00:00:38 给AI的大脑做减法：一种聪明的“偷懒”智慧 00:04:34 你的数据正在“减肥”：人工智能时代的生存新法则 00:09:23 AI的“闭关修炼”：不给它网络，它能有多强？ 00:13:07 给AI装上一个“语法”导航仪 00:16:49 AI的“通感”：为什么“我写的”和“我是作者”是一回事？ 本期介绍的几篇论文： [LG] XQuant: Breaking the Memory Wall for LLM Inference with KV Cache Rematerialization [UC Berkeley & FuriosaAI] https://arxiv.org/abs/2508.10395 --- [LG] SoK: Data Minimization in Machine Learning [ETH Zurich] https://arxiv.org/abs/2508.10836 --- [CL] SSRL: Self-Search Reinforcement Learning [Tsinghua University & Shanghai AI Laboratory] https://arxiv.org/abs/2508.10874 --- [LG] Constrained Decoding of Diffusion LLMs with Context-Free Grammars [ETH Zurich] https://arxiv.org/abs/2508.10111 --- [CL] A Rose by Any Other Name Would Smell as Sweet: Categorical Homotopy Theory for Large Language Models [Adobe Research] https://arxiv.org/abs/2508.10018

当机器在逻辑、计算和效率上无限逼近甚至超越我们时，我们真正的“超能力”，是体验、是感受，是那份允许万物穿过身体的勇气和智慧。

00:00:32 让AI学会“学习”：不再死记硬背的秘密 00:04:30 给AI装个“外挂”：一个叫“记忆解码器”的U盘 00:08:23 开会吵不完？AI教你找共识 00:13:39 大海捞针？不如自己造根针 00:18:41 AI减肥记：如何让聪明的机器不说废话？ 本期介绍的五篇论文： [CL] Learning Facts at Scale with Active Reading [Meta FAIR] https://arxiv.org/abs/2508.09494 --- [CL] Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models [Shanghai Jiao Tong University] https://arxiv.org/abs/2508.09874 --- [CL] The Human-AI Hybrid Delphi Model: A Structured Framework for Context-Rich, Expert Consensus in Complex Domains [Google Research] https://arxiv.org/abs/2508.09349 --- [LG] SYNAPSE-G: Bridging Large Language Models and Graph Learning for Rare Event Classification [Google Research] https://arxiv.org/abs/2508.09544 --- [CL] Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoning [Microsoft Research] https://arxiv.org/abs/2508.09726

无论是人类学习一项新技能，还是AI模型的训练，第一步，永远不是追求“卓越”，而是确保“在场”。

00:00:34 你以为AI是高科技？它的底层逻辑可能几亿年前就有了 00:05:07 秒懂的秘密：如何看透行为背后的“算法” 00:09:28 AI大模型里的“快递小哥”：不只求快，还要求稳 00:14:22 AI的“双档切换”：如何用最聪明的大脑，干最划算的活？ 00:18:24 我们如何思考“如果”：一种全新的思维工具 本期介绍的五篇论文： [LG] Understanding Transformers through the Lens of Pavlovian Conditioning [Meta Platforms, Inc] https://arxiv.org/abs/2508.08289 --- [RO] Rational Inverse Reasoning [MIT CSAIL] https://arxiv.org/abs/2508.08983 --- [LG] Scaled-Dot-Product Attention as One-Sided Entropic Optimal Transport [Stanford University] https://arxiv.org/abs/2508.08369 --- [LG] OverFill: Two-Stage Models for Efficient Language Model Decoding [Cornell University] https://arxiv.org/abs/2508.08446 --- [LG] Topos Causal Models [Adobe Research] https://arxiv.org/abs/2508.08295

大大方方地做自己吧，带着你所有的“不完美”和“异常值”。

00:00:34 AI侦探养成记：如何让机器学会“死磕到底”？ 00:04:22 AI也需要“元学习”：如何打造一把能开万能锁的钥匙？ 00:07:56 拆解AI大脑：它如何学会“绕个弯”解决问题？ 00:12:21 AI学会“举一反三”的秘密：两层楼就够了？ 00:16:34 AI思考的秘密：为什么“少”就是“多”？ 本期介绍的五篇论文： [CL] Beyond Ten Turns: Unlocking Long-Horizon Agentic Search with Large-Scale Asynchronous RL [Tsinghua University] https://arxiv.org/abs/2508.079 --- [LG] AdaptFlow: Adaptive Workflow Optimization via Meta-Learning [Peking University & University of Chinese Academy of Sciences] https://arxiv.org/abs/2508.08053 --- [LG] Multi-head Transformers Provably Learn Symbolic Multi-step Reasoning via Gradient Descent [CMU & UPenn & OSU] https://arxiv.org/abs/2508.08222 --- [LG] What One Cannot, Two Can: Two-Layer Transformers Provably Represent Induction Heads on Any-Order Markov Chains [MIT & EPFL & UC Berkeley] https://arxiv.org/abs/2508.07208 --- [CL] Less Is More: Training-Free Sparse Attention with Global Locality for Efficient Reasoning [Princeton University & CMU] https://arxiv.org/abs/2508.07101

那个能一键“扭转乾坤”的开发者，不是别人，就是你自己。

00:00:30 AI界的“学霸”是怎么炼成的？ 00:04:45 你的下一个AI，为什么必须是个“行动派”？ 00:08:35 AI当上帝：从零开始创造一门语言 00:13:09 如何给AI大模型“摸骨”，看透它的知识边界？ 本期介绍的四篇论文： [CL] UR²: Unify RAG and Reasoning through Reinforcement Learning [Tsinghua University & Hebei University of Economics and Business] https://arxiv.org/abs/2508.06165 --- [CL] GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models [Zhipu AI & Tsinghua University] https://arxiv.org/abs/2508.06471 --- [CL] ConlangCrafter: Constructing Languages with a Multi-Hop LLM Pipeline [Tel Aviv University & UC Berkeley] https://arxiv.org/abs/2508.06094 --- [CL] Efficient Knowledge Probing of Large Language Models by Adapting Pre-trained Embeddings [Georgia Institute of Technology & MIT] https://arxiv.org/abs/2508.06030

你最强大的对手，和你最伟大的盟友，都是你。

00:00:35 我们永远无法根除AI的“幻觉”，但可以学会与它共舞 00:04:28 人工智能的“笨功夫”：一个鸟类识别模型教给我们的事 00:08:44 AI世界的“计分板”，正在悄悄升级 00:12:25 AI如何学会当数学家？三个你也能用的“笨”办法 00:17:15 AI码农进化论：如何“调教”一个更聪明的程序员？ 本期介绍的五篇论文： [CL] A comprehensive taxonomy of hallucinations in Large Language Models [Universitat de Barcelona] https://arxiv.org/abs/2508.01781 --- [LG] Perch 2.0: The Bittern Lesson for Bioacoustics [Google DeepMind] https://arxiv.org/abs/2508.04665 --- [CL] CompassVerifier: A Unified and Robust Verifier for LLMs Evaluation and Outcome Reward [Shanghai AI Laboratory] https://arxiv.org/abs/2508.03686 --- [LG] Goedel-Prover-V2: Scaling Formal Theorem Proving with Scaffolded Data Synthesis and Self-Correction [Princeton University & Tsinghua University] https://arxiv.org/abs/2508.03613 --- [LG] Training Long-Context, Multi-Turn Software Engineering Agents with Reinforcement Learning [Nebius AI] https://arxiv.org/abs/2508.03501

大多数人缺少的，不是天赋和雄心，而是一种人为创造的“紧迫感”。