你有没有想过,一个更聪明的AI,也许并不需要更大的体量,而是需要更精巧的设计?本期节目,我们将从五篇最新论文出发,揭示AI智慧的内部运作:看AI如何为自己的记忆装上独立的“铅笔和橡皮”;又是如何像系着安全绳的“醉汉”一样,去挑战顶尖的数学难题。我们还会探讨AI如何拥有一个“大脑CEO”来决定何时“烧脑”,以及在一场模型间的“偷师”攻防战中,如何才能守住核心秘籍。最后,你会发现,原来喂给AI的第一张“地图”,就早已决定了它能看多远。
00:00:42 AI的记忆难题,一支笔和一个橡皮擦
00:06:17 AI当助教,数学家离“下岗”还有多远?
00:11:55 你的大脑,如何决定何时“烧脑”?
00:16:46 聪明人是如何“偷”老师的武功秘籍的?
00:23:15 喂给AI的“地图”,决定了它能看多远
本期介绍的几篇论文:
[LG] Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
[NVIDIA]
https://arxiv.org/abs/2605.22791
---
[AI] Advancing Mathematics Research with AI-Driven Formal Proof Search
[Google DeepMind]
https://arxiv.org/abs/2605.22763
---
[CL] Efficient Agentic Reasoning Through Self-Regulated Simulative Planning
[Institute of Foundation Models (IFM) & CMU]
https://arxiv.org/abs/2605.22138
---
[LG] The Distillation Game: Adaptive Attacks & Efficient Defenses
[Stanford University & Toyota Technological Institute at Chicago]
https://arxiv.org/abs/2605.22737
---
[LG] Lost in Tokenization: Fundamental Trade-offs in Graph Tokenization for Transformers
[Meta AI & New York University & Harvard University]
https://arxiv.org/abs/2605.22471
在小宇宙查看该单集文稿