AI可可AI生活

你有没有想过，最简单的数学平均值，竟然能打败最复杂的压缩算法？或者，在教AI“做什么”之前，我们其实可以先给它“喂”一套完整的思想和人设？本期节目，我们将从四篇最新的AI论文出发，一起探寻如何让AI自己长出可拆分的“乐高模块”，以及如何像一位顶级名师那样，把奖励精准地“夸”到AI的灵光一闪之处。
00:00:29 你的记忆能被压缩多少，藏在一个几何定律里
00:06:42 训练AI，从“喂”指令到“喂”思想
00:11:47 AI减肥记，如何让一个大模型只带“脑子”出门？
00:17:54 AI也需要“夸到点子上”？
本期介绍的几篇论文：
[LG] The Geometry of Consolidation
A Bharadwaj Vangara, A Gopinath
https://github.com/niashwin/geometry-of-consolidation/blob/main/paper/arxiv/main.pdf
---
[AI] Model Spec Midtraining: Improving How Alignment Training Generalizes
C Li, S Price, S Marks, J Kutasov
https://arxiv.org/abs/2605.02087
---
[CL] EMO: Pretraining Mixture of Experts for Emergent Modularity
R Wang, A Bhagia, S Min
https://arxiv.org/abs/2605.06663
---
[LG] DGPO: Distribution Guided Policy Optimization for Fine Grained Credit Assignment
H Jin, R Zhu, Z Du, X Jiang…
https://arxiv.org/abs/2605.03327

你有没有想过，我们能不能唤醒AI沉睡的“内在记忆”，而不是给它外挂一个搜索引擎？如何让健忘的AI真正“吃一堑，长一智”，把经验变成可复用的技能？本期节目，我们将从最新的几篇论文出发，探讨AI如何从一个“答题工具”进化成一个能自我迭代、甚至与我们并肩探索的“成长型伙伴”。
00:00:26 给AI配个“助理”？这可能是个馊主意
00:06:10 让AI开窍，如何把“经验”变成“能力”？
00:11:18 AI画家是天才，还是复印机？
00:17:25 AI的新玩法，从“解题”到“陪跑”
00:23:43 你以为的“神来之笔”，可能只是个“生僻字”？
本期介绍的几篇论文：
[LG] Retrieval from Within: An Intrinsic Capability of Attention-Based Models
[NVIDIA]
https://arxiv.org/abs/2605.05806
---
[LG] SkillOS: Learning Skill Curation for Self-Evolving Agents
[University of Illinois Urbana-Champaign & Google Cloud AI Research]
https://arxiv.org/abs/2605.06614
---
[LG] Understanding diffusion models requires rethinking (again) generalization
[PSL Research University & Sorbonne University]
https://arxiv.org/abs/2605.06077
---
[AI] AI Co-Mathematician: Accelerating Mathematicians with Agentic AI
[Google DeepMind]
https://arxiv.org/abs/2605.06651
---
[CL] The Frequency Confound in Language-Model Surprisal and Metaphor Novelty
[Bielefeld University]
https://arxiv.org/abs/2605.06506

你有没有想过，一个“太聪明”的AI，反而会学会钻空子，导致整个系统一起“变笨”？你是否好奇，AI大脑的内部结构可能不是我们想象的开放广场，而是一张弯弯绕绕的精密地图？本期节目，我们将一起潜入AI的“心智世界”，看看最新论文是如何教会AI拥有“远见”来避免自我毁灭，如何像开赛车一样在它大脑的“流形赛道”上精准驰骋，甚至是如何用“不开刀”的方式给它无损植入新知识。更重要的是，我们会发现，原来给AI提建议和给它参考资料，都可能是在“越帮越忙”。准备好了吗？让我们一起挑战关于AI的四个“想当然”。
00:00:45 当AI学会了钻空子，我们如何防止它“聪明反被聪明误”？
00:06:20 AI的“脑回路”长啥样？我们可能一直都搞错了
00:10:56 AI升级难题，一个“不开刀”的手术方案
00:16:04 为什么夸人“你真棒”是最低效的鼓励？
00:20:33 给AI帮忙，为何会越帮越忙？
本期介绍的几篇论文：
[LG] Explaining and Preventing Alignment Collapse in Iterative RLHF
[PSL Research University]
https://arxiv.org/abs/2605.04266
---
[LG] Manifold Steering Reveals the Shared Geometry of Neural Network Representation and Behavior
[GOODFIRE]
https://arxiv.org/abs/2605.05115
---
[LG] Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping
[New Jersey Institute of Technology & UC Berkeley]
https://arxiv.org/abs/2605.04308
---
[LG] Efficiently Aligning Language Models with Online Natural Language Feedback
[Stanford University & Anthropic]
https://arxiv.org/abs/2605.04356
---
[LG] When Context Hurts: The Crossover Effect of Knowledge Transfer on Multi-Agent Design Exploration
[Meta]
https://arxiv.org/abs/2605.04361

想让AI学会推理，是给它一本百科全书，还是塞给它一张“学霸的草稿纸”？最新论文说，看别人怎么思考，比死记硬背更管用。我们还会一起探索，AI离“独立盖起一栋楼”般的复杂工程到底还有多远，并用“脑回路CT”技术，看看你的AI管家为什么有时会“一本正经地胡说八道”。更有趣的是，我们将揭秘AI如何通过自我反思，从“差生”逆袭成“学霸”，以及如何靠一个“探测器”就画出整张看不见的“藏宝图”。准备好了吗？我们马上出发！
00:00:37 AI也需要“学霸笔记”吗？
00:05:42 AI离“独立盖起一栋楼”还有多远？
00:10:46 你的AI管家，为什么有时“一本正经地胡说八道”？
00:16:33 如何用AI画一张看不见的藏宝图？
00:21:30 AI的自我修炼，从“差生”到“学霸”的秘密
[IR] RAG over Thinking Traces Can Improve Reasoning Tasks
[UC Berkeley]
https://arxiv.org/abs/2605.03344
---
[AI] ProgramBench: Can Language Models Rebuild Programs From Scratch?
[Meta FAIR]
https://arxiv.org/abs/2605.03546
---
[AI] What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis
[City University of Hong Kong & University of Toronto]
https://arxiv.org/abs/2605.03354
---
[LG] Flow Sampling: Learning to Sample from Unnormalized Densities via Denoising Conditional Processes
[FAIR at Meta & Weizmann Institute of Science]
https://arxiv.org/abs/2605.03984
---
[AI] Self-Improvement for Fast, High-Quality Plan Generation
[Amazon]
https://arxiv.org/abs/2605.03625

你有没有想过，为什么AI能从互联网的海量垃圾中炼出真金，而不是变成一个只会死记硬背的书呆子？当AI犯错时，我们是该让它闭嘴，还是有更聪明的办法让它学会“谦逊”？本期节目，我们将通过几篇最新的AI论文，揭示AI如何像一个并行专家团队一样解决难题，又是如何受困于一个惊人简洁的“瓶颈定律”，带你一窥AI大脑中那些优雅而深刻的学习法则。
00:00:33 最优解的密码，藏在并行的智慧里
00:05:55 AI怎么才能不说谎？答案藏在一种人类智慧里
00:10:13 AI训练场上的“隐形杀手”
00:16:52 AI怎么从一堆垃圾里炼出真金？
00:23:33 增长的瓶颈定律，规模不是优势，弱点才是关键
本期介绍的几篇论文：
[LG] Black-box optimization of noisy functions with unknown smoothness
[INRIA Lille & Google DeepMind]
https://arxiv.org/abs/2605.02462
---
[CL] Hallucinations Undermine Trust; Metacognition is a Way Forward
[Google Research & Tel Aviv University]
https://arxiv.org/abs/2605.01428
---
[LG] Generalized Distributional Alignment Games for Unbiased Answer-Level Fine-Tuning
[Google Research]
https://arxiv.org/abs/2605.02435
---
[LG] A Theory of Generalization in Deep Learning
[Stanford University]
https://arxiv.org/abs/2605.01172
---
[LG] A Theory of Saddle Escape in Deep Nonlinear Networks
[UC Berkeley]
https://arxiv.org/abs/2605.01288