Anima 主张物理世界需要带结构的神经算子而非硬套语言模型缩放：FourCastNet 等在开源气象数据上逼近传统数值预报；傅里叶/球面谐波先验让长 rollout 稳定；聚变等离子体数字孪生可快百万倍；TorchLean 把网络写入 Lean 做形式化边界——目标是多物理场基础模型与反问题设计，并在联合国科学顾问角色推动循证 AI。
- 物理≠语言：无足够 token，且网格上下文可达万亿级，变压器硬上不可行。
- 神经算子：学习函数空间映射，支持任意分辨率。
- 结构先验：傅里叶/球面谐波等古典基仍关键。
嘉宾：Anima Anandkumar（Caltech Bren Professor；Accelerated Understanding）· 主持 Brandon Anderson · 日期：2026-08-26（PT）
播出信息
- 日期：2026-08-26（PT）
- 时长：约 83 分钟（RSS=5011s）
- 章节末段：1:18:13（约 94%）
节目结构（YouTube 章节）
- Cold open and intros — 00:00 🎧 · ▶️
- Anima's thesis: AI for science, not language models — 03:48 🎧 · ▶️
- TorchLean: proving what a neural network will do — 06:44 🎧 · ▶️
- What a neural operator actually is — 16:08 🎧 · ▶️
- Fourier neural operators, and why Fourier — 19:41 🎧 · ▶️
- Old math meets deep learning, and why scaling breaks — 26:53 🎧 · ▶️
- Weather, and the 2021 skeptics — 32:39 🎧 · ▶️
- Why the physical world is forgiving — 42:40 🎧 · ▶️
- Modeling across scales, from atoms to planets — 46:19 🎧 · ▶️
- The data, ECMWF, and calling Hurricane Lee early — 50:33 🎧 · ▶️
- Long rollouts, and why the sphere stays stable — 57:48 🎧 · ▶️
- Career full circle: theory, scale, and back to principled — 1:07:59 🎧 · ▶️
- A foundation model for physics, and inverse design — 1:13:02 🎧 · ▶️
- The UN, regulating AI for science, and a call to action — 1:18:13 🎧 · ▶️
分节详解
1) Cold open and intros — 00:00 🎧 · ▶️
💬 他们说了什么： 开场引入核心命题与嘉宾背景，定下本集问题意识。
💡 为什么重要： 开源与学术分工影响生态与长期信任。
⚡ 争议点： 自动字幕非人工精校；技术细节以论文/官方材料为准。
2) Anima's thesis: AI for science, not language models — 03:48 🎧 · ▶️
💬 围绕「Anima's thesis: AI for science, not language models」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：>> to me broadly like you know my thesis is >> to me broadly like you know my thesis is AI and science how w…）。
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
3) TorchLean: proving what a neural network will do — 06:44 🎧 · ▶️
💬 TorchLean 与可证明界、更有原则的深度学习。 依据字幕：framework, right? So what it really enables is that you can now write neural enables is that you can…
💡 结构先验与算子学习是物理建模相对语言范式的核心差异。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
4) What a neural operator actually is — 16:08 🎧 · ▶️
💬 神经算子可在任意分辨率评估，适配物理场。 依据字幕：>> Yeah. So you know neural operators are in that sense similar to you know it's in that sense…
💡 结构先验与算子学习是物理建模相对语言范式的核心差异。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
5) Fourier neural operators, and why Fourier — 19:41 🎧 · ▶️
💬 神经算子可在任意分辨率评估，适配物理场。 依据字幕：operators as a class of uh models that allow us to have any resolution input allow us to have any re…
💡 结构先验与算子学习是物理建模相对语言范式的核心差异。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
6) Old math meets deep learning, and why scaling breaks — 26:53 🎧 · ▶️
💬 围绕「Old math meets deep learning, and why scaling breaks」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：taken off right. So I think back then people really did think a lot about you people really did think a lot about you kn…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
7) Weather, and the 2021 skeptics — 32:39 🎧 · ▶️
💬 围绕「Weather, and the 2021 skeptics」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：partial differential equations but also more broadly you don't even need to more broadly you don't even need to more bro…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
8) Why the physical world is forgiving — 42:40 🎧 · ▶️
💬 围绕「Why the physical world is forgiving」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：but we've uh tested in our latest uh forecast net 3 model extreme weather forecast net 3 model extreme weather events of…）。
💡 可验证奖励与实验闭环是自动化科学的关键杠杆。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
9) Modeling across scales, from atoms to planets — 46:19 🎧 · ▶️
💬 围绕「Modeling across scales, from atoms to planets」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：them. Uh I mean this one is just kind of them. Uh I mean this one is just kind of showing that you know we have world at…）。
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
10) The data, ECMWF, and calling Hurricane Lee early — 50:33 🎧 · ▶️
💬 围绕「The data, ECMWF, and calling Hurricane Lee early」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：reanalysis data. So this is historical weather data weather data weather data >> that is in a way reanalyzed meani…）。
💡 数据与扰动设计决定因果/预测模型能否可信落地。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
11) Long rollouts, and why the sphere stays stable — 57:48 🎧 · ▶️
💬 围绕「Long rollouts, and why the sphere stays stable」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：the physics. the physics. >> Okay. Okay. You're not assuming the >> Okay. Okay. You're not assuming the >…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
12) Career full circle: theory, scale, and back to principled — 1:07:59 🎧 · ▶️
💬 围绕「Career full circle: theory, scale, and back to principled」讨论。字幕要点大意：嘉宾谈到与此相关的实践与判断（原文片段：right? So when AI was you know in this right? So when AI was you know in this where neural nets were not working where n…）。
💡 该段推进本集核心主张与证据链。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
13) A foundation model for physics, and inverse design — 1:13:02 🎧 · ▶️
💬 强调我们有语言基础模型，却还没有物理基础模型。 依据字幕：for language maybe vision but not for for language maybe vision but not for physics. So you know the…
💡 架构与算力约束决定方法能否规模化。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
14) The UN, regulating AI for science, and a call to action — 1:18:13 🎧 · ▶️
💬 联合国/治理与科学 AI 监管呼吁。 依据字幕：of time, but maybe just can you quickly give a bit of the story behind this and give a bit of the st…
💡 安全、伦理与治理是不可绕过的约束。
⚡ 自动字幕非人工精校；技术细节以论文/官方材料为准。
金句
- 「我们有语言的基础模型，却还没有物理的。」
- 「每一维哪怕几百个格点……上下文可达千亿甚至万亿。别想对这种规模上变压器，全世界的算力都不够。」
- 「把深度学习管用的都带走，但让它们更有原则一点。」
