[人人能懂AI前沿] 从大师的操作手册、模型的微表情,到复杂系统里的简单秩序

OCT 22, 202622 MIN

Description

今天我们将带你拆解四篇极具启发性的最新论文:看新手AI如何靠一份“行为清单”低成本逆袭大师,捕捉大模型算题时暴露的“心虚微表情”来实现自我纠错。我们还要潜入大模型内部,揭秘为什么参数越庞大底层结构反而越简单,最后再聊聊为什么越聪明的AI为了追求信息密度,反而越不爱“说人话”。准备好颠覆你的直觉了吗?让我们马上开始今天的硬核AI之旅! 00:00:35 跨越天赋的鸿沟,如何用最低成本,让新手拥有大师的功力? 00:05:47 别被大模型一本正经的样子骗了,它的“微表情”早已出卖了它 00:10:29 为什么AI越大,底层结构反而越简单?读懂大模型内部的“自组织”密码 00:14:24 读完觉得累?为什么越聪明的AI,越不爱“说人话” 00:17:56 没有老师,如何自我拔高?AI教给我们的“逆向思维” 本期介绍的几篇论文: [AI] D-Skill: Distilling Teacher Knowledge through Agent Skills [UC Berkeley] https://arxiv.org/abs/2610.28914 --- [CL] A Strong MoE Is All You Need for Process Supervision [City University of Hong Kong & CMU] https://arxiv.org/abs/2610.29498 --- [LG] How Does Orthogonalization Adapt to the Neural Network Hessian Structure? A Gradient Self Outer-Product Analysis at Initialization [Dartmouth College & Meta] https://arxiv.org/abs/2610.28975 --- [CL] Why is AI-generated language so hard to understand? [New York University & The University of Texas at Austin] https://arxiv.org/abs/2610.29162 --- [CL] LongZero: Enhancing Long-Context Reasoning in LLMs from Zero Data [Tsinghua University] https://arxiv.org/abs/2610.30132 在小宇宙查看该单集文稿