00:00:35 AI太耗电？可能我们一开始就想错了 00:04:23 AI的“外挂”：我们到底需要什么样的智能？ 00:09:27 如何让AI“带徒弟”：过程，永远比结果更重要 00:13:48 创新的“乐高”：如何像拼积木一样“制造”好点子？ 00:18:26 AI界的“瑞士军刀”：一个模型搞定所有视频“抠图” 本期介绍的几篇论文： [LG] Learning with springs and sticks [University of Toronto] https://arxiv.org/abs/2508.19015 --- [LG] Understanding Tool-Integrated Reasoning [Tencent] https://arxiv.org/abs/2508.19201 --- [LG] StepWiser: Stepwise Generative Judges for Wiser Reasoning [FAIR at Meta] https://arxiv.org/abs/2508.19229 --- [LG] The Ramon Llull's Thinking Machine for Automated Ideation [CMU & OSU & Stanford University] https://arxiv.org/abs/2508.19200 --- [CV] Autoregressive Universal Video Segmentation Model [NVIDIA & CMU] https://arxiv.org/abs/2508.19242

当你停止向世界索取信息时，你的内心世界，才会真正开始对你说话。

00:00:25 高手过招：当“规划大师”遇上“绘画天才” 00:04:28 AI作画：是天才创造，还是高级复印？ 00:08:49 AI界的“大力出奇迹”：模型越大，就越会“搜索”吗？ 00:13:02 你的大脑里，住着一个“项目经理”吗？ 00:17:55 AI养娃指南：机器如何学会“看见”世界？ 本期介绍的几篇论文： [CV] Incorporating Pre-trained Diffusion Models in Solving the Schrödinger Bridge Problem [Tsinghua University & Southeast University & University of Science and Technology of China] https://arxiv.org/abs/2508.18095 --- [LG] On the Edge of Memorization in Diffusion Models [TTIC & UC Berkeley] https://arxiv.org/abs/2508.17689 --- [LG] Retrieval Capabilities of Large Language Models Scale with Pretraining FLOPs [Databricks Mosaic Research] https://arxiv.org/abs/2508.17400 --- [LG] Language Models Coupled with Metacognition Can Outperform Reasoning Models [IBM Research] https://arxiv.org/abs/2508.17959 --- [LG] Disentangling the Factors of Convergence between Brains and Computer Vision Models [Meta AI] https://arxiv.org/abs/2508.18226 --- [CV] Explain Before You Answer: A Survey on Compositional Visual Reasoning [Anonymous / Survey paper] https://arxiv.org/abs/2508.17298

真正决定我们能走多远的，从来不是我们掌握了多前沿的技术，而是我们内心深处，那股驱动自己不断向上的力量。

00:00:29 AI提速秘籍：不开刀，如何给模型换“心脏”？ 00:04:54 AI的“天花板”：它能无限变聪明吗？ 00:08:39 AI的“错题本”：不靠“补课”，如何变聪明？ 00:12:32 AI健忘症？我们找到了便宜的“后悔药” 00:17:28 机器人学会“抬杠”，世界会更好吗？ 本期介绍的几篇论文： [CL] Jet-Nemotron: Efficient Language Model with Post Neural Architecture Search [NVIDIA] https://arxiv.org/abs/2508.15884 --- [CL] A Probabilistic Inference Scaling Theory for LLM Self-Correction [Peking University & Alibaba Group] https://arxiv.org/abs/2508.16456 --- [LG] AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs [UCL & Huawei Noah’s Ark Lab & Jilin University] https://arxiv.org/abs/2508.16153 --- [LG] RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMs [PolyTechnique Montreal & Mila] https://arxiv.org/abs/2508.16546 --- [LG] Do What? Teaching Vision-Language-Action Models to Reject the Impossible [UC Berkeley] https://arxiv.org/abs/2508.16292

请不要低估任何一个微小的决定，更不要被短期的安逸和冲动绑架了你的未来。

00:00:25 如果AI自己就能当科学家，我们该怎么办？ 00:04:46 AI天气预报，猜得到开头，猜不到结局？ 00:08:41 AI变聪明的秘密：死磕难题，还是广见世面？ 00:12:41 像画家一样思考：AI学会了打草稿 00:17:03 AI刷题也“内卷”？换个姿势，自己出题自己做 本期介绍的五篇论文： [LG] Virtuous Machines: Towards Artificial General Science [Explore Science & The University of Queensland] https://arxiv.org/abs/2508.13421 --- [LG] Numerical models outperform AI weather forecasts of record-breaking extremes [University of Geneva & ETH Zurich] https://arxiv.org/abs/2508.15724 --- [LG] Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration [The Hong Kong University of Science and Technology (Guangzhou) & ETH Zurich] https://arxiv.org/abs/2508.13755 --- [CV] Next Visual Granularity Generation [Nanyang Technological University & SenseTime Research] https://arxiv.org/abs/2508.12811 --- [CL] Beyond Pass@1: Self-Play with Variational Problem Synthesis Sustains RLVR [Microsoft] https://arxiv.org/abs/2508.14029

或许，我们真正需要的，不是再一次“刷新”，而是按下“暂停”。不是占有更多信息，而是有勇气选择“不知情”。

00:00:32 AI的“求生欲”：当代码开始为“活下去”而战 00:05:05 预测未来：一条你没想过的新路径 00:10:26 一张图纸值多少钱？AI现在几秒就算清 00:15:17 挑西瓜的智慧：如何在大模型时代做对选择？ 00:21:40 如何给AI装上一个“品味引擎”？ 本期介绍的五篇论文： [LG] Do Large Language Model Agents Exhibit a Survival Instinct? An Empirical Study in a Sugarscape-Style Simulation [The University of Tokyo] https://arxiv.org/abs/2508.12920 --- [LG] Universal Learning of Nonlinear Dynamics [New York University & Princeton University] https://arxiv.org/abs/2508.11990 --- [LG] Machine Learning-Based Manufacturing Cost Prediction from 2D Engineering Drawings via Geometric Features [Boğaziçi University & ArtificaX Technologies & Teknorot Automative Products Inc] https://arxiv.org/abs/2508.12440 --- [CL] Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation [Allen Institute for Artificial Intelligence] https://arxiv.org/abs/2508.13144 --- [LG] Reinforcement Learning with Rubric Anchors [Inclusion AI] https://arxiv.org/abs/2508.12790

如何看待你，是别人的课题；而如何活出自己的人生，才是你唯一的课题。

00:00:27 AI的“一忘皆空”：一种全新的遗忘术 00:04:23 高手思维：如何像AI一样“想得又快又好”？ 00:08:33 高手“停顿”的秘密：如何把机器人的犹豫变成智慧 00:12:15 给机器人一个“梦境模拟器”，世界会怎样？ 00:16:10 从“听话”到“懂你”：人工智能的下一次进化 本期介绍的几篇论文： [LG] Reliable Unlearning Harmful Information in LLMs with Metamorphosis Representation Projection [Peking University & Tsinghua University] https://arxiv.org/abs/2508.1 --- [LG] Deep Think with Confidence [Meta AI & UCSD] https://arxiv.org/abs/2508.15260 --- [RO] Exploiting Policy Idling for Dexterous Manipulation [Google DeepMind] https://arxiv.org/abs/2508.15669 --- [RO] Neural Robot Dynamics [NVIDIA] https://arxiv.org/abs/2508.15755 --- [LG] Open-Universe Assistance Games [MIT CSAIL] https://arxiv.org/abs/2508.15119

资格，从来不是起点处的通行证，而是旅途中被汗水和勇气浸透后，自然风干出的印记。