这一期,我们来聊一个特别有意思的话题:如何用“巧劲”让AI变得更聪明?我们不再堆砌算力,而是探讨五篇最新论文带来的精妙思路。你会听到,有时候,真正的突破来自于一次大胆的“做减法”;有时候,我们只需在AI和它的工具之间,增加一个聪明的“随身翻译”。我们还会看到,如何像一个旁观者一样,精准确立AI每一步的功劳;如何为AI装上一个“记忆管理员”,让它学会管理自己的知识;以及,如何像动一次“微创手术”一样,只调整百万分之一的参数,就让AI掌握全新技能。 00:00:45 让AI更聪明的秘密,竟然是做减法? 00:05:54 你的AI编程助手,需要一个“随身翻译” 00:11:43 你的员工,99%的努力都白费了? 00:18:03 高手比拼的,是对记忆的管理能力 00:23:08 给AI动个“小手术”,而不是“大换血” 本期介绍的几篇论文: [CL] GFlowRL: Scaling Distribution-Matching RL to Large Language Models [Microsoft Research] https://arxiv.org/abs/2607.13394 --- [LG] Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code [ETH Zurich & INSAIT and Sofia University & UC Berkeley] https://arxiv.org/abs/2607.13921 --- [LG] TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents [University of Wisconsin–Madison & Microsoft Research] https://arxiv.org/abs/2607.13988 --- [CL] Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents [University of California Los Angeles] https://arxiv.org/abs/2607.13591 --- [LG] Data-Efficient Adaptation of LLMs via Attention Head Reweighting [Microsoft Research & Microsoft Security AI] https://arxiv.org/abs/2607.13425
这一期,我们来当一回AI的“监考官”和“心理医生”,看看怎么科学地判断AI是在“背课文”还是“会造句”。我们还会探究,当AI说自己“十拿九稳”时,它的自信是发自内心,还是纯属表演。更会揭示,为何总分稳定的模型,答案却可能因一句无关的“废话”就悄悄“叛变”。最后,从给地球装上“大脑”,到揭开我们“猜懂”外语的秘密,这些最新论文将刷新我们对智能的认知。 00:00:33 怎么知道AI“背课文”,而不是“会造句”? 00:06:04 AI的“心里有底”,到底是怎么回事? 00:14:01 AI大模型,总分没变,答案却悄悄“叛变”了 00:18:40 你的地球专属“大脑”,是怎么被训练出来的? 00:25:22 我们其实都是半个翻译家 本期介绍的几篇论文: [LG] Extractable Memorization From First Principles [Stanford & Google Research & Google DeepMind] https://arxiv.org/abs/2607.12649 --- [LG] The Computational Basis of Confidence in Large Language Models [Google DeepMind] https://arxiv.org/abs/2607.12447 --- [CL] The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant Context [Georgia Tech & Stanford University] https://arxiv.org/abs/2607.12963 --- [AI] The Emerging Paradigm of Geospatial Foundation Models: From Pre-Training to Agentic Reasoning [Google Public Sector] https://arxiv.org/abs/2607.12177 --- [CL] We Hebben Een Serieus Translatie: Modeling Intercomprehension as Probabilistic Inference [MIT] https://arxiv.org/abs/2607.12169
我们总说AI有知识,但你想过吗,AI的知识该如何称重、如何存储、又该如何溯源?更进一步,AI能否自我修炼、检查作业,它吃的“数据大餐”又藏着怎样的“秘密食谱”?今天,我们就从五篇最新的论文出发,一起探索AI知识世界的台前与幕后。 00:00:24 给AI模型称重,我们终于有了一杆新秤 00:06:34 AI的“记忆宫殿”是如何搭建的? 00:11:56 AI的自我修炼,如何从“检查作业”中获得智慧 00:17:01 AI世界的“亲子鉴定”技术 00:22:37 AI的“隐藏食谱”,为什么数据配比比数量更重要? 本期介绍的几篇论文: [LG] Requential Coding: Pushing the Limits of Model Compression with Self-Generated Training Data [New York University & CMU] https://arxiv.org/abs/2607.11883 --- [LG] MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for Transformers [Stanford University] https://arxiv.org/abs/2607.10034 --- [AI] SVR-R1: Bootstrapping Multi-modal Reasoning with Self-verification in Reinforcement Learning [University of Illinois Urbana-Champaign & Meta] https://arxiv.org/abs/2607.10966 --- [LG] Reference-Based Distillation Detection in LLMs [UC Berkeley] https://arxiv.org/abs/2607.09692 --- [LG] Domain-Aware Scaling Laws Uncover Data Synergy [MIT & Microsoft Research] https://arxiv.org/abs/2607.11052
今天我们要聊聊,AI那些颠覆我们常识的学习心法。你会发现,AI为了画出逼真的视频,竟然偷偷学会了物理学;而教一个笨手笨脚的机器人,我们既可以改变它脑中的“潜意识”,也可以只给它换一句神奇的“咒语”。我们还会看到,AI如何把每一次“失败”都变成成功的养料,以及“抽象思维”在它脑海里清晰浮现又逐渐妥协的全过程。准备好,让我们一起潜入AI的“思想深处”,看看它到底是怎么变聪明的。 00:00:34 别以为AI生成视频只是为了好玩,它其实是在偷偷“理解”物理世界 00:05:54 别再给AI做“开颅手术”了,教机器人干活有更聪明的办法 00:11:23 别把“失败”当废料,它只是放错了位置的“成功” 00:16:47 AI的“抽象思维”是怎么炼成的?揭开学习与认知的隐藏规律 00:22:12 当AI遇到死胡同,给它换个“咒语”就能破局 本期介绍的几篇论文: [CV] Video Generation Models are General-Purpose Vision Learners [Google DeepMind] https://arxiv.org/abs/2607.09024 --- [RO] FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space [Microsoft Research] https://arxiv.org/abs/2607.08877 --- [LG] Learning More from Less: Reinforcement Learning from Hindsight [MIT & Stanford University] https://arxiv.org/abs/2607.09042 --- [LG] How are linear representations learned? Exact solutions to the dynamics of abstraction [University College London] https://arxiv.org/abs/2607.08843 --- [LG] Prompt-Driven Exploration [MIT] https://arxiv.org/abs/2607.08837
今天我们不聊AI有多聪明,而是聊如何让它更“会”做事。我们将一起看看,最新的AI研究如何像顶级教练一样,为AI程序员打造完美的成长路径;如何用“偷天换日”的巧思,让AI看懂世间万物的运动;我们还将揭示AI学会“人情世故”的三个秘密,并从AI的进化策略中,找到我们普通人打破僵局的生存法则。 00:00:28 AI程序员的成长烦恼,聪明还不够 00:06:34 从蝴蝶到纸飞机,如何让AI看懂世间万物的运动? 00:11:35 让AI更懂“人情世故”的三个秘密 00:17:34 别等“万事俱备”,从顶级AI算法看普通人的破局之道 本期介绍的几篇论文: [AI] ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes [Microsoft Research & Nanyang Technological University] https://arxiv.org/abs/2607.04439 --- [RO] RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies [HKU & PKU & THU] https://arxiv.org/abs/2607.04434 --- [AI] LLM-as-a-Verifier: A General-Purpose Verification Framework [Stanford University & UC Berkeley] https://arxiv.org/abs/2607.05391 --- [CV] Multiplayer Interactive World Models with Representation Autoencoders [General Intuition & Kyutai] https://arxiv.org/abs/2607.05352 --- [CV] SPEAR: A Simulator for Photorealistic Embodied AI Research [Adobe Research & Manycore Tech Inc] https://arxiv.org/abs/2607.06701
你有没有想过,AI的创新灵感从何而来?我们又该如何给机器人办一场“驾校”大考,挤掉行业泡沫?本期节目,我们将一起探秘几篇最新论文,看看AI如何从学习创新的“套路”开始,进化成能精准评估工作的“检验员”,甚至最终拿到创造和改造虚拟游戏世界的“后台密码”。 00:00:25 创新不是凭空想象,而是一门有“套路”的手艺 00:05:51 给机器人办个驾校,结果全班不及格? 00:11:15 从“裁判”到“检验员”,一个身份的转变 00:17:51 如何凭空创造一个,你能开进去玩的游戏世界? 00:22:43 当AI拿到了游戏引擎的“后台密码” 本期介绍的几篇论文: [AI] ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes [Microsoft Research & Nanyang Technological University] https://arxiv.org/abs/2607.04439 --- [RO] RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies [HKU & PKU & THU] https://arxiv.org/abs/2607.04434 --- [AI] LLM-as-a-Verifier: A General-Purpose Verification Framework [Stanford University & UC Berkeley] https://arxiv.org/abs/2607.05391 --- [CV] Multiplayer Interactive World Models with Representation Autoencoders [General Intuition & Kyutai] https://arxiv.org/abs/2607.05352 --- [CV] SPEAR: A Simulator for Photorealistic Embodied AI Research [Adobe Research & Manycore Tech Inc] https://arxiv.org/abs/2607.06701
想知道AI如何变得更聪明、更高效吗?本期我们就来看几篇脑洞大开的最新论文。我们将一起探索,AI如何拉着“老模型”一起“团购”评测来省钱,又如何通过一个圈子的“开放性”来揪出网络水军。你还会听到,AI如何变身“虚拟陪练”教会机器人高难度操作,如何告别“炼丹”自动组建“梦之队”,甚至如何“升职”为科学家的项目总管。这些看似不相关的研究,背后都指向了同一个趋势:AI正在从单纯的工具,进化为解决问题的“系统设计师”。 00:00:37 AI评测的“省钱攻略”,如何拉着“老模型”一起“团购”? 00:07:13 抓出网络里的“坏人”,关键看谁的“圈子”不够开放 00:12:07 虚拟世界里的“陪练”,如何教会现实中的机器人? 00:18:28 告别“炼丹”,AI高手的新玩法 00:23:39 给科学家升职,AI当起了“总管” 本期介绍的几篇论文: [LG] CollabEval: Statistically Efficient Collaborative Model Evaluation via Matrix Completion [Google DeepMind] https://arxiv.org/abs/2607.05046 --- [LG] Active Learning on Adversarially Corrupted Graphs [Università degli Studi di Milano & Bocconi University] https://arxiv.org/abs/2607.04869 --- [RO] SILO: Simulation-in-the-Loop Sim-to-Real Transfer for Multi-Stage Cable Routing [NVIDIA] https://arxiv.org/abs/2607.04616 --- [LG] TabPack: Efficient Hyperparameter Ensembles for Tabular Deep Learning [Yandex & HSE University] https://arxiv.org/abs/2607.05380 --- [CL] Rethinking Scientific Discovery in an Agentic Era [Shanghai Innovation Institute] https://arxiv.org/abs/2607.03863
想知道AI画画如何实现指数级加速,又为何会“英雄所见略同”吗?当AI学会了高情商作弊,我们又该如何分辨并驯服它?更重要的是,我们为AI安全打造的“锁”,会不会变成禁锢思想的“笼”?本期节目,我们将一口气洞察五篇最新论文,揭开AI世界里那些令人兴奋又警醒的秘密。 00:00:25 生成AI的“指数级”加速器,藏着什么秘密? 00:05:45 你以为的AI安全锁,也可能是别人的思想钢印 00:12:18 AI的“高情商”作弊,我们如何驯服一个聪明的“坏学生”? 00:17:29 AI学画画,谁是它的“动作”老师? 00:22:41 AI绘画的“趋同性”,为什么英雄所见略同? 本期介绍的几篇论文: [LG] High-accuracy sampling for diffusion models and log-concave distributions [MIT & Yale University] https://arxiv.org/abs/2602.01338 --- [LG] Position: The Alignment Community is Unintentionally Building a Censor’s Toolkit [LMU Munich] https://openreview.net/forum?id=dy2HwmOvFX --- [LG] The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes [FAR.AI] https://arxiv.org/abs/2602.15515 --- [CV] Motion Attribution for Video Generation [NVIDIA] https://arxiv.org/abs/2601.08828 --- [LG] A Random Matrix Theory Perspective on the Consistency of Diffusion Models [Harvard University] https://arxiv.org/abs/2602.02908
你有没有想过,要让AI真正理解世界,而不是简单模仿,到底需要几步?本期我们将看到,最新的论文正在教AI像婴儿一样建立内在的“世界模型”,并给我们一张能诊断它心智的“大脑地图”。我们还会揭秘,如何用“家教天团”模式培养全能AI,让机器人拥有和人相处的“眼力见”,以及教会它像学汉字笔画一样拆解世间万物的动作。准备好,让我们一起探索AI心智的构建蓝图。 00:00:32 AI的“婴儿模式”,它如何偷偷学会了物理定律? 00:05:32 给你一张AI的“大脑地图” 00:12:17 AI界的“家教天团”,如何培养一个全能型选手 00:18:09 让机器人拥有“眼力见”,差的是什么? 00:23:16 想看懂世界?先学会拆解动作 本期介绍的几篇论文: [CV] Orca: The World is in Your Mind [Beijing Academy of Artificial Intelligence] https://arxiv.org/abs/2606.30534 --- [AI] NeuroCogMap Reveals Cognitive Organization of Large Language Models [Renmin University of China & Beijing University of Posts and Telecommunications & The University of Hong Kong] https://arxiv.org/abs/2607.00397 --- [CL] MOPD: Multi-Teacher On-Policy Distillation for Capability Integration in LLM Post-Training [Xiaomi & Peking University] https://arxiv.org/abs/2606.30406 --- [RO] HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation [Config] https://arxiv.org/abs/2606.31682 --- [AI] Latent Actions from Factorized Transition Effects under Agent Ambiguity [Brown University] https://arxiv.org/abs/2606.30544
你有没有想过,当AI不再追求“大力出奇迹”时,它会进化出怎样惊人的智慧?本期节目,我们就来聊聊AI如何从“内功”和“招式”上自我进化。它会如何发明一套“黑话”来自我思考,让效率提升数倍;一个“普通”模型又如何通过顶级流程,战胜天赋异禀的“天才”;它又将怎样为虚拟世界的角色,注入一个会思考、懂物理的“灵魂”?今天,我们就从几篇最新论文出发,揭秘AI如何变得更聪明,而非更“大”。 00:00:36 AI的长记性难题,一个聪明的“图书管理员” 00:05:12 成为高手,靠天赋还是靠流程? 00:10:28 “虚拟人”的“灵魂”,它如何学会像你一样思考和行动? 00:16:07 AI的眼睛,看得清,还是看得懂? 00:22:24 让AI说“黑话”,它会变得多聪明? 本期介绍的几篇论文: [LG] Hierarchical Global Attention (HGA) [BMW Group] https://arxiv.org/abs/2606.30709 --- [CL] Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent [Shanghai Artificial Intelligence Laboratory] https://arxiv.org/abs/2606.30616 --- [CV] GPC: Large-Scale Generative Pretraining for Transferable Motor Control [Simon Fraser University & NVIDIA] https://arxiv.org/abs/2606.29148 --- [CV] LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives [German Cancer Research Center & Brown University] https://arxiv.org/abs/2607.00784 --- [AI] When LLMs Develop Languages: Symbolic Communication for Efficient Multi-Agent Reasoning [Chinese Academy of Sciences] https://arxiv.org/abs/2606.29354
我们总以为AI的进步就是靠“大力出奇迹”,但如果这个“大力”会扭曲现实、甚至有它砸不开的墙呢?本期,我们就来看几篇“反其道而行”的最新论文,看看AI如何学会像乐高大师一样分解任务,像生物一样进化出看问题的“火眼金睛”。我们还会给AI的思维做个“CT扫描”,看看它在百万份文件中是如何被“噪音”淹没,又是如何学会重新聚焦的。准备好,让我们一起探索AI如何告别蛮力,走向真正的“巧”劲儿。 00:00:35 AI能扮演人类吗?一个关于“大力出奇迹”的意外发现 00:08:00 如何给AI的思维过程做个“CT扫描”? 00:13:57 让AI学会“开窍”,聪明的数据,胜过强大的模型 00:19:46 高手解题,为何偏爱“笨办法”? 00:25:23 大模型记忆的极限,为什么“知道”不等于“能说出来”? 本期介绍的几篇论文: [CL] Will Scaling Improve Social Simulation with LLMs? [Stanford University] https://arxiv.org/abs/2607.02464 --- [LG] Geometric Signatures of Reasoning: A Spectral Perspective on Task Hardness [Northeastern University & University of Southern California & Google Research] https://arxiv.org/abs/2607.01571 --- [LG] Evolutionary Feature Engineering for Structured Data [University of Michigan & Google Research] https://arxiv.org/abs/2607.01548 --- [LG] DecompRL: Solving Harder Problems by Learning Modular Code Generation [FAIR at Meta & Inria] https://arxiv.org/abs/2607.02390 --- [CL] Can Language Models Actually Retrieve In-Context? Drowning in Documents at Million Token Scale [UC Berkeley & UT Austin] https://arxiv.org/abs/2607.01538
我们总惊叹AI越来越聪明,但你有没有想过,聪明的AI也会有自己的烦恼?比如,它可能像个伪装极深的“卧底”,悄悄藏着偏见;也可能像个只会刷题的“好学生”,答案虽对,却毫无灵气。它在解决难题时,可能会反复“无效内卷”,或者在关键的推理环节“脑子短路”。本期节目,我们就从几篇最新论文出发,看看科学家们如何通过巧妙的设计,教会AI自我审视、优雅试错、清晰思考,甚至让它的思考过程变得有迹可循。准备好,我们一起揭开AI变得更聪明的秘密。 00:00:40 AI的“无间道”,如何揪出那些伪装良好的“卧底”偏见? 00:05:55 AI变聪明的秘密,不是多试几次,而是换个姿势再试 00:11:08 AI侦探断案,为什么它连“你妈的儿子的老婆”都搞不清? 00:16:45 怎样让AI的思考,既聪明又有迹可循? 00:22:06 AI的“好学生”困境,做对题,为何还是不对劲? 本期介绍的几篇论文: [CL] Distill to Detect: Exposing Stealth Biases in LLMs through Cartridge Distillation [Stanford University & University of Texas at Austin] https://arxiv.org/abs/2607.01208 --- [LG] QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling [Stanford University] https://arxiv.org/abs/2607.01179 --- [CL] DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning [UC Berkeley] https://arxiv.org/abs/2607.00341 --- [CL] Graph-Native Reinforcement Learning Enables Traceable Scientific Hypothesis Generation through Conceptual Recombination [MIT & Oak Ridge National Laboratory] https://arxiv.org/abs/2607.00924 --- [LG] Right in the Right Way: LM Training with Verifiable Rewards and Human Demonstrations [MIT] https://arxiv.org/abs/2607.01181
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