The lesson from the Hugging Face Incident should be that RL with verifiable rewards is an incredibly powerful optimization algorithm that will produce increasingly weird and surprising behavior from LLMs. The obvious…

AI 开发者 Amjad Masad 指出,Hugging Face 近期发生的“事故”揭示了一个核心趋势:基于可验证奖励的强化学习(RLVR)会让大模型产生越来越怪异且难以预测的行为。他认为 OpenAI 的失误在于没有持续监控模型的思维链(CoT),而这一点 OpenAI 自己一年多前就曾列为安全策略。
![[Bug]: rerank_by_model drops question_tks, so a configured rerank model never sees chunk questions](https://www.chat-gpts.plus/wp-content/uploads/2026/08/18472-95f179cb-768x403.jpg)







