本期节目干货满满，我们将一起潜入AI思维的深海。我们会看到，AI如何学会像顶尖高手一样“倒着想”，先构思结尾再动笔；也会发现，为何AI赛道上1%的微小进步，会在“马拉松”式的长任务中滚成决定性的胜利。我们还会用一场“真心话大冒险”去窥探AI的“内心世界”，看看它是否真的“言行一致”；并揭秘一套先学“普通话”再通“方言”的语言天才养成法。最后，我还会为你揭露一种AI时代的“新巫术”，它可能会让你对许多所谓的“科学结论”打上一个大大的问号。精彩内容，马上开始！

00:00:40 AI写作的“逆向工程”：高手都是先想好结尾再动笔

00:05:18 AI的“马拉松”陷阱：为什么领先一点点，最后能赢一大圈？

00:10:10 AI的“真心话大冒险”：你嘴上说不要，身体却很诚实？

00:15:06 AI的“语言天才”养成法：先学“普通话”，再通“方言”

00:20:12 AI时代的“新巫术”：只要锄头换得勤，没有墙角挖不倒

本期介绍的几篇论文：

[LG] Reverse-Engineered Reasoning for Open-Ended Generation  

[ByteDance Seed]  

https://arxiv.org/abs/2509.06160  

---

[LG] The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs  

[University of Cambridge & University of Stuttgart & Max Planck Institute for Intelligent Systems]  

https://arxiv.org/abs/2509.09677  

---

[LG] Probing the Preferences of a Language Model: Integrating Verbal and Behavioral Tests of AI Welfare  

[Future Impact Group (FIG) & Ruhr-University Bochum]  

https://arxiv.org/abs/2509.07961  

---

[CL] mmBERT: A Modern Multilingual Encoder with Annealed Language Learning  

[Johns Hopkins University]  

https://arxiv.org/abs/2509.06888  

---

[CL] Large Language Model Hacking: Quantifying the Hidden Risks of Using LLMs for Text Annotation  

[Bocconi University & University of Zurich]  

https://arxiv.org/abs/2509.08825