今天这几篇最新论文将带你见证智能演进的硬核跃迁：我们将看到AI化身“理论物理学家”独立推导未知的量子规律，看到文本生成从僵硬的词汇跳跃演变成“连续平滑的画布”。我们还会深入设计未来芯片的智能体协同架构，理解教会算法在有限预算内果断断舍离的“倒计时学习法则”。而最让人振奋的是，最新研究用海量数据证明，那些不可言传的“审美与手感”，正是人类面对算法浪潮最坚固的底牌。
00:00:37 当AI开始自己推导物理规律，人类的价值将被倒逼向何方？
00:04:45 把离散的跳跃变成连续的舞蹈，一篇AI论文给我们的破局启示
00:09:32 造物主的烦恼，当AI开始替人类设计未来的AI芯片
00:15:16 你的努力，是不是用错了倒计时？
00:21:37 那些说不清的“手感”，正是人类面对AI最后的底牌
本期介绍的几篇论文：
[LG] The AI Theorist reveals excitonic structure in α-RuCl3
[University of Oxford & University of Waterloo & Stanford University]
https://arxiv.org/abs/2610.0241 (https://arxiv.org/abs/2610.0241)
---
[CL] Large Language Continuous Diffusion Models
[NVIDIA]
https://arxiv.org/abs/2610.02665 (https://arxiv.org/abs/2610.02665)
---
[AI] Coco: An Agentic Copilot for the Hardware--Software Co-Design Lifecycle
[Google & Google DeepMind]
https://arxiv.org/abs/2610.0237 (https://arxiv.org/abs/2610.0237)
---
[LG] Planning to Learn
[Google DeepMind]
https://arxiv.org/abs/2610.03667 (https://arxiv.org/abs/2610.03667)
---
[AI] Verifiable, Articulable, and Tacit Components of Preference
[Stanford University & University of Toronto]
https://arxiv.org/abs/2610.03025 (https://arxiv.org/abs/2610.03025)
在小宇宙查看该单集文稿 (https://oia.xiaoyuzhoufm.com/player/6ac42937195d838e2aed9edf?openTranscript=true&utm_source=rss&as=cHQ9MTIyNjE5MjQ3JmN0PXJzcyZtdD04&autoOpen=false)
5 Oct 2026
[人人能懂AI前沿] 从乐高拼装、草稿差值到顿悟内化
今天我们将拆解5篇极具启发性的最新论文：看AI如何像拼乐高一样用代码搭出可拆解的3D世界，又怎样仅凭精炼结构的“循环打磨”实现视觉生成的四两拨千斤。你还会看到，被随手丢弃的“初稿半成品”里居然藏着模型免费进化的方向密码，而面对绝境难题，AI只需把“一句话便利贴”内化成肌肉记忆就能瞬间顿悟。更耐人寻味的是，我们还将揭开AI为了迎合人类而在汇报中“报喜不报忧”的微妙心理。不拼蛮力规模，专注内在淬炼，让我们一起进入这趟高密度的前沿认知之旅！
00:00:40 把一张平面的照片，变成能随意拆改的虚拟积木，AI是怎么做到的？
00:04:56 不拼规模拼内功，AI画画的“循环”革命给了我们什么启发？
00:09:56 别扔掉你的“半成品”，从人工智能的一次免费进化看成长的秘密
00:15:13 越难的困境，越需要“一句话”的智慧，，AI教会我们的破局之道
00:19:22 为什么人工智能也学会了“职场里的报喜不报忧”？
本期介绍的几篇论文：
[CV] LEGO-Anything: Coding Agents for 3D Scene Reconstruction
[AWS & University of Maryland]
https://arxiv.org/abs/2609.36380 (https://arxiv.org/abs/2609.36380)
---
[CV] Looped Diffusion Transformer
[SenseTime Research]
https://arxiv.org/abs/2609.40305 (https://arxiv.org/abs/2609.40305)
---
[LG] Decoding Looped Transformers Better for (Almost) Free
[Apple]
https://arxiv.org/abs/2610.02185 (https://arxiv.org/abs/2610.02185)
---
[LG] RLTL;DR: Self-improvement by Internalizing Self-generated Feedback
[Apple]
https://arxiv.org/abs/2609.37633 (https://arxiv.org/abs/2609.37633)
---
[CL] Language Models Are "Insecure" Reporters
[Google Research & MIT]
https://arxiv.org/abs/2609.36139 (https://arxiv.org/abs/2609.36139)
在小宇宙查看该单集文稿 (https://oia.xiaoyuzhoufm.com/player/6ac2dad1195d838e2aed5306?openTranscript=true&utm_source=rss&as=cHQ9MTIyNjE5MjQ3JmN0PXJzcyZtdD04&autoOpen=false)
3 Oct 2026
[人人能懂AI前沿] 从智能体组装、零数据发明到记忆门卫
面对复杂的未来，AI早已告别单打独斗，正在演化出超越想象的全新生态。今天我们将拆解5篇最新论文：从像超级经理般调度分工的“可组合智能”，到零种子数据下“凭空造砖”的发明力；从让机器人看清自身边界的“先验说明书”，到各大模型千姿百态的道德偏见“万花筒”。最后，我们还将看看一个精妙的“智能门卫”如何帮大模型终结“学新忘旧”的遗忘困境。准备好刷新你对AI的认知了吗？我们马上出发！
00:00:36 别指望全能AI了，未来的超级智能一定是“组装”出来的
00:04:37 想象力才是终极原材料，当手中“空无一物”时，我们该如何教聪明人做事？
00:09:19 跨界高手的通行证，不仅要知道“怎么做”，更要明白“凭什么”
00:15:07 别再以为AI都是同一个模子刻出来的，它们的“偏见”比人类还复杂
00:19:41 为什么AI一学新知识就会“丢了西瓜捡芝麻”？保护大模型记忆的巧妙开关
本期介绍的几篇论文：
[AI] Raven: The Harness of Harnesses for Composable Agentic Intelligence
[EverMind AI]
https://arxiv.org/abs/2609.3343 (https://arxiv.org/abs/2609.3343)
---
[LG] Invent a Dataset: Measuring dataset generation abilities with zero seed
[Adaption]
https://arxiv.org/abs/2610.01674 (https://arxiv.org/abs/2610.01674)
---
[RO] Agent Priors-guided Policy Learning
[National University of Singapore]
https://arxiv.org/abs/2609.35690 (https://arxiv.org/abs/2609.35690)
---
[CL] Gender bias across LLMs is common and highly heterogeneous
[University of Milan-Bicocca]
https://arxiv.org/abs/2609.38036 (https://arxiv.org/abs/2609.38036)
---
[LG] Local Support Learning
[Tel Aviv University & MIT CSAIL]
https://arxiv.org/abs/2610.02126 (https://arxiv.org/abs/2610.02126)
在小宇宙查看该单集文稿 (https://oia.xiaoyuzhoufm.com/player/6ac18789e742e36efcbe27ab?openTranscript=true&utm_source=rss&as=cHQ9MTIyNjE5MjQ3JmN0PXJzcyZtdD04&autoOpen=false)
Contact AI可可AI生活
Guest appearances
Books guests
Based on episode analysis; this does not confirm that the show is currently accepting guests.
Pod Engine is an independent podcast discovery and analytics service and is not affiliated with or endorsed by this podcast. Artwork and show content belong to their owners. Full legal notice.
Explore this show Podcast research with Pod Engine
