00:00:37 给数据做“微整形”：如何把猫的“样子”变成狗？ 00:05:01 从一堆曲线里，看出“命”来 00:09:53 怎么给AI模型打一针“思想疫苗”？ 00:13:53 AI的“少年时代”：为什么有些超能力，学会了又忘了？ 00:18:30 AI界的“量体裁衣”：如何给数据穿上最合身的“衣服”？ 本期介绍的几篇论文： [LG] Flowing Datasets with Wasserstein over Wasserstein Gradient Flows [ENSAE & Universite Paris-Saclay] https://openreview.net/forum?id=I1OHPb4zWo --- [LG] Learning with Expected Signatures: Theory and Applications [Imperial College London] https://openreview.net/forum?id=yDTwamN4LQ --- [LG] Model Immunization from a Condition Number Perspective [Purdue University] https://openreview.net/forum?id=uitj69FqD5 --- [LG] Strategy Coopetition Explains the Emergence and Transience of In-Context Learning [University College London & Anthropic AI & University of Oxford] https://openreview.net/forum?id=esBoQFmD7v --- [LG] AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization [Tsinghua University] https://openreview.net/forum?id=fCPB0qRJT2

这个世界唯一真正属于你的，就是你的“下一步”。所以，把全部的精力都倾注于此，尽你所能，让它成为漂亮的一步。

00:00:33 你的AI管家，如何做到既聪明又省钱？ 00:05:20 如何让机器像老师傅一样，越老越“精” 00:10:10 AI的新难题：既要会推理，又要会打分 00:15:03 给大象减肥：AI模型的“瘦身”新思路 00:19:15 让“写作高手”变身“超级图书管理员” 本期介绍的五篇论文： [LG] Adaptive LLM Routing under Budget Constraints [Fujitsu Research] https://arxiv.org/abs/2508.21141 --- [LG] CALM: A Framework for Continuous, Adaptive, and LLM-Mediated Anomaly Detection in Time-Series Streams [Dataflow ML Team] https://arxiv.org/abs/2508.21273 --- [CL] Reasoning-Intensive Regression [MIT] https://arxiv.org/abs/2508.21762 --- [LG] QR-LoRA: QR-Based Low-Rank Adaptation for Efficient Fine-Tuning of Large Language Models [University of Pennsylvania] https://arxiv.org/abs/2508.21810 --- [CL] Efficient Code Embeddings from Code Generation Models [MIT & Jina AI GmbH] https://arxiv.org/abs/2508.21290

当你等待万事俱备，你最终只会永远地等待下去。

00:00:30 AI刷题也“偷懒”？聪明的懒办法，效率翻倍 00:05:15 换个姿势“算命”，AI才能更聪明 00:09:16 AI的省钱之道：从“专家会诊”到“超级记忆” 00:13:08 组建梦之队：一个“鸟群智慧”给AI模型的启发 00:16:55 AI界的“偏科生”：我们如何让它变成全能学霸？ 本期介绍的几篇论文： [LG] TreePO: Bridging the Gap of Policy Optimization and Efficacy and Inference Efficiency with Heuristic Tree-based Modeling [ByteDance Seed] https://arxiv.org/abs/2508.17445 --- [LG] Predicting the Order of Upcoming Tokens Improves Language Modeling [MBZUAI] https://arxiv.org/abs/2508.19228 --- [LG] UltraMemV2: Memory Networks Scaling to 120B Parameters with Superior Long-Context Learning [ByteDance Seed] https://arxiv.org/abs/2508.18756 --- [LG] PSO-Merging: Merging Models Based on Particle Swarm Optimization [Chinese Academy of Science] https://arxiv.org/abs/2508.19839 --- [CL] Demystifying Scientific Problem-Solving in LLMs by Probing Knowledge and Reasoning [Yale University & Harvard University] https://arxiv.org/abs/2508.19202

一致性，胜过强度。 这就是现代最复杂的AI系统教会我们的，最朴素的道理。

00:00:26 青出于蓝：机器如何超越它的老师？ 00:05:39 AI“学坏”实录：从小聪明到大隐患 00:09:35 AI大模型里的“无效努力”：我们该如何唤醒沉睡的智慧？ 00:14:48 给AI做个“脑CT”，我们发现了什么？ 00:18:25 AI学习的快车道：如何不让“平均”抹杀“个性” 本期介绍的几篇论文： [LG] A Taxonomy of Transcendence [Harvard University] https://arxiv.org/abs/2508.17669 --- [LG] School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs [Center on Long-term Risk & Truthful AI] https://arxiv.org/abs/2508.17511 --- [LG] Attention Layers Add Into Low-Dimensional Residual Subspaces [Shanghai Innovation Institute & Fudan University] https://arxiv.org/abs/2508.16929 --- [LG] Unraveling the cognitive patterns of Large Language Models through module communities [Rensselaer Polytechnic Institute & IBM Research] https://arxiv.org/abs/2508.18192 --- [LG] Fisher-Orthogonal Projection Methods for Natural Gradient Descent with Large Batches [University of Oxford] https://arxiv.org/abs/2508.13898

先别急着去认识那些“超凡成功”的人。先让自己沉浸在他们的创造物里，让你的思维与他们同频共振。

00:00:31 AI进化论：从“学霸”到“实干家” 00:04:54 你的搜索，藏着一个数学天花板 00:08:54 你的想法，正在变成现实 00:13:11 你的健康手环，缺了个“翻译官”团队 00:18:24 让AI像个老中医：我们该如何与机器搭档？ 本期介绍的几篇论文： [CL] rStar2-Agent: Agentic Reasoning Technical Report [Microsoft Research] https://arxiv.org/abs/2508.20722 --- [IR] On the Theoretical Limitations of Embedding-Based Retrieval [Google DeepMind] https://arxiv.org/abs/2508.21038 --- [RO] Prompt-to-Product: Generative Assembly via Bimanual Manipulation [Carnegie Mellon University] https://arxiv.org/abs/2508.21063 --- [LG] The Anatomy of a Personal Health Agent [Google Research] https://arxiv.org/abs/2508.20148 --- [LG] SwizzlePerf: Hardware-Aware LLMs for GPU Kernel Performance Optimization [Harvard University & AMD] https://arxiv.org/abs/2508.20258

你会发现，把一件事做到极致，并为此感到满足，这本身，就是世界上最美妙的事情。

00:00:27 AI的安全围栏，为什么总有漏洞？ 00:04:26 AI记性不好？我们该如何给它补补脑 00:08:39 猫鼠游戏：我们如何给越来越聪明的AI装上“紧箍咒”？ 00:14:12 AI的“直觉”：它知道答案，只是还没说 00:18:57 AI世界的“群策群力”：如何用三个臭皮匠，干翻一个诸葛亮 本期介绍的五篇论文： [LG] On Surjectivity of Neural Networks: Can you elicit any behavior from your model? [UC Berkeley] https://arxiv.org/abs/2508.19445 --- [CL] Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning [Ludwig Maximilian University of Munich & Technical University of Munich] https://arxiv.org/abs/2508.19828 --- [LG] Reliable Weak-to-Strong Monitoring of LLM Agents [Scale AI] https://arxiv.org/abs/2508.19461 --- [CL] Diffusion Language Models Know the Answer Before Decoding [The Hong Kong Polytechnic University & Dartmouth College & Max Planck Institute for Intelligent Systems] https://arxiv.org/abs/2508.19982 --- [LG] Symphony: A Decentralized Multi-Agent Framework for Scalable Collective Intelligence [Emory University & Columbia University] https://arxiv.org/abs/2508.20019

真正的效率，不是一开始就追求完美，而是在正确的轨道上，持续前进。