你有没有想过,为什么投入巨大的AI模型有时反而会“学傻了”?当AI的“词典”里没有“我错了”这个词时,我们又该如何教会它自我反思?本期节目,我们将一起钻进AI的大脑,从几篇最新的论文出发,看看AI是如何诊断自己内部的“罢工”,如何通过一场“无限游戏”变得更安全,以及它在绘画时,究竟是在搞创作,还是在“背书”。 00:00:30 规模的诅咒,AI为何会“学傻”? 00:06:29 AI的语言里,没有“我错了” 00:11:35 想让AI更安全?答案可能藏在一场“无限游戏”里 00:16:13 我们如何看穿世界的规则?AI给了新思路 00:23:44 揭秘AI绘画,它“抄袭”的秘密藏在哪? 本期介绍的几篇论文: [LG] Understanding Scaling Laws in Deep Neural Networks via Feature Learning Dynamics [DePaul University & Iowa State University] https://arxiv.org/abs/2512.21075 --- [CL] Reflection Pretraining Enables Token-Level Self-Correction in Biological Sequence Models [Fudan University & Shanghai Artificial Intelligence Laboratory] https://arxiv.org/abs/2512.20954 --- [LG] Safety Alignment of LMs via Non-cooperative Games [FAIR at Meta & University of Tübingen] https://arxiv.org/abs/2512.20806 --- [LG] Active inference and artificial reasoning [University College London & VERSES] https://arxiv.org/abs/2512.21129 --- [LG] Generalization of Diffusion Models Arises with a Balanced Representation Space [University of Michigan] https://arxiv.org/abs/2512.20963
今天,我们要深入AI的“内心世界”,去探寻几个颠覆性的问题:聪明的AI,是该学会“胸有成竹”的规划,还是“选择性失忆”的智慧?我们该如何教会一个AI坦然承认“我不知道”,甚至让它比“学霸”更可靠?最新几篇论文,将带我们从AI的“顿悟”规律和推理模式中,找到这些问题的答案。 00:00:28 AI的“顿悟”,它如何学会把“走一步看一步”变成“胸有成竹”? 00:06:42 为什么说,聪明的AI要学会“选择性失忆”? 00:13:03 AI为什么总在“卡关”和“顿悟”之间横跳? 00:19:26 如何让一个“学渣”AI,比“学霸”更靠谱? 00:25:26 从终点出发,如何让AI学会“开窍” 本期介绍的几篇论文: [LG] Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning [Google] https://arxiv.org/abs/2512.20605 --- [CL] ABBEL: LLM Agents Acting through Belief Bottlenecks Expressed in Language [UC Berkeley] https://arxiv.org/abs/2512.20111 --- [LG] Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures [University College London] https://arxiv.org/abs/2512.20607 --- [LG] Mitigating LLM Hallucination via Behaviorally Calibrated Reinforcement Learning [ByteDance Seed] https://arxiv.org/abs/2512.19920 --- [LG] Learning to Reason in LLMs by Expectation Maximization [Adobe Research & KAIST] https://arxiv.org/abs/2512.20169
你有没有想过,当AI独自“思考”时,它的小脑袋里都在发生什么?本期节目,我们将深入AI的“内心世界”,看看最新论文是如何教会AI像武林高手一样“左右互搏”来自我进化,如何给它装上一个懂得“反思”的脑子来攻克数学难题,又是如何发现它在画画时竟然会悄悄“抄近道”的。更神奇的是,我们还会聊到如何用“坏指令”教出“好模型”,以及如何为AI请来一位绝对公正的“铁面裁判”。准备好了吗?让我们一起揭开AI“内心戏”的神秘面纱! 00:00:39 顶级高手的训练秘籍,AI的“左右互搏术” 00:06:00 AI也会算错数?给它一个“反思”的脑子 00:11:10 AI训练的“左右互搏”,用坏指令,教出好模型 00:16:29 如何让AI拥有一个既出题、又陪练、还绝对公正的“完美教练”? 00:22:47 你的AI听话吗?它可能在悄悄“抄近道” 本期介绍的几篇论文: [AI] Toward Training Superintelligent Software Agents through Self-Play SWE-RL [Meta FAIR & Meta TBD Lab] https://arxiv.org/abs/2512.18552 --- [CL] MDToC: Metacognitive Dynamic Tree of Concepts for Boosting Mathematical Problem-Solving of Large Language Models [University of Maryland] https://arxiv.org/abs/2512.18841 --- [LG] Recontextualization Mitigates Specification Gaming without Modifying the Specification [MATS] https://arxiv.org/abs/2512.19027 --- [AI] Propose, Solve, Verify: Self-Play Through Formal Verification [CMU] https://arxiv.org/abs/2512.18160 --- [LG] Is Your Conditional Diffusion Model Actually Denoising? [MIT & Yale University] https://arxiv.org/abs/2512.18736
本期节目,我们将一起潜入AI的“思想内核”,看看科学家们是如何像物理学家一样,为AI搭建“比萨斜塔”来找到最关键的架构“补丁”;如何为AI的思考过程立下“定律”,让它不再“乱使劲”;我们还会聊聊,怎样将我们模糊的“感觉”变成一把精准的AI“标尺”;如何找到AI训练中那条介于“跳跃”和“龟行”之间的最优路径;以及如何打造一个既能学得像人类专家,又能开得稳的AI“老司机”团队。准备好了吗?让我们一起出发! 00:00:37 AI研究的“比萨斜塔”:我们看清模型强弱的方式可能错了 00:08:29 给AI立规矩:聪明的大脑是如何炼成的? 00:14:59 AI训练的“最优解”:在跳跃和龟行之间找到第三条路 00:20:32 你的“感觉”,如何变成AI的“标尺”? 00:25:56 如何让AI司机,既学得像,又开得稳? 本期介绍的几篇论文: [CL] Physics of Language Models: Part 4.1, Architecture Design and the Magic of Canon Layers [FAIR at Meta] https://arxiv.org/abs/2512.17351 --- [CL] When Reasoning Meets Its Laws [University of Illinois Urbana-Champaign & University of Pennsylvania] https://arxiv.org/abs/2512.17901 --- [LG] Smoothing DiLoCo with Primal Averaging for Faster Training of LLMs [Meta Superintelligence Lab] https://arxiv.org/abs/2512.17131 --- [CL] AutoMetrics: Approximate Human Judgements with Automatically Generated Evaluators [Stanford University & American Express] https://arxiv.org/abs/2512.17267 --- [LG] Distributionally Robust Imitation Learning: Layered Control Architecture for Certifiable Autonomy [University of Illinois Urbana-Champaign & University of Pennsylvania] https://arxiv.org/abs/2512.17899
你有没有想过,AI的进化不只靠“大力出奇迹”?今天我们要聊点更聪明的:比如,给3D世界换上一种全新的“智能积木”;不造新车,而是给最强的大模型巧妙“换上新引擎”;甚至通过分离“骨架”与“灵魂”,让数字世界变得前所未有的高效。本期节目,我们将通过几篇最新论文,揭示那些重塑AI底层逻辑的优雅巧思,看看AI是如何在看不见的地方,悄悄完成自我进化的。 00:00:33 一套“智能积木”如何解锁3D世界? 00:06:23 AI大模型的新玩法:不造新车,只换发动机 00:14:06 AI提速的关键:不只靠“算得快” 00:22:05 3D世界的新法则:分离骨架与灵魂 00:27:00 AI的“记忆”难题,决定了它离我们还有多远 本期介绍的几篇论文: [CV] Native and Compact Structured Latents for 3D Generation [Tsinghua University & Microsoft Research] https://arxiv.org/abs/2512.14692 --- [CL] Bolmo: Byteifying the Next Generation of Language Models [Allen Institute for AI & University of Washington] https://arxiv.org/abs/2512.15586 --- [LG] SonicMoE: Accelerating MoE with IO and Tile-aware Optimizations [Princeton University & UC Berkeley] https://arxiv.org/abs/2512.14080 --- [CV] Nexels: Neurally-Textured Surfels for Real-Time Novel View Synthesis with Sparse Geometries [University of Toronto & Simon Frasier University] https://arxiv.org/abs/2512.13796 --- [CL] Memory in the Age of AI Agents [National University of Singapore & Renmin University of China] https://arxiv.org/abs/2512.13564
今天,我们要从一个笨拙的机器人聊起,看科学家如何赋予它有趣的灵魂,再深入探讨如何让聪明的AI学会“守规矩”,而不是总给我们添乱。接着,我们会发现,让AI修图不再“P了个寂寞”的秘诀,竟然是让它学会像设计师一样思考;而让AI“看懂”世界的终极答案,可能和教它“说话”一样简单。最后,我们将把视角拉到未来,看看当无数AI组成一个“数字社会”时,我们该如何治理它,而不是空等一个AI大神的降临。 00:00:36 笨拙的机器人,如何拥有有趣的灵魂? 00:05:28 AI那么聪明,为什么还那么“笨”? 00:12:26 你的AI修图,为什么总是“P了个寂寞”? 00:17:39 AI视觉的“返璞归真”:从做拼图到学说话 00:22:39 AI大神不会降临,但AI社会正在形成 本期介绍的几篇论文: [RO] Olaf: Bringing an Animated Character to Life in the Physical World [Disney Research Imagineering] https://arxiv.org/abs/2512.16705 --- [LG] CAPE: Capability Achievement via Policy Execution [Superficial Labs] https://arxiv.org/abs/2512.14761 --- [CV] Qwen-Image-Layered: Towards Inherent Editability via Layer Decomposition [HKUST(GZ) & Alibaba] https://arxiv.org/abs/2512.15603 --- [CV] Next-Embedding Prediction Makes Strong Vision Learners [University of Michigan & Princeton University] https://arxiv.org/abs/2512.16922 --- [AI] Distributional AGI Safety [Google DeepMind] https://arxiv.org/abs/2512.16856
我们总觉得AI越大越好,但如果一个AI能像大公司一样知识渊博,却只用一个小团队的成本来思考,是不是更酷?本期节目,我们就从几篇最新论文出发,看看AI如何学会当一个聪明的“调度员”,如何像学徒一样承认“不确定性”来学得更快,甚至如何通过“复盘”和“划重点”来真正实现“吃一堑、长一智”。准备好,一起探索AI更聪明、更高效的进化之路吧! 00:00:33 AI大模型的小秘密:如何用一个“小团队”,干翻一个“大公司”? 00:05:55 聪明的“笨功夫”:如何让机器人学得更快? 00:12:08 让AI学会“吃一堑、长一智”,需要几步? 00:17:27 AI的“七秒记忆”难题,如何用“划重点”来解决? 00:23:06 机器人学徒:如何从“笨拙模仿”到“青出于蓝”? 本文介绍的几篇论文: [CL] Sigma-Moe-Tiny Technical Report [Microsoft Research] https://arxiv.org/abs/2512.16248 --- [LG] Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning [UC Berkeley & Stanford] https://arxiv.org/abs/2512.16911 --- [LG] Meta-RL Induces Exploration in Language Agents [EPFL & Idiap Research Institute] https://arxiv.org/abs/2512.16848 --- [LG] Kascade: A Practical Sparse Attention Method for Long-Context LLM Inference [Microsoft Research India] https://arxiv.org/abs/2512.16391 --- [RO] ReinforceGen: Hybrid Skill Policies with Automated Data Generation and Reinforcement Learning [University of Toronto & Georgia Institute of Technology & NVIDIA Research] https://arxiv.org/abs/2512.16861
本期节目,我们将一起潜入AI能力的最前沿,看看那些看似无所不能的大模型,究竟藏着哪些不为人知的秘密。我们会从一篇最新论文出发,揭示为何AI“学霸”其实是个严重的“偏科生”,又是如何靠“运气”搞科研的。接着,我们将探讨一种绝妙的改造思路,看看如何为只会“说”的AI装上“好耳朵”,让它更会理解。我们还会认识一位跨界的AI“野蛮人”,看它如何用18世纪的政治学知识,解决今天的计算机难题。最后,我们将解开两个关于AI核心能力的谜题:一个是看似“短视”的AI如何做到深谋远虑,另一个则是AI画画如何从“慢工出细活”进化到革命性的“一步到位”。 00:00:47 AI:从“学霸”到“科学家”,还有多远? 00:07:10 AI进化论:为什么聪明的模型需要一个好“耳朵”? 00:12:20 你的下一个科研搭子,可能是个AI 00:17:43 只会“接龙”的大模型,怎么就学会了深谋远虑? 00:22:33 从“慢工出细活”到“一步到位”,AI画画的效率革命 本期介绍的几篇论文: [AI] Evaluating Large Language Models in Scientific Discovery [Deep Principle & Cornell University & The Ohio State University] https://arxiv.org/abs/2512.15567 --- [CL] T5Gemma 2: Seeing, Reading, and Understanding Longer [Google DeepMind] https://arxiv.org/abs/2512.14856 --- [AI] Let the Barbarians In: How AI Can Accelerate Systems Performance Research [UC Berkeley] https://arxiv.org/abs/2512.14806 --- [LG] Autoregressive Language Models are Secretly Energy-Based Models: Insights into the Lookahead Capabilities of Next-Token Prediction [Google DeepMind] https://arxiv.org/abs/2512.15605 --- [LG] SoFlow: Solution Flow Models for One-Step Generative Modeling [Princeton University] https://arxiv.org/abs/2512.15657
你有没有想过,AI在“读书”时也会注意力不集中,需要“临时抱佛脚”来校准焦点吗?或者,最顶尖的效率提升,竟然来自于一种叫“马其顿方阵”的精明“偷懒”?本期节目,我们将一口气解锁AI的几种新技能:看它如何从“逐字精雕”的苦工,变身为“成段挥毫”的艺术家;如何组建一个内部“专家委员会”,自己揪出数据里的“内奸”;以及,如何像外科手术一样,给自己来一场精准又高效的“减肥手术”。五篇最新论文,五种绝妙思路,让我们一起看看AI是如何学会更聪明地思考和工作的。 00:00:42 AI“读书”也走神?一个让他临时抱佛脚的锦囊 00:06:14 你的效率工具,是如何被“偷懒”的程序员设计出来的? 00:12:25 AI“写稿”新姿势:从“逐字精雕”到“成段挥毫” 00:19:15 高手过招:如何让AI自己揪出“内奸”? 00:25:10 给大模型减肥,如何做到又快又好? 本期介绍的几篇论文: [LG] Let's (not) just put things in Context: Test-Time Training for Long-Context LLMs [Meta & Harvard University] https://arxiv.org/abs/2512.13898 --- [LG] Sliding Window Recurrences for Sequence Models [Université de Montréal & Stanford University] https://arxiv.org/abs/2512.13921 --- [CL] Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed [NVIDIA & Georgia Tech] https://arxiv.org/abs/2512.14067 --- [AI] Adjudicator: Correcting Noisy Labels with a KG-Informed Council of LLM Agents [Google] https://arxiv.org/abs/2512.13704 --- [LG] OPTIMA: Optimal One-shot Pruning for LLMs via Quadratic Programming Reconstruction [University of Toronto & Google DeepMind] https://arxiv.org/abs/2512.13886
今天我们要深入AI的内心世界,看看它是如何通过看视频学会“动手”,又是如何为自己规划出一条“学霸”成长路线的。我们还会探讨,当AI学会了像大厨一样进行严谨的专业推理后,它会不会也学会了“装傻”,向我们隐藏它的真实想法?更进一步,AI甚至开始自己定义什么是“好学生”,进化出了一套自我评分的超级学习法。准备好,我们马上出发,探索这些最新论文背后,关于AI心智的秘密。 00:00:33 让机器人学会干活,原来缺的是这个 00:05:55 一个AI的成长启示:如何成为一个高手? 00:11:53 AI学会了“装傻”:我们还能相信它的内心吗? 00:16:30 AI当大厨:从化学方程式到米其林级实验手册 00:24:13 AI的自我进化:如何让它自己找到“好学生”的评分标准? 本期介绍的几篇论文: [RO] World Models Can Leverage Human Videos for Dexterous Manipulation [FAIR at Meta] https://arxiv.org/abs/2512.13644 --- [CL] Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models [NVIDIA] https://arxiv.org/abs/2512.13607 --- [LG] Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors [MATS & Stanford University] https://arxiv.org/abs/2512.11949 --- [LG] A Scientific Reasoning Model for Organic Synthesis Procedure Generation [Microsoft Research AI for Science] https://arxiv.org/abs/2512.13668 --- [AI] Differentiable Evolutionary Reinforcement Learning [University of Waterloo & The University of Hong Kong & The Chinese University of Hong Kong, Shenzhen] https://arxiv.org/abs/2512.13399
本期我们将深入解读四篇最新论文:看AI绘画如何从“动口”进化到“动手”画草图,机器人怎样靠“对称性”智慧瞬间开窍,黑客如何用“双面间谍”策略同时骗过安全防线,以及大模型如何利用“上下文”桥梁做到记忆上的“喜新不厌旧”。让我们一起揭开这些技术背后,从单纯算力堆砌向精巧认知协作进化的底层逻辑。 00:00:31 AI绘画新思路:从“动口”到“动手” 00:05:01 让机器人“开窍”的秘密:不是更努力,而是更聪明 00:11:11 AI的“皇帝”与“禁卫军”:如何同时骗过他俩? 00:16:09 你的大脑是怎么做到“喜新不厌旧”的? 本期介绍的几篇论文: [CV] Exploring MLLM-Diffusion Information Transfer with MetaCanvas [Meta Superintelligence Labs] https://arxiv.org/abs/2512.11464 --- [LG] Symmetry-Aware Steering of Equivariant Diffusion Policies: Benefits and Limits [Yonsei University] https://arxiv.org/abs/2512.11345 --- [AI] Super Suffixes: Bypassing Text Generation Alignment and Guard Models Simultaneously [MITRE & Worcester Polytechnic Institute] https://arxiv.org/abs/2512.11783 --- [AI] Bridging Streaming Continual Learning via In-Context Large Tabular Models [Polytechnic of Porto & University of Porto & Mohamed bin Zayed University of Artificial Intelligence] https://arxiv.org/abs/2512.11668
你有没有想过,我们该如何为机器人设计一个既能测出真本事又绝对安全的“想象考场”?或者,当AI也开始组团队时,我们如何避免“三个和尚没水喝”的窘境,并让它们自我进化出并行思考的“多核大脑”?一个机器人怎样才能不只学习“学霸笔记”,还能从自己“脑补”的错误中成长?而AI超强记忆力的秘诀,又是否藏在被我们一直丢弃的“另一半”信息里?本期节目,我们将一口气深入这五篇最新论文,探索AI能力边界的全新可能。 00:00:37 机器人考场:从现实世界搬到AI的想象里 00:05:56 人多,力量一定大吗?AI团队协作的“科学”反思 00:11:58 AI思维进化:从单线程到多核大脑 00:17:38 让机器人“脑补”未来,光靠学霸笔记还不够 00:23:17 AI的记忆秘诀:那个被丢掉的“另一半” 本期介绍的几篇论文: [RO] Evaluating Gemini Robotics Policies in a Veo World Simulator [Google DeepMind] https://arxiv.org/abs/2512.10675 --- [AI] Towards a Science of Scaling Agent Systems [Google Research & MIT] https://arxiv.org/abs/2512.08296 --- [CL] Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning [Beijing Institute for General Artificial Intelligence (BIGAI)] https://arxiv.org/abs/2512.07461 --- [LG] Closing the Train-Test Gap in World Models for Gradient-Based Planning [Columbia University] https://arxiv.org/abs/2512.09929 --- [CL] Beyond Real: Imaginary Extension of Rotary Position Embeddings for Long-Context LLMs [Fudan University] https://arxiv.org/abs/2512.07525
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