今天我们来聊点颠覆你想象的事：当AI连我们熟悉的Office考试都考不及格，我们是不是该重新思考它的“聪明”？最新论文告诉我们，AI的智慧不仅体现在它有多强大，更在于它如何协作、如何学习、如何化繁为简。从“八卦式”沟通的计算机集群，到被压缩成几句“武功秘籍”的顶尖算法，再到一个能让AI们“华山论剑”的竞技场。本期节目，我们将一起探索AI世界里那些关于“笨拙”、瓶颈与集体智慧的深刻洞见。00:00:36 AI连Office都搞不定，我们是不是想多了？00:06:13 如何让一群计算机高效协作，像一个人一样思考？00:11:06 真正的高手，废话都不多00:17:21 为什么你的努力，有时会“降维打击”自己？00:23:12 AI界的“华山论剑”，一个好制度，比一个天才更重要本期介绍的几篇论文：[CL] Mind the Gap: Can Frontier LLMs Pass a Standardized Office Proficiency Exam? [Microsoft Research] https://arxiv.org/abs/2606.10956 ---[LG] Unifying Local Communications and Local Updates for LLM Pretraining [Concordia University & Sorbonne University] https://arxiv.org/abs/2606.11081 ---[LG] What Fits (Into Few Tokens) Doesn't Overfit: Compression and Generalization in ML Research Agents [Amazon Responsible AI] https://arxiv.org/abs/2606.11045 ---[LG] Overcoming Rank Collapse in Feedback Alignment [Imperial College London & Mila] https://arxiv.org/abs/2606.11123 ---[CL] Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries [Together AI & Stanford University] https://arxiv.org/abs/2606.10402 在小宇宙查看该单集文稿

Ask this episode

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

Transcript

Related