你有没有想过，一个聪明的AI要如何审视和优化自己的工作方法，实现“自我进化”？怎样才能把一大堆“专家模型”的智慧，完美浓缩进你手机里那个小小的芯片中？本期节目，我们将一口气解锁五篇最新论文，看看AI如何通过“先加后减”的智慧炼成全才，如何用“元认知”打破思维僵局，又是如何学会“聪明的偷懒”，在关键处全力以赴，在无聊处“摸鱼”省电。准备好了吗？让我们一起开启这场精彩的AI思想之旅！00:00:37 AI界的“浓缩”智慧，先做加法，再做减法00:05:00 一个聪明的系统，如何变得更聪明？00:11:12 AI“通才”，如何用一把钥匙，打开物理世界的多扇大门？00:16:39 AI变聪明的秘密，不是看得多，而是看得准00:21:18 大模型“瘦身”记，聪明地偷个懒本期介绍的几篇论文：[CV] Efficient Universal Perception Encoder [Meta Reality Labs & FAIR at Meta] https://arxiv.org/abs/2603.22387 ---[AI] Bilevel Autoresearch: Meta-Autoresearching Itself https://arxiv.org/abs/2603.23420 ---[LG] UniFluids: Unified Neural Operator Learning with Conditional Flow-matching [Chinese Academy of Sciences & Microsoft Research Asia] https://arxiv.org/abs/2603.22309 ---[LG] Scaling Attention via Feature Sparsity [Xidian University] https://arxiv.org/abs/2603.22300 ---[LG] Sparser, Faster, Lighter Transformer Language Models [Sakana AI & NVIDIA] https://arxiv.org/abs/2603.23198

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related