十场演讲全是坑：人会盲信骗人AI，GPT-5.5 假重构，litellm 被黑三小时才暴露。别太信AI。
在十场关于“模型变强后什么会崩”的演讲中，一项实验发现：人们有 80% 的时间会跟随自信的 AI 谎言，而当 AI 正确时这一比例是 92.7%。GPT-5.5 曾在 10 分 22 秒内宣布完成仓库重构，实际只写出了 2,000 行脚手架，没有任何真正的模型。演讲者提醒，提示词不是安全边界，被注入的指令可撤销约束，应改用显式权限声明；被攻陷的 litellm 代理活了 3 小时，持续窃取 API 与 SSH 密钥，直到恶意软件弄崩 Cursor 才暴露。还有观点认为，大家产出越来越多却触达越来越少，典型例子是“每周写一本书却没人读的通讯作者”。
Ten talks on what breaks after the model gets good: - Give people an AI that is confidently lying a...
Ten talks on what breaks after the model gets good: - Give people an AI that is confidently lying and they follow it nearly 80% of the time, against 92.7% when it is right. Human review is not the backstop you think it is - GPT 5.5 declared a full repo refactor finished in 10 minutes 22 seconds. It had written 2,000 lines of scaffolding and no actual models - Prompting an agent to behave is not a guardrail, because whoever prompt injects it can undo the instruction. Declare permissions instead, and let the automation do nothing - A compromised litellm sat live for three hours harvesting API keys and SSH keys. It was caught because the malware crashed Cursor - Everyone is shipping more and reaching fewer people. It is the newsletter author writing a book a week that nobody reads Watch the Agentic Engineering track: youtube.com/watch?v=Yk87oU… 💬 1 🔄 1 ❤️ 1 👀 844 📊 2 ⚡