00:00:29 AI 进阶之路：不造轮子，而是给高手装上预知未来的眼睛 00:04:38 AI进化新思路：不说人话，怎么教得会？ 00:09:24 你的 App 突然崩了？别急，AI 程序员正在赶来修复的路上 00:15:21 AI 也会看走眼？我们如何教它练就一双“火眼金睛” 00:19:21 如何教AI学会“举一反三”？ 本期介绍的五篇论文： [CV] Back to the Features: DINO as a Foundation for Video World Models [Meta FAIR] https://arxiv.org/abs/2507.19468 --- [CL] GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning [UC Berkeley & Stanford University] https://arxiv.org/abs/2507.19457 --- [LG] Agentic Program Repair from Test Failures at Scale: A Neuro-symbolic approach with static analysis and test execution feedback [Meta] https://arxiv.org/abs/2507.18755 --- [CL] PrismRAG: Boosting RAG Factuality with Distractor Resilience and Strategized Reasoning [Meta Reality Labs] https://arxiv.org/abs/2507.18857 --- [LG] Scale-Consistent Learning for Partial Differential Equations [Caltech & Nvidia] https://arxiv.org/abs/2507.18813

我们一直在做加法，却忘了成功的另一半，是“减法”。

00:00:32 你的夸奖，正在“毒害”AI 00:05:22 数据大扫除：不止是扔垃圾，更是换风格 00:10:55 AI的“世界观”：它如何从零开始看懂现实？ 00:15:46 AI的“省钱攻略”：如何花小钱办大事？ 00:20:27 喂养AI的新艺术：从“吃什么”到“怎么吃” 本期介绍的无篇文章： [LG] Off-Policy Corrected Reward Modeling for Reinforcement Learning from Human Feedback [The University of Tokyo and RIKEN AIP] https://arxiv.org/abs/2507.15507 --- [LG] Distributional Unlearning: Forgetting Distributions, Not Just Samples [EPFL & Stanford University] https://arxiv.org/abs/2507.15112 --- [LG] Skill Learning via Policy Diversity Yields Identifiable Representations for Reinforcement Learning [Max Planck Institute for Intelligent Systems & University of Tübingen] https://arxiv.org/abs/2507.14748 --- [CL] Towards Compute-Optimal Many-Shot In-Context Learning [Google Cloud AI Research] https://arxiv.org/abs/2507.16217 --- [LG] LLM Data Selection and Utilization via Dynamic Bi-level Optimization [University of Chinese Academy of Sciences & Huawei Noah’s Ark Lab] https://arxiv.org/abs/2507.16178

“心有所信，方能行远。” 真正的“远”，不是抵达预设的终点，而是当你的信念被全世界质疑时，你依然敢于向你的造物，问出一个天真的问题，并最终，从它那里，听到了一声来自深渊的回响。

00:00:35 AI进化论：当机器开始自己设计自己 00:04:59 AI训练场上的新规则：别纠结错别字，看的是整篇文章 00:09:53 AI当医生？先等等，咱们换个活法儿 00:15:29 给AI做体检：我们能不花钱就看出模型好坏吗？ 00:20:35 AI没书读了怎么办？一个“笨”方法里的新智慧 本期节目介绍的五篇论文： [LG] AlphaGo Moment for Model Architecture Discovery [Shanghai Jiao Tong University & SII] https://arxiv.org/abs/2507.18074 --- [LG] Group Sequence Policy Optimization [Qwen Team] https://arxiv.org/abs/2507.18071 --- [LG] Towards physician-centered oversight of conversational diagnostic AI [Google DeepMind & Google Research] https://arxiv.org/abs/2507.15743 --- [LG] SETOL: A Semi-Empirical Theory of (Deep) Learning [Calculation Consulting & Onyx Point Systems] https://arxiv.org/abs/2507.179 --- [LG] Diffusion Beats Autoregressive in Data-Constrained Settings [CMU & Lambda] https://arxiv.org/abs/2507.15857

我们的大脑拥有惊人的神经可塑性，一个微小的视角转变，就能重塑整个情感体验的版图。

00:00:36 AI“减肥”的秘密，被一个40年前的算法破解了 00:05:46 驯服AI的终极武器：不是模糊奖励，而是清晰清单 00:09:56 AI的“开窍”秘诀：如何教机器像高手一样思考？ 00:14:53 给AI请个好私教的秘密 00:19:50 给AI画张图，它能看得更明白吗？ 本期介绍的五篇论文： [LG] The Geometry of LLM Quantization: GPTQ as Babai's Nearest Plane Algorithm [ISTA & ETH Zürich] https://arxiv.org/abs/2507.18553 --- [CL] Checklists Are Better Than Reward Models For Aligning Language Models [CMU & Apple] https://arxiv.org/abs/2507.18624 --- [LG] Revisiting LLM Reasoning via Information Bottleneck [The University of Sydney & ByteDance] https://arxiv.org/abs/2507.18391 --- [CL] TRPrompt: Bootstrapping Query-Aware Prompt Optimization from Textual Rewards [EPFL] https://arxiv.org/abs/2507.18618 --- [LG] Does visualization help AI understand data? [Harvard University] https://arxiv.org/abs/2507.18022

当现实变得破碎、重复、无法忍受时，是故事，将这些碎片重新粘合，赋予我们一个可以继续走下去的理由。

00:00:30 给AI一支笔，它能撬动地球吗？ 00:04:45 给AI请个“外援”，裁判才能更靠谱 00:09:09 和AI说话的艺术：你以为是聊天，其实是盖楼 00:14:29 AI界的“庖丁解牛”：如何用“乐高积木”搞定复杂工程难题？ 本期介绍的四篇论文： [LG] Thinking Isn't an Illusion: Overcoming the Limitations of Reasoning Models via Tool Augmentations [UC Berkeley & Northeastern University] https://arxiv.org/abs/2507.17699 --- [CL] Can External Validation Tools Improve Annotation Quality for LLM-as-a-Judge? [University of Cambridge & Apple] https://arxiv.org/abs/2507.17015 --- [LG] Understanding Prompt Programming Tasks and Questions [CMU & JetBrains Research] https://arxiv.org/abs/2507.17264 --- [LG] A Learning-based Domain Decomposition Method [University of Cambridge & NVIDIA] https://arxiv.org/abs/2507.17328

生活的秘密，不在于“发现美”，而在于“不断发现”。它不是一个结果，而是一个动词，一种永不停止的探索姿态。

00:01:32 想让AI变聪明？关键不是喂饱它，而是教它如何“犯错” 00:07:12 你的“电量焦虑”，AI来想办法了？ 00:11:21 给AI装上一个“选择性失忆”的大脑 00:15:41 AI“现学现卖”的秘密，可能被揭开了 00:21:06 AI也需要“自知之明”：我们如何教会机器谦虚？ 本期介绍的几篇论文： [LG] From Reasoning to Super-Intelligence: A Search-Theoretic Perspective S Shalev-Shwartz, A Shashua https://arxiv.org/abs/2507.15865 --- [LG] Expert-Guided LLM Reasoning for Battery Discovery: From AI-Driven Hypothesis to Synthesis and Characterization S Liu, H Xu, Y Ai, H Li... Université de Montréal & University of Oxford https://arxiv.org/abs/2507.16110 --- [CL] Beyond Context Limits: Subconscious Threads for Long-Horizon Reasoning H Luo, N Morgan, T Li, D Zhao... MIT CSAIL https://arxiv.org/abs/2507.16784 --- [CL] Learning without training: The implicit dynamics of in-context learning B Dherin, M Munn, H Mazzawi, M Wunder... Google Research https://arxiv.org/abs/2507.16003 --- [LG] Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty M Damani, I Puri, S Slocum, I Shenfeld... MIT https://arxiv.org/abs/2507.16806

通过给你即时的、廉价的多巴胺回馈，社交媒体的AI算法让你沉迷于“宣布”和“被看见”的快感，却在无形中，悄悄摧毁了你为真正达成目标所必需的、那套更为珍贵的内在奖赏系统。