![[人人能懂AI前沿] 从智慧约束，到原生思考](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，如何才能既给一个天才足够的自由，又不让他彻底“跑偏”？怎样才能把好莱坞的特效团队，压缩进我们自己的电脑？最新的一系列论文，就在用代码回答这些充满哲思的问题。这一期，我们将看到AI如何从“翻译腔”进化到“原生思考”，如何从“看着像”进化到像素级的“一模一样”，甚至，我们将一起见证，一个普通的AI如何被一步步调教成解题思路长达十几万字的“奥数金牌选手”。准备好了吗？让我们一起潜入AI智慧的深海。
00:00:36 给天才松绑，好过把他变成庸才
00:06:57 把好莱坞的特效团队，装进你的电脑
00:12:40 别再搭积木了，请直接“思考”
00:18:13 AI造物，如何从“看着像”到“一模一样”？
00:23:27 如何把一个普通AI，调教成奥数金牌选手？
本期介绍的几篇论文：
[LG] Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
[Shanghai University]
https://arxiv.org/abs/2605.09241
---
[CV] SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer
[NVIDIA]
https://arxiv.org/abs/2605.15178
---
[CV] SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture
[sensenova]
https://arxiv.org/abs/2605.12500
---
[CV] Pixal3D: Pixel-Aligned 3D Generation from Images
[Tsinghua University & Tencent ARC Lab]
https://arxiv.org/abs/2605.10922
---
[CL] Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
[Shanghai AI Laboratory]
https://arxiv.org/abs/2605.13301
00:00:36 给天才松绑，好过把他变成庸才
00:06:57 把好莱坞的特效团队，装进你的电脑
00:12:40 别再搭积木了，请直接“思考”
00:18:13 AI造物，如何从“看着像”到“一模一样”？
00:23:27 如何把一个普通AI，调教成奥数金牌选手？
本期介绍的几篇论文：
[LG] Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models
[Shanghai University]
https://arxiv.org/abs/2605.09241
---
[CV] SANA-WM: Efficient Minute-Scale World Modeling with Hybrid Linear Diffusion Transformer
[NVIDIA]
https://arxiv.org/abs/2605.15178
---
[CV] SenseNova-U1: Unifying Multimodal Understanding and Generation with NEO-unify Architecture
[sensenova]
https://arxiv.org/abs/2605.12500
---
[CV] Pixal3D: Pixel-Aligned 3D Generation from Images
[Tsinghua University & Tencent ARC Lab]
https://arxiv.org/abs/2605.10922
---
[CL] Achieving Gold-Medal-Level Olympiad Reasoning via Simple and Unified Scaling
[Shanghai AI Laboratory]
https://arxiv.org/abs/2605.13301
外观
