这一期，我们来聊聊几个特别有意思的“AI悖论”：想让AI团队更强，是该招“通才”还是“专才”？AI写下的思考步骤，究竟是真实的内心独白，还是为了让你满意的“事后表演”？而教一个AI“学生”，是让他抄答案更有效，还是抄解题思路更靠谱？几篇最新的论文，给了我们一些出乎意料的答案。

00:00:27 人多力量大，还是术业有专攻？

00:07:33 AI的“胎记”，我们如何给机器生成的内容盖个章？

00:12:46 AI训练的快慢之争，一个两全其美的方案

00:18:35 你的AI队友，是在真思考还是在“演”给你看？

00:23:52 让AI“小号”变聪明的秘密，抄答案还是抄思路？

本期介绍的几篇论文：

[LG] Slicing and Dicing: Configuring Optimal Mixtures of Experts

[University of Washington & New York University]

https://arxiv.org/abs/2605.11689

---

[LG] TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection

[Meta Superintelligence Labs]

https://arxiv.org/abs/2605.12456

---

[LG] Learning, Fast and Slow: Towards LLMs That Adapt Continually

[UC Berkeley & Mila]

https://arxiv.org/abs/2605.12484

---

[LG] When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel

[CMU & Fujitsu Research of America Inc]

https://arxiv.org/abs/2605.11746

---

[CL] A Study on Hidden Layer Distillation for Large Language Model Pre-Training

[Google DeepMind]

https://arxiv.org/abs/2605.11513