[人人能懂] 大模型根本局限、演化策略、认知框架与具身智能
Description
今天,我们不聊AI有多神奇,而是要给它来一次全面的“体检”,看看它那道看不见的“玻璃天花板”究竟在哪。接着,我们会颠覆你对AI训练的认知,看看它除了“上课”,如何像生物一样“演化”,以及它强大的推理能力背后,是否藏着一套需要我们帮它解锁的“思维地图”。最后,我们会发现,无论是教它解奥赛难题,还是教它做家务,最聪明的办法,可能都藏在我们自己的学习和生活经验里。准备好,让我们一起揭开AI光环背后的真实运作逻辑!
00:00:38 AI的玻璃天花板:为什么模型越大,犯的错越“自信”?
00:08:26 训练AI,除了“上课”还能“生娃”?
00:14:31 AI的“聪明”难题:为什么它能解奥数,却像个没头苍蝇?
00:21:39 AI的“题海战术”,跟我们有啥不一样?
00:27:36 一副眼镜,如何成为灵巧机器人的“私教”?
本期介绍的几篇论文:
[LG] On the Fundamental Limits of LLMs at Scale
[Stanford University & The University of Oklahoma]
https://arxiv.org/abs/2511.12869
---
[LG] Evolution Strategies at the Hyperscale
[FLAIR - University of Oxford & WhiRL - University of Oxford]
https://arxiv.org/abs/2511.16652
---
[LG] Cognitive Foundations for Reasoning and Their Manifestation in LLMs
[University of Illinois Urbana-Champaign & University of Washington & Princeton University]
https://arxiv.org/abs/2511.16660
---
[LG] P1: Mastering Physics Olympiads with Reinforcement Learning
[Shanghai AI Laboratory]
https://arxiv.org/abs/2511.13612
---
[RO] Dexterity from Smart Lenses: Multi-Fingered Robot Manipulation with In-the-Wild Human Demonstrations
[New York University & Meta]
https://arxiv.org/abs/2511.16661



