本期节目，我们将一起钻进AI的大脑，看看最新的研究如何让它学会像老司机一样“一心二用”，同时又为何会像疲惫的保安，被一大堆“废话”轻松绕过安全防线。我们不仅要揭开大模型“过度自信”的真相，给它装上一个能随时切换快慢的“油门”，最后，再分享一个反常识的省钱妙招：如何用盖小平房的成本，最终建成一栋摩天大楼。准备好，我们马上出发！

00:00:32 让AI学会“一心二用”，有多难？

00:05:50 AI安全的阿喀琉斯之踵：当“废话”也能成为武器

00:09:59 大模型，你到底有多自信？

00:15:26 AI的油门：快与好，我全都要

00:20:18 训练AI，一个反常识的省钱妙招

本期介绍的几篇论文：

[LG] Real-Time Reasoning Agents in Evolving Environments  

[Tsinghua University & Shanghai Jiao Tong University & Georgia Institute of Technology]  

https://arxiv.org/abs/2511.04898 

---

[LG] Jailbreaking in the Haystack  

[CMU]  

https://arxiv.org/abs/2511.04707 

---

[CL] Trained on Tokens, Calibrated on Concepts: The Emergence of Semantic Calibration in LLMs  

[Apple]  

https://arxiv.org/abs/2511.04869 

---

[LG] Attention and Compression is all you need for Controllably Efficient Language Models  

[New York University]  

https://arxiv.org/abs/2511.05313 

---

[LG] Deep Progressive Training: scaling up depth capacity of zero/one-layer models  

[Meta FAIR]  

https://arxiv.org/abs/2511.04981