[人人能懂] 当AI学会捏橡皮泥、装硬盘、请私教

想让AI更聪明，除了把它造得更大，还有没有更巧妙的办法？本期我们要聊的几篇最新论文，就提供了令人拍案叫绝的答案。我们将看到，AI训练背后藏着一个关乎结构手术的“秘密开关”，可以给它装上按需调用的“外挂硬盘”，甚至不用开刀就能为它配备私人教练。最后，我们还会探讨，如何通过保护一点“笨拙”的火花和调节一个“创意旋钮”，来彻底释放AI的潜能。

00:00:34 人工智能训练的秘密开关：捏橡皮泥还是合并黏土？

00:04:54 给大模型装个“外挂硬盘”，会发生什么？

00:09:44 AI界的私教课：如何不“开刀”就改造一个大模型

00:15:12 为什么最聪明的学生，也需要一点“笨”想法？

00:20:32 解锁AI潜能：你只差一个“旋钮”

本期介绍的几篇论文：

[LG] Topological Invariance and Breakdown in Learning

[University of Michigan & MIT]

https://arxiv.org/abs/2510.02670

---

[CL] Pretraining with hierarchical memories: separating long-tail and common knowledge

[Apple]

https://arxiv.org/abs/2510.02375

---

[LG] How to Train Your Advisor: Steering Black-Box LLMs with Advisor Models

[UC Berkeley]

https://arxiv.org/abs/2510.02453

---

[LG] Low-probability Tokens Sustain Exploration in Reinforcement Learning with Verifiable Reward

[Tencent]

https://arxiv.org/abs/2510.03222

---

[LG] On the Role of Temperature Sampling in Test-Time Scaling

[Stanford University]

https://arxiv.org/abs/2510.02611