[人人能懂AI前沿] 从“笨”办法、摊销智慧到对称破缺

你有没有想过，最聪明的AI，可能也需要最“笨”的办法？本期我们要聊聊，为什么简单的“Ctrl+F”有时能打败高级算法，AI又如何学会“摊销”智慧来走捷关，甚至从古老的物理学中悟出了修炼“内功”的心法。我们还会看到，AI如何像搭积木一样逐层过滤、化繁为简，并通过“反复琢磨”最终获得顿悟。准备好，一起窥探AI思考的“内功”心法吧！

00:00:30 AI大模型，越高级，越需要“笨”办法？

00:07:32 AI求解大师，重复计算是美德，还是偷懒才是？

00:13:46 AI的终极思考，当模型学会了“悟”

00:20:13 AI的超能力，把难题变简单的“过滤器”

00:25:12 AI的“内功”心法，一种来自物理学的古老智慧

本期介绍的几篇论文：

[CL] Is Grep All You Need? How Agent Harnesses Reshape Agentic Search

[PricewaterhouseCoopers]

https://arxiv.org/abs/2605.15184

---

[LG] Local Inverse Geometry Can Be Amortized

[A L. Kachhadiya]

https://arxiv.org/abs/2605.13068

---

[LG] Solve the Loop: Attractor Models for Language and Reasoning

[University of Southern California]

https://arxiv.org/abs/2605.12466

---

[LG] Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning

[EPFL]

https://arxiv.org/abs/2605.13612

---

[LG] Spontaneous symmetry breaking and Goldstone modes for deep information propagation

[University of Amsterdam & Harvard University & Tsinghua University]

https://arxiv.org/abs/2605.14685