合成超级智能：从半导体到超导体——Periodic Labs的Liam Fedus和Ekin Dogus Cubuk

这种多模态的方式确实很有帮助。

And that kind of multimodal really helps.

而且，AI在这里真的很有用，因为人类的上下文理解和计算能力都有限。

And again, AI is really helpful here because humans have limited context and computation power as well.

所以，如果同时给人类10种不同的模态，让他们一致地分析这些数据，那确实有点难。

So, if you give humans 10 different modalities at the same time and say, analyze these all consistently, it's a bit difficult.

但对AI来说，这其实并不算特别难。

But it's actually not superintelligent for an AI to do it.

这真的很棒。

And it's great.

就像超级智能一样，它能做大量的计算。

It's like the superintelligence is it can just do so much calculations.

它能看遍所有东西。

It can look at everything.

这就回到了那个说法，就是在不确定性下做出更好的决策，现在它把来自不同仪器、不同实验、不同重复样本的信号串联起来，然后得出一个更准确的底层模型。是的。

And this just kinda gets to that phrase about, like, better decision making under uncertainty, where now it's stitching together the signal across many different instruments longitudinally across different experiments, across the replicates, and then kind of getting to this, like, underlying better model Yeah.

而不是只盯着一个仪器的一个测量结果。

Rather than just over indexing on one measurement from one instrument.

让我先说一下……

And let me just like...

你之前也提到过，数据量太大，任何合理的计算机或处理流程都装不下。

So, you started, I think, while ago also saying that there's just more data than you can fit in any reasonable computer or a process or anything.

所以，即使你有所有这些重复实验，即使你大概有十种不同的方法来测量同一个东西，有时候还是得扔掉一些数据。

So, at some point, have to throw out data even though you have all these replicates, even though you have presumably like 10 different ways to measure the same thing.

因为万一你的某个温度传感器出错了呢？

It's because what if one of your temperature things is wrong?

那什么时候该扔掉它？

So, when do you throw that out?

你怎么决定怎么做？

How do you decide what to do?

作为一个数据狂热的机器学习爱好者，我就是什么都想要，对吧？

As a data hungry ML guy, I just want everything, right?

然后我就，你知道，随便学一下所有东西。

And then I just, you know, waddle, learn everything.

我不知道自己不知道什么。

I don't know what I don't know.

然后就直接把机器扔过去

And then just throw the machine at

处理。

it.

所以我们不会随便丢弃数据。

So we don't throw in data.

我想我可能的意思是，如果上帝在观察这个实验，那么实验中的数据比人类能放进电脑里的还要多。

I think maybe what I meant was potentially, like if God was observing the experiment, there is more data in the experiment than humans can fit in a computer.

不幸的是，我们这些凡人无法观察到所有的原子。

Unfortunately, us mortals cannot observe all the atoms.

所以即使数据比我们能存储的还要多，我们实际上也无法访问那些数据。

So even though there's more data than we could store, we actually can't access that data.

比如说，我没法追踪所有原子的位置。

For example, I cannot trace all the atoms positions.

所以很不幸，我们仍然处于一个数据非常珍贵的阶段，收集到的任何数据都不会删除。

So unfortunately, we're still in a regime where any data we collect is very precious and we don't delete any of it.

也许换个角度想，跟AI联系起来，就像是在探测基础模型，比如线性探测，对吧？

Maybe way of thinking of it to connect to the AI side is like, think about probing, like linear probing of a foundation model, right?

这个模型包含了所有的数据。

This model has all of this data.

所以上帝对材料的行为有完整的了解。

So God has this entire picture of what the material is doing.

而实验就是一个小小的线性探测器，输出只有八位左右的信息。

And then the experiment is this little tiny linear probe, which has like eight eight bits or something of output.

所以你是把这个庞大的东西粗粒化到一点点信息。

And so you're taking this humongous thing and coarse graining it down to a little bit of information.

然后现在，要么是传统上人类的工作，要么是你代理的工作，从这八位信息中重建出内部实际发生的情况。

And then it's now, either traditionally a human's job or now your agent's job to reconstruct from those eight bits of information what is actually happening inside.

就像我说的，模拟情况下你没法完全看到每个参数，也没法看到每个激活值对于那个输入的情况。

Like I say, analog where you can't have full visibility over every parameter way, over every activation for that input.

你只能得到某种子集，或者某种粗粒度的特征，覆盖整个系统。

You're you're getting, like, some sort of subset on it or some, like, you know, coarse grained features across the whole thing.

但假设你可能会发现……

But suppose you you'll probably find that...

你可能会对麦克斯韦妖的论证很感兴趣，因为那种……

You probably do find Maxwell's daemon argument very interesting because that kind Can

你能背一下吗？

you recite it?

因为我不太……

Because I don't...

他之前提过这个。

He's mentioned it before.

你知道你……

Know you're

其实我忘了。

actually forgetting it.

所以，你知道，在19世纪，人们意识到一个封闭系统的熵必须增加或者保持不变，但不能减少。这就是热力学第二定律。

So, you know, in the 1800s, people realized that entropy of a system has to increase for a closed system or stay the same, but it can't go down It's the second law of thermodynamics.

然后有一个思想实验，问的是如果有一个非常小的、无所不知的恶魔，它能观察原子，每当一个原子能量低的时候，就打开一扇门让它进入另一个房间，并不断这样做来减少系统的熵，这会违反热力学第二定律。

And there was a thought experiment that was asking what would happen if there was a very tiny, old knowing demon that could kind of look at atoms, and then whenever an atom had low energy, kind of open a door and let it into this other room, and keep doing this to reduce entropy of the system, which would break second law thermodynamics.

很长一段时间，我觉得对于为什么会这样，没有一个令人满意的答案。

And for a long time, I think there wasn't like a very satisfying answer for why that was the case.

但后来在20世纪，有一篇更近期的论文，兰道尔提出了一个论点，即你必须消耗能量来删除信息。

But then there was a more recent paper in the 1900s, more recently it was, Landauer came up with the argument that you'd have to spend energy to delete information.

麦克斯韦妖基本上通过做他正在做的事情，会获得如此多的信息，以至于他必须删除一些。

And Maxwell Zimmerman would basically have access to so much information by doing this thing he's doing, that he'd have to delete some.

而删除信息必须消耗能量，这就会确保熵增加。

And deleting information, it had to spend energy, which would then make sure entropy increases.

所以即使我们还没有达到能成为麦克斯韦妖的水平，我认为你的问题非常有远见，指向一个我们存储大量数据、必须删除一些数据的未来。

So even though we're not at the level where we can be Maxwell's daemon, I think your question is like very visionary towards like a future where we store so much data, we have to delete some.

对，不，我显然不是那个意思。

Yeah, no, I wasn't obviously going to that aesthetic sense.

这让我想起空调。

It does remind me of air conditioning.

空调差不多就是这么工作的。

That's kind of how air conditioning works.

但我也觉得，问题就是为什么你不把世界上所有的传感器都买下来，然后疯狂地过度装备、过度复制一切。

But also, think it's just a question of why don't you just buy every sensor in the world and just ridiculously over instrument, over replicate everything.

我们怎么做呢？

How we do that.

好。

Okay.

好。

Okay.

行。

Alright.

我觉得这确实是个问题，对。

Think I think it's a matter of Yeah.

没错。

Exactly.

你能看到单个原子，还是说这根本做不到？

Are you getting visibility to individual atoms versus infeasible?

但我们会加入大量的遥测数据，这基本上能减少那些所谓的隐藏变量。

But, no, we add a huge amount of telemetry, and that basically reduces the number of, like, hidden variables, if you will.

就是好奇问一下，你会不会担心自己只在加州？

And just just out of curiosity, it does, like, do you worry that you're just in like California?

难道你不想在尼泊尔、澳大利亚这些地方也做吗？

Like, don't you wanna do this in like Nepal and like Australia?

对，我们可以。

Yeah, we could.

我的意思是，我们有计划开更多的实验室。

I mean, we have plans to, you know, open more labs.

因为显然重力，还有你面对的方向、你所在的位置都很重要。

Because obviously gravity and like, you know, where you face and like where you are matters.

而且专业知识也可能不同，对吧？

Also like expertise can change, right?

比如我们这里就有某些特定的专业知识。

Like here we have certain kinds of expertise.

比如说，斯坦福的物理系、工程系有某些专业知识，这对我们帮助很大，因为我们可以从那里招人。

For example, Stanford physics department, engineering department has certain kinds of expertise that really helps us because we can hire from there.

我们的研究人员可以合作。

Our researchers can collaborate.

就像你说的，地理位置可能重要，但不同城市在物理、化学方面的专业知识也会有些差异。

Like, as you said, geography can matter, but then different cities can have some differences of expertise within physics, chemistry.

我们不是有Zoom嘛。

I mean, we have Zoom.

所以，对，就是这样。

So, yeah, that.

看起来你的实验会受到你所在位置的影响。

Looks like your experiments are going to be affected by where you are.

我是说，Periodic的未来不是一个单一的实验室。

Mean, the future of Periodic is is not a single lab.

我觉得我需要像你的朋友Jeff那样，去太空，然后……

I mean, I think I need to get up, like, your buddy Jeff and, like, go to space and,

对，没错。

like Right.

我觉得有必要建立这样一个实验室网络。

I mean, I think it's it's necessary to have this sort of network of labs.

而且，就像Derek刚才指出的，每个实验都有一个最佳地点，取决于专业知识或者材料限制之类的。

And and, like, yeah, I think as Derek was just kind of pointing out, like, there's an optimal place for each of these based on expertise or, you know, materials constraint.

有很多不同的限制条件。

There's many different constraints permitting.

好的。

Okay.

所以其中一个就是……

So one of them is...

我一直很好奇的是，你们怎么处理计算工具？

So one thing I've been curious about is how do you deal with computational tools?

或者说，你们怎么把计算工具整合到决策流程里？

Or like how do you integrate computational tools into a decision making workflow?

尤其是当计算和实验不一定能对上号的时候。

Especially when, let's say, computation and experiment don't necessarily meet?

还有理论方面，就是高层次的理解。

It's also theory in terms of high level understanding.

对，这些是怎么协同工作的？

And yeah, how does that sort of work together?

我觉得这里很有帮助的一点是，有些事情在模拟里做更容易，也更准确。

So, I mean, one thing that's really helpful here is there are certain things that are easier to do in simulation and more accurate in simulation.

有些事情在实验里做更容易，也更准确。

There are certain things that are easier to do in experiment and more accurate in experiment.

一般来说，实验当然更准确，但有些测量用实验很难做。

I mean, in general, of course, experiments are more accurate, but certain measurements are very hard to make with experiment.

所以你会做一个不太好的版本，得到的数据就不准确。

So you make a bad version of it, so you get inaccurate data.

我给你举几个例子吧。

Let me just give you some examples.

其中一个就是用密度泛函理论来估算材料的形成焓，还有——

So one of them is we use density functional theory to estimate formation enthalpy of materials and-

抱歉，先停一下。

Sorry, quick pause.

你能解释一下密度泛函理论和形成熵吗？

Can you explain density functional theory and formation in entropy?

对，抱歉，是焓。

Yeah, of Sorry, enthalpy.

密度泛函理论可能是材料领域最常用的模拟方法了。

So, density functional theory is probably the most commonly used simulation method for materials.

这源自科恩-霍恩伯格定理，科恩还因此获得了诺贝尔奖。

It kind of comes from Cohen Hohenberg theorem, and Cohen got a Nobel Prize for this.

他发现，量子力学确实非常昂贵，因为每个波函数都对应一个指数级的希尔伯特空间。

When he realized that, sure, quantum mechanics is very expensive because you basically have this exponential Hilbert space for every wave function.

所以你必须解决这个非常困难的问题，尤其是当系统规模增大时。

So you have to solve this very difficult problem, especially as the system size grows.

科恩和霍恩伯格意识到，其实你不需要在指数空间里做这件事。

Cohen and Hohenenberg realized that turns out you don't need to do it at exponential space.

你只需要考虑电荷密度，至少对于量子力学系统的基态来说是这样。

All you have to do is think about the charge density, at least for the ground state of quantum mechanical systems.

实际上，这个定理很简单。

So it's actually, the theorem is pretty simple.

我觉得甚至物理系的本科生都能理解。

So even I think an undergraduate in physics can understand.

但核心思想是，量子力学系统基态的所有性质都是电荷密度的泛函，这很惊人，因为电荷密度只是一个三维对象，而希尔伯特空间和波函数却是指数级的。

But the idea is all the properties of a quantum mechanical system ground state is a functional of the charge density, which is amazing because charge density is just this three-dimensional object, whereas the Hilbert space and wave functions are exponential.

这个发现真的很有意思。

So that was a really interesting observation.

后来科恩又发表了一篇论文，我记得是和博士后沙姆一起写的，提出了科恩-沙姆波函数，为近似求解材料量子力学性质提供了一种实用的方法。

And then Kohn published another paper, I think with his postdoc, Sham, Kohn Sham wave functions, kind of had a practical way to solving some approximation of quantum mechanical properties of materials.

如今，我们实验室内部和外部都在广泛使用这种方法。

And today we use it very much both in our lab and externally.

信息焓是我们用来理解材料能量的一种测量方式。

And information enthalpy is one of the measurements we can make to understand the energy of the material.

基本上，我们想知道的是，这种材料具有多少能量。

Like basically, we want to ask, what is the energy that this material has?

因为如果这种材料的能量高于这些原子可能形成的其他材料，那么这种材料就不太可能被制造出来。

Because if that energy is higher than other materials these atoms can come into, then it's unlikely that this material will be able to be made.

所以，稳定就意味着它位于其他材料及其能量的凸包上。

So stable just means like it's on the convex hull of other materials and their energies.

这样讲清楚了吗，还是我们需要再解释一下？

Was that clear enough or should we?

我的意思是，也许可以简单概括一下。

I mean, maybe a simple high level point.

密度泛函理论是一种方法，它把原本随着电子数或系统规模呈指数级增长难度的问题，近似地降低到通常 n 的三次方的复杂度，当然会引入一些误差。

Density functional theory is a way of taking something which is exponentially hard in number of electrons or number in your system size and then reducing it approximately with some introduced error to like maybe usually an n cubed approximation.

这样一来，你就能计算那些原本根本无法计算的东西，只是需要付出一些代价。

So now you can compute quantities which are you can now compute things which are just uncomputable but at some cost.

如果能完美计算的话，你就能……

If you could do it perfectly, you would...

我们就不需要实验室了。

We wouldn't need a lab.

但你知道，我们做不到。

But, you know, we can't.

这或许可以呼应我们……

This may be a good callback to our...

实际上，是我们有史以来的第二期节目。

Actually, our second episode ever.

希瑟·库利克曾说过，她有一句名言：材料领域没有AlphaFold，不仅是在计算上，连基本事实都不存在。

Heather Kulik had said, she famously said, there is no alpha fold for materials, not just computationally, but the ground truth just doesn't exist.

就是说，我们其实并不真正了解材料晶体结构是什么样的。

Like, we don't really know what material crystal structures look like.

而DFT通常是我们最接近答案的途径。

And DFT is generally our best route towards.

对，我也这么觉得。

Yeah, I think so.

我想补充两点：DFT从定义上讲并不一定非得是近似方法。

I mean, maybe two things I would add is DFT by definition doesn't have to be an approximation.

定理表明它可以做到精确，但正如你指出的，交换关联泛函我们目前还无法获得。

The theorem shows that it But can be I think as you're pointing out, the exchange correlation functional, we don't yet have access to.

而对于基于电荷密度的模型，我们也不知道动能泛函是什么。

And then for the charge density based models, we don't know what the kinetic energy functional is.

而且即使我们有完美的DFT，我仍然认为我们需要实验室，因为即使DFT完美，我们也没法在计算机里装下10的23次方个原子。

And then even if we had perfect DFT, I still think we'd need a lab because even if we have perfect DFT, we cannot fit 10 to the 23 atoms in the computer.

你对那个特别执着，

You're very obsessed with that,

是不是？

didn't you?

没有。

No.

好吧，行。

Well, okay.

好的。

Okay.

但是随着计算能力呈指数级增长，你可以说，在立方规模的尺寸下，最终，原则上，你可以计算出一些东西。

But but with exponentially growing compute, you can argue that in in cubed scaling size, eventually, you could, in principle, calculate things.

你只需要等到

You just wait till

哦，足够长的时间。

Oh, long enough.

巨大的规模。

Massive meters.

只要立方缩放是个好的扩展方式，你就可以……

As long as inCubed is a good scaling, you can...

作为非科学家，我想提一个观察，不知道你们想不想讨论一下。

As the non scientist, I will make an observation, which I don't know if you want to throw around or anything like that.

我来自金融领域。

So, came from finance.

我们有高斯联结函数，它通过计算相关性，把一切简化成一个核心，来给信用违约互换定价。

We had the Gaussian copula, which was a way of pricing like credit default swaps by doing correlations and reducing everything into a single kernel.

这感觉在精神上也很像VAE，把所有这些压缩成一个单一的参数。

It feels also spiritually similar to the VAE in terms of condensing all these things into like a single sort of parameter.

我在想，是不是到处都有同样的技巧。

I wonder if there's the same trick everywhere.

你的听起来比我的维度更高，VAE只是一个sigma的东西，但听起来是一样的。

Yours sounds a bit more high dimensional than mine, VAE is just one sigma thing, but sounds the same.

是的。

Yes.

电荷密度文件仍然是一个相当大的文件。

The charge density file is still quite a big file.

你有点粗略，但已经足够好了。

You hand wave a bit, but like it's good enough.

这是最重要的部分。

Like this is the highest order bit.

对，没错。

Like, yeah.

我个人不知道有更好的方法来预测新材料的稳定性。

I personally don't know of a better method to predict the stability of a new material.

它肯定不完美，但肯定比我想到的其他方法要好。

It's definitely not perfect, but it's definitely better than other methods I can think of.

对，所以回到你的问题，我们使用它们的原因是某些事情模拟比实验更容易，而且它们相辅相成。

And yeah, so going back to your question, the reason we use them is for certain things easier to simulate than experiment, and they kind of complement each other well.

然后我们可以循环做这件事。

And then we can do this in a loop.

我们觉得单靠模拟永远不够。

We feel like simulations will never be enough by themselves.

但在模拟、AI和实验的循环中，我认为我们可以比以前更快地取得进展。

But in the loop of simulations, AI, and experiments, I think we can make progress much faster than before.

模拟能让你做的一件事是更快地扩展规模。

One thing that simulations let you do is scale up much more quickly.

我的意思是，因为现在计算能力更容易上线，是的。

I mean, because now, like, compute is much easier to bring online Yeah.

比引入……

Than bringing in...

建立新实验室。

Bringing on new labs.

你怎么避免……

How do you avoid...

或者说，如果你把DFT规模扩大很多，你怎么避免你的模型过度依赖这些数据，同时还能保持与现实世界的联系？

Or let's say you scale up a bunch of DFT, how do you avoid in having your models over index on that and still be grounded?

现实世界才是你真正关心的最终真相。

That the real world is ultimately the truth you care about.

如果你，你知道，你怎么……嗯，

If you you know, how do you Well,

我们也会扩大实验室的规模。

we scale up the lab as well.

对，你扩大实验室的规模。

Yeah, you scale up the lab.

但我可以想象，凭借现代计算能力，把DFT扩展到数百万甚至数亿次。

But I could I could imagine scaling up DFT to millions, hundreds of millions given modern compute.

而你的实验室，我猜每天也就几百到几千次。

Whereas your lab, I would imagine your hundreds, thousands a day.

我不确定。

I don't know.

就像，还是有些东西挺……

Like, something still like fairly...

即使在非常高通量的情况下，比如你知道的，你仍然会受到比较率的限制，对吧？

Even in a really high throughput case, like, you know, you're still fairly limited by comparison rate?

我的意思是，一方面是人类，但语言模型其实也挺清楚DFT的局限性的。

I mean, one thing is the humans, but also the LMs are kind of pretty aware of the limitations of DFT.

比如说，即使我们能像你说的那样，用DFT做一亿次试验，我们也永远无法真正尝试它们的微观结构。

For example, even if we can do, as you said, 100,000,000 trials with DFT, we could never actually try their microstructure.

微观结构是指DFT通常模拟完美晶体，但晶体实际上往往并不完美。

And microstructure is this idea that DFT usually simulates the perfect crystal, but crystals usually actually aren't perfect.

而它们的微观结构，也就是在某个尺度上的结构，真的会影响材料的性质。

And their microstructure, which is the structure at a order, can really affect the properties.

当然还有其他问题，比如像超导温度这样的性质，DFT很难模拟。

Other issues, of course, like some properties like superconducting temperature cannot be easily simulated by DFT.

再说，即使你用DFT做一亿次试验，也还是不够。

Again, like even if you do a 100,000,000 trials with DFT, it won't be enough.

所以我们得依靠经验法则、实验等等。

So we have to rely on heuristics, experiments, etcetera.

所以从所有这些计算到实验室里真正执行的，中间有一个巨大的筛选过程。

So there's a huge filter from all those calculations to actually what gets executed in the lab.

那是人工介入，还是大语言模型，还是两者混合？

Is that human mediated or LLM or a mix?

可以是混合的。

It can be a mix.

比如，长期以来，材料项目——那个最好的开源DFT数据库——就是用实验数据来校准DFT的。

Like, for example, for a long time materials project, which was the best open source DFT database, used experimental calibration for DFT.

所以，他们确实会用实验数据来校准DFT。

So, would actually use experimental data to calibrate DFT.

我们内部也是这么做的，每当我们对某个化学体系感兴趣，就会做实验、做模拟，然后进行校准。

We also do this internally here, whenever there's a certain chemical system we're interested in, you get experiments, you get simulations, and you can calibrate them.

这是另一回事。

This is another thing.

我跟很多学生物的朋友聊过，他们总是很困惑，你说材料，你不能只做一个XRD结构就能知道实际的结构是什么。

I've talked to a lot of biologist friends and they are always very confused when you say materials, you can't just do an XRD structure and actually know what the structure is.

因为在生物学领域，你可以看一个蛋白质的晶体结构，通常差不多能达到亚八...

Because if you're in the biology world, you can look at a crystal structure of a protein and typically more or less with like sub eight...

嗯，就是大约埃级别的精度重建。

Well, like order angstrom accuracy reconstructed.

材料的XRD，你知道，X射线衍射，和生物学领域的有什么区别？

What is the difference between the materials XRD, you know, X-ray diffraction.

衍射，抱歉。

Diffraction, sorry.

材料的实验XRD和生物学领域的相比，有什么区别？

Experimental XRD for materials versus, let's say, the biology world.

比如，你丢失了什么信息，为什么这有点像有损投影？

Like, what information do you lose and why is this sort of like a lossy projection?

回答这个问题时很难不显得疯狂，但你不觉得有机化学和生物学几乎源于更低的VC维度，也就是更低的柯尔莫哥洛夫复杂度模型吗？

It's hard not to sound like crazy when answering this question, but do you not feel like organic chemistry and biology almost stem from a lower VC dimension, like lower Kolmogorov complexity model?

你几乎可以把它表示成一维序列，而且几乎可以编译。

Like you can almost represent as a 1D sequence and it can almost be compiled.

所以在复杂度的意义上，它更简单。

So there's a sense in which it's simpler in the convoluted complexity sense.

一维？

1D?

我

I

意思是，DNA就是——对，

mean, DNA is just- Yeah,

或者RNA。

Or RNA.

我觉得大多数人会认为它更像二维，但是——

I think most people would argue it's more two d, but-

好吧，我不是生物学家。

Okay, I'm not a biologist.

我相信他们是对的。

I'm sure they're right.

但在无机化学里，虚拟世界真是疯狂。

But with inorganic chemistry, virtual is crazy.

它显然不是来自这么低量子复杂度的模型，因为三维无机晶体可以承载金属、绝缘体、超导体、钻石。

It's clearly not coming from such a low quantum complexity model because that three d inorganic crystals can host metals, insulators, superconductors, diamond.

这真是非常不一样，太疯狂了。

It's like really crazy different.

比如人们试过，但我们永远找不到一维表示。

Like people tried, but we could never find a 1D representation.

很难为三维无机晶体找到一个简单的字符串。

That is very hard to find a smile string for three d inorganic crystals.

原子之间的相互作用方式也很奇怪，比如共价键、离子键、金属键。

And the way the atoms interact with each other is quite strange, like covalent bonds, ionic bonds, metallic bonds.

但确实，这非常有趣。

But yeah, this is very interesting.

我有时候在想，你有没有研究过固态电解质的问题？

I wonder sometimes like, have you ever looked into issues with solid electrolytes?

就是为什么人们不能用固体来替代电池里的液体电解质？

Like why people cannot replace liquid electrolytes in batteries with solids?

因为会有那些枝晶形成。

There are all these dendrites that form.

基本上，锂会形成这些结构，渗入固态电解质并使其开裂。

Like basically the lithium will form these structures that go into solid electrolyte and crack it.

你看到这个，会觉得这跟生物学相比太原始了。

And you look at that and you think this is so primitive compared to biology.

在生物学里，这些系统能完成如此复杂的纳米技术进程，而我们连让两个界面协同工作都做不到，因为锂会破坏它。

Like in biology, these systems can do such complex nanotechnological progress, whereas we can't even like make two interfaces work with each other because lithium will break it.

对，但我想无机材料可能非常耐高温。

Yeah, but then I guess inorganic stuff can be very resistant to high temperature.

你可以造航天飞机，可以用硅实现摩尔定律的计算。

You can make space shuttles, you can make silicon compute Moore's law.

所以，这确实是个权衡。

So yeah, it's a trade off.

对，我们和其他几位嘉宾也聊过这个话题。

Yeah, we've talked about this with several other guests.

我觉得在我心里一个很大的区别是，生物学里我们有一套工具，基本上可以借用数百万年进化的成果。

I think one of the big differences in my mind is that biology, we have a toolkit that, you know, that you can basically borrow like from millions of years of evolution.

进化已经给了我们所有工具，我们只需要复制它们就行。

That has already given us all the tools and we can just reproduce those.

但进化也限制了蛋白质等的VC维度，使其保持较低。

But also, evolution also constrains the VC dimension to be fairly low for proteins and so on.

像细胞这样大规模系统的涌现行为可能要复杂得多。

The emergent behavior of large scale systems like cells could be much more complicated.

还有意识。

And consciousness.

但也许更具体地说，如果你只看X射线，你有一个完美晶体，用XRD照射它，你会得到光谱。

But maybe even just like more specifically for like if you just look at an X-ray, you had a perfect crystal and you shoot an XRD at it and you get the, you know, you get the spectrum.

那其实还是不能唯一确定晶体结构，对吧？

That still doesn't actually tell you uniquely what the crystal structure is, right?

对。

No.

这就是那个……

And that's something that...

什么？

What?

那只是平均行为。

It's just average behavior.

比如说，歧化反应就是个很复杂的事情，对吧？

Like, for example, disproportionation is a very complicated thing, right?

有可能你有一个完美的晶体，平均来看它的X射线衍射看起来是完美的。

Where it's possible that you're the perfect crystal and on average it looks like it's perfect XRD.

但实际上，可能有一些其他的模式，比如原子稍微向左或向右偏移，但平均来看它们还是在中间。

But in reality, there's some other pattern going on where like atoms slightly shift to the right or left, but on average they're in the middle.

是啊，这日子不好过。

Yeah, it's a difficult life.

但蛋白质就不一样了，你基本知道背后有个遵循一定规则的蛋白质结构。

Whereas with proteins, you basically know there's some sort of underlying protein which follows some rules.

你可以把正向模型和

You can combine a forward model with

然后你就可以慢慢朝着单晶X射线衍射的方向推进，那会很有帮助。

And then you can kind of move and move towards like single crystal XRD and that can help.

我们刚开始做科学播客的时候，到处都是人工智能和科学、人工智能和数学、人工智能和物理这些话题。

When we started the Science Pod, there were all these elements of AI and science, AI and math, AI and physics and all these things.

当时就像在比谁更厉害，哪个更难。

There was this of pissing contest, like which is harder?

我觉得从科学角度来说，可以论证材料科学是最难的。

And I think you can make the scientific argument that material science is the hardest.

其实我们选这个不一定是因为它难。

Well, we didn't pick it necessarily for a hard dose.

其实在很多方面，材料科学反而更容易。

Mean, there's actually a lot of areas where it is easier.

我们有很好的模拟工具，能模拟很多种类的材料，但生物学里，要模拟一个细胞、一个器官或者整个生物体，难度非常大。

I mean, we have these great simulators for large classes of materials, whereas in biology, it's incredibly difficult to model a cell or an organ or a full organism.

所以我觉得这是一个巨大的优势。

So I think that's a huge advantage.

有意思的是，在最小的尺度上，它是一维的还是二维的？

Isn't that interesting that at the smallest level, is one dimensional or two dimensional?

然后它只是有很多个数量级的尺度和结构，这实际上带来了复杂性。

And then it just has many orders, more scale and structure, and that actually introduces the complexity.

这挺奇怪的。

That's weird.

而材料的复杂性可能更多在于微观结构。

Whereas maybe complexity in materials is more the microstructure.

其实材料在不同长度尺度上也有不同的复杂性。

I mean, you still get like different complexity at the different length scales for materials as well.

所以，微观结构以及更宏观的层面。

So, microstructure and beyond.

就像是一个简短的，类似清口的小插曲。

Just sort of like a brief, like sort of palette cleanser.

从科幻的角度来说，假设你可以发明任何你想要的材料。

In terms of science fiction, let's say you can invent whatever materials you want.

什么更有价值？

What is more valuable?

我这里有一个清单。

I have a list here.

显然有室温超导体，但你也提到了电池。

There's obviously room temperature superintonductors, but also, you mentioned batteries.

我一直觉得电池很重要。

I've always thought batteries.

只要比锂离子电池更好，你就赢了。

Just like do better than lithium ion, you're good.

那个东西放了一百年都没动。

That stood around for like a hundred years.

最后一个是碳纳米管，主要是用来建太空电梯的。

And then the last one is carbon nanotubes, mostly for the space elevators.

我不知道是不是……

I don't know if like...

我脑子里想到的就是这三个。

Those are the three that come to mind.

在你们那个领域，有没有大家经常聊的、特别向往的材料？

Are there any that like people talk about in your world that like are the dream?

当然有，超导体、磁体，如果能减少对稀土或者像钴这样难获取的过渡金属的依赖，比如在电池里少用钴，那就太好了。

Definitely, superconductors, magnets, if you can lower the dependence on rare earth or like transition metals that are hard to source, like cobalt, reducing cobalt in batteries.

我觉得有一个非常根本重要的事情，就是如果能让计算能耗接近兰道尔极限就好了。

One of the things I think is very fundamentally important is if we can get close to Landauer limit for compute energy efficiency.

你能解释一下兰道尔极限是什么吗？

Can you explain Landauer limit?

所以你看我们每次计算或每个FLOP消耗的能量，它遵循了摩尔定律式的规律，效率呈指数级提升。

So if you look at how much energy we spend per FLOP or compute, it has followed the Moore's law type behavior where it's gotten exponentially more efficient.

所以我们每个FLOP或每次计算消耗的能量呈指数级减少。

So we've been spending exponentially less energy per FLOP or compute.

我是说，我最近没查过。

I mean, I haven't checked recently.

十年前是这样。

This was the case ten years ago.

对，那很棒。

Yeah, that's great.

我记得我第一次了解它的时候，看了看那个，我们当时离那个极限还差多少个数量级来着，10个还是20个？

I remember when I first learned about it, looked at that, and we were what, like 10, 20 orders of magnitude away from it or something?

接近极限了。

Down the limit.

还有最后一点。

And last

上次我看的时候，已经……

time I looked, it's...

我们离得很近了，挺奇怪的。

We're weirdly close.

大概就差个三四个数量级吧，我也说不准。

Like, we're within a factor of, I don't know, like three or four orders of magnitude away.

我觉得实际上就是指数级的，对。

Think it's it's actually like guess, exponential Yeah.

这么多年一直在指数增长。

Expertual over many years.

挺厉害的。

Pretty impressive.

但是在你这一辈子里，

But like in your lifetime,

你确实看到了，对。

you've seen Yeah.

自从我发现这个方法，大概十年了吧，对。

Since I discovered this way, what ten or Yeah.

十五年了，你知道的。

Fifteen years know.

我们必须要做的一件事就是散热，因为计算过程中会产生热量，这是不可避免的，但之后我们必须把它散掉。

And one of the things we had to do is dissipate heat because like as these computations are happening, it's producing heat, which has to happen, but then we have to dissipate it.

所以，如果我们能接近兰道尔极限，让计算效率尽可能高，这对人类来说太棒了。因为这意味着我们做计算——这可能是人类最基本的活动之一——在能量方面尽可能高效，而能量是宇宙中真正的货币。

So like, if we can get close to the Landauer limit for how efficiently we do computation, that's amazing for humanity, Because that means now we're doing computation, which is probably one of the most fundamental things we do as humanity, but as efficiently as possible from energy perspective, which is a real currency in the universe.

那量子计算会影响这个吗？

That get affected by quantum computing?

我就随便提一下。

I'm just going to throw it out there.

理论上，大规模并行计算，简直让人尴尬。

Theoretically, massively embarrassingly parallel compute.

所以，这肯定受可逆计算的影响。

So, it definitely gets affected by reversible computing.

但说实话，我不是专家，我不太懂。

But honestly, I'm not an expert, I don't understand.

所以，如果你现在能做可逆计算，你其实不需要消耗能量来进行计算，因为你并没有真正删除任何信息，它是可逆的。

So, if you can do reversible computing now, you don't actually have to spend energy to do computation because you're not actually deleting any information, it's reversible.

但我不确定这行不行，不知道。

But I don't know if that works, don't know.

量子计算，我想它肯定也得遵守这些定律，因为它们就像……它们必须遵守汤姆……

Quantum computing, I assume has to obey these laws somehow because they so like, they have to obey Tomb

……内茨定律。

of Nets.

它是幺正的。

It's unitary.

它是可逆的。

It's reversible.

我最难忘的对话之一是和埃拉特·吉尔，他其实对量子计算非常怀疑。

Most want to memorable conversations with Eilat Gil was like, he actually was a very skeptical person about quantum computing.

他说，就算你今天有了量子计算机，也没有实际应用。

I He's like, Even if you had it today, there's no applications.

我当时就想，哇。

I'm like, Woah.

比如它能破解RSA加密。

Like it breaks security, RSA.

就这些了。

Like that's it.

我同意这个观点。

I agree with that.

比如，人们常说，如果你今天有一台量子计算机，你就能模拟很多东西，效果会好得多。

Like, Putin, people talk about how if you're a quantum computer today, you could simulate things so much better.

我就问他们，好，就算这样。

And I asked them, okay, let's accept that.

你想模拟什么呢？

What would you simulate?

你还是在模拟完美晶体。

You're still simulating perfect crystal.

你还是会遇到DFT的那些问题，如果DFT的预测能力是完美的话。

Like you still have the issues that DFT has if DFT was perfect in its prediction ability.

所以我觉得人们确实有些忽视了一些东西，因为量子计算太令人兴奋了。

So I do think people are kind of glossing over some things because quantum computing is so exciting.

比如，如果我们能在量子逻辑空间而不是经典逻辑中计算，这太令人兴奋了，以至于人们忽略了它真正造出来之后能做什么。

Like if we can compute in a quantum logic space instead of classical logic, it's just so exciting that people are glossing over what it would actually do when it's made.

也许这也没关系，因为一旦它造出来，可能会做出我们想象不到的惊人事情。

Maybe that's fine because maybe once it's made, it will do amazing things we can't even imagine.

好。

Good.

好的。

Okay.

我本来想转到自动化实验室那边，就是你已经闭环了你的实验。

I was going to move to the automating the lab side where you've closed the loop on your experimentation.

你可以谈谈机械臂，基本上就是你在实验室做的所有事情。

You can talk about robotic arms, talk about basically it's just everything you've done here.

我最喜欢你说的一句话是，你实验室里的每台设备都会有140的智商。

My favorite quote from you was that every piece of equipment in your lab is going to have 140 IQ.

好的，那是什么意思？

Okay, what does that mean?

早期，当我们扩大规模时，我们意识到操作机器本身就成了巨大的瓶颈。

In the early days, as we were scaling things up, we realized that we had huge bottlenecks imposed just by operating machinery.

所以，在机器操作过程中，技术人员和科学家们会寻找特定的形态，试图确认我们到底在制造什么，看看是否符合我们的预期。

So, for once in the machinery, we would have technicians and scientists looking for, particular morphology and trying to see, okay, what actually were we making, see if this is consistent with our intentions.

随着实验室的扩大，我们很快就遇到了这些瓶颈，机器跟不上进度。

And we really quickly, as we scale up the lab, became into these bottlenecks where it just wasn't keeping up.

于是我们开始用编程的方法来采集扫描电镜的数据。

So we started doing some programmatic approaches to capturing the data on SEM.

这个简单的程序，就是取一个视野，拍照，放大，再拍照，它提供的数据其实没什么用。

And the data that was being surfaced from this like very simple program where it sort of takes a field of view, captures, zooms in, captures, was just not all that useful.

它有点太笨了，达不到我们真正想要的精度。

It was a little too dumb for what we really wanted to to cut.

所以那时候，我们确实有必要直接把AI系统装到机器上，开始控制这些设备。

And so at that point, it was then really pertinent to build AI systems directly onto the machines to start controlling these things.

现在它们完全了解我们想要达到的目标了。

And now they have the full context as to what we're trying to achieve.

那么实验的意图是什么？

So what was the intent of the experiment?

我们想合成什么？

What were we trying to synthesize?

还有哪些其他实验证据？

What were some other experimental evidence?

它实际上是在机器内部观察并捕捉那些数据。

And it's actually looking in the machine and capturing that data.

现在这真的很有价值，因为你知道，我们一直在讨论那些隐藏变量，以及无法捕捉所有信息的问题。

And this is really valuable now because, you know, like we've been talking about these hidden variables and not being able to capture everything.

但如果你能在实验时进行更智能的数据采集，那么你为未来AI系统和计算预测准备的数据就会好得多。

But if you can do more intelligent data capture at the time of that experiment, your data for future AI systems and future computational predictions is that much better.

你会得到更丰富的数据集。

You'll have just a richer set of data.

所以这就是我们所说的，实验室里的每个环节都必须非常智能，才能让数据发挥最大价值。

And so that's sort of what we mean by just, you know, every everything on the lab has to be incredibly intelligent to just make the data as useful as possible.

这就是我们做这件事的动机和灵感来源。

So that was sort of the motivation and inspiration behind that.

目前，在让每个设备都具备智能方面，最先进的技术是什么样的？

What's the state of the art there in terms of, like, putting intelligence on every device?

我觉得一个有趣的方面是延迟问题。对。

Well, I think one interesting aspect is what's the latency Exactly.

控制那台仪器的延迟。

Of controlling that instrument.

因为云端的延迟是一回事，但设备端的计算能力又是另一回事，这

Because there's cloud latency, but then there's also device compute, which

没错。

That that's right.

然后还有那种

And then there's sort of

跟不同物理过程相关的时间尺度。

a timescale associated with different, like, physical processes.

如果调用API之类的太慢，推理量、token数或者工具调用次数跟实际过程的延迟不匹配，那这个方法就行不通。

And if calling out to an API or something is simply too slow, so the amount of reasoning or tokens or tool calls is just not matched to the latency of that actual process, then that that's infeasible.

好的。

Okay.

这挺好的。

That is nice.

听起来像自动驾驶汽车。

It sounds like self driving cars.

很快。

Real quick.

我其实挺惊讶的，因为我觉得推理的延迟相比这些要跑好几个小时的事情来说，应该算小的。

I'm actually surprised to hear that because the latency, I'd imagine, for reasoning seems small compared to a lot of these things take hours to run.

或者也许你的实验快得多，是高通量之类的。

Or maybe your experiments are much faster, high throughput or something.

我确实挺惊讶的。

I'm just surprised to hear that actually.

所以最终，实验确实要花好几个小时甚至几天，但有些特定步骤可能对延迟不敏感。

So ultimately, the experiments do take many hours, days, but there could be particular steps where latency wouldn't matter.

你可能想要更精细的控制。

You might want to have finer control.

另一个问题是成本，因为同样的原因，这些模型分析数据可能很慢，同时也非常昂贵。

Another thing is how expensive it is, because the same reason some of these models can be really slow in analyzing data also makes them very expensive.

最后还有人类患者。

And then finally, the human patients.

如果一个人想分析某个XRD图谱，要等两个小时才能得到好结果，我觉得这和等两分钟就能拿到结果是非常不同的。

If a human wants to analysis over certain XRD pattern, and they have to wait two hours before they get a good result, that's very different, I think, than if they can get it in like two minutes.

哦，好的。

Oh, okay.

然后推理过程，你不是只调用一次那个模型，对吧？

And then the reasoning, you're not doing like a single call to that model, right?

就是说你可能需要生成很多很多的token才能得到这些模式。

Like you're doing potentially many, many tokens to get to these patterns.

就像在循环中运行模拟。

Like run simulations in the

循环。

loop.

没错。

Exactly.

深度研究在

Deep research in the

这不是一次简单的工具调用。

It's not a single tool call.

这不是单一的推理线程。

It's not a single inference thread.

那么延迟可能会大幅增加。

Then the latency can blow up quite a bit.

我明白你的意思。

I see what you mean.

你的延迟中有多少是工具调用造成的？

How much of your latency is tool calls?

我猜工具调用主要是像DFT或计算类的，对吧？

And I assume tool calls is like largely like DFT or computational Yeah.

这些技术占了多少延迟，而真正的推理又占了多少？

Techniques and So how much of your latency is derived from those versus like actual reasoning?

这真的取决于具体过程。

It's really process dependent.

好的。

Okay.

还有一件事我想到的是，从小事做起，逐步做大。我猜一般的倾向或者典型的开发方式是渐进的，一开始都是人工操作，然后你找到自动化的方法，再慢慢扩展。

And then the other thing I think about is also, I guess, building up from small things to bigger things, where I assume the general temptation or the typical development is incremental, where everything is human operated, then and you find ways in which to automate it and then you sort of build out from there.

我担心有时候人们就是这样发展事物的，但那是局部最优解。

I worry that sometimes that that is the way that people evolve things, but that's a local minima, optimal.

就像短视的操作者。

Like short horizon operators.

对。

Exactly.

实际上你应该弄一个机器人，直接把它放进去。

When actually you should get a humanoid and just put them in there.

我觉得我们认为解决人形机器人问题实际上会更慢地达到我们的一些目标。

I think we're of the opinion that solving humanoids would actually be slower to kind of get into some of our goals.

只是确认一下。

Just checking.

再说一次，这很大程度上就是，你每天都在做这件事。

Again, a lot of this is just like, you do this every day.

我们看到了这一点，但我不清楚实际情况。

We see this, but I don't know the reality of the situation.

很多人都在炒作人形机器人。

A lot of people are hyping humanoids.

我觉得一个可能的做法是，通过人和自动化结合，你能很快找出实验流程中的瓶颈，然后一个一个地解决这些瓶颈。

I mean, I think maybe one process is by having a mix of humans and automation, you can identify really quickly what are some of the bottlenecks in the experimental process, and you start alleviating those bottlenecks one at a time.

比如说，如果有一些非常精细的操作任务，人类很擅长，但要花几个月才能实现自动化，那可能就不值得花太多时间在上面。

And so for example, if there's some really tricky dexterity task that humans are excellent at that you'd have to spend months automating machine, maybe don't spend a ton of time there.

也许更有前景的是那些非常常规、占用科学家和技术人员大量时间，但又容易自动化的任务。

And maybe like a more promising thing is something very routine that takes up a huge amount of scientists and technicians' time, and it's easy to automate.

所以我觉得这里面有一种务实的态度。

So I think there's like this pragmatism to it.

但说到底，我们从实验室想要的是海量的数据、高质量的数据、多样化的数据，这些才是我们的目标。

But I mean, ultimately, we want from the lab is a huge quantity of data, high quality data, diverse data, and those are our goals.

而完全自动化并不是我们的目标。

And full autonomy is a non goal.

我们使用自动化，就是为了实现数据方面的目标。

It's sort of in the service that we use automation in the service of achieving the goals on the data.

不过，我觉得自动化的另一个方面是，机器人可能会……

But also, I mean, I think another aspect to automation is robots will just be...

在实验室里犯更少的错误。

Make fewer mistakes potentially in the lab.

而且，AI系统能全面查看我们所有的数据，这真的很有帮助，因为有时候数据里出现异常，它就能发现。

And it's been really helpful having AI systems with a full view over all of our data because sometimes if there's like a permutation in data, then it can identify it.

比如，我们有一个步骤曾经出现循环错误，因为有一台机器装载不正确，导致模式不一致。

So for example, one of our steps at one point had a cyclic error because one of the machines was loaded incorrectly and the patterns were inconsistent.

所以AI在读取这些数据时就说，根据运行情况，这不符合预期。

So the AI was reading through these things and says, well, given what was run, this is not expected.

它纵向查看这一组数据，然后发现，如果我做这种循环置换并反转，一切就一致了。

And it's looking at this basket of data longitudinally and then it realized if I do this like cyclic permutation and reverse it, everything is consistent.

然后我们就能回到物理设施，发现装载过程中出了错。

And then we're able to go back to the physical infrastructure and understand that a mistake had been made in the loading.

我觉得管理数据质量是在物理世界做AI的基础。

And I think like managing data quality is so foundational to doing AI in the physical world.

而且，你知道，我们正在把这些东西融入进去，这样可以改进操作流程。

And, you know, these are the types of things that we're building in, and so you can improve the operating process.

但那时候，你会觉得，好吧，这又是一个自动化的好机会。

But at that point, you're like, okay, this is another great opportunity for automation.

我们怎么才能让它变得如此可靠、如此持久，以至于再也不犯那种错误呢？

How do we make this just so reliable, so durable that we never have those types of mistakes again?

是的，我觉得降低实验数据的噪声底很重要。

Yeah, I mean, reducing the noise floor experimental data I think is very important.

文献中的大多数实验数据噪声底都很高，通常比DFT的精度还差。

Most experimental data in literature have such high noise floor that it's actually usually worse than DFT accuracy.

因为不同的人做实验，不同的实验室，世界不同的地方，不同的时间，都会给结果带来很多波动。

Because when different people do experiments, different labs, different parts of the world, different times, it really introduces so much fluctuation to the result.

所以，我们希望标准化这些流程，最大的好处之一就是降低噪声底。

So, we are hoping that by standardizing these workflows, one of the biggest benefits will be that the noise floor will be lowered.

你一直挺公开地讲你怎么用开源模型，然后针对这类任务做微调。

You've been pretty public about how you use open source models and fine tune them for stuff like this.

这样更好吗？

Is it better?

基本上就是……

Is it basically...

我能想象一种情况，就是让那些能力弱一点的模型专门做某一项任务，然后针对那个任务去优化。

I can imagine a situation where you have maybe the dumber models do one task and then you optimize for that task.

然后呢，再有一个全能型的前沿模型，负责监督所有事情。

And then like a generalist frontier model that, like, supervises everything.

这个理解方式对吗？

Is that a good mental model to have?

还有没有我想不到的更多层次？

Are there more stages to this that I can't think about?

其实我们基本上就是混合使用开源模型和闭源模型。

I mean, we we basically use a mix of open source models and closed source models.

你知道，我们在模拟方面是巨大的受益者，构建了机器学习栈和模拟栈。

You know, we're huge beneficiaries of this from our simulation side, building up the ML stack, the simulation stack.

我们不需要在这方面施加压力。

We don't need to push on on that access.

但在很多领域，延迟太高，成本也太高。

But there's a lot of areas where the latency is too high, the cost is too high.

但在某些情况下，通过获取这些数据，我们实际上可以超越这些系统在最高推理努力下所能达到的边界。

In But some cases, can actually, you know, by having access to this data, we can actually push beyond the frontier of what these systems can do under even some like the highest reasoning efforts.

所以这不仅仅关乎成本或帕累托效率，在某些情况下，当你拥有别人没有的数据时，你可以超越它。

And so it's not necessarily just about cost or, Pareto efficiency, but in some cases you can push beyond it when you have access to data that no one else has.

一种描述方式是从计算效率的角度来看。

One, way to characterize this is in terms of compute efficiencies.

因此，通过获取这些数据，你可以比一些前沿模型更加计算高效。

So by having access to this data, you can be that much more compute efficient compared to some of the frontier models.

而且这已经能够...

And that's been able...

这在我们项目里真的起到了关键作用。

Been like a really instrumental thing in our program.

我，我真是……

I'm I'm so...

数据质量或者数据获取，作为算力的一种权衡，我觉得这里面存在某种汇率，用钱换算力，用钱换数据，人们……

Data quality or data access as a trade off of compute, I think there's some amount of exchange rate of dollars for compute, dollars for data that people...

我觉得现在天平可能正在向数据倾斜。

I feel like the pendulum might be swinging now towards data.

我的意思是，你显然会同意这一点。

I mean, obviously you would agree with that.

是的，确实。

I mean, yes.

特别是说到实验数据的噪声底限，我们需要高质量的数据。

Especially talking about the noise floor of experimental data, we need to have high quality data.

如果你把大量算力扔给一堆噪声，是不会有好结果出现的。

If you throw a huge amount of compute against a bunch of noise, you're not going to have a good thing emerge.

而且这对你来说更难，因为你主动追求稀疏性。

And it's harder for you because you actively try to be sparse.

你实际上是在尝试得到无效结果，并从中学习。

You actually try to get null results and learn from that.

没错。

That's right.

是的，我在其他几个场合也谈过无效结果。

Yeah, talked about null results in several other venues.

比如，对于材料来说，无效结果是什么样的？

Like, what does a null result look like for materials?

以及如何有效利用这些结果，尤其是考虑到我认为你的整体成功率通常可能相当低？

And how do you use that effectively, especially when I think your overall signal is probably quite sparse usually in terms of success?

我

I

的意思是，无效结果可能是我们原本打算制造某种结构，但所有证据都表明我们没有制造出那种结构。

mean, a null result could be we intended to produce some structure and then all of the evidence points to us not producing that structure.

你有没有故意不去生成某种结构的情况？

You ever intend to not produce a structure?

当然有。

Of course.

绝对有。

Absolutely.

好的。

Okay.

你会做阴性对照，绝对会。

You do the negative controls Absolutely.

好的。

Okay.

比如那些会毁掉你性能的杂质相。

Like the impurity phases that will kill your property.

好的。

Okay.

有些是有毒的。

Some be toxic.

没错。

Exactly.

百分之百。

100%.

好的。

Okay.

好的。

Okay.

而且，你知道，在某些情况下，阴性结果实际上——我们本来打算做别的事情，结果反而发现了一个以前没被发现的新结构。

And a negative result, you know, in some cases is actually, you know, we intended to make this the following thing and then we're actually able to identify a new structure that hadn't previously been identified.

所以从某个方面来说是阴性的。

So it's negative in some respect.

就像我们本来想做别的事，但数据里冒出了别的东西。

Like we intended to do something else, but something else emerged from the data.

而且一般来说，当你训练机器学习算法时，尤其是在最基础的层面上，你只是说分类算法。

And also in general, like when you're training a machine learning algorithm, especially like at a very basic level, you just say classification algorithm.

如果没有负样本，而所有样本都是正的，你实际上没法训练。

If you don't have negative samples, you can't really train if everything is positive.

这在材料科学中尤其是个大问题，因为人们通常只发表他们能合成出来的晶体。

And this is a particularly bad problem in material science because people usually publish crystals they could synthesize.

但如果他们合成失败，通常不会发表，有时候可能会。

But they usually don't publish if they fail to synthesize a crystal, sometimes they might.

而且我们真的永远不知道某个晶体是否永远无法合成，对吧？

And also, we never know really if a crystal is not synthesizable ever, right?

可能只是技术问题。

It could just be a skill issue.

没错，尤其是技术问题，

Exactly, Especially a skill with

文献方面的。

the literature.

比如合成方法的问题、技术问题。

Like synthesis method issue, technology.

所以，我觉得当我们自己做实验，在尝试的背景下得到负面结果时，这真的对我们有帮助。

So, I think it truly helps us when we do our own experiments and get negative results in the context of what we tried.

这样我们甚至能训练一个分类算法。

So then we can even train a classification algorithm.

我觉得还有另一个有效的事情，就是当你有一连串的负面结果，然后最终通过过程的迭代，你能够达到那个正面结果。

Think there's also like another valid thing too of when you have this sort of string of negative results, and then finally, through process iteration, you're able to get to that positive result.

这是一组非常有趣的数据，我称之为过程工程类数据。

It's a really interesting set of, I'll call it like, process engineering type data.

所以，通过迭代，你实际上是如何得到正确结果的？

So it's like through iteration, how did you actually get to the correct result?

因为材料科学中有很多关于实际制造过程的模糊性。

Because so much of material science has this ambiguity in actually how something was made.

所以，是的，实际上，即使是一种已知的材料，复制它也可能非常不简单。

And so, yeah, and so there's actually, even if there's a known material, it can be highly nontrivial to replicate that.

有些东西，你知道，就像高中课本里讲的，但其他一些东西真的处于前沿，而且，我们也在积累这方面的经验。

Some things are, you know, in like a high school textbook, but other things are really at the frontier, and, you know, we're building up that know how as well.

然后建立一个系统，你知道，面对一连串的负面结果，你实际上怎么才能达到那个正面的结果？

And then building up a system that, you know, given the string of negative results, how do you actually get to that positive case?

说到分类器的结果，我几乎可以断定，如果你所有事情都在内部做，你的大部分结果实际上会是负面的，而不是正面的，这有点讽刺，因为文献只给你正面的结果。

Going to the classifier result, I would almost assume that if you are doing everything in house, most of your results will actually be negative rather than positive, which is kind of ironic because the literature only gives you positive results.

所以如果你是从零开始，问题似乎是反过来的，你有大量的负面结果，而正面结果不够多。

So if you were sort of cold starting this, it seems like the problem is actually the reverse, that you have an abundance of negative results and not enough positive.

正如你所说，如果你按每个实验来标注，那大部分都会是负面的。

And as you said, like, you were doing per experiment labels, it would be mostly negative.

但如果你按每个项目来标注，并且假设项目成功的话，那可能会更平衡一些。

But if you were doing per campaign and assuming campaigns and when we succeed, then it could be a bit more balanced.

明白了。

I see.

所以你的推理轨迹几乎可以覆盖整个项目周期。

So your reasoning traces could almost be over entire campaigns.

完全同意。

Absolutely.

好的。

Okay.

这是关键所在。

This is the key thing.

这种数据基本上在别的地方找不到。

This type of data basically doesn't exist anywhere else.

我们花了大量时间把科学过程的完整脉络输入到模型中。

And we spend so much of our time getting the full lineage of the scientific process into the model.

所以，追踪所有这些数据，比如对话、直觉、实验室里做了什么、计算是否运行、代码是否编写，把这些都串联起来，价值非常大。

And so, like, tracking all this data, like the conversations, the intuitions, what was executed in the lab, or were the computations run, or was the code written, stitching all this together is so valuable.

而且我觉得，总体目标不是训练科学最终的结果，而是训练做科学的过程。

And I think, like, you know, the overall goal is rather than training on the final output of science, you're training on the process of doing science.

这感觉非常像你……

This feels very much like you...