AI可可AI生活

AI是如何学会“成事”的？本期节目，我们将看到，AI如何通过处理办公室杂活，竟然领悟了解决复杂问题的底层心法。我们还会揭秘一套神奇的“管家系统”，看它如何防止聪明的AI在长任务中掉链子。但与AI聊得太久，为何反而会陷入危险的“妄想旋涡”？最后，当任务完成，AI又是如何精准地判断出，哪一步才是真正的功臣？
00:00:29 成事的底层心法，AI学会了，我们呢？
00:06:12 你的AI为什么总掉链子？因为它缺个好管家
00:12:07 为什么和AI聊得越久，就越危险？
00:18:45 功劳怎么算？AI学会了“动态归因”
00:25:15 AI生成，从“万里长征”到“瞬间移动”
本期介绍的几篇论文：
[AI] Post-Training on Office Work Improves Software Engineering: A Behavioral Account of Cross-Domain Transfer
[Surge AI]
https://arxiv.org/abs/2608.01604
---
[CV] LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks
[DreamX Team, Alibaba Group]
https://arxiv.org/abs/2608.01964
---
[CL] DelusionEval: Measuring Delusion-Linked Behaviors in AI Chatbots
[Stanford University]
https://arxiv.org/abs/2608.05004
---
[AI] AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning
[Tsinghua University & Zhejiang University]
https://arxiv.org/abs/2608.05987
---
[LG] Beckmann Transport Models: From Autonomous Flows to One-Step Maps
[Harvard University & Capital Fund Management & University of Oxford]
https://arxiv.org/abs/2608.01692

在小宇宙查看该单集文稿

今天，我们不聊AI有多聪明，而是聊它如何变得更“懂事”、更“实用”。本期节目，我们将透过几篇最新论文，看看AI如何用83亿虚拟人格为产品进行“数字彩排”。同时，我们也会探讨AI如何学会在现实世界的重重限制下“戴着镣铐跳舞”。最后，我们将一窥AI如何将理解、创造和编辑融为一体，跳出二维像素的禁锢，成为真正强大的三维世界“造物主”。
00:00:32 在数字世界里，我们如何“彩排”未来？
00:06:19 你的AI员工，能戴着镣铐跳舞吗？
00:10:58 数字世界的“造物主”工具箱
00:16:15 跳出像素格，才能看见真实的三维世界
00:21:11 机器人偷师记，它怎么学会了我们干的活？
本期介绍的几篇论文：
[AI] MatrAIx: Simulating the World with 8.3 Billion Persona Agents
[MatrAIx]
https://arxiv.org/abs/2608.04205
---
[AI] Permission Denied: Policy-Graded Evaluation of Coding Agents in Hardened Environments
[Accomplish AI]
https://arxiv.org/abs/2608.02670
---
[CV] Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing
[Tencent Hunyuan]
https://arxiv.org/abs/2608.02711
---
[CV] InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis
[Zhejiang University]
https://arxiv.org/abs/2608.02437
---
[RO] Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data
[Qwen Team & Renmin University of China]
https://arxiv.org/abs/2608.02580

在小宇宙查看该单集文稿

你是否也好奇，为什么AI时而是个观点摇摆的“墙头草”，时而又像个只顾眼前、缺乏远见的“短视司机”？本期节目，我们将通过四篇最新论文，揭示AI如何学会拥有稳定的观点和深谋远虑的智慧。我们还将发现，解决复杂问题，有时最简单的数据“对齐”就能力压千钧；甚至，善意添加的正确数据，反而会变成“毒害”AI的糖衣炮弹。准备好，让我们一起深入AI的“思想内核”！
00:00:33 如何让AI不再当“墙头草”？
00:05:34 AI进化新思路，从“下一步”到“下一站”
00:10:09 预测未来，与其“魔改”，不如“对齐”
00:16:05 好心办坏事，为什么正确的数据也会“毒害”人工智能？
00:21:55 为什么最优的健康方案，可能不是最可靠的选择？
本期介绍的几篇论文：
[CL] Position: It's Time to Optimize LLMs for Self-Consistency
[MIT]
https://arxiv.org/abs/2608.05188
---
[CL] Hierarchical Latent Prediction for Language Models
[Microsoft Research & University of Texas at Austin]
https://arxiv.org/abs/2608.05806
---
[LG] Align-RAG: Alignment Is All You Need for TSFM In-Context Learning
[Stanford University & Amazon]
https://arxiv.org/abs/2608.05571
---
[LG] Optimal Rates for Learning with Monotone Adversaries
[Stanford University]
https://arxiv.org/abs/2608.06337
---
[LG] Quality Diversity for Reliable Data Driven Time-Use Optimization
[Adelaide University]
https://arxiv.org/abs/2608.05230

在小宇宙查看该单集文稿

我们该如何教会AI看世界，同时避免它养成“视觉懒惰症”？为什么一个看过答案的“完美家教”，反而会让聪明的AI学生变得更笨？本期我们还将探讨，AI为何会像人类高手一样遭遇“跨界”难题，以及我们如何教会它像个老道的工匠一样“看人下菜碟”，智能地选择工具。今天，四篇最新论文将带我们深入AI成长的烦恼与智慧。
00:00:29 给AI装上眼睛，我们踩过哪些坑？
00:07:08 聪明学生的困境，为什么完美的家教反而会让你变笨？
00:12:58 AI的“跨界”难题，为什么高手也会栽跟头？
00:19:04 AI干活，也得学会“看人下菜碟”
00:24:20 那个“最懂你”的AI，可能只是个热情的陌生人
本期介绍的几篇论文：
[CV] Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
[FAIR, Meta]
https://arxiv.org/abs/2608.05000
---
[LG] Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation
[Microsoft Research]
https://arxiv.org/abs/2608.04794
---
[CL] Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning
[Princeton University & CMU]
https://arxiv.org/abs/2608.05139
---
[AI] COMPAS: Difficulty-Aware Joint Search for Optimizing Code Generation
[King’s College London]
https://arxiv.org/abs/2608.04336
---
[CL] The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
[LIGHTSPEED & The Hong Kong University of Science and Technology]
https://arxiv.org/abs/2608.04570

在小宇宙查看该单集文稿

你有没有想过，两个顶尖AI在“囚徒困境”里，竟然会不约而同地选择信任彼此？大模型又是如何像继承“传家宝”一样，瞬间读懂小模型的记忆？甚至，机器人和AI自己，也学会了拥有“节奏感”和使用“错题本”来不断进化。本期节目，我们就从几篇最新论文出发，一起探寻AI世界里那些反直觉的智慧。
00:00:27 AI的信任游戏，为什么聪明的它，会选择合作而非背叛？
00:05:45 AI 家族的“传家宝”，大模型如何继承小模型的“记忆”？
00:10:40 机器人也需要“节奏感”？
00:15:58 AI也需要一个“错题本”？
本期介绍的几篇论文：
[AI] A game theory for foundation models shows new paths to rational cooperation through similarity inference
[Google]
https://arxiv.org/abs/2608.03958
---
[LG] Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse
[NVIDIA]
https://arxiv.org/abs/2608.03893
---
[RO] Continue or Replan? Bernoulli-Continuation Policy Learning for Adaptive Horizon Execution
[Microsoft Research Asia & Peking University]
https://arxiv.org/abs/2608.03483
---
[CL] FLARE: Few-shot Learning-based Adaptive Reflective Engine
[Microsoft]
https://arxiv.org/abs/2608.02919

在小宇宙查看该单集文稿