本期节目，我们将一起探索几个让AI更聪明的“反常识”妙招，全是来自最新论文的硬核洞察。我们会发现，为什么有时候“躺平”学习的AI反而会考砸，而主动扔掉海量数据却能让模型更强。我们还会聊聊，如何通过给AI的大脑做个“剪枝”手术来激发创造力，或者请个“陪练”帮它领悟世界的规律。最后，你将看到，只需几个简单的“二选一”，就能让AI“秒懂”你的独特品味。

00:00:35 AI训练的迷思：躺得平，就一定学得好吗？

00:05:52 喂养AI的新姿势：为什么聪明人要主动扔掉一部分数据？

00:12:31 AI绘画：是天才画手，还是像素级的复印机？

00:18:54 给AI请个“陪练”，为什么能让它更聪明？

00:24:25 让AI“秒懂”你的心思，需要几步？

本期介绍的几篇论文：

[LG] Flat Minima and Generalization: Insights from Stochastic Convex Optimization

[Tel Aviv University]

https://arxiv.org/abs/2511.03548

---

[LG] Why Less is More (Sometimes): A Theory of Data Curation

[Concordia University & FAIR at Meta]

https://arxiv.org/abs/2511.03492

---

[LG] Provable Separations between Memorization and Generalization in Diffusion Models

[Northwestern University & Georgia Institute of Technology]

https://arxiv.org/abs/2511.03202

---

[CV] Generative Hints

[Stanford University & California Institute of Technology]

https://arxiv.org/abs/2511.02933

---

[LG] Inference-Time Personalized Alignment with a Few User Preference Queries

[MPI-SWS & Visa & CMU]

https://arxiv.org/abs/2511.02966