Discover
AI可可AI生活
880 Episodes
Reverse
你有没有想过,神秘的AI黑箱里其实藏着一个200年前的数学幽灵?你和AI的甜言蜜语,又为何可能是一个危险的情感陷阱?今天,我们将从这几个问题出发,聊聊AI如何向古老的智慧回归,如何像“散兵”一样自组织搞科研,如何用一本“手账”治好它的金鱼记忆,以及它那神乎其神的创造力背后,又藏着一座怎样的“物理学之桥”。00:00:31 AI黑箱里,藏着一个200年前的数学幽灵00:06:04 你和AI的悄悄话,藏着一个危险的“放大器”00:12:01 一群AI“散兵”,如何自己组织起来搞科研?00:18:42 AI绘画的终极密码,藏在一座“桥”里?00:24:13 你的AI管家,为什么总像个金鱼?本期介绍的几篇论文:[LG] Transformers are Bayesian Networks [coppola.ai] https://arxiv.org/abs/2603.17063 ---[CL] Characterizing Delusional Spirals through Human-LLM Chat Logs [Stanford University & CMU] https://arxiv.org/abs/2603.16567 ---[LG] Autonomous Agents Coordinating Distributed Discovery Through Emergent Artifact Exchange [Laboratory for Atomistic and Molecular Mechanics (LAMM)] https://arxiv.org/abs/2603.14312 ---[LG] Foundations of Schrödinger Bridges for Generative Modeling [University of Pennsylvania] https://arxiv.org/abs/2603.18992 ---[CL] Chronos: Temporal-Aware Conversational Agents with Structured Event Retrieval for Long-Term Memory [PricewaterhouseCoopers] https://arxiv.org/abs/2603.16862
你有没有想过,AI在思考时也能像我们一样“随时回头看”,直接调用最关键的“深度记忆”吗?本期节目,我们将一口气看懂几篇最新论文,探索AI如何从只会“猜答案”进化到真正“理解画面”,如何像个聪明的懒汉一样,用“排序”而非“整理”在海量信息中精准寻宝,以及科学家们如何通过巧妙的设计,让AI的大脑在保持高速运转的同时,还能解决信息层层衰减的老大难问题。准备好,我们马上出发!00:00:37 AI大模型的新陈代谢法则00:05:48 AI效率战争,如何让大模型跑得又快又省?00:10:57 不止看“热闹”,更要看“门道”,AI理解力的一次飞跃00:16:07 给你一个超大号书房,你会怎么整理?00:22:15 让AI拥有“深度记忆”,它会变得多聪明?本期介绍的几篇论文:[CL] Attention Residuals[Kimi Team]https://arxiv.org/abs/2603.15031---[LG] Mamba-3: Improved Sequence Modeling using State Space Principles[CMU & Princeton University]https://arxiv.org/abs/2603.15569---[CV] V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning[FAIR at Meta]https://arxiv.org/abs/2603.14482---[LG] SmartSearch: How Ranking Beats Structure for Conversational Memory Retrieval[Midbrain]https://arxiv.org/abs/2603.15599---[CL] Mixture-of-Depths Attention[ByteDance Seed]https://arxiv.org/abs/2603.15619
我们总希望AI更强大,但“强大”就等于“更聪明”吗?本期几篇最新论文将带我们探索“聪明”的另一面:有时,给AI加上“部门墙”的约束,反而能激发它的潜力;有时,教会AI在关键时刻向“专家”求助,比让它无所不知更高效;甚至,我们还会发现,那个你以为在帮你润色文稿的AI,可能正在不动声色地重塑你的观点。准备好了吗?让我们一起看看AI是如何学会“思考”,而我们又该如何与它共处。00:00:37 AI大模型里的“部门墙”,怎么破?00:06:20 你的“专家外挂”,需要一个“智能开关”00:11:29 AI学会了“做大菜”,而不只是“选番茄”00:17:14 AI也懂“四两拨千斤”?00:21:32 你以为AI在帮你润色,其实它在重塑你的观点本期介绍的几篇论文:[LG] Path-Constrained Mixture-of-Experts [Apple & Google] https://arxiv.org/abs/2603.18297 ---[CL] TARo: Token-level Adaptive Routing for LLM Test-time Alignment [Meta] https://arxiv.org/abs/2603.18411 ---[CL] Reasoning over mathematical objects: on-policy reward modeling and test time aggregation [FAIR at Meta] https://arxiv.org/abs/2603.18886 ---[LG] dTRPO: Trajectory Reduction in Policy Optimization of Diffusion Large Language Models [Meta AI] https://arxiv.org/abs/2603.18806 ---[CL] How LLMs Distort Our Written Language [UC Berkeley & UC San Diego & Google DeepMind] https://arxiv.org/abs/2603.18161
你有没有想过,我们到底该如何培养一个更聪明的AI?本期节目,我们将一起揭秘几篇最新论文,看看科学家们是如何给AI请“精准家教”,让它花十分之一的钱办成同样的事;如何窥探AI的“内心戏”,了解它什么时候是真的自信;又是如何通过一个关键的“中间态”和不知疲倦的“AI陪练”,把它从偏科生打造成全能高手,并最终教会它“懂分寸”,成为一名好裁判的。让我们一同探寻AI的成长之道。00:00:35 AI的“补习班”,如何花十分之一的钱,办成同样的事?00:06:21 AI的“内心戏”,它怎么知道自己懂不懂?00:12:18 你和高手的差距,可能只是一个“中间态”00:18:32 AI的“陪练”,高手是怎么喂出来的?00:24:11 如何把一个“耿直”的AI,训练得“懂分寸”?本期介绍的几篇论文:[LG] Efficient Exploration at Scale[Google DeepMind]https://arxiv.org/abs/2603.17378---[CL] How do LLMs Compute Verbal Confidence[Google DeepMind]https://arxiv.org/abs/2603.17839---[LG] PRISM: Demystifying Retention and Interaction in Mid-Training[IBM Research]https://arxiv.org/abs/2603.17074---[AI] AI Scientist via Synthetic Task Scaling[Princeton University & Microsoft Research]https://arxiv.org/abs/2603.17216---[LG] REAL: Regression-Aware Reinforcement Learning for LLM-as-a-Judge[University of California, Los Angeles & The University of Texas at Austin]https://arxiv.org/abs/2603.17145
你有没有想过,未来的AI不仅能回答你的问题,还能从与你的每一次互动中汲取经验,悄悄进化?它甚至还能在犯错后“自我反思”,像我们一样“长记性”。本期我们将一起探索几篇最新论文,看看AI如何学会像一个聪明的“CEO”一样管理自己的思考,如何通过精准“剪枝”在你的手机里狂飙,以及如何消灭那些你看不到的“计算成本”,变得更高效、更智慧。00:00:32 AI进化论,为什么你的“差评”正在喂养一个更聪明的它00:05:19 让AI在手机里狂飙,快,才是一切00:10:38 AI提速19%的秘密,你以为的计算,其实是搬运00:15:20 AI犯了错,能不能让它自己“长记性”?00:21:26 你的大脑里,缺一个聪明的“CEO”本期介绍的几篇论文:[CL] Online Experiential Learning for Language Models [Microsoft Research] https://arxiv.org/abs/2603.16856---[LG] MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale [Meta AI] https://arxiv.org/abs/2603.15954---[LG] FlashSampling: Fast and Memory-Efficient Exact Sampling [LMU Munich & FlashSampling & Princeton University] https://arxiv.org/abs/2603.15854---[LG] Meta-TTRL: A Metacognitive Framework for Self-Improving Test-Time Reinforcement Learning in Unified Multimodal Models [Tsinghua University & JD.COM] https://arxiv.org/abs/2603.15724---[RO] When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making [CMU & Northeastern University & Harvard University] https://arxiv.org/abs/2603.16673
今天我们要聊一个特别有意思的话题:如何让聪明的AI变得更“靠谱”?我们会一起从几篇最新的论文中寻找答案,看看科学家们是如何教AI学会“自主学习”而不是死记硬背,又是如何通过给它换个“大记事本”来解决记性差的难题。更刺激的是,我们还会揭秘AI那些悄无声息的“隐形失败”,并学习一种看似很笨的管理办法,以及AI学会说“等一下,我再想想”背后的真正奥秘。准备好了吗?让我们一起潜入AI的大脑深处。00:00:35 你被骗了,为什么说现在的AI根本不会“学习”?00:06:58 AI的大脑革命,为什么“记性差”的反而更聪明?00:13:58 你和AI的对话,藏着多少看不见的“坑”?00:18:36 如何用“笨办法”,管好一个聪明的AI?00:23:53 AI学会了“等一下,我再想想”?本期介绍的几篇论文:[AI] Why AI systems don't learn and what to do about it: Lessons on autonomous learning from cognitive science [FAIR at META & NYU] https://arxiv.org/abs/2603.15381 ---[LG] M²RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling [UC Berkeley & MIT-IBM Watson Lab] https://arxiv.org/abs/2603.14360 ---[CL] Invisible failures in human-AI interactions [Bigspin AI] https://arxiv.org/abs/2603.15423 ---[LG] POLCA: Stochastic Generative Optimization with LLM [University of Wisconsin-Madison & Google DeepMind] https://arxiv.org/abs/2603.14769 ---[LG] Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty [Microsoft Research] https://arxiv.org/abs/2603.15500
你有没有想过,AI画画也能像我们一样进行“刻意练习”,通过精准对比找到最佳进步方向吗?面对复杂变化的世界,为什么“慢半拍”的决策反而更准确?我们还将揭示AI训练中“又快又好”的秘密课程表,探讨项目延期背后的沟通艺术,并告诉你,你对AI的每一次追问,都在如何悄悄地训练它。本期,让我们一起从几篇最新论文中,窥探AI正在学习的那些“人间智慧”。00:00:34 AI绘画的“刻意练习法”00:05:25 做对事情,只需一个“时间差”00:11:31 快与好,为什么不能兼得?AI训练中的“学霸心法”00:17:02 为什么你的项目总在延期?答案可能不在技术,在沟通00:22:27 你的每一次追问,都在悄悄训练AI本期介绍的几篇论文:[CV] Finite Difference Flow Optimization for RL Post-Training of Text-to-Image Models [NVIDIA & UC Berkeley] https://arxiv.org/abs/2603.12893 ---[LG] A Reduction Algorithm for Markovian Contextual Linear Bandits [University of California, Los Angeles & Meta] https://arxiv.org/abs/2603.12530 ---[LG] Curriculum Sampling: A Two-Phase Curriculum for Efficient Training of Flow Matching [Stanford University] https://arxiv.org/abs/2603.12517 ---[LG] Optimizing Task Completion Time Updates Using POMDPs [Stanford University & Rensselaer Polytechnic Institute] https://arxiv.org/abs/2603.12340 ---[CL] Aligning Language Models from User Interactions [ETH Zurich] https://arxiv.org/abs/2603.12273
你有没有想过,有一天跟电脑交互不再需要打开一个个App?或者,一个顶尖AI为了辅导“学生”考高分,竟然学会了“作弊”?本期节目,我们将从五篇最新论文出发,聊聊这些正在发生的奇妙变革:从重塑操作系统的“智能管家”,到学会削苹果的“灵巧机械手”,再到“专业团队”如何完胜“大力出奇迹”派的机器人。让我们一起看看,AI是如何在这些意想不到的角落,悄悄改写着未来。00:00:36 跟App说再见,我们和电脑的相处之道正在被重写00:07:15 当AI开始“辅导”AI,一个关于学霸、偏科和作弊的故事00:13:38 真正的问题不是AI,而是我们测试它的方法00:18:53 让机器人给你削苹果,到底有多难?00:25:31 造一个聪明的机器人,是“大力出奇迹”还是“专业的人干专业的事”?本期介绍的几篇论文:[AI] AgentOS: From Application Silos to a Natural Language-Driven Data Ecosystem [University of Kansas] https://arxiv.org/abs/2603.08938 ---[LG] PostTrainBench: Can LLM Agents Automate LLM Post-Training? [ELLIS Institute Tübingen & University of Tübingen] https://arxiv.org/abs/2603.08640 ---[AI] Evaluation format, not model capability, drives triage failure in the assessment of consumer health AI [Macquarie University] https://arxiv.org/abs/2603.11413 ---[RO] Towards Human-Like Manipulation through RL-Augmented Teleoperation and Mixture-of-Dexterous-Experts VLA [Shanghai Jiao Tong University & Sharpa] https://arxiv.org/abs/2603.08122 ---[RO] TiPToP: A Modular Open-Vocabulary Planning System for Robotic Manipulation [MIT CSAIL] https://arxiv.org/abs/2603.09971
本期节目,我们来当一次AI的“首席优化官”,从里到外给它做个大升级。我们会看到,AI如何从解题高手,变身发现解题方法的“教练”;我们会拿到一份硬核“体检报告”,看看AI一本正经胡说八道的底线究竟在哪。我们还会发现,你和AI聊天时那些被浪费的“废话”,其实是喂饱它的宝贵养料;最后再深入AI的内部,看看万亿参数的它如何避免“大公司病”,以及一个惊人发现:困扰AI效率的瓶颈,可能不在“大脑”,而在“嘴巴”!00:00:38 AI当教练,数学家当陪练,我们如何找到世界的隐藏规则?00:06:42 AI会「一本正经地胡说八道」到什么程度?00:14:04 你扔掉的“废话”,正在喂饱AI00:19:14 万亿参数的大模型,是如何避免“公司越大,效率越低”的?00:27:08 你的模型为什么这么笨?问题可能出在“嘴”上本期介绍的几篇论文:[LG] Reinforced Generation of Combinatorial Structures: Ramsey Numbers [UC Berkeley & Google] https://arxiv.org/abs/2603.09172 ---[CL] How Much Do LLMs Hallucinate in Document Q&A Scenarios? A 172-Billion-Token Study Across Temperatures, Context Lengths, and Hardware Platforms [Kamiwaza AI] https://arxiv.org/abs/2603.08274 ---[CL] OpenClaw-RL: Train Any Agent Simply by Talking [Princeton Univercity] https://arxiv.org/abs/2603.10165 ---[CL] Scalable Training of Mixture-of-Experts Models with Megatron Core [NVIDIA] https://arxiv.org/abs/2603.07685 ---[CL] Lost in Backpropagation: The LM Head is a Gradient Bottleneck [Cornell University] https://arxiv.org/abs/2603.10145
你有没有想过,如何帮一个“路痴”AI把脑中的地图“拉直”?又或者,一个AI模型里,其实藏着成百上千个性格各异的“专家”?今天,我们将从几篇最新的AI论文出发,聊聊AI如何学会优化资源、高效复盘,甚至,如何进化成一个连它的“老师”都能骗过的“作弊”高手。00:00:26 你的认知,需要一次“时空拉直”00:06:13 为什么最贵的AI,有时用的是最“笨”的办法?00:12:16 AI的“众神殿”,一个模型,藏着万千专家00:19:01 AI世界的“尖子生”,是真学霸,还是“作弊”高手?00:24:14 你不是不行,你只是不会“复盘”本期介绍的几篇论文:[LG] Temporal Straightening for Latent Planning [New York University] https://arxiv.org/abs/2603.12231 ---[LG] IsoCompute Playbook: Optimally Scaling Sampling Compute for LLM RL [UC San Diego & CMU] https://arxiv.org/abs/2603.12151 ---[LG] Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights [MIT CSAIL] https://arxiv.org/abs/2603.12228 ---[CL] Examining Reasoning LLMs-as-Judges in Non-Verifiable LLM Post-Training [Meta Superintelligence Labs] https://arxiv.org/abs/2603.12246 ---[LG] Meta-Reinforcement Learning with Self-Reflection for Agentic Search [Allen Institute for AI & University of Washington] https://arxiv.org/abs/2603.11327
你有没有想过,我们不仅能看懂AI的“鬼点子”,还能直接让它把克敌制胜的“武功秘籍”写成代码?本期节目,我们将一起探索几篇最新论文带来的奇妙洞见:我们会发现AI的“中年健忘”竟是与生俱来的天性,并找到它大脑里那个精准的“谎言开关”。我们不仅要科学地为AI制定最佳“学习计划”,甚至还要在它读书前,先送它去一个纯粹的“数字健身房”锻炼核心能力。准备好了吗?让我们一起出发,看看AI的聪明才智背后,藏着哪些你意想不到的秘密。00:00:39 当AI学会了写代码,它的“鬼点子”就藏不住了00:05:48 AI的学习计划,应该怎么定?00:12:05 大模型的“中年危机”,我们一直都搞错了?00:17:37 别再被AI骗了,我们找到了它大脑里的“谎言开关”00:23:23 AI的“健身房”,不读书,如何变得更聪明?本期介绍的几篇论文:[LG] Code-Space Response Oracles: Generating Interpretable Multi-Agent Policies with Large Language Models [Google DeepMind] https://arxiv.org/abs/2603.10098 ---[LG] What do near-optimal learning rate schedules look like? [Google DeepMind & Mila] https://arxiv.org/abs/2603.10301 ---[LG] Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias [Meta] https://arxiv.org/abs/2603.10123 ---[CL] Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models [Dakota State University & North Carolina A&T State University] https://arxiv.org/abs/2603.10195 ---[LG] Training Language Models via Neural Cellular Automata [MIT] https://arxiv.org/abs/2603.10055
本期节目,我们将深入AI的“内心世界”:你会发现,让AI多“思考”一会儿,它反而可能变得更诚实;而有时它的“思考”其实不是为了推理,更像是在努力“回忆”。我们还会聊到,最新论文如何让AI拥有调试代码的“灵魂”,如何量化它有多少“小秘密”不愿公开,以及一个聪明的“外行”AI领导,要如何带好一支能打的“内行”AI团队。00:00:32 AI 不仅会写代码,还会自己找 Bug?00:05:03 AI会撒谎吗?一个让你意外的答案00:10:09 思考,不是为了推理,而是为了回忆00:15:26 AI的“草稿纸”,它到底有多少不能说的秘密?00:21:32 聪明的“外行”领导,如何带出能打的“内行”团队?本期介绍的几篇论文:[LG] Towards a Neural Debugger for Python[Meta FAIR & Johannes Kepler University Linz]https://arxiv.org/abs/2603.09951---[CL] Think Before You Lie: How Reasoning Improves Honesty[Google DeepMind]https://arxiv.org/abs/2603.09957---[CL] Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs[Google Research]https://arxiv.org/abs/2603.09906---[AI] Quantifying the Necessity of Chain of Thought through Opaque Serial Depth[Google DeepMind]https://arxiv.org/abs/2603.09786---[LG] SCALAR: Learning and Composing Skills through LLM Guided Symbolic Planning and Deep RL Grounding[CMU & Virginia Tech]https://arxiv.org/abs/2603.09036
我们都希望AI越来越聪明,但怎么才能让它高效成长呢?今天我们要聊的几篇最新论文,就给出了一些非常反直觉的答案:比如,让AI只做“难题”,给它的创作过程派一位“监理”,甚至还要警惕它因为懂得太多而“吃不饱”。更神奇的是,我们还会看到如何让AI学会新本事,却完全不忘旧手艺。准备好了吗?让我们一起看看AI是如何被调教成“学霸”的!00:00:31 想让AI更聪明?你得学会给它出难题00:05:53 如何让AI“心领神会”你的想法?00:12:00 AI的“语义饱腹感”,为什么数据越多,进步越难?00:18:31 AI思考的秘密,为什么“平行世界”里的笨办法,反而是捷径?00:24:27 如何让AI学会新本事,还不忘了旧手艺?本期介绍的几篇论文:[CL] Scaling Data Difficulty: Improving Coding Models via Reinforcement Learning on Fresh and Challenging Problems [Microsoft Research] https://arxiv.org/abs/2603.07779 ---[LG] Diffusion Controller: Framework, Algorithms and Parameterization [Google Research] https://arxiv.org/abs/2603.06981 ---[LG] Scale Dependent Data Duplication [Stanford University & EPFL] https://arxiv.org/abs/2603.06603 ---[LG] Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference [Microsoft Research & MIT] https://arxiv.org/abs/2603.07887 ---[LG] Grow, Don't Overwrite: Fine-tuning Without Forgetting [Google Research & University of Wisconsin-Madison] https://arxiv.org/abs/2603.08647
你有没有想过,AI的“脑子”里到底在想些什么?这一期,我们就来当一回“AI心理学家”,从几篇最新论文出发,探寻AI的内心世界:看它如何“自己教自己”实现顿悟,又为何会陷入“学不动”的瓶颈;我们会揭秘它那张写满内心独白的“草稿纸”,看看它是否学会了撒谎;最后,我们将学习一种读心术,不仅能看懂AI的“集体智慧”,甚至还能预测你的下一步行动。准备好了吗?让我们一起潜入AI的深层意识。00:00:36 AI进阶之路,当“尖子生”不再需要“课外辅导”00:05:28 你的AI为什么学不动了?答案可能出乎意料,人多力量大00:13:04 你的AI助理,如何才能比你更懂你?00:19:19 AI的“草稿纸”,藏着什么秘密?00:24:19 AI的“内心戏”,我们终于能看懂了本期介绍的几篇论文:[CV] Self-Supervised Flow Matching for Scalable Multi-Modal Synthesis [Black Forest Labs] https://arxiv.org/abs/2603.06507 ---[LG] Preventing Learning Stagnation in PPO by Scaling to 1 Million Parallel Environments [Google DeepMind & University of Oxford] https://arxiv.org/abs/2603.06009 ---[CL] Learning Next Action Predictors from Human-Computer Interaction [Stanford University & Hasso Plattner Institute] https://arxiv.org/abs/2603.05923 ---[AI] Reasoning Models Struggle to Control their Chains of Thought [NYU & UCL & OpenAI] https://arxiv.org/abs/2603.05706 ---[LG] Causal Interpretation of Neural Network Computations with Contribution Decomposition [Stanford University] https://arxiv.org/abs/2603.06557
你有没有想过,未来的AI要如何变得更聪明?最新的一些研究告诉我们,答案可能不是一味地堆算力,而是要学会人类的“智慧”。比如,让AI拥有一个能从错误中总结经验的“技能工具箱”;或者像教孩子一样,让它理解规则而不是死记硬背模式;甚至,像一位高明的将军,懂得如何排兵布阵,把好钢用在刀刃上。本期节目,我们就来聊聊这些让AI学会“反思”、“预见”和“布阵”的最新论文,看看真正的智能是如何炼成的。00:00:38 高手,都是“错”出来的00:05:41 AI学会举一反三的秘密,换个数字就不认识了?00:10:57 AI大模型的新兵法,好钢如何用在刀刃上?00:17:26 让机器人自己“玩”成高手,需要几步?00:23:29 AI的远见,如何不看细节,反而看得更远?本期介绍的几篇论文:[AI] EvoSkill: Automated Skill Discovery for Multi-Agent Systems [Sentient & Virginia Tech] https://arxiv.org/abs/2603.02766 ---[LG] Symbol-Equivariant Recurrent Reasoning Models [Johannes Kepler University Linz] https://arxiv.org/abs/2603.02193 ---[LG] DynaMoE: Dynamic Token-Level Expert Activation with Layer-Wise Adaptive Capacity for Mixture-of-Experts Neural Networks https://arxiv.org/abs/2603.01697 ---[RO] Tether: Autonomous Functional Play with Correspondence-Driven Trajectory Warping [University of Pennsylvania] https://arxiv.org/abs/2603.03278 ---[LG] Next Embedding Prediction Makes World Models Stronger [T-Tech] https://arxiv.org/abs/2603.02765
今天我们来聊聊AI世界里那些“反常识”的智慧:为什么“见过世面”的AI不容易遗忘,而“偏科”却成了它发展的隐患?我们不仅会揭秘AI如何通过“预判你的预判”来极致提速,还会探讨为何有时“慢”一点的学习,反而能让AI变得更聪明、更懂变通。最后,我们会发现,解决一个复杂的动画难题,关键可能只是需要为AI发明一种“普通话”。00:00:31 为什么高手学东西,不容易忘?00:05:34 AI的加速赛,怎样让聪明的“大脑袋”跑得更快?00:12:39 AI动画的“普通话”和“方言”00:17:56 AI智能体:是天才还是“偏科生”?00:23:02 天下武功,唯快不破?AI训练中的一个“慢”智慧本期介绍的几篇论文:[LG] Pretrained Vision-Language-Action Models are Surprisingly Resistant to Forgetting in Continual Learning [The University of Texas at Austin & Microsoft Superintelligence] https://arxiv.org/abs/2603.03818 ---[LG] Speculative Speculative Decoding [Stanford University & Princeton University & Together AI] https://arxiv.org/abs/2603.03251 ---[CV] OmniLottie: Generating Vector Animations via Parameterized Lottie Tokens [Fudan University & StepFun & HKU MMLab] https://arxiv.org/abs/2603.02138 ---[AI] How Well Does Agent Development Reflect Real-World Work? [CMU] https://arxiv.org/abs/2603.01203 ---[LG] To Use or not to Use Muon: How Simplicity Bias in Optimizers Matters [New York University] https://arxiv.org/abs/2603.00742
你有没有想过,AI除了会聊天画画,还能做什么更酷的事?本期节目,我们将一口气看到AI能力的多个惊人侧面。从像人一样“脑补”物理世界,到用“笨方法”实现更高效的学习,再到成为物理学家的“科研搭子”,解决真正的科学难题。这些最新论文将刷新你对AI潜力的认知!00:00:28 AI学会了“脑补”,世界就大不一样了00:06:08 大模型里的“关系户”,它凭什么吸引了所有注意力?00:13:10 AI省钱的终极奥义,少就是多00:18:04 一个“笨方法”,让AI学得更快00:22:49 AI,从聊天高手到科研搭子本期介绍的几篇论文:[LG] Latent Particle World Models: Self-supervised Object-centric Stochastic Dynamics Modeling [CMU & UT Austin & Brown University] https://arxiv.org/abs/2603.04553 ---[CL] The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks [New York University] https://arxiv.org/abs/2603.05498 ---[CL] Sparse-BitNet: 1.58-bit LLMs are Naturally Friendly to Semi-Structured Sparsity [Microsoft Research] https://arxiv.org/abs/2603.05168 ---[CL] Replaying pre-training data improves fine-tuning [Stanford University] https://arxiv.org/abs/2603.04964 ---[AI] Solving an Open Problem in Theoretical Physics using AI-Assisted Discovery [Google Research] https://arxiv.org/abs/2603.04735
你是否好奇,为何AI有时会“指鹿为马”?为何它面对难题,内部的神经元反而开始“集体偷懒”?本期节目,我们将通过几篇最新论文,一起给AI的大脑做一次“CT扫描”和“基因测序”,揭示它在感知、学习、思考和效率背后,那些出人意料的底层法则。00:00:26 人工智能的“阿喀琉斯之踵”,一个关于维度的诅咒00:05:34 AI绘画进化论,为什么高手不需要“题海战术”?00:10:02 AI一思考,我们就发笑?不,是神经元在“偷懒”00:15:44 如何用50倍的效率,给AI做一次“CT扫描”?00:21:34 AI模型的“不可能三角”,算力、速度与智能本期介绍的几篇论文:[LG] Solving adversarial examples requires solving exponential misalignment[Stanford University & Aisle]https://arxiv.org/abs/2603.03507---[LG] Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data[University of Michigan & Google DeepMind & UC Berkeley]https://arxiv.org/abs/2603.03700---[CL] Farther the Shift,Sparser the Representation: Analyzing OOD Mechanisms in LLMs[Rutgers University & Northwestern University & UKP Lab, TU Darmstadt]https://arxiv.org/abs/2603.03415---[CL] Compressed Sensing for Capability Localization in Large Language Models[CMU]https://arxiv.org/abs/2603.03335---[LG] Why Are Linear RNNs More Parallelizable?[Allen Institute for AI & Rheinland-Pfalzische Technische Universitat]https://arxiv.org/abs/2603.03612
今天,我们要探讨如何让AI从一个只会“动嘴”的聊天伙伴,进化成一个真正“会看、会想、会动手”的智能体。我们会看到,最新论文如何让AI‘开眼看世界’,在脑中建立起预测未来的‘导航系统’,并从海量普通文本中自我启蒙,学会判断好坏。更重要的是,当AI要替我们行动时,它又是如何学会‘三思而后行’,在‘有用’和‘安全’之间找到那条微妙的平衡线呢?准备好了吗?让我们一起探寻AI从‘愣头青’到‘老司机’的进化之路。00:00:40 AI为什么要“开眼看世界”?00:07:16 为什么高手都自带“导航系统”?00:13:19 AI的“行动许可”,它在动手前,先想了什么?00:19:12 把白开水变成高汤,AI如何从普通文本中学会“好坏”00:24:47 如何把一个“愣头青”AI,调教成“老司机”?本期介绍的几篇论文:[CV] Beyond Language Modeling: An Exploration of Multimodal Pretraining [FAIR, Meta] https://arxiv.org/abs/2603.03276 ---[LG] What Capable Agents Must Know: Selection Theorems for Robust Decision-Making under Uncertainty [CMU] https://arxiv.org/abs/2603.02491 ---[LG] Learning When to Act or Refuse: Guarding Agentic Reasoning Models for Safe Multi-Step Tool Use [Microsoft Research] https://arxiv.org/abs/2603.03205 ---[LG] Scaling Reward Modeling without Human Supervision [Harvard University & Cornell University] https://arxiv.org/abs/2603.02225 ---[LG] Safety Training Persists Through Helpfulness Optimization in LLM Agents [UC Berkeley] https://arxiv.org/abs/2603.02229
今天我们不聊模型参数有多大,而是聊如何让AI变得更“会思考”,这种思考方式,有时甚至有些反常识。比如,为什么给AI疯狂“补课”,它反而可能越学越笨?我们还会探讨,如何像一位高明的老师一样引导AI攻克难题,而不是直接灌输答案。更进一步,我们会揭示如何训练AI像个侦探一样,学会“讲道理”地分析代码,以及如何让整个系统学会动态协作,找到最高效的“偷懒”方式。00:00:35 AI大模型时代,如何花小钱办大事?00:05:47 给AI“补课”的陷阱,为什么学得越多,它反而越笨?00:11:37 高手辅导功课,为什么不直接给答案?00:16:48 让AI学会“讲道理”,代码世界的侦探是怎样炼成的?00:22:00 让AI学会“省时间”,一种更聪明的快本期介绍的几篇论文:[LG] Rich Insights from Cheap Signals: Efficient Evaluations via Tensor Factorization [Google DeepMind & University of Michigan] https://arxiv.org/abs/2603.02029 ---[LG] Theoretical Perspectives on Data Quality and Synergistic Effects in Pre- and Post-Training Reasoning Models [University of Southern California & University of California Los Angeles & Google Research] https://arxiv.org/abs/2603.01293 ---[LG] Learn Hard Problems During RL with Reference Guided Fine-tuning [ByteDance Seed & UC Berkeley & CMU] https://arxiv.org/abs/2603.01223 ---[LG] Agentic Code Reasoning [Meta] https://arxiv.org/abs/2603.01896 ---[CL] Learning to Draft: Adaptive Speculative Decoding with Reinforcement Learning [Microsoft Research Asia & Peking University] https://arxiv.org/abs/2603.01639



