你有没有想过，为什么AI会越练越笨，甚至被人类“算计”？本期节目，我们将一口气解锁五篇最新论文带来的颠覆性思考：从用香农定律解释AI的“U型学习曲线”，到仅用一块显卡就能“调教”出听话的大模型；从赋予AI“底线思维”来应对未知风险，到让它写出“绝对正确”的关键代码；最后，我们还会看到AI如何从“看图识字”进化到真正“读懂规矩”。准备好，一场关于AI思维升级的风暴即将来临！

00:00:37 你的模型为什么越练越笨？

00:07:11 AI大模型调教指南，如何用一块显卡搞定？

00:12:44 面对“深不可测”的世界，AI如何做出明智选择？

00:18:45 AI开始写“绝对正确”的代码了

00:23:43 AI的新玩法，从“看图识字”到“读懂规矩”

本期介绍的几篇论文：

[LG] LLMs as Noisy Channels: A Shannon Perspective on Model Capacity and Scaling Laws

[ByteDance Seed]

https://arxiv.org/abs/2605.23901

---

[LG] Convex Optimization for Alignment and Preference Learning on a Single GPU

[Stanford University]

https://arxiv.org/abs/2605.23244

---

[LG] Infra-Bayesian Reinforcement Learning Agents Outperform Classical RL For Worst-Case Robustness

[Purdue University & CMU & WorldQuant University]

https://arxiv.org/abs/2605.23146

---

[AI] Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems

[UC Berkeley]

https://arxiv.org/abs/2605.23109

---

[CV] General Hazard Detection

[Swinburne University of Technology]

https://arxiv.org/abs/2605.23304