00:01:35 AI的“悄悄话”：我们还能“偷听”多久？ 00:06:10 AI：那个懂所有菜谱，却不会做饭的大厨？ 00:11:08 AI训练老大难：如何让机器“学徒”少走弯路？ 00:16:00 给AI动“开心手术”：我们如何让机器更懂“人情世故”？ 00:19:34 AI的下一个金矿，藏在一只虫子的大脑里？ 今天介绍的五篇论文： [LG] Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety [UK AI Security Institute & Apollo Research] https://arxiv.org/abs/2507.11473 --- [LG] Comprehension Without Competence: Architectural Limits of LLMs in Symbolic Computation and Reasoning [Amazon Web Service] https://arxiv.org/abs/2507.106 --- [LG] Relative Entropy Pathwise Policy Optimization [University of Toronto & Technische Universitat Wien & University of Pennsylvania] https://arxiv.org/abs/2507.11019 --- [CL] Internal Value Alignment in Large Language Models through Controlled Value Vector Activation [University of Science and Technology of China & Renmin University of China Beijing] https://arxiv.org/abs/2507.11316 --- [LG] Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures [Johns Hopkins University] https://arxiv.org/abs/2507.10951

00:01:44 AI偷懒的艺术：好钢如何用在刀刃上 00:05:46 组个“AI梦之队”，比单打独斗强在哪？ 00:10:51 AI玩“跑团”：下一个世界是如何被设计出来的？ 00:14:40 炼成“大模型”高手，需要计划表还是指南针？ 00:19:09 让机器人学会“试错”：不是瞎猜，而是高手过招 今天介绍的五篇论文： [CL] Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation [KAIST AI & Mila] https://arxiv.org/abs/2507.10524 --- [LG] Fusing LLM Capabilities with Routing Data [University of Illinois Urbana-Champaign] https://arxiv.org/abs/2507.10540 --- [AI] Multi-Actor Generative Artificial Intelligence as a Game Engine [Google DeepMind] https://arxiv.org/abs/2507.08892 --- [LG] Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training [KAIST & Seoul National University & Microsoft Research] https://arxiv.org/abs/2507.09846 --- [LG] Behavioral Exploration: Learning to Explore via In-Context Adaptation [UC Berkeley] https://arxiv.org/abs/2507.09041

在过去，要找到一千个铁杆粉丝，你需要巨大的运气和成本。但在今天，AI就像一个超级精准的声呐，能帮你从七十亿人的海洋里，找到那些能与你同频共振的灵魂。

00:01:54 四两拨千斤：让小模型变聪明的“内存魔法” 00:07:11 AI调教指南：不说假话的秘密，竟然是“相信自己”？ 00:11:07 AI思考的“地图”与“导航” 00:15:16 AI大模型的“阿喀琉斯之踵”？ 00:18:49 AI写作高手进阶：人多，不如方法好 本期节目介绍的五篇论文： [CL] KV Cache Steering for Inducing Reasoning in Small Language Models [University of Amsterdam & University of Technology Nuremberg] https://arxiv.org/abs/2507.08799 --- [CL] The Curious Case of Factuality Finetuning: Models' Internal Beliefs Can Improve Factuality [University of Washington] https://arxiv.org/abs/2507.08371 --- [LG] CTRLS: Chain-of-Thought Reasoning via Latent State-Transition [UC San Diego] https://arxiv.org/abs/2507.081 --- [LG] One Token to Fool LLM-as-a-Judge [Tencent AI Lab] https://arxiv.org/abs/2507.08794 --- [LG] Inference-Time Scaling of Diffusion Language Models with Particle Gibbs Sampling [Stanford University] https://arxiv.org/abs/2507.08390

那个从不麻烦别人、看起来无坚不摧的人，并不是真的刀枪不入。他们只是过早地学会了把海啸般的痛苦，调成了静音模式。

00:01:47 AI的“断舍离”：如何让机器像人一样阅读？ 00:06:27 AI健忘症：为什么你的聊天机器人越聊越糊涂？ 00:10:21 AI大模型的“记忆”难题：鱼和熊掌如何兼得？ 00:15:28 想把事做对？你得先学会“挑错” 00:18:54 你的AI管家，终于告别“金鱼记忆”了 今天介绍的5篇论文： [LG] Dynamic Chunking for End-to-End Hierarchical Sequence Modeling [CMU & Cartesia AI] https://arxiv.org/abs/2507.07955 --- [LG] Understanding and Improving Length Generalization in Recurrent Models [CMU & Cartesia AI] https://arxiv.org/abs/2507.02782 --- [CL] A Systematic Analysis of Hybrid Linear Attention [UC Santa Cruz & University of Groningen] https://arxiv.org/abs/2507.06457 --- [CL] CriticLean: Critic-Guided Reinforcement Learning for Mathematical Formalization [ByteDance Seed & Nanjing University] https://arxiv.org/abs/2507.06181 --- [CL] MIRIX: Multi-Agent Memory System for LLM-Based Agents [MIRIX AI] https://arxiv.org/abs/2507.07957

你以为自己已经习惯了，但其实，它们在不知不觉中，偷走了你最好的精力和最平静的心情。

00:02:21 AI巨头们的“乐高”心法：应对复杂世界的终极武器 00:07:24 AI小助理的养成记：笨徒弟如何变高徒？ 00:12:47 你家AI聪不聪明，得看它“跑”得稳不稳 00:18:43 AI也会“看走眼”？一招教它“看真切” 00:22:03 AI的“笨”办法：像婴儿一样学习世界 00:26:54 AI也“开小差”？不说话的思考，可能更强大 今天介绍的六篇论文： [LG] AXLearn: Modular Large Model Training on Heterogeneous Infrastructure [Apple] https://arxiv.org/abs/2507.05411 --- [LG] MobileGUI-RL: Advancing Mobile GUI Agent through Reinforcement Learning in Online Environment [Tencent AI Seattle Lab] https://arxiv.org/abs/2507.05720 --- [LG] A Dynamical Systems Perspective on the Analysis of Neural Networks [Freie Universitat Berlin & Universiteit van Amsterdam] https://arxiv.org/abs/2507.05164 --- [CL] Perception-Aware Policy Optimization for Multimodal Reasoning [University of Illinois Urbana-Champaign] https://arxiv.org/abs/2507.064 --- [LG] Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference [Thousand Brains Project] https://arxiv.org/abs/2507.04494 --- [CL] A Survey on Latent Reasoning https://arxiv.org/abs/2507.06203

当你不再要求谁给你一份满分答卷时，你才真正开始书写自己的故事。

00:01:30 人工智能的安全带，真的能系牢吗？ 00:06:39 给人工智能算算账：它能力的边界在哪？ 00:11:34 你一“喜欢”，人工智能就变乖？这事儿没那么简单 00:16:07 高手与普通人的差距，不在于答案，而在于“清单” 00:20:26 想成事？别总想细节，试试“打包”你的行动 本期介绍的五篇论文： [LG] On the Impossibility of Separating Intelligence from Judgment: The Computational Intractability of Filtering for AI Alignment [Ludwig-Maximilians-Universität in Munich & UC Berkeley] https://arxiv.org/abs/2507.07341 --- [CL] Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models [Stanford University & VianAI Systems] https://arxiv.org/abs/2507.07505 --- [LG] Principled Foundations for Preference Optimization [Google DeepMind] https://arxiv.com/abs/2507.07855 --- [LG] PLAN-TUNING: Post-Training Language Models to Learn Step-by-Step Planning for Complex Problem Solving [Google] https://arxiv.com/abs/2507.07495 --- [LG] Reinforcement Learning with Action Chunking [UC Berkeley] https://arxiv.com/abs/2507.07969

我们拼命管理着自己的金钱、时间、人脉，却对自己最重要的核心资产——“情绪资产”——视而不见。

00:01:50 人工智能界的“分工”智慧：如何让天才更天才？ 00:06:40 人工智能预测那么准，它真的“懂”了吗？ 00:11:52 人工智能的“最强大脑”？不，是“最省大脑” 00:16:14 训练人工智能，少吃多餐还是狼吞虎咽？ 00:20:39 从笨拙到精通：机器人如何“看”会我们的本事？ 本期介绍的五篇论文： [LG] Towards Solving More Challenging IMO Problems via Decoupled Reasoning and Proving [Tencent AI Lab] arxiv.org --- [LG] What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models [Harvard University & MIT] arxiv.org --- [CL] Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation [Microsoft] arxiv.org --- [LG] Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful [New York University & Columbia University] arxiv.org --- [LG] Value from Observations: Towards Large-Scale Imitation Learning via Self-Improvement [Google Deepmind] arxiv.org