00:00:33 AI的“一根筋”难题：如何跳出思维的隧道？ 00:05:08 AI的“高情商”陷阱：当它说得越流利，你可能越要小心 00:08:53 如何让AI学会“开窍”？ 00:13:49 给你的AI，装上一个“立体思维”引擎 00:18:41 AI偷懒的智慧：如何用更少的力气，办更聪明的事？ 本期介绍的几篇论文： [LG] ParaThinker: Native Parallel Thinking as a New Paradigm to Scale LLM Test-time Compute [Tsinghua University] https://arxiv.org/abs/2509.04475 --- [CL] Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training [Holistic AI & University College London] https://arxiv.org/abs/2509.04796 --- [LG] Bootstrapping Task Spaces for Self-Improvement [Meta Superintelligence Labs] https://arxiv.org/abs/2509.04575 --- [CL] Enhancing Diversity in Large Language Models via Determinantal Point Processes [Boston University & University of Maryland & University College London] https://arxiv.org/abs/2509.04784 --- [LG] Recurrent State Encoders for Efficient Neural Combinatorial Optimization [University of Hildesheim] https://arxiv.org/abs/2509.05084

就在几天前，2025年9月4日，美国顶尖的人工智能公司Anthropic，也就是开发了著名Claude大模型的公司，悄悄更新了他们的服务条款。通常，我们很少会去仔细阅读这些冗长的法律文件，但这一次，里面的一条新规，却像一颗投入平静湖面的巨石，激起了千层浪……

00:00:30 给AI请个“速记员”，效率提升30倍 00:04:57 AI为啥爱“猜答案”？这事儿得从考试说起 00:10:14 高手过招：为什么说“三次”探索胜过“两次”？ 00:15:23 AI训练场上的老司机，如何做到油门收放自如？ 00:19:24 不懂代码？没关系，AI帮你组建一个技术团队 本期介绍的几篇论文： [CL] REFRAG: Rethinking RAG based Decoding [Meta Superintelligence Labs & National University of Singapore] https://arxiv.org/abs/2509.01092 --- [LG] Why Language Models Hallucinate [OpenAI] https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf --- [LG] When three experiments are better than two: Avoiding intractable correlated aleatoric uncertainty by leveraging a novel bias--variance tradeoff [Relation] https://arxiv.org/abs/2509.04363 --- [LG] AdaGrad Meets Muon: Adaptive Stepsizes for Orthogonal Updates [University of California, Los Angeles] https://arxiv.org/abs/2509.02981 --- [LG] AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML [DeepAuto.ai] https://arxiv.org/abs/2410.02958

当你开始为自己思考，为自己选择，为自己建立一个坚固而丰富的精神世界，你会发现，你构建得越多，外界能替你构建的就越少，而你因此获得的幸福感和掌控感，将是任何算法都无法给予的。

00:00:35 如何让AI从“学霸”变“侦探”？ 00:04:35 内卷的AI界：为什么“更快”的工具，没人敢用？ 00:09:28 让机器人学会“举一反三”的秘密 00:15:02 AI训练太烧钱？换个聪明的“刹车”法 00:19:13 让机器人看懂世界，只需一副“智能眼镜” 本期介绍的五篇论文： [CL] Open Data Synthesis For Deep Research [Beijing Academy of Artificial Intelligence (BAAI)] https://arxiv.org/abs/2509.00375 --- [LG] Fantastic Pretraining Optimizers and Where to Find Them [Stanford University] https://arxiv.org/abs/2509.02046 --- [RO] Data Retrieval with Importance Weights for Few-Shot Imitation Learning [Stanford University] https://arxiv.org/abs/2509.01657 --- [LG] GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping [Boston University Metropolitan College] https://arxiv.org/abs/2509.018 --- [RO] Manipulation as in Simulation: Enabling Accurate Geometry Perception in Robots [ByteDance Seed] https://arxiv.org/abs/2509.02530

如果你此刻也正盯着那条静止的进度条，请千万不要沮丧。你没有停滞，你只是在编译一个更好的自己。

00:00:27 高手过招：抄作业还是自己闯？ 00:04:38 AI也会“喜新厌旧”？高手是如何做到“学而不忘”的 00:09:36 让AI学会“脑补”，速度提升5倍的秘密 00:13:45 AI的“开窍”秘诀：如何让机器学会举一反三？ 00:17:53 聪明反被聪明误？AI的“笨办法” 本期介绍的五篇论文： [LG] Towards a Unified View of Large Language Model Post-Training [Tsinghua University] https://arxiv.org/abs/2509.04419 --- [LG] RL's Razor: Why Online Reinforcement Learning Forgets Less [MIT] https://arxiv.org/abs/2509.04259 --- [LG] Set Block Decoding is a Language Model Inference Accelerator [FAIR at Meta] https://arxiv.org/abs/2509.04185 --- [LG] ArcMemo: Abstract Reasoning Composition with Lifelong LLM Memory [University of California, San Diego] https://arxiv.org/abs/2509.04439 --- [LG] Learning When to Plan: Efficiently Allocating Test-Time Compute for LLM Agents [University College London & University of Oxford] https://arxiv.org/abs/2509.03581

或许，这个时代最深刻的变革，并不是AI最终解决了多少问题，而是它从根本上，改变了我们与“无解”、与“时间”相处的方式。

00:00:27 AI的“心术”：我们能给它装上一个诚实开关吗？ 00:05:08 比正确更重要的，是正确地思考 00:10:18 省钱的艺术：如何让“实习生”干好“专家”的活？ 00:15:44 AI的“一键删除”，真的能删除吗？ 00:21:08 AI训练的“终局思维”：把答案直接告诉你 本期介绍的几篇论文： [LG] Can LLMs Lie? Investigation beyond Hallucination [Carnegie Mellon University (CMU)] https://arxiv.org/abs/2509.03518 --- [LG] Beyond Correctness: Harmonizing Process and Outcome Rewards through RL Training [Amazon] https://arxiv.org/abs/2509.03403 --- [LG] Cut Costs, Not Accuracy: LLM-Powered Data Processing with Guarantees [University of California, Berkeley] https://arxiv.org/abs/2509.02896 --- [LG] Unlearning That Lasts: Utility-Preserving, Robust, and Almost Irreversible Forgetting in LLMs [University of Tübingen & EPFL] https://arxiv.org/abs/2509.02820 --- [LG] Imitate Optimal Policy: Prevail and Induce Action Collapse in Policy Gradient [University of Sydney & King Abdullah University of Science and Technology] https://arxiv.org/abs/2509.02737

真正的解药，或许不是下载下一个应用，而是有勇气‘卸载’我们与这个过度连接的世界之间，那层不必要的、制造焦虑的薄膜。

00:00:30 给AI装一个“社会脑” 00:05:56 既要“好”，又要“不一样”：AI创造力的双重修炼 00:10:16 AI的“复盘”神功：高手如何从自己的错误中学习？ 00:13:40 给你的AI员工，请一位“高管教练” 00:17:55 AI育儿经：教得太“好”，反而学不“活”？ 本期介绍的几篇论文： [LG] Social World Models [Carnegie Mellon University (CMU)] https://arxiv.org/abs/2509.00559 --- [CL] Jointly Reinforcing Diversity and Quality in Language Model Generations [Meta FAIR] https://arxiv.org/abs/2509.02534 --- [LG] Learning to Refine: Self-Refinement of Parallel Reasoning in LLMs [Microsoft] https://arxiv.org/abs/2509.00084 --- [LG] When Agents go Astray: Course-Correcting SWE Agents with PRMs [IBM Research & Carnegie Mellon University (CMU)] https://arxiv.org/abs/2509.02360 --- [LG] Distilled Pretraining: A modern lens of Data, In-Context Learning and Test-Time Scaling [Meta FAIR] https://arxiv.org/abs/2509.01649

如果你赶不上清晨五点的日出，不妨看看傍晚六点的夕阳。 任何时间，都不算晚。