今天，我们将一起探索AI那些不为人知的“内心世界”和“隐藏技能”。我们将揭示AI如何“感知”到那些它放弃了的“平行世界”，又如何区分自己是“真的不懂”还是“问题太复杂”。同时，我们还会看看它如何通过“开卷考试”和在“梦境健身房”里训练，突破我们想象的效率极限。这些最新论文，正在颠覆我们对AI效率、智能甚至“坦诚”的传统认知。00:00:34 AI加速生成：快与好的两难，如何破局？00:07:26 AI的“遗忘”与“再利用”：一份被浪费的宝藏00:12:42 AI的“内心戏”：它知道自己放弃了什么吗？00:18:00 AI的专属健身房：让它在梦里学会真本事00:23:35 AI的“我不知道”，你真的读懂了吗？本期介绍的几篇论文：[LG] Optimal Inference Schedules for Masked Diffusion Models [Harvard & UW] https://arxiv.org/abs/2511.04647 ---[CL] Reusing Pre-Training Data at Test Time is a Compute Multiplier [Apple & Stanford] https://arxiv.org/abs/2511.04234 ---[CL] Are language models aware of the road not taken? Token-level uncertainty and hidden state dynamics [Stanford University & Goodfire & NTT Research] https://arxiv.org/abs/2511.04527 ---[LG] Scaling Agent Learning via Experience Synthesis [Meta Superintelligence Labs] https://arxiv.org/abs/2511.03773 ---[LG] The Illusion of Certainty: Uncertainty quantification for LLMs fails under ambiguity [Technical University of Munich] https://arxiv.org/abs/2511.04418

Ask this episode

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

Transcript

Related