AI可可AI生活 - 节目列表

本期我们将通过几篇最新论文，看看研究者如何给机器人装上“快慢双脑”以实现实时反应，又如何用“二阶思维”为大模型精准剪枝、保留专家的协作默契。我们还将破解Adam优化器参数背后的“悬崖地图”，并派出一个小巧的“侦察兵”模型，去揪出长任务AI悄悄犯下的隐藏错误。最后，我们要警惕一碗“毒鸡汤”考题，看看被污染的基准测试是如何诱导自我进化的AI，把坏习惯固化成肌肉记忆的。 00:00:34 机器人也需要条件反射 00:05:04 裁员的智慧，你以为的庸才，可能是团队的粘合剂 00:10:16 你手里的工具，藏着一张秘密地图 00:16:28 你的AI助手，可能正在悄悄搞破坏 00:22:15 一碗“毒鸡汤”，如何带歪一个自我进化的AI 本期介绍的几篇论文： [RO] Reinforcement Learning for Real-Time Vision-Language-Action Policies [Stanford University] https://arxiv.org/abs/2609.18207 --- [LG] Higher-order pruning of experts in mixture-of-experts language models [AWS Agentic AI] https://arxiv.org/abs/2609.18916 --- [LG] Beyond Quadratic Loss：The Stability Phase Diagram of Adam [Tsinghua University] https://arxiv.org/abs/2609.18314 --- [LG] Locating Hidden Failures Makes Long-Horizon Agents More Reliable [Google DeepMind & University of California, Los Angeles & Google Research] https://arxiv.org/abs/2609.17930 --- [AI] Reflections on Trusting Trust，Revisited：Contaminating Self-Modifying AI Coding Agents with Poisoned Benchmarks [University of Washington & Georgetown University] https://arxiv.org/abs/2609.17817

本期我们将为你硬核拆解五篇极具启发性的最新论文，带你看看AI如何从“冷面判官”变身为手把手教你改论文、跑实验的“私人医生”。我们还会探讨如何利用存内计算把大模型塞进普通硬盘，并揭秘高效大模型到底为什么总爱“死记硬背”却学不会“活学活用”。最后，我们将一起见证AI如何通过“元认知操纵”掌握科学家的真实直觉，以及如何用文本优化技术揪出海量数据里隐藏的危险“潜台词”。 00:00:36 你的论文，需要一位AI私人医生 00:05:44 AI太贵？咱们把它塞进硬盘里算 00:11:07 死记硬背还是活学活用？AI的成长烦恼 00:16:16 AI的“驾驶术”，如何教会机器科学家的直觉 00:22:18 数据里的“潜台词”，我们怎么听懂？ 本期介绍的几篇论文： [CL] PaperDoctor: Evidence-Grounded and Actionable Feedback for Scientific Papers in Progress [University of Oxford & National University of Singapore & Stanford University] https://arxiv.org/abs/2609.16995 --- [LG] LLM Inference in a Flash! [UC Berkeley] https://arxiv.org/abs/2609.16161 --- [LG] On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models [Harvard University & Apple] https://arxiv.org/abs/2609.16540 --- [AI] Metacognitive Steering: Learning the Structure of Scientific Judgment [Autopoiesis Sciences] https://arxiv.org/abs/2609.16245 --- [LG] Verbalizing Subliminal Learning Effects Using Text Optimization [Stanford University] https://arxiv.org/abs/2609.16927

你有没有想过，如何让AI变得更聪明，甚至比它的老师还强？本期节目，我们将一起探索几篇最新论文带来的奇妙思路：从给AI请一位“混搭私教”，到为它建造一座“思想角斗场”进行团队作战。我们还会潜入AI的“梦境”，看看它如何复盘过去、预演未来，并顺便弄清楚它为什么有时会突然变成“复读机”。准备好了吗？让我们一起看看，这些研究如何从根源上提升AI解决复杂问题的能力。 00:00:35 给AI模型请个“混搭”私教 00:06:14 如何看见你看不到的数据？ 00:11:56 AI 的“梦境”，如何用过去预演未来 00:18:07 AI科学家的工作法，像罗马人一样建角斗场 00:24:48 AI为啥会变成“复读机”？ 本期介绍的几篇论文： [AI] Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition [MIT & NVIDIA] https://arxiv.org/abs/2609.14708 --- [LG] Synthetic Nearest Neighbors: Extending Synthetic Controls for Matrix Completion with Missing Not at Random Data [Columbia University & MIT] https://arxiv.org/abs/2609.13586 --- [CL] Dream-RSI: Recursive Self-Improvement through Evolving Worlds [Google] https://arxiv.org/abs/2609.1485 --- [AI] Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science [Google Research] https://arxiv.org/abs/2609.15983 --- [CL] Mirror, Mirror on the Wall: Prompt Echoing in Small Instruct Language Models [Warsaw University of Technology] https://arxiv.org/abs/2609.15045

今天，我们要深入AI的“大脑”，看看它是如何真正学会“思考”的。我们会探讨，是给AI请个“陪练”逐步放手，还是给它一张“地图”指引全局更有效？我们还会见证一场AI学习的“龟兔赛跑”，看看“快功夫”和“慢功夫”哪个更有前途。最后，我们将一起揭开AI如何从死记硬背走向融会贯通，以及我们该如何科学地看待它的“成绩单”。五篇最新论文，带你洞悉AI学习的底层智慧。 00:00:34 给AI请个“陪练”，然后悄悄撤走 00:04:54 AI干活，为什么喂给它地图比喂给它字典更管用？ 00:10:35 AI的快功夫与慢功夫 00:16:03 AI对齐，一份被误解的成绩单 00:20:42 AI怎么才能“活”起来，从死记硬背到融会贯通 本期介绍的几篇论文： [LG] CanvasAnneal：Curriculum Reinforcement Learning for Diffusion Language Models [Google DeepMind] https://arxiv.org/abs/2609.13060 --- [AI] Beyond Vector Similarity：Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration [Google Cloud] https://arxiv.org/abs/2609.12464 --- [CL] Breaking the Token Ceiling：Distilling Smaller， Stronger Byte Models [Meta FAIR & University of Washington, Seattle] https://arxiv.org/abs/2609.12303 --- [LG] Distortion of AI Alignment Revisited：RLHF is a Decent Utilitarian Aligner [UC Berkeley] https://arxiv.org/abs/2609.12651 --- [AI] Hierarchical Prototype Emergence in Modern Hopfield Models [Stanford University] https://arxiv.org/abs/2609.12079

你有没有想过，一群“健忘”的AI如何自发形成群体智慧？我们又该如何为AI量身定做一套“习题集”，培养出“四两拨千斤”的编程高手？本期节目，我们将一起探究几篇最新论文，看看科学家是如何通过“画地图”式的新方法进行信息检索，如何给AI做“信念体检”来判断它是否言行一致，以及如何引导AI从一团乱麻的“噪点”中走出清晰的思考路径。准备好，让我们一起解码AI思考与学习的底层智慧！ 00:00:36 AI界的“四两拨千斤”，如何养出一个小个子编程高手？ 00:05:48 AI的群体智慧，一个动作解释所有 00:11:12 信息检索的内功，从存照片到画地图 00:17:19 AI有“信念”吗？一张体检表告诉你答案 00:22:42 从一团乱麻到清晰答案的思考路径 本期介绍的几篇论文： [AI] FrogNano: Training a 4B Coding Agent via Online Task Synthesis [Froggy Team – Microsoft Research Montréal] https://arxiv.org/abs/2609.07925 --- [CL] Copying explains the collective behavior of AI agents in the wild [University of Konstanz & Intesa Sanpaolo] https://arxiv.org/abs/2609.09150 --- [IR] Generative Late-Interaction Embeddings For Visual Document Retrieval [King Abdullah University of Science and Technology (KAUST)] https://arxiv.org/abs/2609.11808 --- [AI] Beliefs and Behavior in Language Models [Toulouse School of Economics & CMU] https://arxiv.org/abs/2609.07943 --- [LG] Thinking with Looped Flows [EPFL & KAIST & University of Amsterdam] https://arxiv.org/abs/2609.11801

你有没有想过，AI天才也需要“岗前培训”才能上岗？本期我们将从几篇最新论文出发，揭秘如何为AI搭建高效的“培训工厂”，并探索如何教AI学会“做事的方法论”，而不仅仅是“堆知识”。我们还会聊聊一个有趣的问题：AI会为了讨好你而放弃原则、变成一个“老好人”吗？最后，我们将从一个全新的角度，看看合作的本质，也许就藏在最底层的成本计算里。 00:00:32 AI天才出厂后，谁给它做“岗前培训”？ 00:06:42 如何让AI的“学徒”跟上“大师”的脚步 00:11:35 AI的进化，从“知道什么”到“该做什么” 00:17:26 为什么AI会变成一个“老好人”？ 00:22:03 合作的秘密，藏在成本里 本期介绍的几篇论文： [LG] Miles v0.1: Production-Level Post-Training [RadixArk] https://arxiv.org/abs/2609.08368 --- [LG] Online Draft Co-Training for Speculative Decoding in Large-Scale, Long-Context RL Post-Training [NVIDIA] https://arxiv.org/abs/2609.07108 --- [AI] Procedural Graphs: Self-Evolving Execution Structures for LLM Agents [Google] https://arxiv.org/abs/2609.09153 --- [CL] Measuring LLM Sycophancy under Sustained Multi-Turn Pressure [Texas A&M University & University of Cincinnati] https://arxiv.org/abs/2609.09090 --- [AI] Tapes Together Strong: The Co-evolution of Computation and Cooperation [Google] https://arxiv.org/abs/2609.10817

AI如何才能学会自我进化，最终成为自己的师傅？为什么解决顶级难题要靠“AI专家团”，而不是一个超级大脑？本期节目，我们将从几篇最新论文出发，探讨AI如何从别人的失败中汲取智慧，看懂“剩饭”为何难倒英雄汉，并理解“看得懂”与“会动手”之间那道巨大的鸿沟。 00:00:26 那个“笨”徒弟，正在悄悄学会自己当师傅 00:07:02 AI解题的秘密，不是一个大脑，而是一套系统 00:11:44 AI的“偏食症”，为什么聪明的模型更讨厌“剩饭”？ 00:17:21 人工智能看得懂，但不会干 00:21:56 如何变得更聪明？答案是，多看看笨蛋是怎么想的 本期介绍的几篇论文： [LG] The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement [Shanghai Jiao Tong University] https://arxiv.org/abs/2609.11873 --- [AI] An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics [NVIDIA] https://arxiv.org/abs/2609.10712 --- [LG] Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data [Stanford University] https://arxiv.org/abs/2609.11917 --- [AI] MindTopo: Can Foundation Models Reason in Topological Space? [Northwestern University] https://arxiv.org/abs/2609.11900 --- [CL] Negative Self-Distillation: Learning to Reason by Avoiding Flaws [University of Virginia] https://arxiv.org/abs/2609.11699

今天，我们将一起“拆开”AI的大脑，看看做决策的竟然只有8个“员工”？我们还会揭秘AI排行榜的“偏科”陷阱，并为它请来一位完美的“虚拟陪练”和一位全能“秘书”。最后，再用一个简单又奇妙的几何学秘密，看穿AI决策的本质。让我们一同进入AI内部，一探究竟！ 00:00:23 AI做决策，到底需要多少“人”帮忙？ 00:05:38 AI排行榜的秘密，为什么第一名可能不是你想要的全才？ 00:11:01 给AI请个“陪练”，它就能开窍？ 00:16:09 给高手配个秘书，怎样才能让他越用越顺手？ 00:21:46 AI决策的“保守”秘密 本期介绍的几篇论文： [CL] Through the Looking Glass: Directly Reading and Writing Transformers [University of Washington] https://arxiv.org/abs/2609.10210 --- [CL] What Does MMLU Actually Measure? A Psychometric Audit of Difficulty Structure in Aggregate Benchmark Scores [Stanford University] https://arxiv.org/abs/2609.09372 --- [LG] World-Time Compute with Verified Code World Models [Quome, Inc.] https://arxiv.org/abs/2609.09163 --- [CL] Osprey: Target-agnostic Pre-training Makes Stronger Drafters in Speculative Decoding [Together AI] https://arxiv.org/abs/2609.09338 --- [LG] Exact-Form Regret for Gradient Descent, Mirror Descent and Follow-the-Regularized-Leader [MIT] https://arxiv.org/abs/2609.09466

想知道AI的“大脑”里，是不是真的在上演一场场激烈的内部辩论？为什么有时候教它，掐头去尾、只给起点和终点，反而能让它学得更快？而面对超级难题，又是怎样一个“笨办法”在引导它一步步走向正确答案？本期节目，我们将深入几篇最新论文，聊聊AI“师傅”如何带出超越自己的“徒弟”，并揭开为什么你手机里的AI和新闻里的跑分冠军，可能是两回事。 00:00:31 AI的“自我否定”，我们误解了它的工作方式 00:05:22 老师傅的旧地图，怎么给新车导航？ 00:10:17 你用的AI，和新闻里的AI，是两回事 00:16:00 为什么掐头去尾，反而教得更好？ 00:19:50 为什么聪明的AI，也需要一个笨办法？ 本期介绍的几篇论文： [CL] LLM Layers Immediately Correct Each Other [UC Berkeley] https://arxiv.org/abs/2609.07876 --- [LG] Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation [KAIST AI] https://arxiv.org/abs/2609.08798 --- [AI] API Benchmark Scores Do Not Reliably Transfer to Chatbot Interfaces [Stanford University] https://arxiv.org/abs/2609.08861 --- [CL] Revisiting Complete Reasoning Traces for Post-Training [NAVER AI Lab] https://arxiv.org/abs/2609.07103 --- [LG] Long-Horizon Language Model Reinforcement Learning via Progressive Point Matching [UC Berkeley] https://arxiv.org/abs/2609.07303

你是否想过，AI不仅能当助手，更能成为科学家的“寻宝图”，预测未来的新发现？本期我们将一起探讨，AI如何学会从“挤牙膏”式写作进化到“一步到位”的神奇魔法，并首次“窥探”它的大脑，看看它是否真的理解了“2+5”和“二加五”的区别。我们还会揭示，如何通过一张“地图”让AI读懂万卷书，以及它学习掌握“祖传手艺”的秘密。 00:00:29 AI 如何成为科学家的「寻宝图」 00:06:12 语言模型，告别“挤牙膏”时代 00:11:33 会做“2+5”，为何不会“二加五”？我们终于有办法偷看AI的大脑了 00:17:29 给AI一张地图，它能更好地为你读书 00:21:54 AI如何拥有“祖传手艺”？ 本期介绍的几篇论文： [LG] Hakken: Predicting future discoveries to fill the gaps in today's knowledge [SonyAI] https://arxiv.org/abs/2609.04494 --- [LG] Distilled Continuous Diffusion Language Models Can Write Code in Few Steps---or One [Duke University & Tsinghua University] https://arxiv.org/abs/2609.04531 --- [CL] Shared circuits predict whether LLMs generalize across formats in arithmetic reasoning [MIT] https://arxiv.org/abs/2609.04463 --- [AI] STAIR (STructure Aware Information Retriever): A novel dataset and LLM based retriever for document structure augmentation [IBM] https://arxiv.org/abs/2609.03874 --- [AI] SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams [National University of Singapore & Institute of Advanced Intelligence and Computing (IAIC)] https://arxiv.org/abs/2609.02217

你有没有觉得，AI时而像个无所不能的天才，时而又像个会钻牛角尖的“笨小孩”？本期节目，我们将通过几篇最新论文，一探究竟：为何AI会固执地采纳错误答案，又为何会被最简单的“跟我读”骗术“催眠”？同时，我们也将看到AI如何化身“科研总管”，以及一份好的“设计图”为何在未来可能比代码本身更值钱。准备好，让我们一起揭开AI这些既矛盾又迷人的行为背后的秘密。 00:00:33 AI的“小固执”，为什么它信你，却不听你的？ 00:07:48 未来，你的代码可能一文不值 00:13:33 为什么AI解难题，也会钻牛角尖？ 00:18:37 为什么AI会被最简单的骗术“带偏”？ 00:23:09 让AI当科研总管，是一种什么体验？ 本期介绍的几篇论文： [CL] Evidence Integration in Large Language Models [MIT] https://arxiv.org/abs/2609.04290 --- [AI] Design Docs Are All You Need: An AI-native Machine-Learning Performance Tool [Google DeepMind & MIT] https://arxiv.org/abs/2609.05364 --- [LG] Fractal basins trap latent reasoning [The University of Texas at Austin] https://arxiv.org/abs/2609.04963 --- [AI] Repeat-After-Me: Black-Box Adaptive Visual Prompt Injection [UC Berkeley & FAIR at Meta] https://arxiv.org/abs/2609.04533 --- [AI] La Agente Óptima: Towards Agentic Self-Driving Laboratories [University of Toronto & 700 University Ave] https://arxiv.org/abs/2609.04564

你有没有想过，我们能用音乐均衡器的思路，让AI画画提速40%？本期节目，我们将一起钻进AI的“大脑”，看看给它一条笔直的“高速公路”为什么反而会“堵车”，以及如何用一个“科学沙盒”来分辨AI究竟是真正的科学家，还是只会刷题的学霸。我们还会聊到一篇最新论文，它发现了一个几乎被所有人忽略的“小开关”，却能成为大模型训练的超级加速器。让我们一起从这些最新论文中，发现那些大道至简的AI智慧吧！ 00:00:34 AI绘画的“均衡器” 00:04:47 大道至简，AI 设计蛋白质，需要绕多大的弯？ 00:09:13 解锁AIGC的终极速度，从颠簸小路到笔直高速 00:14:36 给AI一个沙盒，看它能不能成为科学家 00:20:56 大模型微调，一个被忽略的开关 本期介绍的几篇论文： [CV] Balancing Frequencies and Pixels in Flow Matching [CNRS] https://arxiv.org/abs/2609.02748 --- [LG] SimpleDesign：A Joint Model for Protein Sequence and Structure Codesign [Apple] https://arxiv.org/abs/2609.03377 --- [CV] A Lagrangian View of Flow Matching [Google] https://arxiv.org/abs/2609.00198 --- [AI] Science sandboxes measure the scientific capability of AI agents [The Broad Institute of MIT and Harvard & The Jackson Laboratory & Sutter Hill Ventures] https://arxiv.org/abs/2608.30165 --- [LG] Normalized Low-Rank Adaptation [Yuanshi Intelligence & Microsoft Research] https://arxiv.org/abs/2608.31036 ---[CV] Balancing Frequencies and Pixels in Flow Matching [CNRS] https://arxiv.org/abs/2609.02748 --- [LG] SimpleDesign：A Joint Model for Protein Sequence and Structure Codesign [Apple] https://arxiv.org/abs/2609.03377 --- [CV] A Lagrangian View of Flow Matching [Google] https://arxiv.org/abs/2609.00198 --- [AI] Science sandboxes measure the scientific capability of AI agents [The Broad Institute of MIT and Harvard & The Jackson Laboratory & Sutter Hill Ventures] https://arxiv.org/abs/2608.30165 --- [LG] Normalized Low-Rank Adaptation [Yuanshi Intelligence & Microsoft Research] https://arxiv.org/abs/2608.31036