我们都知道AI越来越强大，但你有没有想过，我们该如何让它跑得更快、更稳，甚至更“多才多艺”？本期节目，我们将一起探索几篇最新论文，看看科学家们是如何给AI的训练过程装上一个更稳健的导航系统，并揭开AI绘画高手背后“民间偏方”的科学原理。我们还会聊到，如何像培养一个“通才”一样，让一个AI同时学会两百件事。最后，我们将见证两种神奇的“魔法”：如何在没有数据的情况下给大模型高效“瘦身”，以及如何对一个黑箱模型进行精准的“微创手术”。00:00:41 如何给AI装上一个更聪明的“导航系统”00:05:19 AI绘画高手，背后藏着什么训练秘诀？00:11:06 AI通才养成记：如何让一个机器学会200件事？00:17:12 AI模型“瘦身”，如何做到无米之炊？00:25:14 给AI模型做微创手术，需要几步？本期介绍的几篇论文：[LG] ROOT: Robust Orthogonalized Optimizer for Neural Network Training [Huawei Noah’s Ark Lab] https://arxiv.org/abs/2511.20626 ---[LG] Demystifying Diffusion Objectives: Reweighted Losses are Better Variational Bounds [Google DeepMind] https://arxiv.org/abs/2511.19664 ---[LG] Learning Massively Multitask World Models for Continuous Control [University of California San Diego] https://arxiv.org/abs/2511.19584 ---[LG] CafeQ: Calibration-free Quantization via Learned Transformations and Adaptive Rounding [Google] https://arxiv.org/abs/2511.19705 ---[LG] ModHiFi: Identifying High Fidelity predictive components for Model Modification [CSA, IISc & HP Inc. AI Lab & Google] https://arxiv.org/abs/2511.19566

Ask this episode

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

Transcript

Related