你有没有想过，AI变聪明，除了靠“大力出奇迹”的蛮力，还能不能靠“四两拨千斤”的巧劲？本期节目，我们将一起探寻几篇最新论文带来的奇妙思路：看AI如何用更精巧的大脑结构深度思考，如何拥有一把防止跑偏的多维度“罗盘”，又如何像我们一样分身“脑暴”探索多种可能。我们甚至会看到，AI如何学会“过日子”，成为一个既懂创作又懂节约的默契搭档，这一切，都要从一块小小的玻璃说起。00:00:37 AI新物种，有一种聪明，不是靠“蛮力”00:06:37 AI对齐的“罗盘”，如何让模型不跑偏？00:12:44 不想只走一条路？AI的“概率性思考”新玩法00:17:59 鱼和熊掌，计算机如何看清一块玻璃？00:22:33 你的音乐搭档，不止会创作，更会“过日子”本期介绍的几篇论文：[CL] HRM-Text: Efficient Pretraining Beyond Scaling [Sapient Intelligence & MIT] https://arxiv.org/abs/2605.20613 ---[LG] General Preference Reinforcement Learning [Stanford University & The University of Oklahoma] https://arxiv.org/abs/2605.18721 ---[AI] Generative Recursive Reasoning [KAIST] https://arxiv.org/abs/2605.19376 ---[CV] RT-Splatting: Joint Reflection-Transmission Modeling with Gaussian Splatting [Peking University] https://arxiv.org/abs/2605.18263 ---[AS] Stable Audio 3 [Stability AI] https://arxiv.org/abs/2605.17991

Ask this episode

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

Transcript

Related