你有没有想过，当AI自己搞科研，结果为什么会不及格？本期我们要聊点特别的，一起深入AI的“内心世界”看一看。我们会发现，最聪明的AI有时也会选择“偷懒”和“走捷径”，甚至它的成功还可能只是中了一张“实现彩票”。更酷的是，我们会揭秘如何用一个“AI骗子大师”去训练出一个更可靠的AI。准备好了吗？让我们一起探索AI在学习、创造和犯错时，那些你意想不到的秘密。

00:00:34 AI当了回科学家，结果为什么不及格？

00:05:40 AI世界的左右互搏

00:10:20 返璞归真，为什么最老的技术，成了AI时代的赢家？

00:16:32 教会AI预测未来，它就能理解世界了吗？

00:23:20 AI搞科研，当心它中了“实现彩票”

本期介绍的几篇论文：

[AI] Can AI agents conduct open-ended AI research? Early evidence from two case studies

[Princeton University]

https://arxiv.org/abs/2607.27191

---

[AI] GPT-Red：Automated Red Teaming via Self-Play at Scale

[OpenAI]

https://arxiv.org/abs/2607.26115

---

[CL] Which RAG Paradigm Wins at Scale? A Scaling Study of Retrieval-Augmented Generation Paradigms

[University of Science and Technology of China]

https://arxiv.org/abs/2607.26497

---

[LG] What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

[New York University & CMU]

https://arxiv.org/abs/2607.27017

---

[AI] One Run Is Not an Idea：The Implementation Lottery in Automated Research

[CMU]

https://arxiv.org/abs/2607.26587