00:01:15 AI变“聪明”的代价：那根看不见的绳子 00:05:47 给AI“长考”时间，它怎么反而变笨了？ 00:11:35 AI的“潜移默化”：看不见的信号，看得见的风险 00:15:45 AI教育的秘诀：一句“再试试”的力量 00:19:28 高手记忆的秘密：不是记得多，而是会“扔” 本期介绍的几篇论文： [LG] The Invisible Leash: Why RLVR May Not Escape Its Origin [Stanford University & University of Tokyo] https://arxiv.org/abs/2507.14843 --- [LG] Inverse Scaling in Test-Time Compute [University of Edinburgh & EPFL & University of Texas at Austin] https://arxiv.org/abs/2507.14417 --- [LG] Subliminal Learning: Language models transmit behavioral traits via hidden signals in data [Anthropic Fellows Program & Truthful AI] https://arxiv.org/abs/2507.14805 --- [LG] A Simple "Try Again" Can Elicit Multi-Turn LLM Reasoning [Imperial College London & Northwestern University & University of Washington] https://arxiv.org/abs/2507.14295 --- [LG] LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models [Georgia Tech] https://arxiv.org/abs/2507.14204

“人生的境界是在不断探索中拓展的，而非在既有框架中满足。” - 王国维《人间词话》

00:01:07 AI当学徒：如何把“笨学生”调教成“优化大师”？ 00:05:48 AI专家速成指南：换个姿势，我们能造出“神医”吗？ 00:10:48 从“大力出奇迹”到“精准打击”：AI的进化新思路 00:14:59 AI的“偏见”：它如何看透纷繁信息，抓住本质？ 00:19:50 让AI学会“多想一步”：笨办法里的真智慧 本期介绍的几篇论文： [LG] CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning [DeepReinforce Team] https://arxiv.org/abs/2507.14111 --- [CL] Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need [Princeton University] https://arxiv.org/abs/2507.13966 --- [LG] Reframing attention as a reinforcement learning problem for causal discovery [University of Tübingen & University of Amsterdam] https://arxiv.org/abs/2507.13920 --- [LG] Provable Low-Frequency Bias of In-Context Learning of Representations [University of Michigan & Harvard University] https://arxiv.org/abs/2507.13540 --- [LG] Change of Thought: Adaptive Test-Time Computation [Google Research & Georgia State University & Maynooth University] https://arxiv.org/abs/2507.13569

他一直向往着远方的冰岛，却没发现，真正持续为他‘充电’的，是方圆五百米内的这些微小互动和日常之美。

00:01:37 如何让你的机器人管家，不再健忘？ 00:05:43 你家的AI，可能有个“邪恶双胞胎” 00:09:30 AI：一面照见我们自己的镜子 00:13:20 AI训练场上的“复古”新潮流 00:18:14 AI的“复盘”神功：如何想得更深？ 00:21:26 驯养AI：我们是如何让机器“猜”出你心思的？ 本期介绍的几篇论文： [RO] Enter the Mind Palace: Reasoning and Planning for Long-term Active Embodied Question Answering [Field AI] https://arxiv.org/abs/2507.12846 --- [CL] Jailbreak-Tuning: Models Efficiently Learn Jailbreak Susceptibility [FAR.AI & Mila] https://arxiv.org/abs/2507.11630 --- [CL] Emergence of Hierarchical Emotion Organization in Large Language Models [Harvard University] https://arxiv.org/abs/2507.10599 --- [LG] ShiQ: Bringing back Bellman to LLMs [Cohere] https://arxiv.org/abs/2505.11081 --- [LG] AbbIE: Autoregressive Block-Based Iterative Encoder for Efficient Sequence Modeling [University of Cambridge] https://arxiv.org/abs/2507.08567 --- [LG] Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities [Not explicitly stated, likely academic tutorial paper] https://arxiv.org/abs/2507.13158

你不是宇宙中的一粒微尘，整个宇宙，都是你思想中的一件‘玩具’。

00:01:45 你的走法，决定了你是谁 00:05:57 AI翻译哪家强？我们终于有了个“懂行”的裁判 00:10:03 成长的秘密：如何让机器像高手一样思考和探索 00:14:13 AI的“口是心非”：我们如何看穿它？ 00:18:11 让 AI 既能读书破万卷，又能下笔如有神 今天介绍的五篇论文： [LG] Optimizers Qualitatively Alter Solutions And We Should Leverage This [Google DeepMind] https://arxiv.org/abs/2507.12224 --- [CL] TransEvalnia: Reasoning-based Evaluation and Ranking of Translations [Sakana.ai] https://arxiv.org/abs/2507.12724 --- [LG] Spectral Bellman Method: Unifying Representation and Exploration in RL [Technion & Georgia Institute of Technology] https://arxiv.org/abs/2507.13181 --- [CL] LLMs Encode Harmfulness and Refusal Separately [Northeastern University & Stanford University] https://arxiv.org/abs/2507.11878 --- [CL] Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs https://arxiv.org/abs/2507.09477

毕竟，人生不是一段需要严格执行的代码，而是一首允许即兴、允许跑调，却依然动人的歌。

00:01:36 你的勤奋，是在“拼宽度”还是在“拼深度”？ 00:06:00 AI 变聪明的秘密：不止是“喂饱”，更要“喂好” 00:11:19 AI的“学霸修炼手册”：如何跳出成长平台期？ 00:16:15 给AI一个“提示”，解锁更高阶的智慧 00:20:23 你怎么用“临时工”大脑，摆平全世界？ 00:05:30 喂AI，也是一门技术活 今天介绍的六篇论文： [LG] The Serial Scaling Hypothesis [UC Berkeley] https://arxiv.org/abs/2507.125 --- [LG] Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved) C Qin, J T Springenberg https://arxiv.org/abs/2507.12856 --- [LG] Scaling Up RL: Unlocking Diverse Reasoning in LLMs via Prolonged Training M Liu, S Diao, J Hu, X Lu... [NVIDIA] https://arxiv.org/abs/2507.12507 --- [CL] QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation J Li, H Lu, K Wen, Z Yang... [Ant Research & Shanghai Qi Zhi Institute & Stanford University & Tsinghua University] https://arxiv.org/abs/2507.13266 --- [CL] Modeling Open-World Cognition as On-Demand Synthesis of Probabilistic Models L Wong, K M. Collins, L Ying, C E. Zhang... [Stanford University & MIT] https://arxiv.org/abs/2507.12547 --- [CL] A Survey of Context Engineering for Large Language Models https://arxiv.org/abs/2507.13334

清晰，不是一种天赋，而是一种可以被刻意练习的纪律。

00:01:37 你的天气APP，正在经历一场认知革命 00:05:44 机器人学做家务，最快的老师原来是我们自己 00:09:50 AI的“偏科”秘籍：如何精准定制一个“特长生”？ 00:14:10 揭秘大脑“黑箱”：我们是如何学会说话的？ 00:20:16 AI“一目十行”的秘密 今天介绍的五篇论文： [LG] FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale [NVIDIA & Lawrence Berkeley National Laboratory] https://arxiv.org/abs/2507.12144 --- [RO] EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos [UC San Diego, UIUC & MIT] https://arxiv.org/abs/2507.12440 --- [CL] Language Models Improve When Pretraining Data Matches Target Tasks [Apple] https://arxiv.org/abs/2507.12466 --- [CL] Simulated Language Acquisition in a Biologically Realistic Model of the Brain [MIT & Columbia University] https://arxiv.org/abs/2507.11788 --- [CL] Your LLM Knows the Future: Uncovering Its Multi-Token Prediction Potential [Apple] https://arxiv.org/abs/2507.11851

在一个数据和算法无处不在的时代，决定我们生活品质和事业高度的，恰恰是那些无法被量化的东西。