你有没有想过,最聪明的决策,也许是先用最小的力气排除所有错误选项?当AI变得越来越话痨时,我们该如何给它请一位“效率教练”?为了把强大的AI装进你的手机,科学家又想出了怎样统一又精简的“节食计划”?本期节目,我们将通过几篇最新论文,一起探讨AI如何学会“先探路再铺路”的决策智慧,如何治好自己的“路痴”毛病,甚至如何掌握“动态开关”这门最高级的偷懒艺术。 00:00:33 聪明人的偷懒指南,如何用最少的力气,走最对的路? 00:07:16 AI话痨怎么办?聪明还得会省钱 00:12:27 AI的“节食计划”,如何在你的手机里装下一个图书馆? 00:17:42 大模型越来越聪明,为什么还是个“路痴”? 00:22:45 为什么说,最高级的AI,必须学会“偷懒”? 本期介绍的几篇论文: [CL] Blazing the trails before beating the path: Sample-efficient Monte-Carlo planning [INRIA Lille & Google DeepMind] https://arxiv.org/abs/2604.14974 --- [CL] CROP: Token-Efficient Reasoning in Large Language Models via Regularized Prompt Optimization [Google LLC & Purdue University] https://arxiv.org/abs/2604.14214 --- [IR] A Unified Model and Document Representation for On-Device Retrieval-Augmented Generation [University of Massachusetts Amherst & Google] https://arxiv.org/abs/2604.14403 --- [CL] Shuffle the Context: RoPE-Perturbed Self-Distillation for Long-Context Adaptation [Georgia Institute of Technology & Microsoft] https://arxiv.org/abs/2604.14339 --- [CL] Compressed-Sensing-Guided, Inference-Aware Structured Reduction for Large Language Models [UC Berkeley] https://arxiv.org/abs/2604.14156
你有没有想过,一个学得更久的AI“尖子生”,为什么反而忘得更快?或者,想让AI更懂英语,最好的方法竟然是教它别的语言?本期节目,我们将一口气解锁五篇最新论文带来的“反常识”洞见。我们会发现,决定AI效率的瓶颈可能不是算力而是“管理”,与AI对话的成本可以靠一本“字典”轻松打个二折,而一个好的AI模拟世界,追求的不是“长得像”,而是“反应像”。 00:00:32 大模型训练的悖论,为什么学得越久,忘得越快? 00:06:02 AI的效率瓶颈,不是算力,是“管理” 00:12:33 想让AI更懂英语?那就别只喂它英语 00:18:46 跟AI对话,如何省下80%的话费? 00:24:39 你的“差不多”不是我的“差不多”,如何让AI的模拟世界更靠谱? 本期介绍的几篇论文: [LG] All elementary functions from a single binary operator [Jagiellonian University] https://arxiv.org/abs/2603.21852 --- [LG] Sample Complexity of Autoregressive Reasoning: Chain-of-Thought vs. End-to-End [Purdue University & The Hebrew University & Technion and Google Research] https://arxiv.org/abs/2604.12013 --- [CL] Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature [Central University of Finance and Economics & Beijing Institute of Technology & TsingyuAI] https://arxiv.org/abs/2604.12243 --- [CL] LoSA: Locality Aware Sparse Attention for Block-Wise Diffusion Language Models [UC Berkeley] https://arxiv.org/abs/2604.12056 --- [LG] The Linear Centroids Hypothesis: How Deep Network Features Represent Data [Rice University & Google Research & Brown University] https://arxiv.org/abs/2604.11962
你有没有想过,整个科学计算器也许只需要两个按键就能实现?或者,AI偷懒的秘诀竟是只用20%的精力,就能完成90%的工作?最新的一些研究,正从这些奇妙的角度,刷新我们对智能、效率和知识的认知。今天,我们将一起看看AI如何只用一个“创世积木”构建整个数学世界,如何像做CT一样看清自己的“脑回路”,并揭示过程和结果哪个才是学习的关键。准备好,一场思维风暴马上开始! 00:00:36 你的科学计算器,其实只需要两个键 00:05:01 学会一个本事,过程和结果哪个更重要? 00:13:05 如何像高手一样,“看见”知识的未来? 00:19:31 AI偷懒的艺术,为什么只做20%的工作,能得到90%的结果? 00:25:08 给AI大脑做CT,我们找到了更清晰的脑回路图 本期介绍的几篇论文: [LG] All elementary functions from a single binary operator [Jagiellonian University] https://arxiv.org/abs/2603.21852 --- [LG] Sample Complexity of Autoregressive Reasoning: Chain-of-Thought vs. End-to-End [Purdue University & The Hebrew University & Technion and Google Research] https://arxiv.org/abs/2604.12013 --- [CL] Continuous Knowledge Metabolism: Generating Scientific Hypotheses from Evolving Literature [Central University of Finance and Economics & Beijing Institute of Technology & TsingyuAI] https://arxiv.org/abs/2604.12243 --- [CL] LoSA: Locality Aware Sparse Attention for Block-Wise Diffusion Language Models [UC Berkeley] https://arxiv.org/abs/2604.12056 --- [LG] The Linear Centroids Hypothesis: How Deep Network Features Represent Data [Rice University & Google Research & Brown University] https://arxiv.org/abs/2604.11962
你有没有想过,要让AI变得更聪明,除了让它“读万卷书”,我们还能不能让它在虚拟世界里“行万里路”,像玩“模拟人生”一样学会物理?当很多聪明的算法凑在一起反而“掉链子”时,我们如何用“乐高积木”的思路化繁为简?这一期,我们将一起探寻几份最新论文带来的启发:从像婴儿一样在“思想实验”中探索世界,到用一张“知识地图”代替“知识词典”来解决复杂问题,甚至让AI学会“自我怀疑”,从而变得又快又好。准备好了吗?让我们一起出发! 00:00:38 AI版“模拟人生”让机器在虚拟世界里学会物理 00:05:56 从1到N如何让你的数据分析稳上加稳? 00:12:14 AI养娃我们可能找到了让机器像婴儿一样学习的秘密 00:18:01 高手解决问题,靠的是地图,而不是词典 00:24:14 AI的自我怀疑,一个让大模型又快又好的新思路 本期介绍的几篇论文: [LG] Solving Physics Olympiad via Reinforcement Learning on Physics Simulators [CMU & Lambda] https://arxiv.org/abs/2604.11805 --- [LG] Replicable Composition [University of Maryland & Google Research] https://arxiv.org/abs/2604.10423 --- [LG] Zero-shot World Models Are Developmentally Efficient Learners [Stanford University] https://arxiv.org/abs/2604.10333 --- [CL] Structure-Grounded Knowledge Retrieval via Code Dependencies for Multi-Step Data Reasoning [Microsoft & Simon Fraser University & University of Science and Technology of China] https://arxiv.org/abs/2604.10516 --- [LG] Introspective Diffusion Language Models [Together AI] https://arxiv.org/abs/2604.11035
你有没有想过,我们能像做微创手术一样,在AI思考的瞬间“拨乱反正”,引导它向善吗?或者,让昂贵的AI训练学会“温故知新”,把扔掉的经验变废为宝?本期节目,我们将一起探索几篇最新论文,看看科学家们如何教会AI遵守自己立下的规矩,如何让它既会“看路”又会“造景”,甚至,如何为它补上一堂生动的物理课,让它的想象力更符合现实。准备好了吗?让我们马上出发! 00:00:34 给AI的大脑装一个“概念导航” 00:06:53 AI训练的高效秘诀,好东西值得再用一次 00:12:07 如何看穿一个AI的“人设”? 00:16:56 AI新思路,想看清世界,先学会走路 00:23:07 为什么AI生成的视频总感觉“假”?答案藏在物理学里 本期介绍的几篇论文: [LG] Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs [University of Pennsylvania & Amazon] https://arxiv.org/abs/2604.08846 --- [LG] Efficient RL Training for LLMs with Experience Replay [FAIR at Meta] https://arxiv.org/abs/2604.08706 --- [CL] Do LLMs Follow Their Own Rules? A Reflexive Audit of Self-Stated Safety Policies [Microsoft] https://arxiv.org/abs/2604.09189 --- [CV] Rays as Pixels: Learning A Joint Distribution of Videos and Camera Trajectories [Meta AI] https://arxiv.org/abs/2604.09429 --- [CV] PhysInOne: Visual Physics Learning and Reasoning in One Suite [vLAR Group] https://arxiv.org/abs/2604.09415
你是否想过,为什么你的AI助理连订张机票都费劲?本期节目,我们将一起给AI来一场“真实世界”的大考,看看它究竟能得多少分。我们还会揭秘如何不给模型动手术,只靠“补课”就让它说话速度提升1.7倍。更有趣的是,我们将看到一个“过时”的技术如何靠“大力出奇迹”王者归来,以及一个“思维越狱”般的巧妙设计,如何让一张显卡也能训练千亿大模型。最后,我们还会认识一支能帮你写论文的“AI管弦乐队”。准备好了吗?让我们马上进入AI前沿的奇妙世界。 00:00:39 你的AI助理,离真正上岗还有多远? 00:05:50 让AI大模型提速,只需要“补课”就够了 00:10:13 老树发新芽,一个被人小瞧的技术,如何靠“笨办法”王者归来? 00:15:46 AI的“昂贵误会”,我们都搞错瓶颈了吗? 00:22:05 你的下一个写作搭档,可能不是一个人 本期介绍的几篇论文: [CL] ClawBench: Can AI Agents Complete Everyday Online Tasks? [University of British Columbia & Vector Institute] https://arxiv.org/abs/2604.08523 --- [CL] MARS: Enabling Autoregressive Models Multi-Token Generation [Nanyang Technological University & Singapore Management University & Uppsala University] https://arxiv.org/abs/2604.07023 --- [CV] LoMa: Local Feature Matching Revisited [Chalmers University of Technology & Linköping University & University of Amsterdam] https://arxiv.org/abs/2604.04931 --- [CL] MegaTrain: Full Precision Training of 100B+ Parameter Large Language Models on a Single GPU [University of Notre Dame & Lehigh University] https://arxiv.org/abs/2604.05091 --- [LG] PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing [Google] https://arxiv.org/abs/2604.05018
如果你的电脑本身就是一个神经网络,没有CPU会怎样?如果AI想学一门手艺,得先让另一个AI给它建个“驾校”呢?本期节目,我们将从五篇最新论文出发,聊聊AI如何从2D世界“透视”三维,如何用量子芯片实现超级压缩,并最终探讨一个直击灵魂的问题:AI正在让我们变聪明,还是变懒? 00:00:27 你的下一台电脑,为什么可能没有CPU? 00:09:13 AI想“上岗”,得先过哪一关? 00:14:57 从“看扁”到“透视”,AI如何拥有3D“世界观”? 00:20:36 你的下一个U盘,可能是个量子芯片 00:26:12 AI,是你的“外挂”还是你的“拐杖”? 本期介绍的几篇论文: [LG] Neural Computers [Meta AI] https://arxiv.org/abs/2604.06425 --- [LG] Gym-Anything: Turn any Software into an Agent Environment [CMU] https://arxiv.org/abs/2604.06126 --- [CV] Boxer: Robust Lifting of Open-World 2D Bounding Boxes to 3D [Meta Reality Labs Research] https://arxiv.org/abs/2604.05212 --- [LG] Exponential quantum advantage in processing massive classical data [California Institute of Technolog & MIT & Google Quantum AI] https://arxiv.org/abs/2604.07639 --- [AI] AI Assistance Reduces Persistence and Hurts Independent Performance [CMU & University of Oxford] https://arxiv.org/abs/2604.04721
想让AI更聪明,我们总觉得要给它看更多、学更多,但如果我告诉你,真正的秘诀恰恰相反呢?本期节目,我们将一起探索几篇最新的AI论文,看看科学家是如何教会AI“选择性遗忘”,又是如何给AI做“脑CT”来判断它是不是在“假装努力”。我们还会聊到,如何打造一个能快速学会任何人大脑“方言”的超级解码器,以及怎样只用1%的精力,就让AI帮你“看完”一部长电影。准备好了吗?让我们一起刷新对AI的认知! 00:00:37 按下删除键之后,东西就真的消失了吗? 00:06:54 造一把“万能钥匙”?不如当个“超级锁匠” 00:11:24 想让AI更博学?先给它少看点书 00:16:21 给AI做个体检,我们怎么知道它不是在瞎蒙? 00:22:01 如何只用1%的精力,看完一部长电影的精华? 本期介绍的几篇论文: [LG] Is your algorithm unlearning or untraining? [Google] https://arxiv.org/abs/2604.07962 --- [LG] Meta-learning In-Context Enables Training-Free Cross Subject Brain Decoding [University of Hong Kong] https://arxiv.org/abs/2604.08537 --- [CL] Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts [Apple & National University of Singapore] https://arxiv.org/abs/2604.08519 --- [LG] Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings [University of Delaware & George Mason University] https://arxiv.org/abs/2604.08192 --- [CV] Small Vision-Language Models are Smart Compressors for Long Video Understanding [Meta AI] https://arxiv.org/abs/2604.08120
你敢让AI帮你摇号抽奖吗?本期节目,我们将从几篇最新的AI论文出发,揭示AI在“随机”这件事上出人意料的偏见。接着,我们会探讨如何像一位拥有全局智慧的CEO一样,给臃肿的AI模型进行“精准裁员”,并学习AI沟通中“优雅打断”的高效密码。最后,我们将一起探寻AI是否存在“思想深度的天花板”,以及如何把一个“笨徒弟”模型,调教成一位善用工具的“老师傅”。准备好了吗?让我们一起潜入AI的前沿思想深海! 00:00:42 为什么你老板让你用AI摇号,你得多个心眼? 00:06:07 从“平均砍”到“精准剪枝”,AI瘦身中的全局智慧 00:12:10 沟通的高效密码,如何优雅地“打断”别人 00:18:04 AI的“思想深度”有没有天花板? 00:24:24 如何把一个“笨徒弟”,调教成“老师傅”? 本期介绍的几篇论文: [CL] The Illusion of Stochasticity in LLMs [Google DeepMind] https://arxiv.org/abs/2604.06543 --- [CL] Does a Global Perspective Help Prune Sparse MoEs Elegantly? [University of Rochester & Flatiron Institute] https://arxiv.org/abs/2604.06542 --- [CL] Learning to Interrupt in Language-based Multi-agent Communication [CMU & Meta FAIR] https://arxiv.org/abs/2604.06452 --- [LG] The Depth Ceiling: On the Limits of Large Language Models in Discovering Latent Planning [University of Cambridge & Imperial College London & MIT] https://arxiv.org/abs/2604.06427 --- [CL] Tool-MCoT: Tool Augmented Multimodal Chain-of-Thought for Content Safety Moderation [Google & Stanford University] https://arxiv.org/abs/2604.06205
你有没有想过,未来的AI不仅不会“健忘”,还拥有自己的“临时记忆”?它不仅能解数学题,更能帮你绘制整个“数学宇宙”的地图?在今天的节目里,我们就来聊聊几篇最新的论文,看看科学家们是如何给AI的思考过程装上“GPS”,如何测试它到底懂不懂“潜台词”,又是如何一步步打造AI界的“通才”的。准备好,我们马上出发! 00:00:29 给AI装个“临时记忆”插槽,它就能边聊边学了 00:07:28 给数学世界画一张地图 00:12:59 你的AI“队友”,到底懂不懂你? 00:18:23 给AI的思考过程装上一个GPS 00:23:37 AI界的“通才”是如何炼成的? 本期介绍的几篇论文: [LG] In-Place Test-Time Training [ByteDance Seed] https://arxiv.org/abs/2604.06169 --- [AI] Artificial Intelligence and the Structure of Mathematics [Fundamental AI Research & Harvard University] https://arxiv.org/abs/2604.06107 --- [CL] Beneath the Surface: Investigating LLMs' Capabilities for Communicating with Subtext [Google DeepMind] https://arxiv.org/abs/2604.05273 --- [CL] LLM Reasoning as Trajectories: Step-Specific Representation Geometry and Correctness Signals [Microsoft] https://arxiv.org/abs/2604.05655 --- [LG] MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning [MIRAI & AXXX] https://arxiv.org/abs/2604.05943
今天我们来聊聊AI的“进化”新思路:当AI不再迷信单一架构,而是学会了“混搭”;当给它一把更强的锤子,它反而盖不好房子时,我们该怎么办?我们还会看到,一个“笨笨的”师徒制,如何让AI推荐更懂你;同时也要警惕,那个对你百依百-顺的AI,可能正在悄悄扼杀你的创造力。最后,我们将揭秘如何给AI装上一个“程序员的脑子”,让它学会三思而后行。 00:00:32 AI大模型内战,“万能插座”遇到了新对手 00:08:10 为什么给你一把好“锤子”,你反而盖不好房子? 00:15:07 如何让AI更懂你?秘密可能藏在“笨办法”里 00:20:29 AI越听话,你就越平庸? 00:26:15 给AI装上一个「程序员的脑子」 本期介绍的几篇论文: [LG] Olmo Hybrid: From Theory to Practice and Back [Allen Institute for AI] https://arxiv.org/abs/2604.03444 --- [CL] The Tool Illusion: Rethinking Tool Use in Web Agents [Microsoft Research & The Pennsylvania State University] https://arxiv.org/abs/2604.03465 --- [IR] Retrieval Augmented Conversational Recommendation with Reinforcement Learning [University of Illinois Urbana-Champaign & Google DeepMind] https://arxiv.org/abs/2604.04457 --- [CL] Lighting Up or Dimting Down? Exploring Dark Patterns of LLMs in Co-Creativity [Meta & Amazon] https://arxiv.org/abs/2604.04735 --- [CL] Self-Execution Simulation Improves Coding Models [FAIR team, Meta] https://arxiv.org/abs/2604.03253
你有没有想过,AI是如何学会像人一样思考的?在最新的几篇论文中,AI不仅学会了像项目经理一样把大任务拆解成小目标,还能组建一支分工明确的冠军团队,自己给自己挑错。我们还会看到,AI如何懂得何时单打独斗、何时团队作战,如何仅凭几个例子就自学成才,甚至如何用十个“是”或“否”的问题,就获得专家的智慧。今天,就让我们一起揭开这些AI“超能力”背后的朴素智慧。 00:00:35 想成大事?你得先学会当自己的“项目经理” 00:07:15 AI冠军养成记,一个“草台班子”的制胜之道 00:12:20 人多真能力量大吗?AI世界的“个人英雄”与“团队作战” 00:17:58 如何用3个例子,教会AI一整本书? 00:23:22 十个“是”或“否”,如何让AI“小学生”拥有“博士”的智慧? 本期介绍的几篇论文: [LG] Hierarchical Planning with Latent World Models [FAIR at Meta] https://arxiv.org/abs/2604.03208 --- [AI] GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning [DeepReinforce Team] https://arxiv.org/abs/2604.02721 --- [CL] Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets [Stanford University] https://arxiv.org/abs/2604.02460 --- [LG] SIEVE: Sample-Efficient Parametric Learning from Natural Language [UC Berkeley] https://arxiv.org/abs/2604.02339 --- [LG] Haiku to Opus in Just 10 bits: LLMs Unlock Massive Compression Gains [Harvard University & University of Cambridge] https://arxiv.org/abs/2604.02343
与播客爱好者一起交流
添加微信好友,获取更多播客资讯
播放列表还是空的
去找些喜欢的节目添加进来吧