23分15秒，是一个普通人，在注意力被一次微小的干扰打断后，重新回到之前深度专注状态所需要付出的平均时间。

00:01:33 “差生”配对，如何“炼”出优等生？ 00:06:09 AI的“刻意练习”：怎样探索才最高效？ 00:10:19 让AI学会顶尖“手艺活”，这事儿靠谱吗？ 00:14:35 黑箱里的光：我们好像找到了AI学习的秘密开关 00:19:47 AI 程序员的“心事”：它真的懂你的需求吗？ 今天介绍的五论文： [LG] The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains [University of Washington] https://arxiv.org/abs/2507.06187 --- [LG] Epistemically-guided forward-backward exploration [ETH Zurich & University of Tübingen] https://arxiv.org/abs/2507.05477 --- [LG] AutoTriton: Automatic Triton Programming with Reinforcement Learning in LLMs [Tsinghua University] https://arxiv.org/abs/2507.05687 --- [LG] FACT: the Features At Convergence Theorem for neural networks [MIT & UCSD & UC Berkeley] https://arxiv.org/abs/2507.05644 --- [CL] Coding Triangle: How Does Large Language Model Understand Code? [Shanghai AI Laboratory] https://arxiv.org/abs/2507.06138

我们真正要警惕的，不是犯错，而是那种‘万一错了怎么办’的恐惧。

00:01:37 你的 AI，是“记性好”还是“真会学”？ 00:06:19 管好AI的“注意力”：一个“分而治之”的智慧 00:10:22 造个“世界”给AI：为什么光看视频还不够？ 00:16:51 AI团队管理的终极难题：既要、又要、还要，怎么办？ 00:21:45 想把AI训练好？别等“秋后算账” 今天介绍的五篇论文： [LG] Memory Mosaics at scale [New York University & FAIR, Meta Inc] https://arxiv.org/abs/2507.03285 --- [CL] RAT: Bridging RNN Efficiency and Attention Accuracy in Language Modeling [CLAIRE, EPFL & Google DeepMind] https://arxiv.org/abs/2507.04416 --- [LG] Critiques of World Models [CMU] https://arxiv.org/abs/2507.05169 --- [LG] Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards [Stanford University] https://arxiv.org/abs/2507.03041 --- [LG] Discrete Diffusion Trajectory Alignment via Stepwise Decomposition [Stanford University & Caltech] https://arxiv.org/abs/2507.04832

“我此刻的行为，是在点燃一支烟火，还是在种下一棵树？是在透支未来的我，还是在投资未来的我？”

00:01:26 AI的“心里话”，为什么也可能是装出来的？ 00:06:38 AI的“记忆力”，如何才能赶上一本小说？ 00:12:26 AI陪练：笨功夫如何“点化”天才？ 00:16:27 AI搞装修：如何盖一座又高又省钱的“摩天楼”？ 00:20:34 小模型，大智慧：AI的“学徒”进化论 本期介绍的五篇论文： [LG] Chain-of-Thought Is Not Explainability [Oxford & WhiteBox] https://arxiv.org/abs/2025.02 --- [LG] Arctic Long Sequence Training: Scalable And Efficient Training For Multi-Million Token Sequences [Snowflake AI Research] https://arxiv.org/abs/2506.13996 --- [CL] ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models [NVIDIA] https://arxiv.org/abs/2505.24864 --- [LG] Don't be lazy: CompleteP enables compute-efficient deep transformers [Cerebras Systems & ETH Zurich] https://arxiv.org/abs/2505.01618 --- [CL] Distilling LLM Agent into Small Models with Retrieval and Code Tools [KAIST & KRAFTON] https://arxiv.org/abs/2505.17612

迷路，不是一种状态，它是一种必要的暂停。它强迫我们从“自动驾驶”切换到“手动模式”，用自己的意图，去行走，去感受。

00:01:29 AI的“人设”：谷歌“腹黑”，OpenAI“傻白甜”？ 00:06:09 给AI“减肥餐”：为什么数据越多，模型可能越笨？ 00:10:32 AI训练场上的新策略：先当“神算子”，再做“阅读家” 00:14:53 功劳簿到底该怎么写？ 今天介绍的四篇论文： [LG] Strategic Intelligence in Large Language Models: Evidence from evolutionary Game Theory K Payne, B Alloui-Cros [King’s College London & University of Oxford] https://arxiv.org/abs/2507.02618 --- [LG] Data Uniformity Improves Training Efficiency and More, with a Convergence Framework Beyond the NTK Regime Y Wang, S Gu [Johns Hopkins University & UC Berkeley] https://arxiv.org/abs/2506.24120 --- [CL] Should We Still Pretrain Encoders with Masked Language Modeling? H Gisserot-Boukhlef, N Boizard, M Faysse, D M. Alves... [Artefact Research Center & Diabolocom & Illuin Technology] https://arxiv.org/abs/2507.00994 --- [LG] Disentangled Feature Importance J Du, K Roeder, L Wasserman [CMU] https://arxiv.org/abs/2507.00260

时间本身就是最高形式的财富，而你，需要成为第一个收款人。

00:01:33 AI育儿经：聪明的大脑，是“喂”出来的还是“练”出来的？ 00:07:02 换个方向看世界：当AI学会“倒着想” 00:10:52 AI变强有公式？别太天真了 00:14:53 抓住那个“不对劲”：比记住更多更重要的，是懂得如何“修正记忆” 本期介绍的四篇论文： [CL] NaturalThoughts: Selecting and Distilling Reasoning Traces for General Reasoning Tasks [FAIR at Meta] https://arxiv.org/abs/2507.01921 --- [CL] LEDOM: An Open and Fundamental Reverse Language Model [Peking University & University of California, Santa & BarbaraNational University of Singapore] https://arxiv.org/abs/2507.01335 --- [CL] Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check [New York University] https://arxiv.org/abs/2507.00885 --- [LG] xLSTMAD: A Powerful xLSTM-based Method for Anomaly Detection [AGH University of Krakow] https://arxiv.org/abs/2506.22837

你应该容忍、甚至鼓励那些可能带来巨大回报的、高失败率的实验。因为在商业世界里，一次本垒打的得分，远超你的想象。

00:01:45 AI 炼丹术：发现模型训练的“万能公式”？ 00:07:01 AI军备竞赛，弹药快打光了怎么办？ 00:12:16 让 AI 学会“反思”，会发生什么？ 00:16:48 AI“高考”也偏科？换个考法，学霸变学渣 00:22:08 AI的“外挂”：当你的短板，成了我的跳板 00:27:18 让AI“看人下菜碟”：下一代人工智能的效率革命 本期介绍的六篇AI前沿论文： [LG] Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks [Google DeepMind & New York University] https://arxiv.org/abs/2507.02119 --- [LG] Fast and Simplex: 2-Simplicial Attention in Triton [Meta & University of Texas at Austin] https://arxiv.org/abs/2507.02754 --- [LG] Energy-Based Transformers are Scalable Learners and Thinkers [Amazon & UVA] https://arxiv.org/abs/2507.02092 --- [CL] Answer Matching Outperforms Multiple Choice for Language Model Evaluation [Max Planck Institute for Intelligent Systems & Tübingen AI Center] https://arxiv.org/abs/2507.02856 --- [RO] MultiGen: Using Multimodal Generation in Simulation to Learn Multimodal Policies in Real [UC Berkeley] https://arxiv.org/abs/2507.02864 --- [LG] Reasoning on a Budget: A Survey of Adaptive and Controllable Test-Time Compute in LLMs https://arxiv.org/abs/2507.02076