AI可可AI生活

[人人能懂] 从异步思考、复古架构到教练式学习

这一期,我们不谈AI的“体型”,只聊它的“智商”。最新的几篇论文告诉我们,让AI变聪明的关键,可能不是堆数据,而是教它如何像项目经理一样组织思考,像学霸一样按课程表学习。我们还会发现,AI的记忆可能不是死记硬背,而是在脑中绘制地图,甚至需要一位“贴身教练”来指导它成长。准备好,一起窥探AI更聪明的大脑是如何炼成的!

00:00:31 AI也懂分工了?聪明人是怎样“外包”思考的

00:06:49 想让AI变聪明?先送它去上数学课

00:11:55 AI世界的“复古”潮流:被冷落的武功秘籍,为何突然又香了?

00:17:31 如何把学霸“掰开揉碎”,喂给一个普通模型?

00:23:54 AI的记忆之谜:它是在死记硬背,还是在脑中画地图?

本期介绍的几篇论文:

[LG] The Era of Agentic Organization: Learning to Organize with Language Models  

[Microsoft Research]  

https://arxiv.org/abs/2510.26658 

---

[CL] Reasoning Curriculum: Bootstrapping Broad LLM Reasoning from Math  

[Salesforce AI Research & University of California, Los Angeles]  

https://arxiv.org/abs/2510.26143 

---

[LG] Encoder-Decoder or Decoder-Only? Revisiting Encoder-Decoder Large Language Model  

[Google DeepMind]  

https://arxiv.org/abs/2510.26622 

---

[CL] Supervised Reinforcement Learning: From Expert Trajectories to Step-wise Reasoning  

[Google Cloud AI Research]  

https://arxiv.org/abs/2510.25992 

---

[LG] Deep sequence models tend to memorize geometrically; it is unclear why  

[Google Research & CMU]  

https://arxiv.org/abs/2510.26745