这一期，我们来聊聊几个特别有意思的“AI悖论”：想让AI团队更强，是该招“通才”还是“专才”？AI写下的思考步骤，究竟是真实的内心独白，还是为了让你满意的“事后表演”？而教一个AI“学生”，是让他抄答案更有效，还是抄解题思路更靠谱？几篇最新的论文，给了我们一些出乎意料的答案。00:00:27 人多力量大，还是术业有专攻？00:07:33 AI的“胎记”，我们如何给机器生成的内容盖个章？00:12:46 AI训练的快慢之争，一个两全其美的方案00:18:35 你的AI队友，是在真思考还是在“演”给你看？00:23:52 让AI“小号”变聪明的秘密，抄答案还是抄思路？本期介绍的几篇论文：[LG] Slicing and Dicing: Configuring Optimal Mixtures of Experts [University of Washington & New York University] https://arxiv.org/abs/2605.11689 ---[LG] TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection [Meta Superintelligence Labs] https://arxiv.org/abs/2605.12456 ---[LG] Learning, Fast and Slow: Towards LLMs That Adapt Continually [UC Berkeley & Mila] https://arxiv.org/abs/2605.12484 ---[LG] When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel [CMU & Fujitsu Research of America Inc] https://arxiv.org/abs/2605.11746 ---[CL] A Study on Hidden Layer Distillation for Large Language Model Pre-Training [Google DeepMind] https://arxiv.org/abs/2605.11513

Ask this episode

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

Transcript

Related