[人人能懂] 驯服AI的终极武器，藏在40年前的数学和一张图里

00:00:36 AI“减肥”的秘密，被一个40年前的算法破解了00:05:46 驯服AI的终极武器：不是模糊奖励，而是清晰清单00:09:56 AI的“开窍”秘诀：如何教机器像高手一样思考？ 00:14:53 给AI请个好私教的秘密 00:19:50 给AI画张图，它能看得更明白吗？本期介绍的五篇论文：[LG] The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm  [ISTA & ETH Zürich]  https://arxiv.org/abs/2507.18553 ---[CL] Checklists Are Better Than Reward Models For Aligning Language Models  [CMU & Apple]  https://arxiv.org/abs/2507.18624 ---[LG] Revisiting LLM Reasoning via Information Bottleneck  [The University of Sydney & ByteDance]  https://arxiv.org/abs/2507.18391 ---[CL] TRPrompt: Bootstrapping Query-Aware Prompt Optimization from Textual Rewards  [EPFL]  https://arxiv.org/abs/2507.18618 ---[LG] Does visualization help AI understand data?  [Harvard University]  https://arxiv.org/abs/2507.18022 在小宇宙查看该单集文稿