviewer Has the Same Blind Spots You Do

A regex pattern was designed to catch adversarial token repetition, aiming for over 95% precision. However, it was discovered that the pattern utilized a backreference, which Rust's regex engine does not support, leading to compilation failure and panic at startup. The plan was then sent to five AI model families for independent review. GPT spent five minutes in the actual Rust codebase, identified the compilation call, and reported that backreferences are not supported. Separately, Qwen flagged the same pattern in two seconds through a completely different lens, indicating untested concerns.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等