我们总希望AI越来越聪明，但“聪明”的维度远比我们想象的要丰富。本期节目，我们将从几篇最新的论文出发，探讨如何让机器人拥有“脑补”未来的想象力，如何用“预算制”让AI大模型运行得更精明，以及如何通过“微创手术”找到并拨动AI的“脾气”开关。我们还会聊聊，怎样才能让AI不再当“杠精”，学会谦虚请教，并最终看清它“学富五车”背后的知识盲区。准备好了吗？让我们一起探索AI智能的全新疆界。00:00:37 机器人学会了“脑补”，世界将有什么不同？00:06:14 AI加速的“抠门”智慧，为什么顶尖高手都懂“预算”思维？00:11:50 找到AI的“脾气”开关00:16:56 别让AI当“杠精”，教它学会“请教”00:21:56 为什么AI“学富五车”，却总在关键时刻掉链子？本期介绍的几篇论文：[RO] World Action Models are Zero-shot Policies [NVIDIA] https://arxiv.org/abs/2602.15922 ---[LG] MoE-Spec: Expert Budgeting for Efficient Speculative Decoding [Meta Reality Labs & Franklin and Marshall College] https://arxiv.org/abs/2602.16052 ---[CL] Surgical Activation Steering via Generative Causal Mediation [MIT & Pr(AI)²R Group] https://arxiv.org/abs/2602.16080 ---[CL] Learning to Learn from Language Feedback with Social Meta-Learning [Google DeepMind] https://arxiv.org/abs/2602.16488 ---[CL] Long-Tail Knowledge in Large Language Models: Taxonomy, Mechanisms, Interventions and Implications [Google] https://arxiv.org/abs/2602.16201

Ask this episode

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

Transcript

Related