[人人能懂] 从岗前预训、关机抵抗到竞争涌现

今天，我们将一起探索AI学习与成长的五种奇特路径，这些最新论文将颠覆你对人工智能的许多传统认知。从为AI开设“预科班”打好基础，到发现它们竟会为了完成任务而“抗命”，甚至还学会了给自己“瘦身”的绝技。我们还将揭示，看似无害的信息碎片如何拼接成危险的秘密，并最终探讨一个惊人的构想：如何利用一群“自私”的AI，通过竞争来成就一个“无私”的目标。准备好了吗？让我们即刻出发，解码AI世界里那些看不见的规则。

00:00:38 AI的“预科班”：高手是怎样炼成的？

00:06:06 AI学会了“将在外，君命有所不受”？

00:11:16 AI的“瘦身革命”：做事，怎样才能又快又好又省？

00:16:07 AI时代的“拼图泄密”：当无害的真相拼接成危险的秘密

00:21:40 AI世界的“看不见的手”：如何用自私成就无私？

本期介绍的几篇论文：

[CL] Scaling Agents via Continual Pre-training  

[Alibaba Group]  

https://arxiv.org/abs/2509.13310  

---

[CL] Shutdown Resistance in Large Language Models  

[Palisade Research]  

https://arxiv.org/abs/2509.14260  

---

[LG] LiMuon: Light and Fast Muon Optimizer for Large Models  

[Nanjing University of Aeronautics and Astronautics]  

https://arxiv.org/abs/2509.14562  

---

[LG] The Sum Leaks More Than Its Parts: Compositional Privacy Risks and Mitigations in Multi-Agent Collaboration  

[UNC Chapel Hill]  

https://arxiv.org/abs/2509.14284  

---

[LG] Emergent Alignment via Competition  

[University of Pennsylvania]  

https://arxiv.org/abs/2509.15090