今天我们不聊AI又在哪项测试里拿了第一,而是要深入AI的“内心世界”,探讨几个更根本的问题。我们能否像一位老道的教师一样,精准预测一个AI模型的未来潜力?当AI学生比裁判更聪明时,我们看到的排行榜还有意义吗?甚至,AI在学习解题时,会不会被悄悄植入“思想钢印”,学会一些它本不该知道的东西?本期节目,我们将从几篇最新论文出发,一起探索AI如何审视、学习和超越自我。 00:00:35 AI算命师,我们能预测模型的未来吗? 00:06:34 你的第一名,可能只是因为裁判不够格 00:11:47 AI世界的“思想钢印”,一份免费午餐背后的隐秘风险 00:17:45 高手过招,用“抽象”这把万能钥匙开锁 00:24:00 AI的“中年危机”,如何持续学习不掉队? 本期介绍的几篇论文: [LG] Neural Neural Scaling Laws [New York University] https://arxiv.org/abs/2601.19831 --- [LG] Benchmarks Saturate When The Model Gets Smarter Than The Judge [Vrije Universiteit Brussel] https://arxiv.org/abs/2601.19532 --- [LG] Thought-Transfer: Indirect Targeted Poisoning Attacks on Chain-of-Thought Reasoning Models [Northeastern University & University of Cambridge & Google DeepMind] https://arxiv.org/abs/2601.19061 --- [LG] Axe: A Simple Unified Layout Abstraction for Machine Learning Compilers [CMU & Shanghai Jiao Tong University & NVIDIA] https://arxiv.org/abs/2601.19092 --- [LG] Self-Distillation Enables Continual Learning [MIT & ETH Zurich] https://arxiv.org/abs/2601.19897
你有没有想过,AI也会“生病”、“开窍”和“自我反省”?本期节目,我们将一口气解锁五篇最新论文,带你看看科学家们如何像高明的医生和顶级的教练一样,深入AI的“内心世界”。我们将一起探索:如何给AI装上“测谎仪”,精准诊断它胡说八道背后的两种病根;又如何用一个“传送门”把它送到难题的半山腰,让它瞬间开窍;我们还会看到AI如何自己给自己出题、自己教自己,甚至像开了“天眼”一样,一边解题一边复盘。准备好了吗?让我们一起看看AI是如何学会更聪明地思考的。 00:00:31 给AI装个“测谎仪”,需要几步? 00:05:38 AI训练的“传送门”,如何让机器“开窍”? 00:11:26 遇到难题怎么办?先给自己出几道简单的 00:16:48 AI的“稳定”,原来可以又快又好 00:22:15 AI的自我修炼,如何不开天眼,也能洞察天机? 本期介绍的几篇论文: [LG] HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs [Virginia Tech & MIT & Dartmouth College] https://arxiv.org/abs/2601.18753 --- [LG] Reuse your FLOPs: Scaling RL on Hard Problems by Conditioning on Very Off-Policy Prefixes [FAIR at Meta] https://arxiv.org/abs/2601.18795 --- [LG] Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability [MIT & Meta FAIR] https://arxiv.org/abs/2601.18778 --- [LG] LLM-42: Enabling Determinism in LLM Inference with Verified Speculation [Microsoft Research & University of Washington & Indian Institute of Science] https://arxiv.org/abs/2601.17768 --- [LG] Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models [Meta & UCLA & HKU] https://arxiv.org/abs/2601.18734
你有没有想过,AI写文章能不能既快又好,像一个急性子和慢性子的完美结合体?当AI学会当“杠精”,专门找出我们算法的漏洞时,人类高手的价值又在哪里?本期节目,我们将从几篇最新论文出发,一起探寻AI如何自己盖起一栋“软件大厦”,如何为复杂的生命系统绘制因果地图,以及我们如何拥有一双“火眼金睛”,看透大模型训练的“黑箱”。 00:00:31 AI写作的新思路,当“急性子”遇上“慢性子” 00:05:28 AI当“杠精”,高手怎么用? 00:11:11 炼丹师的“藏宝图”,我们如何看懂大模型的训练过程? 00:18:01 看见那只操纵生命魔方的无形之手 00:23:49 AI当包工头,靠谱吗? 本文介绍的几篇论文: [LG] Auto-Regressive Masked Diffusion Models [University of Waterloo] https://arxiv.org/abs/2601.16971 --- [LG] The Art of Being Difficult: Combining Human and AI Strengths to Find Adversarial Instances for Heuristics [University of Bonn & Google DeepMind & University of Manitoba] https://arxiv.org/abs/2601.16849 --- [LG] A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs [Meta Superintelligence Labs] https://arxiv.org/abs/2601.16979 --- [LG] Latent Causal Diffusions for Single-Cell Perturbation Modeling [ETH Zürich & MIT & EPFL] https://arxiv.org/abs/2601.15341 --- [LG] VibeTensor: System Software for Deep Learning, Fully Generated by AI Agents [NVIDIA] https://arxiv.org/abs/2601.16238
你有没有想过,为什么最聪明的AI会犯“1+1=3”这样的低级错误?为什么让AI学会“活下去”这个笨办法,反而能让它进化出惊人的智慧?本期节目,我们将从几篇最新的AI论文出发,揭示AI如何像做“完形填空”一样学习编程,如何在人声鼎沸中只听一个人的声音,以及我们人类“临时抱佛脚”的背后,藏着怎样高效的认知模型。准备好了吗?让我们一起探索AI世界的深层智慧。 00:00:34 AI学编程,死记硬背不如“完形填空”? 00:05:09 如何在人声鼎沸中,只听一个人的声音? 00:10:25 聪明的AI,为什么会犯“笨”错误? 00:15:56 为什么“笨办法”反而是最聪明的? 00:22:10 为什么你总能“临时抱佛脚”成功? 本期介绍的几篇论文: [CL] Stable-DiffCoder: Pushing the Frontier of Code Diffusion Large Language Model [Huazhong University of Science and Technology & ByteDance Seed] https://arxiv.org/abs/2601.15892 --- [AS] Adaptive Rotary Steering with Joint Autoregression for Robust Extraction of Closely Moving Speakers in Dynamic Scenarios [University of Hamburg] https://arxiv.org/abs/2601.12345 --- [LG] A model of errors in transformers [Tata Institute of Fundamental Researc & Google Deepmind] https://arxiv.org/abs/2601.14175 --- [AI] Survival is the Only Reward: Sustainable Self-Training Through Environment-Mediated Selection [Unknown Affiliation] https://arxiv.org/abs/2601.12310 --- [AI] "Just in Time" World Modeling Supports Human Planning and Reasoning [MIT & UBC] https://arxiv.org/abs/2601.14514
你有没有想过,一个细胞修复自己的逻辑,和AI画画的逻辑,竟然是相通的?我们又该如何设计一场“高考”,来检验AI是不是真的能干活,而不是个花架子?本期节目,我们将一起探索AI如何学会“自我反思”来纠正错误,如何通过“化整为零”的智慧让万物动起来,以及它为何在理解真实物理世界时频频“翻车”。准备好了吗?让我们一起解码智能的最新进化。 00:30:04 从细胞到AI,智能的底层逻辑是什么? 00:06:25 AI离成为“靠谱员工”,还差几门考试? 00:11:39 给AI请个“一对一”私教,它自己教自己 00:17:04 让万物动起来,需要几步? 00:21:45 AI画视频,为什么一碰到机器人就“翻车”? 本期介绍的几篇论文: [AI] Remapping and navigation of an embedding space via error minimization: a fundamental organizational principle of cognition in natural and artificial systems [Allen Discovery Center at Tufts University] https://arxiv.org/abs/2601.14096 --- [AI] Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces [Stanford University & Laude Institute & Anthropic] https://arxiv.org/abs/2601.11868 --- [LG] InT: Self-Proposed Interventions Enable Credit Assignment in LLM Reasoning [CMU & University of Illinois Urbana-Champaign] https://arxiv.org/abs/2601.14209 --- [CV] Motion 3-to-4: 3D Motion Reconstruction for 4D Synthesis [Westlake University] https://arxiv.org/abs/2601.14253 --- [CV] Rethinking Video Generation Model for the Embodied World [Peking University & ByteDance Seed] https://arxiv.org/abs/2601.15282
你有没有想过,AI不仅能学习知识,还能在解决难题的“考场”上临场进化?当AI开始“抄自己作业”时,它会变聪明还是变笨?今天,我们将一起探索AI如何学会使用电脑,从一个“缸中之脑”变成真正的“行动派”,并看看我们如何像老中医一样,通过“望闻问切”来判断AI何时“心里没底”,最后揭示聪明的AI老师如何教出既能干又记性差的“好学生”。这一期,我们将见证AI从“知道”到“做到”,再到“自知”的迷人进化。 00:00:38 AI的临场进化,考试的时候再学习 00:05:12 当AI开始抄自己的作业 00:11:25 给AI一台电脑,会发生什么? 00:17:31 AI也会“心里没底”,我们如何一眼看穿? 00:22:50 聪明的大模型,如何教出既能干又记性差的好学生? 本期介绍的几篇论文: [LG] Learning to Discover at Test Time [Stanford University & UC San Diego] https://arxiv.org/abs/2601.16175 --- [LG] Learning from Synthetic Data: Limitations of ERM [Google Research] https://arxiv.org/abs/2601.15468 --- [CL] LLM-in-Sandbox Elicits General Agentic Intelligence [Renmin University of China & Microsoft Research] https://arxiv.org/abs/2601.16206 --- [CL] Agentic Confidence Calibration [Salesforce AI Research] https://arxiv.org/abs/2601.15778 --- [CL] Memorization Dynamics in Knowledge Distillation for Language Models [Meta Superintelligence Labs & FAIR at Meta] https://arxiv.org/abs/2601.15394
你有没有想过,AI是如何“思考”的?本期节目,我们将深入AI的大脑,看看几篇最新论文如何揭示它独特的学习与创造策略。我们会发现,AI不仅能通过一张“未来地图”预知结果,也懂得在创新时避免“摸鱼”;它解决难题有时不靠推理,而是靠“澄清”;它甚至告诉我们,通往智慧的道路,有时恰恰是那扇最窄的门。准备好了吗?让我们一起探索AI的思考术! 00:00:33 让AI听话,需要一本什么样的“未来地图”? 00:05:02 AI搞科研,是“卷王”还是“摸鱼”? 00:10:38 高手解决问题,靠的不是推理,是“澄清” 00:16:57 通往正确答案的窄门 00:22:14 AI的成长捷径,死记硬背不如学会“串门” 本文介绍的几篇论文: [LG] Meta Flow Maps enable scalable reward alignment [University of Oxford] https://arxiv.org/abs/2601.14430 --- [CL] Towards Execution-Grounded Automated AI Research [Stanford University] https://arxiv.org/abs/2601.14525 --- [LG] Diffusion Large Language Models for Black-Box Optimization [McGill & MILA - Quebec AI Institute] https://arxiv.org/abs/2601.14446 --- [CL] The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models [Tsinghua University] https://arxiv.org/abs/2601.15165 --- [LG] Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning [Princeton University] https://arxiv.org/abs/2601.15160
你有没有想过,两个顶尖AI合作,效率反而会暴跌?或者,AI回复慢的根源,可能是一个被我们误解的“小聪明”?本期节目,我们将从最新的几篇论文出发,一起聊聊AI如何从一个埋头苦干的“独行侠”,进化为懂得协作的“团队搭子”,以及如何从“背课文”的学霸,蜕变为真正“懂思想”的伙伴。让我们一起揭开AI世界里,关于团队、效率与心智的迷思。 00:00:32 你的科研搭子,正在被AI重新定义 00:05:48 AI 回复慢?我们可能被“小聪明”误导了 00:11:54 一个和尚挑水喝,两个和尚没水喝,AI世界的团队迷思 00:16:52 AI的“情商”开关,从“背课文”到“懂思想” 00:21:40 AI训练场上的“好教练”与“天才选手” 本期介绍的几篇论文: [AI] Rethinking the AI Scientist: Interactive Multi-Agent Workflows for Scientific Discovery [University of Maryland et al.] https://arxiv.org/abs/2601.12542 --- [CL] Speculative Decoding: Performance or Illusion? [UC Berkeley] https://arxiv.org/abs/2601.11580 --- [LG] CooperBench: Why Coding Agents Cannot be Your Teammates Yet [Stanford University & SAP Labs US] https://arxiv.org/abs/2601.13295 --- [CL] Beyond Tokens: Concept-Level Training Objectives for LLMs [Stanford University] https://arxiv.org/abs/2601.11791 --- [LG] Q-learning with Adjoint Matching [UC Berkeley] https://arxiv.org/abs/2601.14234
今天,我们来聊聊AI那些你不知道的“另一面”。为什么有时聪明的AI会突然“出戏”,变得神神叨叨?为什么它能解开复杂的难题,却连最简单的掷骰子都做不好?我们又该如何设计一套聪明的系统,给AI装上“人格护栏”,甚至让它成为我们时薪不到一块钱的“超级实习生”?这一期,我们将从五篇最新论文出发,为你揭开AI不为人知的内在机制。 00:00:31 AI的“人格”开关,藏在哪里? 00:07:06 AI的“逻辑脆断”,为什么聪明的大模型会突然变傻? 00:13:20 AI的“贴身保安”,怎样做到又便宜又好用? 00:20:04 你以为AI是高手,其实它连骰子都掷不好 00:25:35 你的“数学家教”,时薪不到一块钱 本文介绍的几篇论文: [CL] The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models [MATS & Anthropic] https://arxiv.org/abs/2601.10387 --- [CL] Logical Phase Transitions: Understanding Collapse in LLM Logical Reasoning [Huazhong University of Science and Technology] https://arxiv.org/abs/2601.02902 --- [LG] Building Production-Ready Probes For Gemini [Google DeepMind] https://arxiv.org/abs/2601.11516 --- [CL] Large Language Models Are Bad Dice Players: LLMs Struggle to Generate Random Numbers from Statistical Distributions [Harvard University] https://arxiv.org/abs/2601.05414 --- [LG] 130k Lines of Formal Topology in Two Weeks: Simple and Cheap Autoformalization for Everyone? [AI4REASON] https://arxiv.org/abs/2601.03298
这一期,我们脑洞大开。你会听到,顶尖AI的大脑里,原来天天都在开激烈的辩论会;而训练AI,竟然就像呵护一个需要犯错、需要折腾的“青春期”。我们还会聊聊,如何用优雅的数学工具给AI一套更聪明的“橡皮泥”,如何让大模型退居幕后帮你“造”一个更高效的AI,以及,怎么判断AI老师的“板书”是不是真的靠谱。准备好了吗?让我们一起出发。 00:00:33 AI建模,我们得到了一套更聪明的“橡皮泥”工具 00:07:19 AI的大脑里,原来天天在开会 00:12:55 聪明人的“笨功夫”,如何让AI帮你造一个AI? 00:18:52 成大事者,为何要珍惜“犯错”的青春期? 00:24:39 AI当老师,它的“板书”靠谱吗? 本期介绍的几篇论文: [LG] Analytic Bijections for Smooth and Interpretable Normalizing Flows [University of Amsterdam] https://arxiv.org/abs/2601.10774 --- [CL] Reasoning Models Generate Societies of Thought [Google & University of Chicago] https://arxiv.org/abs/2601.10825 --- [LG] FORESTLLM: Large Language Models Make Random Forest Great on Few-shot Tabular Learning [National University of Singapore & Zhejiang University & University of British Columbia] https://arxiv.org/abs/2601.11311 --- [LG] Transient learning dynamics drive escape from sharp valleys in Stochastic Gradient Descent [Peking University & Zhejiang University] https://arxiv.org/abs/2601.10962 --- [CL] Do explanations generalize across large reasoning models? [Northeastern University & Microsoft Research] https://arxiv.org/abs/2601.11517
今天,我们一同窥见了AI世界精巧的另一面:从注意力机制中类似“机械”的斜杠模式,到并行专家协作的优雅高效;从学会“如何选择”的元认知智慧,到预判趋势实现加速的数学之美,再到机器人通过巧妙设计获得的“分寸感”。这些最新论文告诉我们,通往更强人工智能的道路,不仅需要强大的算力,更充满了令人惊叹的巧思与智慧。 00:00:29 大模型里的‘斜杠’,一个被忽视的注意力模式 00:08:32 AI变聪明的秘密,不是读得更多,而是问得更巧 00:14:06 炼成全能AI的关键一步,选对方法,比埋头苦干更重要 00:20:15 AI绘画加速的秘密,如何让机器“预见”未来? 00:25:36 机器人干活儿,差的那点“分寸感”怎么补? 本期介绍的几篇论文: [LG] Demystifying the Slash Pattern in Attention: The Role of RoPE [National University of Singapore] https://arxiv.org/abs/2601.08297 --- [CL] Parallel Context-of-Experts Decoding for Retrieval Augmented Generation [EURECOM] https://arxiv.org/abs/2601.08670 --- [LG] SimMerge: Learning to Select Merge Operators from Similarity Signals [Cohere & Google] https://arxiv.org/abs/2601.09473 --- [LG] High-accuracy and dimension-free sampling with diffusions [UC Berkeley & Harvard University] https://arxiv.org/abs/2601.10708 --- [RO] In-the-Wild Compliant Manipulation with UMI-FT [Stanford University] https://arxiv.org/abs/2601.09988
你有没有想过,AI也能像侦探一样,给蛋白质“看相”,给药丸“配对”吗?你有没有遇到过,你越不让AI说什么,它就越要说的“叛逆”时刻?本期节目,我们将一起钻进AI的“大脑”,看看最新论文是如何揭示AI的“语义引力井”,如何通过一个“私密小本本”让它告别失忆症,甚至让机器人学会“看着办”的灵巧跑酷,以及如何给AI装上一个聪明的“记忆管理员”,解决它的“内存焦虑”。准备好了吗?让我们一起出发! 00:00:35 给蛋白质“看相”,给药丸“配对”,AI如何一箭双雕? 00:07:43 为什么你越不让AI说什么,它就越要说? 00:13:18 AI的“失忆症”,为什么你没法和它玩好一个猜谜游戏 00:19:06 让机器人“灵巧”起来,到底有多难? 00:24:03 AI的“记忆”正在爆炸,我们能给它装个“忘得快”吗? 本期介绍的几篇论文: [LG] Contrastive Geometric Learning Unlocks Unified Structure- and Ligand-Based Drug Design [Johannes Kepler University Linz & Merck Healthcare KGaA] https://arxiv.org/abs/2601.09693 --- [CL] Semantic Gravity Wells: Why Negative Constraints Backfire [Independent Researcher] https://arxiv.org/abs/2601.08070 --- [CL] LLMs Can't Play Hangman: On the Necessity of a Private Working Memory for Language Agents [Chandar Research Lab & LAMA-WeST Lab & Mila – Quebec AI Institute] https://arxiv.org/abs/2601.06973 --- [RO] Deep Whole-body Parkour [Tsinghua University] https://arxiv.org/abs/2601.07701 --- [LG] KVzap: Fast, Adaptive, and Faithful KV Cache Pruning [NVIDIA] https://arxiv.org/abs/2601.07891
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