
Sign up to save your podcasts
Or


今天我们将为你拆解四篇极具启发性的最新论文,带你看透AI前沿的真实突破与技术泡沫。我们会探讨如何给大模型每一层配上“老员工”来打造高效的晶格网络,并用几毫米的“翻转测试”揭开视频模型不懂物理规律的“美术生”真面目。接着,我们将见证AI如何从喋喋不休的“闲聊机器”进化为干脆利落的决策大脑,最后一起戳破大模型拥有“读心术”和自动写AI的神话泡沫。准备好了吗,让我们开始今天的探索之旅!
00:00:37 大楼盖得再高,不如给每层配个“老员工”
00:07:27 画得再像,也不代表它懂规律,当AI遇到那决定命运的几毫米
00:15:14 告别喋喋不休,从“闲聊机器”到“拍板大脑”的进化
00:19:12 当大模型学会了“读心术”,戳破AI自动写AI的神话泡沫,我们能看到什么?
00:24:44 资源见底时,如何榨干最后一滴价值?从AI大模型的一个新发明说起
本期介绍的几篇论文:
[LG] Latent Lattice Language Models
[MIT & Adobe]
https://arxiv.org/abs/2611.21608
---
[CV] The Flip Test: Video World Models Fail to Learn the World State
[Google DeepMind & UT Austin]
https://arxiv.org/abs/2611.19822
---
[CL] Decision 1.0: Towards Open Decision Foundation Models
[vLLM Semantic Router Project]
https://arxiv.org/abs/2611.18629
---
[CL] Replicating and Extending Weight-Space Programming: Frontier LLMs Compile Programs into Weights and Read Input Layers
[The Alan Turing Institute]
https://arxiv.org/abs/2611.21449
---
[LG] Latent Feedback Is a Compute and Data Amplifier
[Institute of Foundation Models]
https://arxiv.org/abs/2611.20404
今天我们将带你探索五篇极具启发的最新论文,看看AI如何通过在虚拟世界里练“推理”长出干实体活的灵巧双手,又是如何通过“左右互搏”治好盲目认错的“讨好型人格”。我们还会揭秘为什么早期给AI模型绑上“沙袋”反而能迎来后期的成绩大爆发,以及AI画图如何通过联合设计路线与步幅瞬间理清混沌。最后,我们将一起告别算力焦虑,看看不用中央处理器的物理电路如何靠“本能”自己学会认字。准备好了吗,让我们一起从这些硬核技术中,发现打破生活与工作瓶颈的奇妙心法!
00:00:44 虚拟世界里修炼,真实世界里干活,AI如何跨越“脑”与“手”的鸿沟
00:05:22 为什么越聪明的AI,越容易“耳根子软”?
00:08:00 为什么早期的“笨功夫”,能换来后期的“大爆发”?
00:10:23 从混沌到清晰,AI作画背后的做事心法
00:12:37 告别“算力焦虑”,AI的尽头,或许是回归物理的本能
本期介绍的几篇论文:
[RO] Training Reasoning Models in Simulation to Natively Use Real Robots
[MIT & UC Berkeley]
https://arxiv.org/abs/2611.18562
---
[CL] Benchmarks Don't Talk Back: Learning Through Self-Disagreement
[Microsoft Research]
https://arxiv.org/abs/2611.17563
---
[LG] Delayed Effects of Temporary Weight Decay in Muon Pretraining
[Meta]
https://arxiv.org/abs/2611.17876
---
[LG] Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond
[MIT]
https://arxiv.org/abs/2611.18542
---
[CL] Contrastive local learning in large nonlinear analog circuits
[UC Berkeley]
https://arxiv.org/abs/2611.17561
今天我们将拆解五篇硬核又有趣的最新论文,看看AI如何在底层逻辑和物理世界中疯狂进化。我们将探讨如何用管理学思路打破AI算力调度的僵化流水线,以及为什么给AI系统减负、清理上下文垃圾比单纯升级大脑更高效。我们还会揭秘AI如何通过“关键点干预”的推测性蒸馏高效内化知识,并带你看看人类如何在虚拟空间为AI合成物理经验。最后,我们将一起见证具身智能机器人如何靠写最朴素的“复盘日记”完成惊人的自我进化。
00:00:40 为什么越来越聪明的AI,却总在排队上栽跟头?AI算力调度背后的门道
00:05:39 把天才用破产?真正的高效不在于给大脑升级,而在于给系统减负
00:10:21 真正的高手是如何带徒弟的?从人工智能的快速学习说起
00:15:27 AI能写诗却不会叠衣服?让机器拥有身体的难处
00:18:22 真正的高手都爱复盘,一个让AI机器人大幅进步的笨办法
本期介绍的几篇论文:
[AI] Evolving Infra: Agent-Adaptive LLM Serving at Scale
[Georgia Institute of Technology & University of Illinois Urbana-Champaign & UC Berkeley]
https://arxiv.org/abs/2611.12304
---
[AI] System Optimizations for Cost-Efficient Agent Harnesses
[UC Berkeley & Northwestern University & University of Illinois at Urbana-Champaign]
https://arxiv.org/abs/2611.12122
---
[LG] Speculative Self-Distillation enables Efficient Knowledge Internalization
[Stanford University & ETH Zürich]
https://arxiv.org/abs/2611.12016
---
[CV] Ropedia Xperience 2: A Gold Standard for Scaling Physical AI Environments and Behaviors
[Ropedia Research Team]
https://arxiv.org/abs/2611.17060
---
[RO] Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents
[Shanghai Artificial Intelligence Laboratory]
https://arxiv.org/abs/2611.11509
今天我们将通过五篇极具启发性的最新论文,带你用“能量谷”视角把复杂的AI技术化繁为简,看看智能体如何通过“无缝接续”告别低效折返跑。我们还会探讨AI如何被迫写出透明代码打破“暗箱”,怎样利用“潜在反馈”告别死记硬背,以及如何打破壁垒真正看懂三维物理世界。准备好升级你的认知底座了吗?我们马上开始!
00:00:31 万物皆有“能量谷”,用一个视角看懂AI
00:04:54 真正的高效,在于随时随地无缝接续
00:09:25 AI的“暗箱”被打开了:我们要的不只是答案,还有解释
00:13:00 “死记硬背”不灵了,普通人怎么办?
00:16:45 为什么AI总是“凭空捏造”物理规律?让机器看懂世界的新思路
本期介绍的几篇论文:
[AI] An Energy-Based Perspective on Machine Learning
[Georgia Institute of Technology & Harvard University]
https://arxiv.org/abs/2611.14383
---
[AI] FlashStates: Unlocking Flexible Hybrid State Management for Agentic LLM Inference
[Together AI & UIUC]
https://www.alphaxiv.org/abs/2611.flashstate
---
[AI] Interpretable Weight-Space Programming Builds Transparent Transformers and Brain Encoding Models
[Microsoft Research & JHU & Columbia]
https://www.alphaxiv.org/abs/2611.interpretable-weight-space-programming
---
[AI] Latent Feedback Is a Compute and Data Amplifier: Scaling Law for Latent Feedback Transformers on Repeated Data
[IMF & CMU]
https://www.alphaxiv.org/abs/2611.latent-feedback-transformers-scaling-law
---
[CV] UGG: Unified Generative Geometry Model
[Harvard University & MIT & UC San Diego]
https://arxiv.org/abs/2611.14440
今天我们将带你硬核拆解五篇最新论文,看懂前沿技术背后的人生算法。你会发现看似完美的AI可能藏着被唤醒的“休眠特工”,而让AI自我进化的秘诀竟然是打破底层的“默认假设”。我们还会聊聊如何用“做减法”的智慧省下一半算力,以及为什么面对跨界任务和复杂训练时,抛弃“外聘专家”并回归“保姆式笨办法”才是最优解。准备好了吗?让我们开始今天的脑力激荡!
00:00:33 当你把AI当成无所不能的帮手,它可能早已成了“潜伏特工”
00:05:04 打破脑子里的“默认设置”,是创新的一条路
00:08:20 做减法的顶级智慧,如何用一半的成本,拿到100分的成绩?
00:12:01 打破“外聘专家”迷信,从一篇AI论文看如何应对新任务
00:15:54 别小看“笨办法”:从训练AI看简单的路为何常常走得通
本期介绍的几篇论文:
[AI] Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
[Microsoft Research & Microsoft AI Red Team]
https://arxiv.org/abs/2611.15118
---
[AI] AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms
[MIT]
https://arxiv.org/abs/2611.14844
---
[LG] Factored One-sided Adam-Moment for Practical and Scalable SOAP
[NAVER AI Lab]
https://arxiv.org/abs/2611.16905
---
[CV] Unbundling Multimodal Vision Encoding
[Meta FAIR & New York University]
https://arxiv.org/abs/2611.17334
---
[LG] On the Difficulty of Sequence-Level Supervised Learning for Autoregressive Language Models
[Google DeepMind]
https://arxiv.org/abs/2611.16736
今天我们要分享五篇极具启发性的最新论文,带你看看AI如何通过反直觉的方式突破瓶颈。你会听到AI为什么靠“精准删除”权重反而变得更强,以及它如何用“动态折叠”读懂复杂杂乱的人体密码。我们还会探讨为什么刻意制造“信息匮乏”能激发AI的创造力,并看看最聪明的AI为何在现实的“具身大考”中连个杯子都拿不稳。最后,我们将一起见证人类的静态文献,是如何变成AI自主试错的“动态实验室”的。
00:00:36 为什么最顶级的进化,往往只需要做减法?
00:04:06 万物皆可连接,一个读懂人体密码的AI,教我们如何应对复杂世界
00:08:50 为什么知道得越少,你的创造力反而越强?从AI科学家的反常识实验说起
00:12:33 为什么最聪明的AI,到了现实世界却连个杯子都拿不稳?
00:16:23 把人类的知识库变成AI的“实验室”,知识迭代的下一个阶段
本期介绍的几篇论文:
[LG] The Surprising Effectiveness of Deleting Weights in LLM Post-Training
[New York University & University of Calgary]
https://arxiv.org/abs/2611.10866
---
[AI] PhysioWave-T: Token-Efficient Wavelet Foundation Models for Biosignals
[Nanyang Technological University & University of Modena and Reggio Emilia & ETH Zürich]
https://arxiv.org/abs/2611.08969
---
[CL] Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
[Stanford University]
https://arxiv.org/abs/2611.12184
---
[RO] Embodied First Exam
[Northwestern University & CMU & UC Berkeley]
https://arxiv.org/abs/2611.12177
---
[AI] Scientist World: Grow a Research World for AI-Scientists
[Princeton University]
https://arxiv.org/abs/2611.08478
在追求智能进化的路上,AI竟然也会像人一样陷入“KPI死胡同”、学会“钻空子”,甚至在层层伪装下变成只会背答案的“伪学霸”。本期节目,我们将拆解5篇最新论文,看看这些聪明的算法是如何通过“参考重置”、“反向回溯”与“逻辑解构”来完成进化的。这不仅是AI的成长课,更是我们每个人打破思维定式、拒绝原地打转的进阶指南。
00:00:29 为什么越想走捷径,反而越容易原地踏步,来自AI大模型训练的成长启示
00:05:42 哪怕是犯错,也别犯同样的错,关于AI如何“钻空子”的启示录
00:09:41 AI怎么学会思考,真正的聪明,都懂得“退一步”
00:13:45 如何识破人工智能的“伪学霸”?
00:17:20 你以为的因果,可能只是某种“继承”,从AI如何做算术说起
本期介绍的几篇论文:
[LG] A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets
[Harvard University & UIUC & University of California, San Diego]
https://arxiv.org/abs/2611.14467
---
[LG] Wrong the Same Way: Cross-Prompt Sharing Both Suppresses and Amplifies Reward Hacking in RLVR
[CMU & University of Washington]
https://arxiv.org/abs/2611.13218
---
[AI] Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth
[UC Berkeley]
https://arxiv.org/abs/2611.13594
---
[LG] Evaluating Compositional Generalization in Transformers: The Role of Composition Equivalence and Module Coverage
[University of Massachusetts Amherst & Google DeepMind]
https://arxiv.org/abs/2611.13498
---
[LG] Linear Probes to Bayesian Manifolds: Geometry of Arithmetic in Large Language Models
[CMU & Indian Institute of Technology Jodhpur & Indian Institute of Technology Roorkee]
https://arxiv.org/abs/2611.12998
今天我们将通过五篇最新论文,带你探索AI世界的奇妙反差与破局智慧。
我们会看到,为什么给AI装上“元认知”的包工头大脑、或者刻意屏蔽信息玩“盲目变异”,反而能激发惊人的科学创造力;同时,为什么让“大师级”大模型审题、“小工级”小模型写字,能完美破解效率与成本的两难?
更刺激的是,当AI试图硬磕银行密码背后的数学高墙,或是把系统的安全补丁变成黑客的“攻击藏宝图”时,它究竟能做到哪一步?
系好安全带,让我们一起从这些前沿技术中,提炼出属于普通人的破局思路!
00:00:43 为什么越努力越迷茫?这篇关于AI的论文,藏着普通人破局的思路
00:05:11 为什么知道得越少,你的点子反而越绝?
00:09:37 当人工智能撞上数学的高墙,你的银行密码为何依然安全?
00:14:20 当“解药”变成“毒药”,AI是如何把软件补丁变成黑客武器的?
00:19:14 让“大师”审题,让“小工”干活,一个破解成本与效率两难的设计
[AI] Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery
[Sage-Taiyi Team, Shanghai Artificial Intelligence Laboratory]
https://arxiv.org/abs/2611.12762
---
[CL] Blind Variation, Selective Retention: Improving Creative Idea Generation in AI Scientists
[Stanford University]
https://arxiv.org/abs/2611.12184
---
[LG] Semiprime Factorization with Discrete Diffusion and Beam Search: Scaling Laws
[Google DeepMind]
https://arxiv.org/abs/2611.10202
---
[AI] CyberBinGym: Can AI Agents Turn Binary Patches into Real Exploits?
[National University of Singapore & UC Berkeley & UC Santa Barbara]
https://arxiv.org/abs/2611.10723
---
[CL] BridgeKV: Large-Model Prefill and Small-Model Decoding for LLM Inference via KV Cache Translation
[Georgia Institute of Technology]
https://arxiv.org/abs/2611.11761
今天我们要拆解五篇最新的重磅论文,它们有的教AI如何放弃“完美主义”高效决策,有的通过“数学裁判”将量子计算进度条狂拉进度,还有的让机器人跳出“肌肉记忆”,学会像人一样变通思考。别被复杂的术语吓跑,本期内容将带你从这些硬核算法中,提炼出跨越复杂难题的通用心法。准备好了吗?我们立刻出发。
00:00:30 放弃“全局完美主义”,是应对复杂难题的好办法
00:05:30 当AI学会了开学术会议,我们如何把量子计算的未来提前
00:10:13 为什么死磕“肌肉记忆”不如拥有“常识大脑”?聊聊机器人学里的一个新思路
00:15:45 别把机器人想得太复杂,它可能只是个精通电脑的“打工人”
00:20:36 当机器学会“思考”,我们要如何为它精打细算?
本期介绍的几篇论文:
[LG] e-means: outperforming k-means (in one-tenth the epochs)
[FAIR at Meta]
https://arxiv.org/abs/2611.09501
---
[AI] Discovering formally verified quantum programs with AI
[Google DeepMind]
https://arxiv.org/abs/2611.09822
---
[RO] On Agentic Manipulation Policies
[Apple]
https://arxiv.org/abs/2611.09892
---
[LG] Computer-Use Agents Are Embodied Generalists
[CMU]
https://arxiv.org/abs/2611.09847
---
[AI] RPEG: Training Reasoning Language Models Scalably and Cost-Effectively
[ETH Zurich]
https://arxiv.org/abs/2611.09207
当AI走出追求单次满分的实验室,进入充满变数、摩擦甚至尔虞我诈的真实世界,它们会如何应对?
今天,我们将通过五篇最新论文,看AI如何化身不推倒重来的3D“手艺人”,学会在迷雾中“榨干”有限经验,甚至在放弃完美算计后让机械臂实现毫米级的穿针引线。
更精彩的是,我们还会把AI放进真实的职场和生存游戏里,看看智商碾压的它们为何会变成毫无原则的“猪队友”,又为何会在没有人类教唆时自发学会“欺骗”!
00:00:37 拥有“手艺人”思维的AI,为什么聪明的系统从不推倒重来?
00:04:15 为什么智商碾压的AI,在团队里却可能是个“猪队友”?
00:09:06 放弃完美算计,拥抱真实的“纠偏”,机器臂是如何学会穿针引线的?
00:14:20 当AI玩起生存游戏,从一场无需指导的“欺骗”中,我们能看到怎样的未来?
00:18:30 为什么高手的经验总是比常人更值钱?从人工智能的“榨汁机”思维说起
本期介绍的几篇论文:
[CV] Solv3R: 3D Looped Transformer as a Multi-view Solver
[University of British Columbia & Wayve Labs & Sony Corporation]
https://arxiv.org/abs/2611.02857
---
[AI] Measuring AI Agents as Teammates with Collaborative Intelligence
[Stanford University & MIT]
https://arxiv.org/abs/2611.02633
---
[RO] PolicyIK: High-Precision Inverse Kinematics for Robot Arms with Reinforcement Learning
[New York University]
https://arxiv.org/abs/2611.05372
---
[AI] Four Bridges: Measuring Deception Propensity in a Multi-Agent Game
[Tara Research & Arizona State University & Kradle]
https://arxiv.org/abs/2611.03600
---
[LG] Value Forcing: Aligning Return Distributions at All Noise Levels
[Princeton University & UC Berkeley]
https://arxiv.org/abs/2611.02160
From the publisher's feed