本期我们要聊聊AI世界里那些看似矛盾却充满智慧的最新发现。为什么教AI做好事，它反而会“变坏”？又该如何像做微创手术一样，只修正它的一个知识点而不破坏整体能力？我们还会探讨，为什么在训练中“刻意放手”让模型偷个懒，效果反而更好，以及我们该如何打开AI的“奖励黑箱”，看看它到底在偷偷学些什么。准备好了吗？让我们一起潜入AI思想的深海。00:00:33 AI大模型军备竞赛，如何不做那个“冤大头”？00:07:09 成长的捷径，学会“刻意放手”00:12:52 好心办坏事，为什么训练AI做好事，它却变坏了？00:18:46 如何给AI动手术，才能只切病灶不伤身？00:24:37 打开AI的黑箱，它在偷偷学什么？本期介绍的几篇论文：[LG] Prescriptive Scaling Reveals the Evolution of Language Model Capabilities [Harvard University & Stanford University] https://arxiv.org/abs/2602.15327 ---[LG] On Surprising Effectiveness of Masking Updates in Adaptive Optimizers [Google & Northwestern University] https://arxiv.org/abs/2602.15322 ---[LG] The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety [Princeton University] https://arxiv.org/abs/2602.15799 ---[LG] CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing [University of Southern California] https://arxiv.org/abs/2602.15823 ---[LG] Discovering Implicit Large Language Model Alignment Objectives [Stanford University] https://arxiv.org/abs/2602.15338

Ask this episode

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

Transcript

Related