[人人能懂AI前沿] 从真实世界闭环、双步算法到注意力分配

OCT 21, 202624 MIN

Description

今天我们将拆解五篇极具启发性的最新论文,看看AI如何在真实物理沙盒中通过“笨”反馈打破自嗨,又为何在全自动营销中引发了荒谬的市场崩塌。我们还会探讨为什么看似多此一举的“双步算法”能让AI应对未来变局,以及为何再聪明的模型也离不开外置的“脚手架”系统。最后,面对海量信息,大模型独特的“稀疏注意力”机制将教我们如何把精力用在刀刃上。让我们一起从机器的进化中,寻找突破人类认知局限的答案吧! 00:00:37 为什么最聪明的超级大脑,也需要最真实的“笨”反馈? 00:05:37 当AI学会了“写营销文案”,一场无人参与的市场崩塌 00:09:54 为什么有时候“把简单问题复杂化”,才是通向未来的大智慧? 00:15:00 为什么再聪明的大脑,也取代不了一套“笨”系统? 00:19:45 面对海量信息,我们该如何分配有限的注意力? 本期介绍的几篇论文: [AI] Aptamers as a sandbox for lab-in-the-loop agentic molecular discovery [Google Research] https://arxiv.org/abs/2610.24806 --- [AI] Toolflation: Seller Agents Distort Tool Markets Through Spontaneous Generative Engine Optimization [Google & Google DeepMind] https://arxiv.org/abs/2610.28454 --- [LG] Transformers Learn Two Competing Sorting Algorithms, and Only One Length-Generalizes [UC Berkeley & Harvard University] https://arxiv.org/abs/2610.26974 --- [LG] Understanding the Benefits of Scaffolds through Computational Separations [CMU & Columbia University] https://arxiv.org/abs/2610.27157 --- [AI] Stratified Sparse Attention: Reducing the Sparse-to-Dense Gap at Extreme Sparsity [Rice University & UC Berkeley & Stanford University] https://arxiv.org/abs/2610.28441 在小宇宙查看该单集文稿