当AI变得越来越强大，我们还能从哪些地方挖掘它的潜力呢？本期我们聚焦几篇思路极其巧妙的最新论文，它们不约而同地告诉我们：真正的飞跃，不一定来自更大的模型，而来自更聪明的工作方式。我们将一起探讨，AI如何学会为自己省下90%的训练开销，如何免费装上“直觉”来审时度势，又是如何通过“抓重点”实现一目十行。更重要的是，我们将看到科学家们如何努力为整个AI行业的发展，打造一把统一的“度量衡”。00:00:38 AI调参省钱术：从“大力出奇迹”到“聪明省力气”00:07:44 AI绘画，如何从“慢跑”变“冲刺”？00:13:11 给AI发展装上一个统一的度量衡00:19:25 如何免费给AI装上“直觉”？00:24:56 AI“一目十行”的秘密：不靠算力，靠“会抓重点”本期介绍的几篇论文：[LG] Efficient Hyperparameter Search for Non-Stationary Model Training[Google DeepMind & Google Research]https://arxiv.org/abs/2512.01258---[CV] Improved Mean Flows: On the Challenges of Fastforward Generative Models[CMU & THU & Adobe]https://arxiv.org/abs/2512.02012---[AI] A Rosetta Stone for AI Benchmarks[Google DeepMind]https://arxiv.org/abs/2512.00193---[LG] ZIP-RC: Zero-overhead Inference-time Prediction of Reward and Cost for Adaptive and Interpretable Generation[UC Berkeley & MIT]https://arxiv.org/abs/2512.01457---[LG] Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding[UC Berkeley & MIT & University of Washington]https://arxiv.org/abs/2512.01278

Ask this episode

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

Transcript

Related