Why the Next AI Breakthrough May Come from Physics with Max Welling

00:00

I want to send a big thanks to Blitzie for supporting the podcast and sponsoring this episode.

我想对 Blitzie 表示大大的感谢，感谢他们支持播客并赞助本期节目。

00:06

Once we accelerate software development velocity by 5X, you need Blitzie, which brings

一旦我们把 software development velocity 提升 5X，你就需要 Blitzie，它能

00:10

autonomous software engineering to your enterprise.

将 autonomous software engineering 带到你的企业。

00:14

Blitzie starts by reverse engineering your codebase, building a dynamic understanding of your

Blitzie 首先通过 reverse engineering 你的 codebase，建立对你

00:18

entire application ecosystem.

整个 application ecosystem 的动态理解。

00:20

Your engineer simply declares intent, and once approved, Blitzie autonomously executes

你的工程师只需声明 intent，一旦获批，Blitzie 就会自主执行

00:25

entire software epics, delivering validated end-to-end tested code, more than 80% of

整个 software epics，交付经过 end-to-end 测试并验证的代码，超过 80% 的

00:31

the work completed in a single run.

工作在 single run 中完成。

00:34

Blitzie is not just generating code, it's developing software at the speed of compute.

Blitzie 不只是生成 code，而是以 compute 的速度在开发 software。

00:39

Experience Blitzie firsthand at Blitzie.com-slash2mall.

访问 Blitzie.com-slash-2mall，亲身体验 Blitzie。

00:43

That's B-L-I-T-Z-Y.com-slash-T-W-I-M-L.

就是 B-L-I-T-Z-Y.com-slash-T-W-I-M-L。

00:49

Max is so great to be on the line with you again.

Max，能再次和你连线真是太好了。

00:52

It's been a while.

有一阵子了。

00:53

It's great to be back, Sam.

Sam，能回来真是太好了。

00:54

I'm looking forward to our discussion.

我很期待我们的讨论。

00:57

I am as well.

我也是。

00:59

Our audience can look up our conversations from, I think, 2019 and 2020, where we covered

我们的听众可以查看我们的对话，大概是2019年和2020年的，当时我们聊了

01:07

what you were working on at the time, and I think still echoes into your work today,

你那时候在做的工作，而且我觉得那些内容至今仍然影响着你的研究，

01:12

geometric neural networks, and gauge, aquaverience, neural networks, and the like.

几何神经网络，还有 gauge、aquaverience、神经网络之类的。

01:20

But I'd love to have you kind of catch us up on what you've been up to since, it's been

但我想请你跟我们更新一下你从那之后的进展，已经过了

01:25

quite a while.

好一阵子了。

01:26

Yeah, actually, the aquaverience theme has definitely continued, so, in fact, I found out

嗯，其实 aquaverience 这个主题确实一直在延续。事实上，我发现

01:33

that aquaverience was used very fruitfully in chemistry and material science.

aquaverience 在化学和材料科学中得到了非常成功的应用。

01:39

So in chemistry and material science, people train neural network models to predict

所以在化学和材料科学中，人们训练神经网络模型来预测

01:46

the forces on atoms, because people want to evolve atoms forward in time in order to

原子所受的力，因为人们想要让原子在时间上向前演化，以便

01:51

compute their properties, which is called molecular dynamics.

计算它们的性质，这被称为molecular dynamics。

01:56

And typically, you need to use quantum mechanics to compute these forces, because a large

而通常，你需要用quantum mechanics来计算这些力，因为很大一部分

02:03

contribution comes from the electrons, and electrons are very light.

贡献来自电子，而电子非常轻。

02:08

And so you need to treat them with quantum mechanics.

所以你需要用quantum mechanics来处理它们。

02:10

But, you know, if you get ten electrons or more, at big as it becomes completely unfeasible

但是，你知道，如果电子达到十个或更多，当体系变得足够大时，就完全不可行

02:15

to solve the so-called Schrodinger equation, so people have come up with approximations

去求解所谓的Schrodinger equation，所以人们提出了近似方法

02:20

like density functional theory, known as DFT, and people, you know, the inventors of that

比如density functional theory，也就是DFT，然后人们，你知道，那些发明者

02:27

got the Nobel Prize for that.

就因为这个拿了 Nobel Prize。

02:29

But what now people do is they train surrogate, so they provide data, computing this, using

但现在人们做的是训练 surrogate，提供数据，用这种昂贵的 quantum mechanics 近似来计算。

02:35

this expensive approximation to quantum mechanics.

然后他们用 neural networks 来走捷径，预测那个计算的结果，但速度要快得多。

02:39

And then they use neural networks to shortcut the computation, so to predict the outcome of

所以换句话说，相对于这些 quantum mechanics 近似，加速了三到四个数量级，效率更高。

02:45

that computation, but at a much more accelerated pace.

而在那些模型中，因为世界是三维对称的，所以如果我旋转一个分子，

02:48

So in other words, three orders or four orders of magnitude, acceleration, more efficiently

所以换句话说，三到四个数量级的加速，更高效，

02:54

relative to these quantum mechanical approximations.

是相对于这些 quantum mechanical approximations 而言的。

02:58

And in those models, because the world is three-dimensional symmetric, so if I rotate a molecule,

而在那些模型里，因为世界是 three-dimensional symmetric 的，所以如果我旋转一个分子，

03:05

all the forces will rotate with it, and so we could now put the same ideas that we use

所有的力都会随之旋转，所以我们就能把用在图像上的那套想法，放到这些分子、这些预测力的模型上，而且我们也可以用 equivariants。所以这也是为什么我——也因为我背景是科学，我博士读的是理论物理——我当时想，好吧，这简直是我旧热情和新热情的完美结合，我可以把它们放在一起，也就是从那时起我开始对 AI for science 感兴趣。所以这也就很直接地引向了 Cusp AI 的创立。

03:10

for images, we could put them in these molecules, these models that predict the forces, and

比如，我们可以把它们放进这些模型里，这些预测 forces 的模型里，

03:15

we could use agrivarians.

我们可以用 equivariants。

03:18

And so that's why I kind of, also because my background is in science, I did my PhD in

所以这就是为什么我有点——也因为我的科学背景，我读 PhD 时做的是

03:23

theoretical physics.

theoretical physics。

03:24

I thought, okay, this is a perfect unification of my old sort of passion, and my new passion,

我想，好吧，这完美统一了我以前的热情和现在的热情。

03:33

I can put it together, and that's when I started to be interested in AI for science.

我能把这些整合起来，也正是从那时起，我开始对AI用于科学产生了兴趣。

03:36

So that leads it pretty directly to the founding of Cusp AI.

所以这几乎直接促成了Cusp AI的创立。

03:41

Actually, first, I spent two years at Microsoft Research as a VP, because they were building

其实，首先，我在Microsoft Research做了两年VP，因为他们在组建AI for science实验室，我回应了他们，然后帮忙推进了那件事。

03:47

their AI for science lab, and I answered them, and so I helped that along.

但两年后，我想自己搞一个startup。

03:52

But after two years, I wanted to start a startup.

我之前也做过一个startup，但那家后来被Qualcomm收购了，之后我在Qualcomm待了一段时间。

03:56

I already did a startup a while ago, but this was actually the startup I got acquired

我真的很喜欢startup，那种充满活力的环境，以及你能产生的影响力。

04:01

by Qualcomm, and then I spent some time at Qualcomm.

而且我想用Silicon Valley的方式来做这件事，和我的联合创始人Chad Edwards一起，所以我们在2024年创立了Cusp AI。

04:05

And I really like startups, the dynamical environment, and the impact you can make.

我真的很喜欢startups，那种充满活力的环境，以及你能产生的影响力。

04:10

And I wanted to do it the Silicon Valleyway, together with my co-founder, Chad Edwards,

我想以Silicon Valley的方式来做这件事，和我的联合创始人Chad Edwards一起，

04:16

and so we started Cusp AI in 2024.

所以我们在2024年创立了Cusp AI。

04:19

Talk a little bit about the progress that you've made since then.

跟我们聊聊你从那以后的进展吧。

04:23

What is the kind of shape of the company today?

公司现在大概是什么形态？

04:25

This is being a huge ride, actually, it's a roller coaster.

这其实是一段很疯狂的旅程，就像坐过山车一样。

04:29

So we started, I think about two years ago, so may spring 24, yeah, we started with

所以我们大概两年前开始的，也就是2024年春天，对，我们一开始拿到了

04:39

a good initial investment of about 30 million from which we could hire an excellent team.

一笔大约3000万的初始投资，让我们能组建一支很棒的团队。

04:45

So the team has grown to about 50 people right now across different geographies, so

现在团队已经发展到大约50人，分布在不同地区，

04:53

there's a half-quarter, both in Amsterdam and in Cambridge.

在阿姆斯特丹和剑桥都有总部。

04:58

Actually, the half-quarter officially is in Cambridge, but the two initial labs were

实际上，总部正式设在剑桥，但最初的两个实验室是在

05:02

Amsterdam and Cambridge because Chad is from Cambridge, you know, from Amsterdam.

阿姆斯特丹和剑桥，因为Chad是剑桥人——你知道，阿姆斯特丹来的。

05:07

We now also have labs in London and Berlin, and we're also expanding into Asia and North

我们现在在伦敦和柏林也有实验室，同时也在扩展到亚洲和北美。

05:14

America.

明白。

05:15

Got it.

毫不意外，你的顾问名单有点像名人录，Jeff Hinton和Jan LeCoon排在最前面？

05:16

And no surprise, your list of advisors is a bit of a who's who, with Jeff Hinton and

是的，这些顾问确实非常棒。

05:23

Jan LeCoon at the top of the list?

所以我们有Jeff Hinton和Jan LeCoon。

05:25

Yes, the advisors are actually fantastic.

是的，我们的顾问们其实都非常棒。

05:29

So we have Jeff Hinton and Jan LeCoon.

我们有Jeff Hinton和Jan LeCoon。

05:33

We added to that also Martin Vandembrink and Lord Brown.

我们还加上了Martin Vandembrink和Lord Brown。

05:39

So Lord Brown is the former CEO of BP, and Martin Vandembrink is the former president and

所以Lord Brown是BP的前CEO，而Martin Vandembrink是前总裁，也是

05:46

CTO of ASML, they're both retired, and they like to spend their time with new startups

ASML的CTO，他们俩都退休了，喜欢把时间花在新的startups上

05:53

and help them along.

并帮助他们发展。

05:56

And then there's a variety of harding, she's working for the UK government and also deep

然后还有一个叫Harding的人，她在英国政府工作，同时也在DeepMind

06:00

mind, I think.

我想。

06:01

And then, or maybe formerly a debind.

然后，或者说，也许以前是在DeepMind。

06:04

And then Kristen Person who has sort of initiated the materials project, and she's also advising

然后是Kristen Person，她算是发起了Materials Project，而且她也在做顾问。

06:11

along that side.

沿着那一侧。

06:12

Got it.

明白了。

06:13

So let's dig into the technology and the approach that you're taking to kind of apply

那我们就来深入探讨一下你采用的技术和方法，如何应用

06:21

your, this original set of work that you developed to materials.

你开发的这套原始工作到材料上。

06:28

How did you get started with that effort?

你是怎么开始这项工作的？

06:31

So I think the opportunity is, to say to me, there's a deep fascination with the fact

所以我认为机会在于，对我来说，有一个事实让我深感着迷，

06:37

that there isn't sheer infinite amount of possibilities in which you can put together

那就是，并不是有无穷无尽的可能性，让你能够组合

06:41

atoms.

原子。

06:42

And the universe has only figured out so many of them, because you know, they stay, they

而宇宙目前也就演化出了这么多种，因为你知道，它们会留在宇宙中，自然形成，我猜是这样。

06:47

form naturally, I guess, in the universe.

但还有更多你可以自己设计的材料，带有各种奇异的性质。

06:50

But there's many more that you can design yourself with all sorts of exotic properties.

我们当初创办这家公司的时候，我和Chad都挺担心气候问题的，现在也依然如此。

06:55

When we started this company, both Chad and I were kind of concerned about the climate,

所以我们觉得，很有必要加速能源转型，转向更可持续的能源来源，同时也要试着，你知道，把大气中的二氧化碳给弄出来。

07:02

we're still are.

我们仍然如此。

07:04

And so we felt there's a strong need to accelerate the energy transition to more sustainable

所以我们觉得很有必要加速 energy transition，转向更可持续的能源来源，同时也要尝试，你懂的，把大气里的 carbon dioxide 弄出来。

07:09

energy sources, as well as trying to, you know, take the carbon dioxide that's in

那是极其庞大的量。

07:16

the atmosphere out.

那大概是，我觉得，有 Lake of Geneva 那么大，装满液态 carbon dioxide，一个巨大的量。

07:17

So not many people know that by the time it's, but 2015, of course, we really like to

其实没多少人知道，到2015年的时候，我们当然真的很想做到完全carbon neutral。

07:23

be completely carbon neutral.

但在这之后，还有50到100年的时间，我们每年都得把目前排放量的大约一半再移除掉。

07:26

But after that, there is still 50 to 100 years where we have to take out every year, about

所以那就是每年20 gigatons。

07:32

a half of what we currently put in.

这个量极其巨大。

07:33

So that's 20 gigatons a year.

我觉得那差不多就是整个Lake of Geneva装满liquid carbon dioxide的规模——一个庞大的量。

07:35

That's an enormous amount.

所以如果还没有现成的合适方案，它实际上会完全从头设计它们。

07:36

That's the kind of the, I think the size of the lake of Geneva filled with, sort of,

那大概就是，我想，有Geneva湖那么大，装满了，怎么说呢，

07:41

liquid carbon dioxide, a gigantic amount.

liquid carbon dioxide，巨量。

07:44

And we don't have the technology for that because it's actually very hard to take it out.

而且我们还没有那个技术，因为实际上把它提取出来非常难。

07:49

Because it's so dilute in the atmosphere, it's very expensive too.

因为它在大气中太稀薄了，所以成本也非常高。

07:53

And, you know, the two, the two factors which are more expensive are energy and the material,

而且，你知道，两个更贵的因素是能源和材料，

08:00

the sort of material that you use in order to take the carbon dioxide out of the atmosphere.

就是你用来把 carbon dioxide 从大气中提取出来的那种材料。

08:06

And so we've been working the first project we've been doing was improving the materials

所以我们一直在做的第一个项目，就是改进这些材料，

08:11

that take out this carbon dioxide from the atmosphere.

也就是把 carbon dioxide 从大气中提取出来的材料。

08:14

And these are called metal organic frameworks.

这些被称为 metal organic frameworks。

08:17

And this is the material that this year won the Nobel Prize in Chemistry.

而且这种材料今年获得了 Nobel Prize in Chemistry。

08:23

And yeah, we, so we build a platform and I can go into much more detail, but we build

嗯，对，我们建了一个 platform，我可以更详细地讲，就是我们用 AI 来设计这些材料。

08:26

a platform that designs these material with the use of AI.

所以你可以大致把它想成一个 search engine，但它不是搜索现有的文档。

08:30

So you can sort of think of it as a search engine, but it's not searching over existing

它搜索的是已知和未知的材料。

08:34

documents.

所以如果没有现成的合适材料，它实际上会从零开始完全设计它们。

08:35

It's existing, it searches over known and unknown materials.

而且它有很多组件。

08:39

So it actually completely designs them from scratch if, if there's no suitable ones

所以它实际上会从头开始完全设计它们，如果没有合适的

08:43

already available.

现成可用的话。

08:45

And there's many components.

而且有很多组成部分。

08:46

It's a genetics.

这是genetics。

08:47

There is an agent sitting at the core of it who orchestrates a long computation.

它核心有一个agent，在编排一个很长的computation。

08:53

And it searches through existing databases.

它会搜索现有的databases。

08:55

It generates entirely new molecules.

它生成全新的molecules。

08:59

And it also evaluates all these molecules with all sorts of tools.

它也会用各种tools来评估所有这些molecules。

09:02

And in this generation and evaluation, you know, ingredients and all sorts of methods

而且在这个generation和evaluation中，你知道，我们在我的academic lab里这些年开发出的ingredients和各种methods，你知道，扮演了非常重要的角色。

09:08

that, you know, we have developed over the years in my academic lab are play a very important

这些，你知道，是我们在我的学术实验室里多年来开发的，起着非常重要的

09:13

role.

作用。

09:14

To be clear, you mentioned a molecule that won the Nobel Prize, were you involved in

说清楚一下，你提到一个得了 Nobel Prize 的 molecule，你参与发现那个 particular molecule 了吗？

09:22

the discovery of that particular molecule?

没有没有，我倒希望是。

09:24

No, no, I wish.

那是化学家做的，是他们发现的。

09:28

Those were chemists who did that, who discovered it.

我想是 Professor Gita Gawa、Professor Yagi 和 Professor Robson。

09:32

I think Professor Gita Gawa, Professor Yagi, Professor Robson.

我想就是这三位，最后一位我不太确定。

09:37

I think those are the three I'm not quite sure of last one.

不过这类 molecule 是，你知道，在 metal node 里有一个 metal complex，有各种 atoms 沿着它，还有在 graph 的 vertices 上。

09:41

But this is a class of molecules where there is a metal complex in a metal node with all

但这是一类分子，其中一个metal complex在一个metal node里，带有所有...

09:47

sorts of, you know, atoms along it and at the vertices of a graph.

就是，你懂的，原子沿着它以及 graph 的 vertices 排列。

09:53

And then there is so called linkers, which are also organic complexes, which are connecting

然后还有所谓的 linkers，它们也是 organic complexes，用来连接 graph 上的这些 vertices。

09:59

these vertices on the graph.

而且它们 extremely porous。

10:02

And they're extremely porous.

所以中间有非常大的 holes，surface area 也巨大。

10:04

So they have a very large holes in the middle with an enormous surface area.

所以如果你让空气 atmosphere 吹过去，空气中的 molecules，也就是 water、nitrogen 和 carbon dioxide——carbon dioxide 只占很小一部分。

10:11

And so if you blow sort of air atmosphere through it, the molecules in the air, which

它们 tend to stick to the sides。

10:17

is water and nitrogen and carbon dioxide, the carbon dioxide is only a small fraction

是水、氮气和二氧化碳，二氧化碳只是

10:22

of that.

其中的一小部分。

10:23

They tend to stick to the sides.

它们往往会粘在边上。

10:26

And so then what you need to make is a molecule, you need to design a molecule where actually

所以接下来你需要制造的是一个 molecule，你需要 design 一个 molecule，也就是说，实际上

10:33

only the carbon dioxide sticks inside these pores and the rest, you know, goes through.

只有 carbon dioxide 会吸附在这些 pores 里面，其他的，你懂的，就直接通过。

10:37

So that's a design question.

所以那是一个 design 问题。

10:39

And also when it's full, you want to shake it or heat it to get it out so that you can

而且当它满了之后，你要摇晃它或者加热它，把它弄出来，这样你才能

10:46

actually reuse that particular material.

真正重复使用那个特定的材料。

10:49

What we added to this is a way to fine tune or design a molecule for a particular purpose.

我们在此基础上加入的是一种方法，可以针对特定目的去 fine-tune 或者 design 一个 molecule。

10:58

So people have, you know, actually made maybe around 100,000 of these molecules and labs

所以人们，你知道，实际上可能已经在 labs 里制造了大概 100,000 个这样的 molecules，

11:06

and and verify their structure.

然后，然后验证了它们的 structure。

11:08

And so what we can do is we can now basically come up with an entirely new molecule for a

所以现在我们可以做的，基本上就是针对一个非常具体的任务，设计出一种全新的 molecule，然后在实验室里把它做出来，再用于那个特定的任务。

11:14

very specific task and then make it in the lab and then use that for that, for that particular

我要说的是，这项工作不仅仅是在 molecule 上。事实上，这只是我们合作的第一批 molecule。我们后来已经扩展到了 semiconductor 领域。我们做了很多关于 perovskite 的工作，perovskite 是用于太阳能面板的材料，可以改进太阳能面板。

11:19

task.

任务。

11:20

I should say cost was not only working on months.

我要说的是，我们不只是研究分子。

11:23

In fact, this was just the first set of molecules that we worked with.

事实上，这只是我们接触的第一批分子。

11:27

We have hence expanded to semiconductors.

因此，我们已经扩展到半导体领域。

11:31

We're doing a lot of work on and perovskides, which is materials for solar panels improved

我们在perovskites上做了很多工作，这是一种用于改进太阳能电池板的材料。

11:39

improved solar panels.

改进的太阳能电池板。

11:41

We look at semiconductors for new chip materials.

我们研究 semiconductors，寻找新的 chip materials。

11:47

We look at, you know, battery materials, fuel cells.

我们也会看，你知道，battery materials、fuel cells。

11:53

We also look at removing PFAS from water and we in the current set of molecules we use

我们也研究从水中去除PFAS，而我们目前用的那组 molecules

11:59

for that is again, metal organic frameworks.

用于这个的，还是 metal organic frameworks。

12:02

Do you partner with other companies that have an interest in these particular molecules

你们会和那些对这些特定 molecules 感兴趣的其他公司合作吗？

12:09

or are you out exploring and then if you find something, you will find partners and maybe

还是你们自己去探索，然后如果发现了什么，再找合作伙伴，也许

12:14

license to them.

授权给他们。

12:15

What's the thinking around the business model?

关于商业模式，你们是怎么考虑的？

12:18

So we like to work with partners because there's a very broad class of materials and

所以我们喜欢和合作伙伴一起工作，因为材料类别非常广泛，

12:23

every class has its own super experts that focus on those particular areas.

每一类都有自己的超级专家，专注于那些特定领域。

12:27

And they're either in academia or they're in companies.

他们要么在学术界，要么在公司里。

12:32

And of course, we also like to partner on the actual synthesis of these materials.

当然，我们也喜欢在这些材料的实际 synthesis 上进行合作。

12:35

So we partner with academic labs, but also with the labs inside of these companies.

所以我们既与学术实验室合作，也和公司内部的实验室合作。

12:40

And so we build an ecosystem or a network where our engine can actually help in all of these

这样我们就建立了一个 ecosystem 或 network，让我们的 engine 能够在所有这些

12:49

different material classes design the materials and then we work with those companies to actually

不同的材料类别中帮助设计材料，然后我们与那些公司合作，实际

12:54

make it.

把它做出来。

12:55

That's a partnership model, but there's also internal projects that we run.

那是一个合作模式，但我们也有内部开展的项目。

13:03

So for instance, the project on metal organic frameworks for carbon capture, we ran self-funded

比如，那个关于metal organic frameworks用于carbon capture的项目，我们就是自筹资金在内部做的。

13:08

internally.

然后我们现在在semiconductor方面也有另一个项目，也是我们在做。

13:11

And then we also have another project now in the semiconductor side where we run it.

所以如果我们发现了很棒的东西，因为是我们自己出资的，我们就拥有IP，然后可以为这个IP找客户。

13:16

And so if we discover something fantastic, we self-funded, so we got the IP and then

最重要的是，我觉得最重要的是我们发现了某些东西，能让人们对我们有信心，相信我们能做这件事。

13:20

we can find customers for that IP.

我们可以为这些IP找到客户。

13:24

Most important, I think the most important thing is that we discover something that gives

最重要的是，我认为最重要的是，我们发现了一些东西，它能带来

13:28

people certain confidence that we can do this.

人们对这件事有一定的信心。

13:31

We actually really own this process and we know how to do it.

我们其实完全掌控这个过程，而且我们知道怎么做。

13:35

And so then the customers to design materials for their specific needs.

然后客户就可以针对他们的特定需求来设计 materials。

13:40

You talked a little bit about the generative or agentic nature of the scanning scientific

你之前提到了一点，关于扫描 scientific literature 的 generative 或 agentic 特性，以及如何用它来识别潜在的 molecules。

13:45

literature and using that to identify potential molecules.

但我们也聊过你工作里的一些 geometric 含义。

13:51

But we've also talked about some of the geometric implications of your work.

你暗示过可能会用到 simulation。

13:58

You alluded to potentially the use of simulation.

你能再多讲讲识别这些 molecules 的 end-to-end 过程吗？

14:03

Can you talk a little bit more about the end-to-end process of identifying these molecules?

好的，我很乐意。

14:09

Yeah, happy to.

嗯，很乐意。

14:10

So I guess there's a sequence of things that happens, right?

所以我猜事情是有一个先后顺序的，对吧？

14:14

The first thing is like in a search engine, you actually type a request.

第一件事就像在 search engine 里，你实际上输入一个请求。

14:19

So you basically say, I want a material and these are all the properties that it should

所以你基本上会说，我想要一种材料，这些是它应该

14:23

have.

具备的所有性质。

14:24

And these are the things it should not have.

还有这些是它不应该有的东西。

14:26

These are the properties it should not have.

这些是它不应该有的性质。

14:29

And so you give this as a query.

然后你就把这个作为 query 提交。

14:32

And then you could also tell if you have prior knowledge about how you want this particular

然后你还可以告诉它，你是否拥有 prior knowledge，关于你想要这个特定的

14:36

search to happen.

为了让搜索发生，

14:37

You could tell the system, maybe use these tools and maybe sequence it in this way.

你可以告诉系统，也许使用这些 tools，也许以这种方式来安排顺序。

14:43

So you can also give it some instructions.

所以你也可以给它一些 instructions。

14:46

And then it goes through a process of steps.

然后它会经历一个多步骤的 process。

14:48

So the first step is it will look through its database.

所以第一步是它会在它的 database 里查找。

14:51

So we have a very large database of materials, which we have all ingested into this database.

所以我们有一个非常大的 materials database，我们把这些材料都 ingest 进了这个 database。

14:58

Lots of it is sort of exclusive licenses from the big publishing houses.

其中很多是从大型出版社获得的 exclusive licenses。

15:04

And then it will start to look through all of the literature, whether something exists

然后它会开始浏览所有的 literature，看看是否存在相关内容。

15:08

out there that has these properties, or which is close to having these properties.

外面已经有的，能具备这些性质的，或者接近具备这些性质的。

15:14

And so if it doesn't, then it will have to go into a new phase.

如果没有的话，那就得进入一个新阶段。

15:18

So you can hold the conversation with this agent and talk about it.

所以你可以和这个 agent 对话，跟它讨论这件事。

15:21

So that's already quite useful.

那这已经挺有用了。

15:23

But typically, then the next step is that you go to a generative model.

但通常来说，下一步就是转向 generative model。

15:27

So in this case, it's the same generative model that generates images or video.

所以在这种情况下，它就是同一个能生成图像或视频的 generative model。

15:33

And so but in this case, it will generate molecules for you.

但在这里，它会为你生成分子。

15:38

And so you tell it, I want molecules with these very specific properties, the conditioning

然后你告诉它，我想要具有这些非常特定性质的分子，这就是 conditioning。

15:42

statement.

陈述。

15:43

It's cold.

天气很冷。

15:44

And then it will start to generate these molecules, often, you know, hundreds of thousands

然后它就会开始生成这些 molecules，通常，你知道，会有成千上万个。

15:48

of them.

因为在计算机里这相当便宜。

15:49

Because it's quite cheap in the computer.

然后就到了下一个阶段，就是从这些生成的 molecules 中。

15:52

And then comes the next phase, which is out of these generated molecules.

我们现在必须筛出那些看起来，看起来非常有前景的。

15:57

We now have to sieve out the ones which are look, look very promising.

而这可能是一个非常昂贵的步骤。

16:01

And this can be a very expensive step.

而这可能是非常昂贵的一步。

16:05

In some sense, you're built a multi-scale digital twin of the process that you really want

从某种意义上说，你构建了一个多尺度的 digital twin，模拟你真正想让这些分子在其中运作的那个过程。所以第一步基本上就是把这个分子 relax 到它的 ground state，确保它处于最佳能量状态。然后你会做一堆检查，比如它带不带电荷？如果我晃它一下，它会不会散掉？如果孔洞大小重要的话，孔洞有多大，你知道的，各种容易快速算出来的东西，用来筛掉一大堆看起来没什么希望的候选。

16:12

to these molecules to operate in.

让这些 molecule 可以在其中运作。

16:14

And so the first step is basically you relax the molecule to its ground state to make

所以第一步基本上就是把这个 molecule relax 到它的 ground state，确保它处于最佳的 energy state。

16:20

sure that it's the best energy state.

然后你会做一堆检查，比如它是不是 charged？如果我摇一摇它，会不会散架？

16:24

Then you do a bunch of checks, like is it charged?

然后你做一堆检查，比如它有没有带电？

16:29

If I shake it, will it fall apart?

如果我摇一摇它，会不会散架？

16:31

How big are the pores inside if that's important, you know, all sorts of things that are easy

如果这很重要的话，里面的pores有多大，你知道，有各种各样容易快速compute然后扔掉的东西，你懂的，一大堆看起来不太有前景的。

16:37

to compute fast to throw away, you know, a whole bunch of things that do not look promising.

然后我们把这个force field提炼成一个对这个特定类别的问题非常高效的force field。

16:43

And so then filtering it down.

然后就把它过滤下来。

16:45

Yes, definitely a filtering step, yes.

对，肯定有一步是 filtering。

16:47

And then go to the next step.

然后再进入下一步。

16:48

So we have a pipeline that fine tunes or distills machine learning force fields for that

所以我们有一个 pipeline，专门针对那种特定材料去 fine-tune 或 distill machine learning force fields。

16:57

particular material.

这个过程就是，我们收集关于那种材料的所有数据。

16:58

So this is a process by which we take all the data there is about that material.

我们有一个 foundation model，它在更广泛的数据上训练过。

17:03

We have a foundation model that's trained on a much wider range of data.

然后我们把这个 force field distill 成一个——对这种材料来说非常高效的 force field。

17:08

And then we distill this force field into this is a very efficient force field for this

然后我们在MD loop，通常是molecular dynamics loop里用它来模拟分子，分子会扭来扭去、动来动去，从中你通常能计算出那个分子的非常关键的properties。

17:14

particular class of problem.

这些properties之后通常会进入更高尺度上的partial differential equation。

17:16

And we use that in an MD loop, typically the molecular dynamics loop, to simulate the molecule

然后我们在 MD loop 里用到这个，通常是 molecular dynamics loop，来模拟这个分子

17:22

as it wiggles around and moves around from which you can often compute very key properties

当它扭来扭去、动来动去的时候，你通常能从中算出非常关键的性质。

17:27

of that particular molecule.

那个特定分子的。

17:30

Those properties then often go into partial differential equation at a higher scale.

这些属性随后通常会进入更高尺度的partial differential equation。

17:35

And or in a process that actually models the device in which you want this up this material

或者说，在一个实际模拟你要让这种材料运行的设备的流程中。

17:41

to operate.

所以那又更贵了。

17:43

And so that's again, more expensive.

然后你得到一个candidate，我们目前，我会说，这个领域正在发生的革命是self-driving labs，你能做的实验数量要快得多得多。

17:45

And so again, you want to do this with fewer and fewer candidate materials.

所以再说一遍，你希望用越来越少的候选材料来做这件事。

17:50

And then at the very end, you go to, you know, to an experiment, right, you go so now

然后到最后关头，你会去，你知道，做实验，对吧，就是说现在真的去做那个实验。

17:55

actually do the experiment.

那更贵，而且更慢。

17:56

Now that's more expensive and even slower.

而且你应该只用大概10种材料，用老方法来做。

17:59

And that's you should do this only with order 10 materials post in the old way.

所以然后你会得到一个候选物，我们目前——当前这个领域正在发生的革命，我会说，就是self-driving labs，在那里你做实验要快得多，快得多。

18:05

And so then, so and then you get a candidate, what we are currently, the current, I would

所以对于这些来说，experimental loop也要快得多。

18:09

say, revolution that's happening in this space is self-driving labs where the amount of

这是一个非常有趣的发展，我们现在正把我们的platform和它整合。

18:15

experiments you can do is much, much faster.

你能做的实验要快得多得多。

18:17

So you could do maybe a hundred experiments a day.

所以你可能一天大概能做一百个 experiments。

18:21

And then the game is more like the agent figures out what the settings of the experiment

然后这个游戏更像是 agent 去搞清楚 experiment 的 settings 应该是什么。

18:26

should be.

Experiments 做完了。

18:27

The experiments are done.

Data 会传回来。

18:28

The data comes back in.

然后你就有来自 experiment 的 data 和来自 computations 的 data。

18:30

And then you have the data from the experiment and the data from your computations.

你把这些结合起来，去设定 experiments 的下一阶段。

18:35

You combine them to set the next stage for the experiments.

所以 experimental loop 对这些来说会快非常多。

18:39

And so the experimental loop is much, much faster for those.

所以对于这些来说，experimental loop要快得多得多。

18:44

And that's a very interesting development that we are now integrating our platform with.

这是一个非常有趣的进展，我们现在正在将我们的平台与之整合。

18:49

Got it.

明白了。

18:50

And so how many molecule classes and individual molecules have you kind of gone, you know,

那你们大概接触过多少种 molecule classes 和 individual molecules，你知道，

18:59

all the way through this cycle with?

完整走完这个 cycle 的？

19:01

Yeah, we are engaged in in a few of those, but you know, the question is a little bit,

是啊，我们确实参与了其中一些，但你知道，这个问题有点，

19:07

what do you mean by all the way?

你说的 all the way 是什么意思？

19:09

The different materials are at different stages of maturity.

不同的 materials 处于不同的成熟阶段。

19:12

So one of them we went all the way to, you know, actually doing, you know, doing the lab experiments.

所以其中一个我们真的走到了，你知道，真正去做 lab experiments 那一步。

19:18

Another way on semiconductors is on its way.

另一种关于 semiconductors 的方法也快出来了。

19:22

And probably in a few months, we'll start doing the experiments.

大概几个月后，我们就会开始做实验。

19:26

Met all the way to lab experiments.

一路做到了 lab experiments。

19:28

And I was curious if you have enough data to say that you're hit rate with this process

而且我很好奇，你们是否有足够的数据可以说你们在这个 process 上的 hit rate

19:35

is higher, you know, versus lower or, you know, there's anything you can say qualitatively

是更高，你知道，还是更低，或者，你知道，有没有什么能定性地说的，

19:42

or quantitatively about the candidates that you produce relative to, you know, the traditional

或者定量地，关于你们产生的 candidates，相对于，你知道，传统的

19:47

approach.

approach。

19:48

Yeah.

嗯。

19:49

So I definitely think that, you know, there's definitely evidence that these things are

所以我确实认为，你知道，确实有证据表明这些东西是

19:55

much more efficient.

效率高太多了。

19:56

So some of our scientists, they have said things like, we can do now in a few days, what took

所以，我们有些科学家说过类似的话：以前一个博士要花很久才能做完的事，我们现在几天就能搞定，但这更多是在digital领域。

20:02

a PhD before, but that's more in the digital domain.

这些科学家其实更多是做simulation方面工作的。

20:07

These are actually scientists that are more simulation-based work.

我们有两个正在进行中的项目，里面有实验性的部分。但我们有一个——你知道——我们和亚洲一个大型国家实验室签了合同，目前还不能说太多细节，我们在那里规划了更多类似的实验。

20:12

We have two projects going, which have experimental pieces to it, but we have a, you know, we have

所以基本上，安排就是我们带一些客户过来，然后我们可以在那里做实验。

20:21

a contract with a big national lab in Asia, which cannot quite say the details of yet, where

与亚洲一家大型国家实验室的合同，我们还不能透露太多细节，在那里

20:28

we have many more of these experiments planned out.

我们已经规划了更多这样的实验。

20:31

So basically, the setup is that we bring some customers and we can do experiments in that

所以基本上，我们的安排是，我们带来一些客户，然后我们可以在那里做实验。

20:37

lab, but they're lab scientists and they can also bring their partners and they can

实验室，但他们是实验室科学家，他们也可以带上他们的合作伙伴，他们可以

20:42

use our platform.

使用我们的platform。

20:44

And then we can collect data that way.

然后我们就可以通过那种方式收集data。

20:46

And I think, oh, and then there is one other one that we're currently doing in Amsterdam

而且我想，哦，还有一个我们目前在Amsterdam做的

20:53

on perovskites, which is running right now.

关于perovskites的，现在正在进行中。

20:56

So that's also a lab engagement.

所以那也是一个实验室合作项目。

20:58

There is one in...

还有一个在……

20:59

Perovskites is one.

Perovskites就是其中一个。

21:00

Sorry.

抱歉。

21:01

Yeah.

嗯。

21:02

Perovskites is a material class that's a semiconductor crystal structure that you would put

Perovskites 是一种材料类别，它是一种 semiconductor 晶体结构，通常你会把它放在……放在 silicon 上，也就是普通 solar cells 的上面。

21:08

on a... on silicon, typically, on top of the normal solar cells.

这可以帮助你过滤掉，或者说基本上把可见光中的大量能量转化为能量。

21:14

And that can help you filter out or basically convert a large amount of the energy in visible

对吧？

21:21

light to energy.

所以我们和 DTU，也就是 Danish Technical University，也有一些关于 catalysis 的合作。

21:23

Right?

对吧？

21:24

So we also have something with catalysis with DTU, which is the Danish Technical University.

所以我们和DTU也有关于catalysis的合作，也就是Danish Technical University。

21:29

So that's on currently running on catalysis, and we have a project that's about to start

所以那目前是在做催化(catalysis)，我们有一个项目即将开始

21:36

them off.

启动它们。

21:37

And quite a few engagement with labs that are either running or are starting to run,

并且有不少实验室的合作，有的已经在运行，有的即将开始运行，

21:44

but they haven't finished completely.

但它们还没有完全结束。

21:46

And is your lab, or does CUSP publish?

那你的实验室，或者CUSP，会发论文吗？

21:52

Are you active?

你还活跃吗？

21:54

Are you still active in academic publishing around materials science now?

你现在还活跃在材料科学(materials science)的学术发表领域吗？

21:59

So myself, I have one day at the university, and of course I still work in that day with

我自己嘛，在大学里有一天的时间，当然那天我还是会和

22:03

students and there we publish everything.

学生们，而且我们在那里发布所有内容。

22:06

I'm very interested in all sorts of things we have to do with AI for science.

我对我们为科学所做的各种AI相关的事情非常感兴趣。

22:10

So definitely yes.

所以当然是肯定的。

22:12

But also CUSP actually publishes.

但CUSP实际上也有发布。

22:15

So we have interns that work with our scientists, where we publish the results.

所以我们有实习生和我们的科学家一起工作，我们会发布结果。

22:22

So we've, Francis recently had results on machine learning force fields with uncertainty,

所以，我们——Francis最近在带uncertainty的machine learning force fields方面有了结果，

22:26

prediction, property predictors.

prediction、property predictors。

22:30

And the most important thing, I think, that we've recently released open source and published

而且最重要的是，我认为，我们最近发布了open source并发表了。

22:35

a blog post about is our new molecular dynamics framework called Cups, confusingly.

有篇博客文章讲的是我们的新 molecular dynamics framework，叫 Cups，挺让人困惑的。

22:43

So that's universal, it's universal particle simulation simulator.

所以它是 universal 的，是一个 universal 的 particle simulation simulator。

22:49

And so the reason why we did this is actually together with Nvidia is that, you know, now

而我们做这个的原因，实际上是和 Nvidia 一起的，就是，你知道，现在

22:56

with this new development of these machine learning force fields that I talked about

随着我之前提到的这些 machine learning force fields 的新发展

23:00

is neural networks that replace quantum mechanical calculations for the forces.

也就是用 neural networks 来替代用于计算 forces 的 quantum mechanical calculations。

23:05

You cannot run them very efficiently in the current sort of MD simulators, the simulators

你不能在现有的 MD simulators 中很高效地运行它们，也就是那些 simulators

23:12

that evolve a material or, you know, a molecule forward in time.

那些让材料或者，你知道，molecule 随时间向前演化的 simulators。

23:17

Because you need to, you need to run these neural networks on GPUs.

因为你需要，你需要把这些 neural networks 跑在 GPUs 上。

23:22

And typically, you know, these, these simulators, they don't run GPUs or on CPUs.

而且通常，你知道，这些 simulators 并不运行在 GPUs 或 CPUs 上。

23:29

And you want to run things in parallel.

而且你想要并行运行。

23:32

And so what was built by our team is a framework where you can compile these force fields into

所以我们团队构建的是一个 framework，你可以将这些 force fields 编译成

23:41

jacks, which is Python based.

jacks，它是基于 Python 的。

23:44

And then it will actually very efficiently run these MD simulators.

然后它实际上会非常高效地运行这些 MD simulators。

23:50

And you can run them in parallel as well so that you use your GPUs, the utilization of

而且你也可以并行运行它们，这样你使用你的 GPUs，

23:54

your GPUs is high.

你的 GPUs 的利用率会很高。

23:55

And so we released this open source a few days ago, actually, where we could go at

所以我们实际上几天前发布了这个 open source，然后我们可以去

23:59

I clear.

我清楚了。

24:01

And yeah, we got a lot of excited responses to that are the representations of the molecules,

而且，关于分子的representations，我们收到了很多热烈回应。

24:09

like are there standard formats for representing these things that someone working in the

比如，在这个领域工作的人是不是已经有了一些standard formats来表示这些东西？

24:14

space would would already have.

然后你的machinery就直接在这些已有的格式上运行。

24:15

And then your machinery just works on those existing ones.

你可以把representations看作是一种化学领域的foundation model。

24:20

The representations, you can think of that as maybe a foundation model for chemistry.

所以你要做的是——这也是我们和Meta一起做的工作。

24:25

So what you do is, and this is also work that we did together with meta.

所以在那种情况下，你就自己训练一个这样的force field。

24:31

So what you do in that case is you train yourself one of these force fields.

所以在这种情况下，你要自己训练一个这样的force field。

24:37

So you take a molecular structure, so that is basically the position of the atoms and

所以你取一个 molecular structure，那基本上就是 atoms 的位置，

24:43

their and their and their class, like if it's oxygen or hydrogen.

以及它们的，它们的，它们的 class，比如是 oxygen 还是 hydrogen。

24:49

And of course, you want this to be aquivarian because rotations and translations don't

当然，你会希望这个是 equivariant 的，因为 rotations 和 translations 并不

24:53

matter.

重要。

24:54

And then you map this into a latent space where they get represented by some code.

然后你把它映射到一个 latent space，在那里它们用某种 code 来表示。

24:59

That is, that is meaningful.

也就是说，这是有意义的。

25:01

And if you do a machine learning force field, you will then actually predict the forces

如果你用 machine learning force field，你就能真正预测 forces

25:04

and the energies from that.

以及从那里得到的 energies。

25:06

But you can stop at that intermediate level.

但你可以就停在那个中间层级。

25:08

And then you have a, you know, a representation for that particular material.

然后你就有了，你懂的，针对那种特定材料的 representation。

25:14

And you can train this on a very wide range of materials and chemistry.

而且你可以在非常广泛的材料和化学上训练这个。

25:19

And from that representation, you can build property predictors.

然后从那个 representation，你就可以构建 property predictors。

25:23

You can use it to condition your generative models.

你可以用它来 condition 你的 generative models。

25:25

There's also some uses for that, you know, foundation model.

你懂的，那个 foundation model 也还有一些用途。

25:30

But it's also the starting point for distilling, let's say models for, for specific material

但它也是，比如说，针对特定材料类别来 distilling 模型的起点。

25:37

classes.

类别。

25:38

And to be clear, are these foundation models trained, you know, per molecule or per class

而且说清楚一点，这些foundation model是每个分子训练，还是按类别训练？

25:43

or are they, you know, very broad?

还是说它们非常宽泛？

25:47

Yeah.

对。

25:48

So that's a great question.

这是个好问题。

25:49

A foundation model, almost by definition, is trained on a very broad class of materials.

一个foundation model，几乎从定义上讲，就是在很宽泛的材料类别上训练的。

25:57

The best data set for that is the materials project data set and, and omol from, from

最好的数据集是materials project数据集，还有Meta的omol。

26:03

meta.

所以他们生成了非常大量的DFT计算，这些量子...

26:04

So they have, they have generated very large number of dft calculations, these quantum

所以他们生成了非常大量的DFT计算，这些quantum……

26:08

mechanical calculations to create that data set.

用力学计算来生成那个数据集。

26:12

And those calculations are being used to train these force fields that I talked about

而这些计算正被用来训练我提到的那些 force fields。

26:16

or those representations, those foundation models, right?

或者那些 representations，那些 foundation models，对吧？

26:19

And then if, if you say, but I'm actually interested in this particular material, right?

然后如果你说，但我实际上对这个特定材料感兴趣，对吧？

26:24

What you can then do is start from that very broad representation, this foundation model

那你要做的就是从那个非常宽泛的 representation，也就是这个 foundation model 出发。

26:30

and then fine tuned model for that particular material class.

然后再对那个特定材料类别做 fine-tuned 的模型。

26:34

That way it is specialized for the material class, but that way it's also very fast because

这样一来它对这类材料是专精的，但同时也非常快，因为

26:38

you need these things to do very fast because in a molecular dynamics simulation, you have

你需要这些东西运行得非常快，因为在 molecular dynamics simulation 里面，你要面对的是……

26:43

to call them many, many times, right?

要 call 它们很多很多次，对吧？

26:45

Which is one, one little step in an MD simulator is a femtosecond.

在 MD simulator 中，一小步就是一 femtosecond。

26:49

It's a tiny step.

这是非常小的一步。

26:51

And so you have to call them many times to make any progress.

所以你必须 call 它们很多次才能有所进展。

26:54

And you mentioned distillation earlier, is that where distillation comes in?

你之前提到了 distillation，distillation 就是在这里派上用场的吗？

26:57

You're trying to get a smaller model that's more focused on the molecules that you care

你想得到一个更小的 model，让它更专注于你关心的那些分子。

27:03

about.

完全正确。

27:04

Absolutely.

没错。

27:05

That's, that's, that's what distillation is called.

这，这，这就是所谓的distillation。

27:06

You distill the big model into a small model.

你把大模型distill成一个小模型。

27:09

In this case, the use of fine tuning, is it, how analogous is it to the process of fine

在这种情况下，使用fine tuning，它跟LLM里的fine tuning过程有多相似？你知道，reinforcement fine tuning，用traces，就是language-based traces？

27:17

tuning in LLM, you know, reinforcement fine tuning with, you know, traces, you know,

比如，你的dataset是什么？你在这个过程里用来做fine tuning的dataset是什么？

27:23

language-based traces?

所以实际上它跟LLM很相关。

27:24

Like what is the, your dataset that you're fine tuning in your process that you're fine

所以有，有两种做法，通常你可以用graph neural network，它实际上看三维结构，或者你可以把它当作LLM，比如……

27:29

tuning with here?

用这里来 tuning 吗？

27:30

So actually it's quite related to an LLM.

所以实际上这和 LLM 关系很大。

27:34

In fact, the models we use are very related to an LLM, so you can basically think of the

事实上，我们用的模型和 LLM 非常相关，所以基本上你可以把现有的信息看作一个 sequence，然后把它 sequentialize。

27:39

sequence, the information that you have, you can sequentialize it.

然后从那里，你就可以，呃，把它映射到这个 latent space 里。

27:45

And then from there, you can actually, you know, map it into this latent space.

所以有，有两种做法，通常你可以用 graph neural network，它实际上看三维结构；或者你把它当 LLM 用，像 sequence 版本，比如分子经常被表示成一种叫 smile string 的字符串，这样它就变成了一维 string，你也可以用那个来表示分子。

27:49

So there's, there's either, and typically you can do like a graph neural network, which

所以有，有一种，通常你可以用 graph neural network，它

27:58

actually looks at the three-dimensional structure, or you can use it as an LLM, like

实际上查看三维结构，或者你可以把它用作 LLM，比如

28:01

a sequential version, like often molecules are represented as, you know, as a, as a

一种 sequential 的版本，就像分子经常被表示成，你知道，作为一个叫 SMILES string 的 string，然后就变成了一个 one-dimensional string，你也可以用那个来表示分子。

28:08

string called a smile string, so then it becomes a one-dimensional string, and you can

它甚至更进一步，因为你基本上还可以看 language、graphs 和 molecules 的组合。

28:12

use that to represent the molecule as well.

所以你可以某种程度上，从你知道的，比如说 literature 里学习，那里有很多 text，然后每当提到一个 molecule 的时候，你可以表示，你

28:15

So you can use both LLM style models as well as a graph neural network style models

所以你可以同时使用LLM风格的模型以及graph neural network风格的模型来做这个。

28:19

for this.

所以你的基础模型可能和LLM类似，但在表示molecules的时候，你有一个非常独特的tokenizer？

28:20

So your base model might be similar to an LLM, but you've got a very unique tokenizer

这更进一步，因为你基本上还可以同时考虑语言、graphs和molecules的组合。

28:25

in the case of representing molecules?

所以你可以有点像是，你可以从，你知道的，比如说文献中学习，那里有大量文本，然后你可以表示，每当提到一个molecule的时候，你...

28:29

It goes even further because you can basically also look at combinations of language and

我认为那是一个非常令人兴奋的 proof point，会帮助整个 industry 向前推进。

28:35

graphs and molecules.

图和分子。

28:36

So you can sort of have, you can learn from, you know, let's say the literature, where

所以你可以有点，你可以从，你懂的，比如说，文献中学习，那里

28:42

there's a lot of text, and then you can repress, every time a molecule is mentioned, you

有大量的文本，然后你可以表示，每次提到一个分子时，你

28:47

can actually use the molecular representation, which is then the little graph neural network

实际上可以用 molecular representation，然后它就是那个小型的 graph neural network

28:52

that represents, so that becomes the token.

它表示出来的东西就成了 token。

28:54

You turn this into a bunch of tokens using your graph neural network.

你用 graph neural network 把它转换成一大堆 token。

28:58

And so you can, you can really start to combine these things actually.

所以实际上你真的可以开始把这些东西组合起来。

29:01

What do you see the work that you're doing yet, cusp going in the future, like what are

你觉得你现在做的这项工作，未来会怎么发展？比如

29:07

the kind of near and midterm and longer term things that you're most excited about with

你对近期、中期和长期的哪些事情最感兴趣，尤其是

29:13

that work?

这项工作？

29:14

Okay, so you're constantly expanding the tools that we need for the different class, material

好的，所以你是在不断扩展我们为不同类别材料所需的工具。

29:21

classes.

类别。

29:22

So we know how to train these tools, but every new material class has to be retrained.

所以我们知道如何训练这些工具，但每一种新的材料类别都需要重新训练。

29:27

We also going up the stack, so we're going into more coarse, grained representations,

我们也在顺着stack往上走，进入更粗粒度的表示，

29:34

like larger scale, digital twins, all the way up to modeling the actual reactor or device,

比如更大规模的digital twins，一直到建模实际的反应堆或设备，

29:41

which is material sitting.

也就是材料系统。

29:45

The most, you know, the effort that's most important for us right now is connecting

目前对我们来说最重要的，你懂的，是连接

29:50

the platform to self-driving labs, so to really be very, you know, generate large amounts

平台和self-driving labs，从而真正地，你知道，生成大量的

29:56

of data from the self-driving lab.

来自self-driving lab的数据。

30:01

That's important.

这很重要。

30:02

And I think ultimately it's very important to go through the entire process of predicting

而且我认为，最终最重要的是走完整个流程，从预测

30:08

the molecule, making it in the lab, scaling the material and then actually putting it in

molecule，在 lab 里把它做出来，对这个 material 做 scaling，然后真正把它放进

30:12

an actual device and getting a customer excited about that particular material and willing

一个实际的 device，让客户对那种特定的 material 感到兴奋，并且愿意

30:19

to pay money for it, right?

为它付钱，对吧？

30:21

But this basically means that, you know, this last part of scaling is something that

但这基本上意味着，你知道，最后的 scaling 这部分是

30:27

is still needs to be done, but it's, yeah, you know, getting there as quickly as possible

还需要去做的，但，是的，你知道，尽快做到那一步

30:32

as I think absolutely key.

我认为绝对是关键。

30:36

And so I'm really looking forward to discovering a material that is unique, it could not

所以我真的很期待能发现一种独特的材料——它不可能由AI单独完成——而且能真正做成semiconductor device或者类似的东西，或者是一种新的solar cell。

30:43

have been done by AI, and that makes it into an actual semiconductor device or something

然后我们可以说，你知道，AI确实帮了忙，确实在这种材料的设计中是非常关键的一环。

30:49

like this or a new solar cell.

我觉得那会是一个非常令人兴奋的证明点，能推动整个行业向前发展。

30:51

And we can say, you know, AI actually helped, you know, was a very important piece of the

你还在写一本书，书名特别有意思。

30:57

design of this particular material.

这本书讲的是generative AI和stochastic thermodynamics。

30:59

I think that's a very exciting proof point that will help the entire industry forward.

我觉得那是一个非常令人兴奋的 proof point，能推动整个行业向前发展。

31:04

You also have a book that you're working on that has a really interesting title.

你还有一本正在写的书，书名特别有意思。

31:10

It's talking about generative AI and stochastic thermodynamics.

它讲的是generative AI和stochastic thermodynamics。

31:16

What is that?

那是什么？

31:18

What's the connection between generative AI and thermodynamics and tell us a little bit

generative AI和thermodynamics之间有什么联系，给我们稍微讲讲

31:23

about the direction you're taking with the book?

你写这本书的方向又是什么？

31:25

Right.

对。

31:26

So maybe first a bit of a history on this.

那也许先简单讲讲这件事的背景。

31:28

So I started this almost two years ago, not even before CAUSEP.

我差不多两年前开始做这件事，甚至都没早于CAUSEP。

31:34

So I had a bit of a, sort of a law between, you know, when I stopped working for Microsoft

所以我有那么一点，算是，一个间歇期，你知道，就是在我从微软离职之后

31:39

and the start of CAUSEP, I started, so there was about six months between that.

到CAUSEP开始，中间大概有六个月吧。

31:45

And I started to, we were on, me and my wife were on vacation in Italy and see the, you

然后我开始，我们当时在，我和我妻子在意大利度假，然后你看，你

31:50

know, I cannot do absolutely nothing.

知道，

31:53

And so in the mornings, I mean, I'm just going to write a book.

我完全没法什么都不做。

31:56

Yeah.

所以到了早上，我是说，我就打算写一本书。

31:57

I had my cappuccino, you know, in the sun, and I would start a book.

对。

32:01

So wonderful, the best vacation you can have.

我端着我的卡布奇诺，你知道，在太阳底下，然后就开始写书。

32:04

And then the afternoon, we would hike through the mountains and enjoy life.

太棒了，你能拥有的最好的假期。

32:08

And so, and then I was teaching a course at, in a town called Mausenberg, which is

然后到了下午，我们就在山里徒步，享受生活。

32:17

a Dutch word, but it probably people said Musenburg, but it's close to Cape Town in South

一个荷兰词，不过人们可能说的是Musenburg，它在南非开普敦附近。

32:22

Africa.

那里有African Institute for Mathematical Sciences。

32:23

There is the African Institute for Mathematical Sciences.

所以我非常荣幸在那里待了几周，围绕我书里写的那些想法去教学。然后，然后我找了两个学生来帮我把它写完，因为你知道，开始容易，完成难，确实需要帮手。

32:27

And so I had the pleasure of spending a couple of weeks there teaching from these kind

所以我很高兴能在那里待上几周，教这些

32:32

of ideas that I was writing a book about.

我当时写书时涉及的那些想法。

32:36

And then, and then I recruited two students to help me out, actually, you know, finish it

然后，然后我找了两个学生来帮我，其实，你知道，就是完成它

32:40

because it's, you know, it's, you know, starting as easy finishing as hard, and you need some

因为，你知道，开始容易，完成难，所以你需要一些

32:45

help.

帮助。最后说到information theory，对吧，物理的核心是information。

32:46

And so it took two more years to actually finish it as a huge amount of work.

所以又花了两年时间才真正完成，工作量非常大。

32:50

But what is the book about the book is about actually the mathematics that describes modern

但这本书讲的是什么呢？这本书实际上讲的是描述现代

32:59

generative AI, including probabilistic models, including diffusion models and many other

generative AI，包括probabilistic models，包括diffusion models以及许多其他

33:05

things.

东西。

33:06

Their mathematics turns out to be equivalent to the mathematics that describes modern

它们的数学实际上等同于描述现代

33:12

non-equilibrium, statistical mechanics or thermodynamics.

non-equilibrium statistical mechanics或thermodynamics。

33:18

And how specific is that statement, is that statement, you know, everything is a PDE

这个说法具体到什么程度呢？这个说法，你知道，是不是说一切归根结底都是PDE

33:26

at the end of the day, or is that statement, you know, more kind of specific and concrete?

还是说这个说法，你知道，更偏向具体和明确？

33:33

It's surprisingly specific, and I say surprisingly because I don't think, maybe many people think

它具体得惊人，我之所以说“惊人”，是因为我不觉得——也许很多人会想它是否有类比。

33:40

if it has an analogy.

所以我认为这是一个非常好的问题，因为我觉得这要深入得多。

33:41

So that's what I think is a really good question because I think this runs much deeper.

所以我觉得，在我看来，这不仅仅是一个类比。

33:45

So I think the, it's more than an analogy in my mind.

我认为它们之间有着非常非常深的联系，而这归根结底与 information theory 有关，对吧？物理学的核心就是 information theory。

33:49

I think there is a very, very deep connection and this has to do with the fact that we are

而 machine learning 的核心也是 information theory。

33:54

talking about information theory in the end, right, that the core of physics is information

最后在说 information theory，对吧，物理学的核心就是信息。

34:01

theory.

理论。

34:02

And in the core of machine learning is information theory.

而 machine learning 的核心是 information theory。

34:06

And both of these are described.

而这两者都是这样描述的。

34:09

So if you talk about information theory with loss of information, in other words, processes

所以，如果你谈论的是带 information loss 的 information theory，换句话说，就是那些随着时间演化而丢失信息的过程。

34:15

where you lose information as you are evolving over time.

所以有一个 observer，试图去描述一个过程，但有太多 degrees of freedom 需要追踪，你根本做不到。

34:19

So there's an observer that tries to, you know, describe a process, but there's so many

所以发生的情况就是，你在丢失 information，而你需要用 probabilities 来捕捉这一点。

34:25

degrees of freedom to keep track of, you can't.

而这门数学正是 thermodynamics 和 machine learning 两者的核心。

34:28

And so what happens is that you're losing information and you need to capture that by probabilities.

而我认为这就是核心观点。

34:36

And that mathematics is the core of both thermodynamics as well as machine learning.

而这一数学是 thermodynamics 和 machine learning 两者的核心。

34:43

And I think that's the core statement.

而我认为这就是核心陈述。

34:47

And this thing, I think, you know, there's this concept of entropy, which is maybe interesting

而且这个东西，我觉得，你知道，有一个叫 entropy 的概念，可能在两个理论里都挺有意思的。

34:52

in both of these theories.

所以也许我们值得稍微聊一下这个。

34:54

And so maybe it's good to talk a little bit about that.

所以 entropy 基本上就是你对某个 particular state 所感受到的 surprise。

34:56

So entropy is basically the surprise that you find for a particular state.

你知道，你清楚那个 precise state 吗？还是你只有一个非常模糊的 probability distribution 来描述那个 particular state？

35:03

You know, do you know the precise state or is you only have a very vague probability distribution

确实如此。

35:08

for that particular state?

但在物理学里，你会把 entropy 看作一个可以对 system 进行测量的东西。

35:11

And that's true.

确实如此。

35:14

But in physics, you think of entropy as maybe a thing you can measure of the system.

但在物理学中，你会把 entropy 视为系统可测量的某个东西。

35:23

Because you have to subtract entropy from energy actually to get free energy, which is the

因为实际上你得从能量中减去熵才能得到自由能，也就是你可以自由用来在这个世界上做有用功的那部分能量。

35:27

energy you're free to use to do useful work in the world.

所以它感觉是一个真实的东西，而不是某种我对世界一无所知的东西。

35:32

And so it feels like a real thing, not like something that I don't know about the world.

而在贝叶斯统计学里，情况正好就是这样，你谈论的是主观统计，因为它描述的是我对这个世界不了解的部分。

35:39

And in a Bayesian statistics, this is exactly what it's like, you talk about a subjective

我需要，你知道，对那些我不了解的事情去描述概率。

35:47

statistics because it describes what I do not know about the world.

但在我看来，很多物理学家也这么认为，尤其是E.T. James很出名，熵在物理学中也精确地描述了你所缺失的所有信息。

35:50

And I need to, you know, describe probabilities to the things that I do not know about the world.

而我需要，你知道，对我不知道的世界上的事物赋予 probabilities。

35:56

But in my view, and the number of physicists as well, in particular, E.T. James is a famous

但在我看来，也有不少物理学家，特别是 E.T. James，是一位著名的...

36:01

one, entropy and physics also precisely describes all the information you're missing about

第一，entropy 和物理学也精确地描述了所有你缺失的信息——关于

36:08

the world.

世界。

36:09

And so there's this very deep connection between these two fields.

所以這兩個領域之間有非常深的連結。

36:13

Maybe the last thing I want to say about this, there is this famous theorem by Landauer who

關於這個，我想說的最後一件事是，Landauer 有一個著名的定理，他

36:18

basically said that if you erase a bit from a device, you must radiate K L and T, which

基本上說，如果你從一個裝置抹除一個 bit，你必須輻射出 K L 和 T，而這

36:30

is just a number, amount of energy as heat to the environment.

只是一個數字，是以熱的形式釋放到環境中的能量。

36:36

So here you can see the direct relationship.

所以你在這裡可以看到直接關係。

36:38

I'm deleting a bit, which is a piece of Shannon information from a chip.

我正在刪除一個 bit，也就是從晶片刪掉一塊 Shannon 資訊。

36:45

And heat determined at that, when you do that, you must radiate heat to the environment.

而熱量就取決於此——當你這麼做時，你必須向環境輻射熱量。

36:50

And so you can see the directness.

所以你可以看到这种直接性。

36:51

That's just kind of a conservation of information along the lines of conservation of mass, which,

这就有点像 conservation of information，类似于 conservation of mass，而，

36:56

you know, is very foundational in physics.

你知道，这在物理学中是非常基础的东西。

36:59

Yeah.

对。

37:00

In fact, the statement is the second law of thermodynamics says that the entropy must

事实上，这个说法就是 second law of thermodynamics，说的是 entropy 必须

37:06

always increase, which means, you know, you either keep the level of information the same,

总是增加，也就是说，你知道，你要么保持 information 的水平不变，

37:12

which is when the entropy stays the same, or you lose information in the process.

也就是 entropy 保持不变的时候，要么在这个过程中丢失 information。

37:17

And that's when the entropy goes up.

而那就是 entropy 增加的时候。

37:19

So of course, what happens is the information in the universe doesn't go away, but it's transferred

所以当然啦，实际情况就是，宇宙中的信息并没有消失，而是从一个你熟悉的系统转移到了heat bath里面，在那里你已经彻底失去了这些信息，完全不可恢复。

37:24

from something you know, the system to the heat bath, where you've completely lost this

所以，嗯，总之，这有点技术性，但核心思想是，information theory 是这两个理论共同的基础，基本上让它们的数学形式变得非常非常相似。

37:29

information, it unrecoverable.

而且，很多在一个领域里发展出来的工具，在另一个领域里都有完全对应的东西。

37:32

And so yeah, anyway, it's a bit technical, but the idea is that information theory is

因为很多时候，找到这些不同的——你知道——就像是两者之间的一本词典。

37:36

behind both of these theories, which basically makes the math very, very similar.

这两个理论背后的东西，这基本上让数学变得非常非常相似。

37:42

And many of the tools which have been developed in one field have their exact analog in the

而且，很多在一个领域里开发出来的工具，在另一个

37:47

other one.

领域里都有精确的对应。

37:48

Because a lot about finding these different, you know, it is kind of a dictionary between

因为很多关于寻找这些不同对应关系的事情，你知道，它有点像一本词典，在两者之间。

37:55

the two fields, right?

这两个领域，对吧？

37:57

There's something called stochastic normalizing flows in one in the machine learning and

在 machine learning 那边有一种叫 stochastic normalizing flows 的东西，

38:03

then there is escorted free energy estimation in the other field and there turn out to be

然后另一个领域里则有 escorted free energy estimation，结果它们

38:08

exactly the same methods.

其实是完全一样的方法。

38:11

And so ultimately, who is the book for and how do you see the book changing the way they

那么说到底，这本书是写给谁的？你觉得它会如何改变他们

38:17

view the world or the way they're able to do the things they do?

看待世界的方式，或者他们做事的方式？

38:22

For actually, you know, for both sides, it's written for the machine learner who is interested

实际上，你知道，对两边来说，这本书是写给那些有兴趣的 machine learner——

38:28

to learn a little bit about, you know, thermodynamics and non-equilibrium thermodynamics.

想稍微了解一下，嗯，thermodynamics 和 non-equilibrium thermodynamics 的人。

38:34

And I think, and in reverse, right, it's for also for the physicists who wants to get into

而且我觉得，反过来也一样，对吧，这也是对那些想进入 machine learning 的物理学家说的，他们可以基于自己已经知道的东西继续构建。

38:39

machine learning and build on the things they already know.

我只想指出一点：关于 fusion models 的第一篇论文，标题里其实就有 non-equilibrium thermodynamics 这个词。

38:42

And I just want to point out that the first paper that was written about the fusion models

所以那篇论文的作者其实早就知道这层联系。

38:46

actually had the word non-equilibrium thermodynamics in the title.

所以 diffusion model 本质上就是一个过程：你先有结构，然后摧毁它——这通常就是世界上发生的事情。

38:49

So the authors of that paper actually already knew about this connection.

熵不断增加，然后我们试图在时间上逆向反转这个过程，也就是从...

38:54

So it's a diffusion model really are a process by which you take structure and you destroy

所以 diffusion model 其实就是一个过程，你取出结构然后摧毁它，

39:00

it, which is typically what happens in the world.

这通常是世界上发生的事。

39:03

The entropy goes up and then we try to reverse that backward in time, which is to start with

entropy 会上升，然后我们试图在时间上把这个过程倒回去，也就是从

39:09

noise and create structure, which is our generative models.

从噪声中创造结构，这就是我们的 generative models。

39:14

And the way I think it will evolve into the future is, of course, there's, you know,

而我认为它未来会如何演变，当然，你知道的，

39:18

or the way it can be used fruitfully is, first of all, it can help people in physics to

或者说它能被有效利用的方式，首先是它可以帮助物理学领域的人

39:27

use these tools for machine learning.

使用这些工具去做 machine learning。

39:29

And let me give you one example.

我给你们举个例子。

39:31

So it's a very important problem to compute the free energy difference between, let's

所以，计算 free energy 之间的差异是个非常重要的问题，比如

39:37

say a unbound system, which is like a protein and a drug that you want to neutralize the

说一个 unbound system，就像一个 protein 和一个 drug，你想要 neutralize 这个

39:45

protein.

protein。

39:46

And so you want this drug to attach or bind to the particular protein.

所以你想要这个药物去attach或者bind到特定的protein上。

39:50

And so you want to figure out how much does it want to bind to this particular pocket

所以你要搞清楚它到底有多想bind到这个protein上的那个特定的pocket。

39:56

on the protein.

所以你需要计算unbound state和bound state之间的free energy difference，unbound state就是这两个东西分开的状态，bound state就是它们在一起的状态。

39:57

And so you need to compute the free energy difference between the unbound state where

这是一个非常重要的问题。

40:01

these two things are for a part and the bound state where they're together.

因为这是化学家通常都会想要做的事情。

40:04

That's a very important problem.

但现在你可以用现代的machine learning方法，比如diffusion models，来加速这个过程。

40:07

Because it's something that a chemist would typically want to do.

因为那是化学家通常会想做的事情。

40:10

But now you can use modern machine learning methods like diffusion models in order to accelerate

但现在你可以用现代 machine learning 方法，比如 diffusion models，来加速

40:15

those particular calculations.

那些特定的计算。

40:17

So that's a clear way in which you can start from a machine learning method and help

所以这是一个清晰的方式，你可以从一种 machine learning 方法出发，去帮助

40:22

the chemists do their calculations better.

化学家们把他们的计算做得更好。

40:25

But then the other direction is also true, right, because there is things which have been

但反过来也一样，对吧，因为有一些东西已经

40:29

developed in stochastic thermodynamics, like some like a concept called counter diabetic

在 stochastic thermodynamics 中发展出来，比如有一个概念叫 counterdiabatic

40:34

driving.

driving。

40:35

It's something that they have figured out on how to do very efficient control of particular

这是他们已经搞明白的，就是如何对特定的

40:41

physical systems.

物理系统进行非常高效的控制。

40:44

And those concepts can now be used in diffusion models in order to do a better job at generating

而这些概念现在可以用于 diffusion models，以便更好地生成图像，因为事实证明，这非常有助于降低 diffusion 过程中的统计误差或噪声。

40:51

images because it turns out that that's very helpful in bringing down sort of statistical

图像，因为事实证明这其实非常有助于降低某种 statistical

40:59

error or noise in the diffusion process.

所以这实际上可以帮助你构建更好的 diffusion models。

41:03

And so it can actually help you to build better diffusion models.

所以它确实可以帮助你构建更好的 diffusion models

41:06

So this is a cross fertilization between these two fields.

所以这是这两个领域之间的交叉融合。

41:10

And do you see on the machine learning side, the primary target of the analogy or relationship

那在 machine learning 方面，你认为这个类比或关系的主要目标是什么？

41:19

as being diffusion models, or is it more fundamental than that?

那么从 machine learning 的角度来看，你认为这种类比或关系的主要对象是 diffusion models，还是比这更根本？

41:25

It's more fundamental than that.

这比那更根本

41:27

In fact, the book goes into variation of autoencoders.

实际上，这本书深入讲解了autoencoders的各种变体。

41:32

It goes into MCMC methods, Markov chain Monte Carlo methods, which are used to sample,

它涵盖了MCMC方法，也就是Markov chain Monte Carlo方法，用来从特定分布中采样。

41:39

sort of sample from particular distributions.

它还涉及free energy estimation，以及很多不同的应用和交叉融合。

41:42

It goes into free energy estimation, it goes, there's many different applications and cross

所以我觉得作为一个领域，我们其实有点挣扎——我们在创造越来越复杂的foundation models，但对它们的理解仍然停留在表面，比如在它们的mechanics层面。

41:48

fertilizations that can happen.

可能发生的 fertilizations

41:50

So I think as a field, we struggle a little bit with, we're creating these ever more

所以我觉得作为一个领域，我们有点挣扎，我们在创造这些越来越复杂的 foundation models，但对它们的理解仍然停留在表面，比如在它们的 mechanics 方面。

41:57

complex foundation models, our understanding of them, still somewhat surface level, like

复杂 foundation models，我们对它们的理解仍然有些表面，比如

42:06

in terms of their mechanics.

就它们的机制而言。

42:08

We've thermodynamics as a field is much more mature.

Thermodynamics 作为一个领域要成熟得多。

42:14

And do you see us being able to use this relationship to better understand the models

你觉得我们能不能利用这种关系，从核心上更好地理解这些 models？

42:23

kind of at their core?

当然能。

42:25

Absolutely.

至少它提供了一个不同的视角来理解事物。

42:26

At least it gives a different perspective in trying to understand things.

所以我会说，stochastic thermodynamics 这个领域本身其实相当新，很有意思。

42:30

So I would say the field of stochastic thermodynamics actually itself quite new, interestingly.

所以它仍然是一个非常活跃的研究领域。

42:36

So there's still very actively resourge.

所以仍有一股非常活跃的复兴势头。

42:38

So thermodynamics is systems in equilibrium.

所以热力学是研究处于平衡态的系统。

42:41

Stochastic thermodynamics is systems out of equilibrium.

Stochastic thermodynamics 讲的是非平衡态的系统。

42:43

And that's actually quite new field.

这其实是个相当新的领域。

42:47

But it gives you a different way to try and understand how diffusion models work.

但它给了你一种不同的方式来试着理解 diffusion models 是如何工作的。

42:54

Because concepts like heat and work and entropy production, these things we don't use when

因为像 heat、work 和 entropy production 这些概念，我们在谈论 machine learning models 的时候

43:05

we talk about machine learning models.

是不会用到这些东西的。

43:08

But I think they give you a very new and interesting perspective on how to think about what's

但我认为它们提供了一个非常新颖有趣的视角，去思考正在发生什么，

43:13

going on and how to also improve them.

以及如何改进它们。

43:17

And so, yeah, I think in reverse, it's also true, right?

所以，是的，我觉得反过来也同样成立，对吧？

43:21

Machine learners have their own way of thinking about how to improve models.

做machine learning的人有他们自己关于如何改进模型的思考方式。

43:25

And that could also help the physicists to think about problems.

而这也可以帮助物理学家去思考问题。

43:28

And I've found both parties actually to be very interesting, interested in the other

而且我发现双方实际上都对彼此的工作非常感兴趣，也很有意思。

43:33

parties work.

所以这很好。

43:34

So that's good.

也许为了更进一步地抽象，你最近做了一个主题演讲，那个我很清楚。

43:35

Maybe to take the next step in getting even more abstract, you recently did a keynote

而且除了谈到你在CUSP的工作和那本书之外，你还提出了另一个类比。

43:42

that I clear.

这一点我说清楚了。

43:44

And in addition to talking about your work at CUSP and the book, you put forth another analogy,

而且除了谈到你在CUSP的工作和这本书之外，你还提出了另一个类比。

43:54

let's call it, or opportunity to kind of learn from the physical world, and that is

我们姑且称之为，或者说是一种从物理世界学习的机会，那就是

43:59

in looking at waves and applying that to machine learning and stochastic systems.

观察waves，并将其应用到machine learning和stochastic systems上。

44:06

Talk a little bit about that work.

稍微聊聊这方面的工作吧。

44:08

Yeah, so this is, I think, very exciting in the sense that the thing that there's two

是的，我觉得这非常令人兴奋，因为有两个

44:15

reasons why we thought we needed to think more about waves in science.

原因让我们觉得需要在科学中更多地思考waves。

44:21

And that is because first of all, waves are actually seen in the brain now because we've

首先是因为，我们现在确实能在大脑中看到waves，因为我们已经

44:27

gone from single electrode measurements to maybe hundreds or thousands of electrode measurements.

从single electrode measurements发展到了可能有几百或几千个electrode measurements。

44:33

And of course, in a single electrode, you cannot see a moving wave, a traveling wave.

当然，在single electrode上，你看不到移动的wave，也就是traveling wave。

44:36

But if you put hundreds or thousands in them, you can actually see these traveling waves.

但如果你把几百上千个放进去，你确实能看到这些 traveling waves。

44:39

And people have now observed them everywhere.

现在人们已经在各处都观察到它们了。

44:44

And then the question becomes, is this function always just a side effect or something?

然后问题就变成了，这个 function 是不是总只是一个 side effect 之类的？

44:49

That's the first reason.

这是第一个原因。

44:50

The other reason is that in neural networks, you have this phenomenon called over smoothing,

另一个原因是，在 neural networks 里，有一种现象叫 over smoothing，

44:55

which means that you start with a piece of information and you have thousands of layers.

也就是说，你从一条信息开始，后面有成千上万层。

45:00

And this information basically is exponentially suppressed as you go through it.

而这条信息在穿过这些层时，基本上会被 exponentially suppressed。

45:06

So it's the input and the output of the neural network become independent of each other.

所以这个 neural network 的 input 和 output 就变得互相独立了。

45:12

Of course, it's something that you really do not want because you want the output to say

当然，这是你绝对不想要的，因为你希望output能说出关于input的某些信息，比如这个特定input的class是什么。

45:16

something about the input, like maybe what's the class of this particular input.

但是一直以来，要确保信息从input层经过成千上万个layers传到output层，总是有点困难。

45:19

But it's always been a little hard to make sure that the information travels all the

人们用了很多技巧来实现这一点。

45:25

way from the input to the output layers through thousands of layers.

但我觉得我们在deep learning中开发的很多技巧，恰恰就是为了做到这一点。

45:29

People have used many tricks to do that.

另一个应用则是reasoning。

45:32

But I feel many of the tricks we have developed in deep learning is exactly to try to do that.

所以如果你需要对问题进行reasoning，有时候就得在memory里保留一些东西。

45:38

And in the other application is in reasoning.

而另一个应用是在推理方面。

45:41

So if you have to do reason about the problem, you have to sometimes keep things in memory.

所以如果你要推理一个问题，有时候你得把一些东西记在memory里。

45:47

Of course, we kind of write it maybe to shorter memory.

当然，我们可能会把它写到更短的 memory 里。

45:51

And then we read and write from this kind of memory.

然后我们从这种 memory 里读写。

45:53

It's a quite a more stable piece where things don't get mixed up so fast.

这是一个更稳定的部分，东西不会那么快就搞混。

45:57

So you need some kind of memory.

所以你需要某种 memory。

46:02

And basically the question becomes how do you communicate very deep into time or reason

基本上问题就变成了，你怎么和很深的时间交流，或者推理

46:11

very deep into time or how do you communicate with things that are very far away in your brain

很深的时间，或者你怎么和你大脑里离得很远的东西交流

46:16

or take information from a material that actually has very long range interactions.

或者从一种实际上有 very long-range interactions 的材料中获取信息。

46:25

And also in materials, materials use waves for long range interaction, which are called

而且在材料中，材料用波来进行 long-range interaction，这些波叫做

46:30

phonons.

46:31

Phonons are lattice vibrations and they carry information over long distances in a material.

Phonons 是 lattice vibrations，它们能在材料中长距离传递信息。

46:40

In fact, if you think about the universe, the only reason we can see very far and deep

事实上，如果你想想宇宙，我们能看得很远很深、深入宇宙的唯一原因，

46:46

into the universe is because light waves hit our instruments in our eyes.

就是 light waves 击中我们眼睛里的仪器。

46:53

And that is, again, waves that are doing the communication over these long distances.

而这又是 waves 在长距离中完成信息传递。

47:00

But it's now interesting is that in neural networks, we don't use this tool at all.

但有趣的是，在 neural networks 里，我们完全不使用这个工具。

47:03

So we create maps and the maps, you know, map numbers to new numbers.

所以我们创建 maps，而这些 maps，你知道的，把数字映射到新的数字。

47:11

And so the question becomes can we actually start using this phenomenon, this wave phenomenon

所以问题就变成了，我们能不能真正开始利用这个现象，这个 wave phenomenon。

47:16

that we see in the brain, can we also start to use it in neural networks.

我们在大脑里看到的那些 waves，能不能也开始用在 neural networks 里呢？

47:19

And so this has been the work that we've been developing and especially for, so we've

所以呢，这就是我们一直在做的工作，尤其是，我们...

47:24

basically designed neural networks that naturally create these waves, it's moved this information

基本上就是设计了能自然产生这些 waves 的 neural networks，它把这些信息移动

47:30

around in what we call channels or memory channels, or you can call them capsules where

到我们叫做 channels 或 memory channels 的地方，你也可以叫它们 capsules，在那里

47:36

this information gets sort of moved around and stays stable.

这些信息会被移动来移动去，但保持稳定。

47:41

And we've been very successful in, for instance, task that requires memory.

而且我们非常成功地处理了一些需要 memory 的任务，比如

47:46

So a task where a neural network or an RNN, basically, you know, the task is you take

就是让 neural network 或 RNN 去做，基本上，你知道，任务是拿一个

47:52

a number, you hold it in memory, and then as a random other point, another number arrives,

数字，把它记在 memory 里，然后在某个随机的时间点，另一个数字就来了。

47:56

you hold it in memory, and then you add them, and then another random point you're asked

你在记忆里存着它，然后你把它们加起来，接着又有一个随机点需要你

48:01

to produce the sum.

给出总和。

48:03

And that point could be very far into the future.

而那个点可能出现在很远的未来。

48:05

So you have to do the computation and you have to hold things in memory.

所以你必须做计算，而且要在记忆里保持这些东西。

48:09

And these models that naturally work with these waves, they can actually, they can do

而那些天然处理这些波的模型，它们实际上能做

48:13

these tasks much, much better than RNNs.

这些任务比RNN好得多，好太多了。

48:16

And I hear RNNs and long-running memory, I think, about challenges like exploding gradients

而我一听到RNN和长期记忆，就想到像是exploding gradients这样的挑战

48:24

and then I think about kind of the oscillatory nature of waves and that that has some inherent

然后我又想到波的振荡性质，那里面有某种固有的

48:32

dampening properties that allow you to kind of access memory further into the future

这种 dampening 特性让你可以某种程度上在更远的未来 access 到 memory

48:37

without, you know, these exploding gradient challenges, is that part of the mechanism

而且你知道，没有 exploding gradient 这些挑战，这是不是 mechanism

48:42

at work here?

在这里起作用的一部分？

48:43

That's a very good point.

这个观点非常好。

48:45

So if you think of this a neural network as a dynamical system, so that's basically,

所以如果你把 neural network 看作一个 dynamical system，基本上就是，

48:50

you think of the layers as time, and you're trying to get a signal from the beginning and

你把 layers 看作 time，然后你试图从开头获取一个 signal，并且

48:54

you're trying to propagate it through the layers over time.

你试图让 signal 随着 time 通过各个 layers 进行 propagate。

48:57

So a couple of things can happen, just different types of dynamical system.

所以可能会有几种情况，也就是不同类型的 dynamical system。

49:03

The first one has a stable system where basically, if you take two different inputs, they

第一个有一个稳定的系统，基本上，如果你拿两个不同的输入，它们

49:10

map to this, you know, after a while the collapse onto a same point and then from there

映射到这里，你知道，过一段时间后，它们会collapse到同一个点，然后从那里

49:14

on that one point moves forward.

开始，那个点继续向前移动。

49:16

So basically every converges onto what's called a point, a table point, and then from there

所以基本上每个都会收敛到一个所谓的点，一个stable point，然后从那里

49:23

on basically that's what propagates.

开始，基本上那就是传播下去的东西。

49:26

That's not what you want because, you know, information gets lost because everything gets

这不是你想要的，因为，你知道，信息会丢失，因为所有东西都

49:31

mapped to a point.

被映射到一个点。

49:32

Now that's an imploding gradient if you wish, that's something where the information gets

现在，如果你愿意的话，这就是一个imploding gradient，这是一种信息会

49:36

lost.

丢失了。

49:38

There's another extreme where you take two input points, I say two images that are similar,

还有另一种极端情况，你取两个 input points，比如说两张相似的图像，

49:44

and then they start to sort of wildly move around, and that's called chaos.

然后它们开始有点疯狂地移动，这就叫做 chaos。

49:49

So if it's a non-linear system, so this starts to wildly move around, and they also, you

所以如果是一个 non-linear system，它就会开始疯狂地移动，而且它们也，你

49:55

also lose the information, but for a different reason.

也会丢失信息，但原因不同。

49:59

The information is still there, but you cannot track it numerically.

信息仍然存在，但你不能 numerically 追踪它。

50:03

Your numerical precision is too small, and that's actually what we mean when we say entropy

你的 numerical precision 太小了，这实际上就是我们说 entropy

50:08

is increasing.

在增加时的意思。

50:09

That means it basically becomes a random, it's a random process, and we all know that

那基本上就变成随机了，它是一个 random process，而且我们都知道，

50:15

a mark of chain is a random process that depends on the previous step, it will lose information

Markov chain 是一个依赖于前一步的 random process，它会丢失与输入相关的信息，

50:21

about the input because we can track it.

因为我们可以追踪到它。

50:24

That's also not what you want.

那也不是你想要的。

50:26

So what you really want is sitting somewhere in the middle, somewhere that's not too stable

所以你真正想要的是处在中间的某个位置，一个不是太稳定，

50:31

and not too unstable, and that's called edge of chaos.

又不过于不稳定，这就是所谓的 edge of chaos。

50:36

And people have found that, in fact, neural networks that perform best operate at this edge

而且人们发现，事实上，表现最好的 neural networks 就是在 edge of chaos 上运行的。

50:43

of chaos, and people have also found that, unsurprisingly, the brain operates at the edge

人们还发现，毫不意外的是，大脑也运行在这个 edge 上。

50:48

of chaos.

混沌的。

50:50

And we found that if you include these waves in your neural network, you, very naturally

而且我们发现，如果你把这些waves纳入你的神经网络，你非常自然地

50:59

operate in this regime edge of chaos, and you don't have to find you in the system to

在这个edge of chaos的regime里运作，而且你不需要在系统中找到自己才能在那里。

51:05

be there.

要进入这个edge of chaos regime非常容易。

51:06

It's very easy to get to this edge of chaos regime.

然后，如果时间允许，我们可以谈谈我们是怎么做到这一点的——这是通过spontaneous symmetry breaking实现的，但我觉得这是相当技术性的讨论。

51:12

And then maybe if we have time, we can talk about how we do this, and this is done by

那么什么是spontaneous symmetry breaking呢？

51:16

spontaneous symmetry breaking, but that's a rather technical discussion, I think.

spontaneous symmetry breaking就是一个具有特定对称性的系统——比如完美的translation symmetry，就像一种气体具有完美的translation symmetry那样。

51:22

So what is spontaneous symmetry breaking?

那么什么是 spontaneous symmetry breaking？

51:24

Okay, this is a very interesting phenomenon in physics.

好的，这是物理学中一个非常有趣的现象。

51:28

So here's another example of something quite deep from physics that we can start to use

那么，这里还有一个来自物理学的相当深奥的例子，我们可以开始用它来理解 neural networks，这也是你之前问过的问题之一。

51:33

to understand neural networks, which is one of your previous questions.

所以 spontaneous symmetry breaking 就是：一个系统具有某种特定的对称性，比如完美的 translation symmetry。你可以想象气体就有完美的 translation symmetry，或者液体更合适。然后液体在温度降低时，实际上会凝结成固体，就像水变成冰一样，对吧？

51:38

And so spontaneous symmetry breaking is where a system that has a particular symmetry,

所以 spontaneous symmetry breaking 就是，一个系统在具有特定对称性时，

51:44

let's say a perfect translation symmetry, just think of a gas as perfect translation symmetry

比如说完美的 translation symmetry，就把气体想象成完美的 translation symmetry。

51:50

or a liquid better.

或者液体更好。

51:52

And then the liquid at lower temperature, the liquid will then actually condense it into

然后温度更低的那个液体，就会真的把它凝结成

52:00

as a solid, you can go from water to ice, right?

固态，你可以从水变成冰，对吧？

52:07

And if you look at ice, you know, it has a crystal structure, so it has less symmetry

你看冰，你知道，它有crystal structure，所以它的symmetry更低。

52:12

now, because it's now the continuous translation group has now been turned into a discrete translation

现在，因为原来的continuous translation group现在已经变成了一个discrete translation

52:19

group where you can only translate over the lattice spacing.

group，你只能沿着lattice spacing来平移。

52:23

And so you've actually broken the symmetry into something smaller.

所以你实际上是把symmetry破缺成了更小的东西。

52:27

And there is a very deep theory from physics that says when that happens, when a continuous

物理学里有一个非常深刻的理论，它说当这种情况发生，也就是当一个continuous

52:32

symmetry breaks into a discrete symmetry or breaks into a smaller symmetry, then you'll

symmetry破缺成discrete symmetry，或者破缺成更小的symmetry时，那么你就会有

52:39

have new wave-like modes, which can propagate without using any information or with very,

新的wave-like modes，它们传播时不依赖任何信息，或者只需要非常，

52:50

very little information.

非常少的信息。

52:51

And these are these phonons or light, or these are the waves that can travel very long distances.

而这些就是这些 phonons 或光，或者说这些是能传播非常远距离的波。

52:59

They don't decay, they don't disperse, they hold the shape, you start with a certain

它们不会 decay，不会 disperse，能够保持形状，你从一个特定的

53:04

shape, then the shape can just translate and just move.

形状开始，然后这个形状就可以直接 translate 和移动。

53:09

And so we thought that's a really good candidate for these traveling waves.

所以我们觉得那真的是 traveling waves 的很好的候选者。

53:13

And so what we did is we created the neural network with the large symmetry, we just

所以我们做的就是，我们创建了一个具有 large symmetry 的 neural network，我们只是

53:17

bake it in, it's not like rotation or translation that we typically use, we just give every neuron

把它 bake in，这不像我们通常用的 rotation 或 translation，我们只是给每个 neuron

53:25

some extra dimensions and we say there is some symmetry.

一些额外的 dimensions，然后我们说存在一些 symmetry。

53:30

And then we do something, we generate random weights, we generate random activations,

然后我们做一件事，我们生成 random weights，我们生成 random activations，

53:38

and if you make the weights large enough, the distribution from which you sample the

如果你把 weights 设得足够大，你从中 sample weights 的 distribution 也足够大，原本的 symmetry 就会被打破，这个解释起来很复杂，但总之 symmetry 被打破了。

53:44

weights large enough, the original symmetry gets broken, and that's complex to explain,

现在你就处于这个 broken symmetry phase 里，有这些 Goldstone modes，或者说这些 waves，它们可以在不消耗任何能量的情况下传播，而且确实如此。

53:50

but the symmetry gets broken.

所以当你创造出那种情况时，你就会发现它们，你会得到这些 waves，你能非常自然地看到东西在周围 oscillate。

53:52

And now you're in this broken symmetry phase where you have these goldstone modes or

然后这个想法是，你不能 train neural network 去利用这些 oscillations。

53:56

these waves which can travel without any using any energy, and they do.

这些波可以在不消耗任何能量的情况下传播，而且确实会这样。

54:01

So you find them when you create that, you get these waves where you can just see things

所以当你创造那种条件时，你得到这些波，你能看到事物

54:05

oscillate around, very naturally.

非常自然地振荡来振荡去。

54:09

And then the idea is, you cannot train the neural network to make use of these oscillations

然后这个想法是，你没法训练 neural network 去利用这些振荡，

54:15

which are basically for free, they're running there without costing any energy, and they're

基本上是免费的，它们就在那里运行，不耗费任何能量，而且非常稳定。

54:19

very stable.

所以它们也从 neural network 的开头一直加到 neural network 的末尾。

54:20

So they also add all the way from the beginning of the neural network to the end of the

而且你无法设计系统让它拥有这些 waves，并且拥有这种 edge of chaos 的行为。

54:23

neural network.

所以信息从 neural network 的 input 开始，通过所产生的这些 wave-like patterns，一路传播到 neural network 的 output。

54:25

And you cannot design the system so that it has these waves and it has this edge of chaos

你也没法设计系统，让它具有这些波和这种 edge of chaos。

54:30

behavior.

行为。

54:31

So information from the beginning, the input of the neural network propagates all the way

所以从一开始，neural network 的 input 就一路传播

54:36

to the output of the neural network through these wave-like patterns that are created

到 neural network 的 output，通过这些产生的 wave-like patterns

54:42

by this phenomenon called spontaneous symmetry breaking.

通过这个叫做 spontaneous symmetry breaking 的现象。

54:45

And so here we use something a deep result from physics as a design principle for neural

所以我们在这里使用了物理学中一个很深刻的结果，作为 neural networks 的设计原则。

54:50

networks.

嗯，我很喜欢你能玩转所有这些物理工具，把它们引入 machine learning，试图找到 correlations，以及我们能利用它们的方法。

54:51

Well, I love that you're getting to play with all of these tools from physics and bringing

我觉得后面还有很多。

54:56

them into machine learning and trying to find the correlations and ways that we can take

数学和物理学中有一个非常丰富的领域，都可以被 machine learning 利用。

55:01

advantage of them.

利用它们。

55:03

There's many more to come, I think.

我觉得还会有更多。

55:04

There's a very rich field of mathematics and physics that can all be leveraged by machine

数学和物理学中有一个非常丰富的领域，都可以被机器利用。

55:09

learning.

学习。

55:10

But it's beautiful because the cross-fertilization goes in two directions, right?

但这很美好，因为这种交叉融合是双向的，对吧？

55:14

Because with our cost platform, we are using AI tools to help the scientists discover new

因为通过我们的 cost platform，我们用 AI tools 来帮助科学家发现新的

55:23

materials and accelerate their simulation tools by using AI.

材料，并通过使用 AI 来加速他们的 simulation tools。

55:30

And so it's really a healthy cross-fertilization between these two communities, which are

所以这真的是这两个社区之间一种健康的交叉融合，而这些社区非常

55:35

very exciting.

令人兴奋。

55:36

Well, Max, it's been great catching up with you.

嗯，Max，和你聊聊近况真是太好了。

55:40

You have a lot going on, you're working on a lot of different angles.

你有很多事情在进行，在从很多不同的角度开展工作。

55:44

It's been enjoyable to learn about them, looking forward to keeping in touch.

了解它们很有趣，期待保持联系。

55:49

Well, thank you, Sam.

嗯，谢谢你，Sam。

55:50

It was a pleasure again to talk to you.

再次和你交谈很愉快。

55:52

Thanks so much.

非常感谢。

55:53

See you next time.

下次见。

55:54

See you next time.

下次见。

Why the Next AI Breakthrough May Come from Physics with Max Welling - #774

Play Queue