00:00:28 AI的“学霸”和“教练”，如何合二为一？ 00:04:33 想让AI更聪明？关键不是砸钱，是“会花钱” 00:08:26 AI的“朋友圈”：单个脑补，不如组团思考 00:13:12 如何让AI既是“通才”，又是“专才”？ 00:16:07 数据世界的“蝴蝶效应”：我们如何揪出那个扇动翅膀的“坏数据”？ 本期介绍的五篇论文： [LG] Your Reward Function for RL is Your Best PRM for Search: Unifying RL and Search-Based TTS [Rutgers University & Nanyang Technological University] https://arxiv.org/abs/2508.14313 --- [LG] Compute-Optimal Scaling for Value-Based Deep RL [UC Berkeley] https://arxiv.org/abs/2508.14881 --- [LG] Graph Concept Bottleneck Models [Stony Brook University & University of California, San Diego & IBM Research] https://arxiv.org/abs/2508.14255 --- [LG] Amortized Bayesian Meta-Learning for Low-Rank Adaptation of Large Language Models [Princeton University] https://arxiv.org/abs/2508.14285 --- [LG] Understanding Data Influence with Differential Approximation [University of Hong Kong & Chinese University of Hong Kong] https://arxiv.org/abs/2508.14648

我们总因为心疼已经付出的代价，而选择在错误的路途上追加投入，最终走向一个代价更高的终点。

00:00:27 AI也懂“拿不准”：靠谱的智能来自何处？ 00:04:51 AI的“乐高”工厂：如何让机器兼具直觉与逻辑 00:09:00 AI作画不听话？高手是怎么把它变聪明的 00:12:54 如何把AI助理，调教成真正的干活高手？ 本期介绍的四篇论文： [LG] BLIPs: Bayesian Learned Interatomic Potentials [University of Amsterdam & CuspAI] https://arxiv.org/abs/2508.14022 --- [LG] The DeepLog Neurosymbolic Machine [KU Leuven] https://arxiv.org/abs/2508.13697 --- [CV] DiffIER: Optimizing Diffusion Models with Iterative Error Reduction [Shanghai Jiao Tong University & The Chinese University of Hong Kong] https://arxiv.org/abs/2508.13628 --- [LG] ComputerRL: Scaling End-to-End Online Reinforcement Learning for Computer Use Agents [Tsinghua University & Zhipu AI & University of Chinese Academy of Sciences] https://arxiv.org/abs/2508.14040

在这个人人都有麦克风，信息超载的时代，我们或许都该向AI学习这种“沉默的算法”。

00:00:34 AI也会“想多了”？聊聊机器的“顿悟”与“纠结” 00:05:23 AI的“顿悟”时刻：为什么高手总在“瓶颈期”后迎来爆发？ 00:10:29 AI的大脑换食谱：不吃“山珍海味”，改吃“压缩饼干” 00:16:16 给复杂世界建模：AI如何成为科学家的“超级学徒” 00:20:44 顶尖高手的心法：如何一眼看出谁是“好学生”？ 本期介绍的五篇论文： [LG] MDPO: Overcoming the Training-Inference Divide of Masked Diffusion Language Models [University of Tübingen] https://arxiv.org/abs/2508.13148 --- [LG] Learning In-context n-grams with Transformers: Sub-n-grams Are Near-stationary Points [EPFL] https://arxiv.org/abs/2508.12837 --- [CV] Separating Knowledge and Perception with Procedural Data [MIT] https://arxiv.org/abs/2508.11697 --- [LG] Simulation-Based Inference: A Practical Guide [University of Tübingen] https://arxiv.org/abs/2508.12939 --- [CV] Dextr: Zero-Shot Neural Architecture Search with Singular Value Decomposition and Extrinsic Curvature [Friedrich-Alexander-Universität Erlangen-Nürnberg & Universität Ulm] https://arxiv.org/abs/2508.12977

那些痛苦、焦虑、倦怠，正是你的“内在警报系统”在发出信号，告诉你：当前的路径可能偏离了你内心真正重视的价值。

00:00:32 AI的“统一场”：它怎么看懂世界？ 00:04:10 AI的“半成品”，竟然是个宝库？ 00:07:43 AI学会“听话”的秘密：回到耳朵本身 00:11:39 给数据“画个像”：一种全新的学习“渔”法 本期介绍的四篇论文： [CL] Beyond the Rosetta Stone: Unification Forces in Generalization Dynamics [Google DeepMind] https://arxiv.org/abs/2508.11017 --- [CL] Diffusion is a code repair operator and generator [Microsoft] https://arxiv.org/abs/2508.11110 --- [CL] Representing Speech Through Autoregressive Prediction of Cochlear Tokens [MIT & Stanford University] https://arxiv.org/abs/2508.11598 --- [LG] Compressive Meta-Learning [Stanford University & University of California, Santa Cruz & Polytechnic University of Catalonia] https://arxiv.org/abs/2508.11090

与人之间的差距，很多时候不是天赋，而是认知的精度。

00:00:30 你的手环，藏着一位私人教练 00:05:21 AI当学徒，画家成总监：一门“偷懒”的艺术 00:09:30 AI的大脑，和我们想的一样吗？ 00:13:58 AI的“偏心眼”：它为何总觉得世界和美国一样？ 00:18:16 AI的“偏心眼”：它为何总觉得世界和美国一样？ 本期介绍的五篇论文： [LG] A personal health large language model for sleep and fitness coaching [Google] https://www.nature.com/articles/s41591-025-03888-0 --- [CV] ToonComposer: Streamlining Cartoon Production with Generative Post-Keyframing [The Chinese University of Hong Kong & Tencent PCG] https://arxiv.org/abs/2508.10881 --- [LG] Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning [University of Amsterdam] https://arxiv.org/abs/2508.10057 --- [CL] Entangled in Representations: Mechanistic Investigation of Cultural Biases in Large Language Models [University of Copenhagen & KAIST] https://arxiv.org/abs/2508.08879 --- [LG] Wavelet Mixture of Experts for Time Series Forecasting [Shanghai University of Engineering Science] https://arxiv.org/abs/2508.08825

真正的力量，也许并非来自更聪明的机器，而是源于重新发现，如何做一个简单而坚韧的，人。

00:00:29 解锁AI推理能力：少即是多的秘密 00:05:17 AI巨兽的“成长配方”：如何让小个子智慧指导大块头？ 00:10:03 AI作画的“临门一脚”：如何让快马跑得又快又好？ 00:14:05 AI 创造 3D 模型：从“盖房子”到“捏泥人” 00:07:44 AI界的“小灶”，如何喂出绝顶高手？ 本期介绍的五篇论文： [LG] Part I: Tricks or Traps? A Deep Dive into RL for LLM Reasoning [Alibaba Group] https://arxiv.org/abs/2508.08221 --- [LG] μ-Parametrization for Mixture of Experts [University of Warsaw] https://arxiv.org/abs/2508.09752 --- [LG] Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models [Technical University of Munich & Inceptive] https://arxiv.org/abs/2508.09968 --- [CV] VertexRegen: Mesh Generation with Continuous Level of Detail [Meta Reality Labs Research & UC San Diego] https://arxiv.org/abs/2508.09062 --- [LG] Beyond Scaling Law: A Data-Efficient Distillation Framework for Reasoning [Zhongxing Telecom Equipment (ZTE)] https://arxiv.org/abs/2508.09883 --- [CL] Speed Always Wins: A Survey on Efficient Architectures for Large Language Models [Unknown / Survey paper] https://arxiv.org/abs/2508.09834

真正的强大，不是拥有一个不会犯错的记忆，而是拥有重写故事的勇气。