AlphaFold 3：解锁生命分子的结构

欢迎回到 Google DeepMind 播客，我是主持人 Hannah Fry 教授。从节目一开始就关注我们的听众会知道，利用人工智能加深对世界的科学理解一直是团队的核心目标。我们有天气预报——那是先进的气象学，有核聚变——为物理学家服务，而最引人注目的或许是 AlphaFold，它彻底改变了生物学。但如果这一切听起来有点学术、有点小众，似乎只应该让实验室里的少数科学家兴奋，那么希望今天的节目能让你相信这些想法对人类产生的巨大影响。因为我再次邀请到了 Google DeepMind 科学研究负责人 Pushmeet Kohli。Pushmeet，非常感谢你再次来到播客。自从我们上次交谈以来，发生了很多事情。你现在有多少篇《自然》论文了？

Welcome back to Google DeepMind: The Podcast with me, your host, Professor Hannah Fry. Now, okay, those of you who've been following us from the beginning will know that using artificial intelligence to enhance a scientific understanding of the world has always been a key goal of the team here. You've got weather forecasting that's advanced meteorology, you've got nuclear fusion for physicists, and perhaps most notably is AlphaFold, which has been an absolute game-changer for biology. But if all of that seems a little bit academic, a bit niche, perhaps the kind of stuff that should only really be exciting to a handful of scientists in their labs, then well, hopefully today's episode will persuade you of the substantial impact that these ideas can have on humanity. Because I am joined once again by the guy who is in charge of scientific research here at Google DeepMind, Pushmeet Kohli. Pushmeet, thank you so much for joining us back on the podcast. Quite a lot has happened since we last spoke to you. How many Nature papers have you got now?

我不太确定。我们没在数。多到数不清。

I'm not really sure. We're not counting. Too many to count.

好的。我想和你聊聊 AlphaFold，因为自从上次交谈以来，它似乎又取得了令人振奋的进展。我得说明一下，想了解更多关于 AlphaFold 的工作原理和功能的人可以回顾上一系列的第一集。但我想，你能简单总结一下 AlphaFold 是什么吗？

Okay. So I want to talk to you about AlphaFold because it feels as though this is one of the most exciting things that has advanced even further since I last spoke to you. And I should say actually that anybody who wants to know more about how AlphaFold works and what it does can go back to episode one of the previous series for more detail. But I guess just briefly, can you summarize what AlphaFold is?

是的，AlphaFold 是一个系统，给定一个蛋白质——蛋白质构成了我们周围的一切，它们是生命的基石——本质上它们可以表示为氨基酸序列。而任何给定蛋白质的结构是一个巨大的谜团。所以如果你得到这些氨基酸的序列，也就是一个蛋白质，它的结构是什么？这非常重要，因为结构决定了蛋白质的功能。AlphaFold 的作用是输入蛋白质序列并预测其结构。这对科学家理解蛋白质的功能至关重要。这就是 AlphaFold 所做的：它解决了科学界一个 50 年来的重大挑战，让科学家能够在几秒钟内找到任何给定蛋白质的结构。

Yeah, so AlphaFold is a system that, given a protein—proteins make everything around us, they are the building blocks of life—and essentially they can be represented as a sequence of amino acids. And it's a great mystery as to what is the structure of any given protein. So if you are given a sequence of these amino acids, a protein, what is its structure? And that's really important because that informs the function of that protein. What AlphaFold does is it takes a sequence of a protein and predicts its structure. And that's really important for scientists to understand what is the function of that protein. So that's what AlphaFold did: it solved that 50-year-old grand challenge in science by making scientists able to find the structure of any given protein in a matter of seconds.

所以我想更准确的说法是，它曾经是个谜，但现在已经不是了。

So I guess actually a more accurate statement would be it was a mystery but is no longer.

是的，我的意思是关于蛋白质的行为方式、蛋白质动力学等方面，仍然有一些已知的未知。但蛋白质理解中这个标准的大问题——蛋白质的结构是什么——我们现在已经知道了。

Yeah, I mean there's still sort of a few things that are known unknowns about how proteins sort of behave, the dynamics of the proteins and so on. But this one standard big problem in protein understanding—what is the structure of a protein—is now, we now know about it.

那么从上一次到现在有什么变化呢？上一系列节目时还是 AlphaFold 2，AlphaFold 3 能做什么之前版本做不到的事情？

So what's changed since last time then? Since the last series when there was AlphaFold 2, what can AlphaFold 3 do that the previous versions couldn't?

这是个很好的问题。我说过蛋白质是生命的基石，但它们并不是我们体内唯一的分子。我们的身体还有许多其他生物分子。我们有核酸：RNA、DNA，它们是生命的配方，是构成你我的东西，Hannah。还有小分子、药物，它们与这些蛋白质相互作用。还有抗体。在任何一个生命体中，有如此多种类的分子在相互作用，我们想了解所有这些分子的结构。这就是 AlphaFold 3 解锁的能力。它不仅给出蛋白质的结构，还给出这些蛋白质如何形成复合物、如何相互连接、如何与小分子或抗体等相互作用的结构。

Yeah, that's a very good question. Now, I said proteins are the building blocks of life, but they are not the only molecules in our body. Our body has many other biomolecules. We have nucleic acids: RNA, DNA, which are the recipe of sort of life, what makes you and me, Hannah. Then there are small molecules, drugs that sort of interact with these proteins. There are antibodies. There are so many different types of molecules that are interacting in any living being, and we want to understand the structure of all these molecules. And that's what AlphaFold 3 unlocks. It not only gives you the structure of a protein, but it also gives you the structure of how these proteins sort of form complexes, or they connect to each other, or how do they interact with small molecules or antibodies and so on.

那么，科学家们的反应如何？毕竟 AlphaFold 已经问世一段时间了。

What's been the reaction then from scientists now that this has had a bit of time to sort of be out there in the world?

对我来说——我的背景是计算机科学家——实际上很难理解 AlphaFold 的影响。我第一次意识到这一点是在参加一个生物学会议时，一位生物学家对我说：‘有一种蛋白质我研究了十年，收集了大量数据，但它的结构仍然是个谜，因为在 AlphaFold 之前，要弄清楚结构太难了。’这位科学家用了 AlphaFold，得到了结构。然后他的问题是：‘好吧，我现在该做什么？我有结构了。’所以他必须彻底重新思考生物学现在需要什么，下一步是什么。一个十年的项目，放进 AlphaFold，几分钟或一夜之间就解决了。

So it was very hard for me—for my background, I'm a computer scientist by training—it was very hard for me to actually understand the impact of AlphaFold. And the first time I realized, I went to a conference, a biology conference, where a biologist spoke to me and said, 'Well, there was this protein that I had been trying to study for the last 10 years, and I have collected so much data about it, but still the structure of that protein was a mystery because it was so hard to work out what the structure might be before AlphaFold.' And the scientist had used AlphaFold and had the structure. And then the question he had was, 'Well, what do I do now? I have the structure.' So he had to completely rethink what biology requires now in terms of the next steps. A 10-year project, and it just put it into AlphaFold and it sorted it out for him in a few minutes, or maybe overnight.

他是兴奋，还是在某种程度上有点沮丧？

Was he excited or was that in some ways a bit demoralizing?

不，我认为既有惊讶也有兴奋。这是你无法想象的事情。就像第一次电话出现时，你可以和几英里外的人通话，这是难以想象的。而你可以输入任何蛋白质的序列，可视化它的 3D 结构——这给了这些科学家一种他们之前无法想象的超能力。对于能用这种超能力做什么，大家非常兴奋。

No, I think there was a mixture of both surprise but also excitement. Something that you wouldn't imagine. It's like giving people an ability that did not exist. Imagine the first time telephones came about, right? And you could now talk to people who are miles away. This was unimaginable. And the fact that you can take any protein, put in the sequence of that protein, and visualize what the 3D structure is—that just gave these scientists a superpower that they had not imagined earlier. And there was a lot of excitement as to what we could do with that superpower.

人们现在就是这种感觉吗？就像拥有了超能力？

And is that how people are feeling? Like they've got a superpower now?

是的。所以我们看到的一件事是 AlphaFold 数据库。我们创建了这个 AlphaFold 数据库，找到了几乎所有已知蛋白质的结构，并将这些结构放在由我们的合作伙伴欧洲分子生物学实验室（EMBL）托管的数据库中。这 2.5 亿个结构免费提供给全世界任何人。这个数据库已被 140 个国家的 180 万科学家使用。有 180 万科学家在寻找蛋白质结构——我的意思是，如果这还不能说明世界的积极状态，那我就不知道什么能了。我们总想着社会的阴郁，但看看科学的进步，有 180 万人在研究蛋白质。

Yeah. So one of the things that we have seen is the AlphaFold database. We created this AlphaFold database where we found the structures of almost all known proteins and we put the structures for those in a database hosted by the European Molecular Biology Laboratory, our partners EMBL. And these 250 million structures were available to anyone in the world for free. And that particular database has been used across 140 countries by 1.8 million scientists. And the fact that there are 1.8 million scientists who are looking for protein structures—I mean, if that is not a positive statement about the state of the world, then I don't know what is. Right? I mean, we think about the doom and gloom in society, but if you look at how science has progressed, there are 1.8 million people who are studying proteins.

我认为当你和人们谈论 AlphaFold 时，真正值得注意的是，理解它的人——那 180 万人——和其他人之间的巨大差距。你也有同样的感觉吗？我的意思是 AlphaFold 有点技术性，很难真正理解它的重要性。你兴奋的应用是什么，那些其他人也能理解的？

I think the really notable thing when you talk to people about AlphaFold is just how big of a difference there is between the people who understand it—that 1.8 million people—and sort of everybody else. Is there something that you notice as well? I mean AlphaFold feels a bit technical to really understand the magnitude of it. What are the applications that you are excited about that other people will understand?

是的，所以我认为对于大部分群体来说，他们把 AlphaFold 视为一个人工智能的突破。

Yeah, so I think for a large section of the community, they sort of see AlphaFold as an AI breakthrough.

但从事这个领域的科学家们非常深刻地理解它的影响，对吧？他们知道 AlphaFold 对于极其重要的事情的意义，比如药物发现、设计新疫苗、考虑抗菌素耐药性、考虑分解塑料的新酶。我可以一直说下去。但这不是我们上次谈话以来这里唯一进行的项目。所以请告诉我们更多关于你一直在做什么。

But the scientists who work on this topic understand the implications of it in a very deep way, right? They know the implications of AlphaFold for extremely important things like drug discovery, designing a new vaccine, thinking about antimicrobial resistance, thinking about new enzymes to decompose plastics. I could just go on and on and on. But that's not the only project that has been going on here since we last spoke to you. So tell us a little bit more about what you've been working on.

我们一直在研究各种不同的主题，从材料科学到聚变，再到计算机科学、数学、天气预报、气象学的新发现。所以我们正在关注的领域非常广泛。

We have been working on a whole spectrum of different topics, from material science to fusion, to working on new discoveries in computer science, mathematics, weather prediction, meteorology. So there's a whole spectrum of areas that we are looking at.

在这些领域中，你认为过去几年哪些方面取得了真正显著的进展？

And of those, where do you think there's been really significant progress in the last couple of years?

我们可以谈论的一个时刻是我们在天气预报方面的工作。天气和气候是我们目前都在思考的问题。但如果你看看 DeepMind 之前所做的，我们有临近预报模型，它能够做出非常好的预测，但时间尺度很短。通过我们去年发布的名为 GraphCast 的新模型，我们现在可以处理 10 天预报的问题。我们已经证明，这个新模型可以比一些正在使用的模型更准确地做出这些 10 天预测，这些模型是英国气象局使用的 I 类模型，是在超级计算机上运行数小时的超级计算机。我们在准确性上可以超越它们，并且可以在单个芯片上在一分钟内做出预测。从某种意义上说，这确实开辟了该领域的研究。许多其他实体现在可以进行天气预报研究。结果令人惊叹。一个特别引人入胜的例子是去年飓风李在 Nova Scotia 登陆。我们的模型 GraphCast 能够提前 9 天预测登陆事件，而经典模型只能提前 6 天预测。所以他们可以多提前三天发出警报。

One of the moments we can talk about is our work on weather prediction. Weather and climate is something we are all thinking about at the moment. But if you look at what DeepMind had done earlier, we had our nowcasting model which was able to make very good predictions but at very short time scales. With our new model called GraphCast, which we released last year, we can now look at the problem of 10-day forecast. We have shown that this new model can make these 10-day predictions more accurately than some of the models that are being used, the class I models that are being used by the Met Office, the supercomputers which run on supercomputers for many hours. We can outperform them in terms of accuracy and make predictions in a matter of a minute on a single chip. In some sense, this really opens up research in this area. A lot of other entities can now conduct research in weather prediction. The results are amazing. One particular example that was fascinating is that there was this cyclone, Hurricane Lee, last year which made landfall in Nova Scotia. Our model GraphCast was able to make the prediction of the landfall event 9 days earlier, while the classical models were only able to do it 6 days earlier. So they could give a three-day additional heads up.

那么这是一种不仅仅适用于生物学的超能力。

That's a superpower that doesn't just apply to biology then.

绝对如此。人工智能和机器学习在所有这些学科中产生的影响是惊人的。如果你仔细想想，这感觉很自然，因为在任何这些科学领域，我们都在收集大量数据，并且我们正在使用的模型的复杂性确实在扩大。单个人类思维很难理解这些数据中的真实模式，这是很自然的。机器学习和人工智能正好赋予你能力，去弄清楚许多这些问题中需要什么。

Absolutely. The amount of impact that AI and machine learning is having across all these disciplines is amazing. If you think about it, it feels natural because in any of these areas of science, we are collecting a lot of data, and the complexity of models that we are playing with is really expanding. It's just very natural that a single human mind has difficulty comprehending what are the real patterns in this data. Machine learning and AI just give you that ability to figure out what is needed in many of these problems.

好的，你提到了材料科学。请为我们描述一下，在人工智能出现之前，材料科学中的一般问题是什么。

Okay, you mentioned material science. Frame for us the general problem, as it were, in material science pre-AI.

材料发现的问题是要发现具有某些特性的材料。我们经历了从石器时代到铁器时代再到青铜时代等所有时代。在每个时代，我们都在使用新材料，新材料赋予我们新的能力。到目前为止，这是如何做到的？这是以一种非常实验性的方式完成的。人们在实验室里尝试不同的东西。有时事情如预期的那样，有时事情不如预期。有一些经验法则，有一些理论，但我们不知道材料中可能性的范围。事情经常是偶然发现的，比如硫化橡胶，或者石墨烯，我们都知道这个神奇的材料是如何用胶带分离出来的，通过取碳片并用胶带反复使其变薄。它具有惊人的特性，最终成为可以制造的最薄物质，单原子厚，并且在导电性等方面具有非常有趣的特性。在人工智能甚至计算方法出现之前，这基本上是常态。即使在今天，从某种意义上说，材料科学仍然是一门非常实验性的科学，人们试图发现用于制造电池、光伏电池或超导体等的新材料。今天计算系统的使用方式是事后合理化为什么该材料会以这种方式表现。但我们离能够根据属性合理设计材料还很远。比如说，找到一种材料，发明一种可以最大化这些特性的新材料。这就是人工智能的问题：它能否在计算机中，从头开始，发明任何具有某种特性的材料，这样你就可以进去说，“我想要一种非常灵活、非常轻、易于开采的东西”，然后它就会说，“这是物理结构，原子结构，将产生那种材料。”这就是愿景。

The problem of material discovery is to discover materials which have certain properties. We have gone through all these ages from the Stone Age to the Iron Age to the Bronze Age and so on. At every age, we are working with a new material, and new material gives us new abilities. How was that done till now? It was done in a very experimental way. People tried different things in the lab. Sometimes things worked out as expected, sometimes things didn't work out as expected. There was some rules of thumb, there was some theory, but we did not know the extent of what was possible in materials. Often things get discovered by accident, like vulcanized rubber, or graphene, where we all know the story about how this magical material was isolated by cellotape, by taking pieces of carbon and repeatedly making it thinner with sticky tape. It has amazing properties, ending up being the thinnest substance that can be manufactured, single atom thick, and it has very interesting properties in terms of conductivity and so on. Before AI and before even computational methods, that was essentially the norm. Even today, in some sense, material science is a very experimental science where people are trying to discover new materials for constructing a battery, or for a photovoltaic cell, or for a superconductor, and so on. How computational systems are used today is to then post hoc rationalize why that material is behaving in a way that it is behaving. But we were very far from the place where we could rationally design a material given a property. Say, find me a material, invent a new material which can maximize these properties. That's what the problem is for AI: can it somehow in silico, de novo, from scratch, invent any given material with some property, so that you can go in and say, 'I want something that is extraordinarily flexible, extraordinarily light, easy to mine,' and then it's like, 'Here's the physical structure, the atomic structure that will result in that material.' That's the vision.

那么让我们以例子来说明。你提到了电池。我们现有的电池有什么问题？

Let's anchor this to an example then. You mentioned batteries. What's wrong with the batteries we have?

我认为它们表现不错。它们是锂离子电池。但存在一些问题。首先，它们基于某些难以获取的资源。例如，今天使用的电池使用钴，这很难开采。未来我们可能希望提高这些电池的能量密度。我们可能希望使它们更热稳定。我们可能希望使它们更便宜。如果你能以某种方式获得一种神奇的材料，具有更高的能量密度、更稳定、易于制造，那么你当然会想要过渡到它并找到它，而不是仅仅等待实验室中的意外发生。

I think they are doing well. They are lithium-ion batteries. There are a number of issues with that. First, they are based on certain resources which are difficult. For example, the batteries used today use cobalt, which is difficult to mine. We might want to increase the energy density of these batteries in the future. We might want to make them more thermally stable. We might want to make them cheaper. If you could somehow have a magic material which can have higher energy density, is more stable, is easy to manufacture, then of course you would want to transition to it and find it without just waiting for an accident to happen in a lab.

但你确定存在这样的材料吗？你怎么知道锂不是宇宙能提供的最好的？

Are you sure there is one though? How do you know that lithium isn't the best that the universe has to offer?

有可能锂是最好的，我们非常幸运。

It could be that lithium is the best and we were extremely lucky.

我们找到了，但这似乎非常遥远。我的意思是，有很多材料——举个例子，人们现在摆弄的无机材料大约有 2 万种。通过计算方法，又发现了 2.8 万种。所以大约有 4 万到 5 万种已知材料，它们在零开尔文或某些理论条件下是稳定的，不会分解成其他材料。

We found it, but that seems very remote. I mean, there are a number of materials—just to give you an example, there were around 20,000 inorganic materials that people sort of play with now. Using computational methods, 28,000 have been found out. So there are roughly 40,000 to 50,000 known materials that were known to be stable in the sense that at zero Kelvin, or under some theoretical situations, they will not decompose to other materials.

对，如果你把它们冷冻到绝对零度，它们不会分裂。没错。所以它们是稳定的材料。这就是材料科学界已知的情况。我们去年推出的新 AI 模型 GNoME 将这个集合扩展到了 220 万种新的稳定无机材料——从 5 万种起步。

Right, so if you freeze them down to absolute zero, they don't split apart. Exactly. So they are stable materials. That was what was known in the materials science community. Our new AI model, GNoME, last year expanded that set and said there are 2.2 million new inorganic materials that are stable—from 50,000.

哇，好吧，哇，这太庞大了，对吧？而且我们说的是石墨烯——这个集合里有 5.2 万种单链层状材料。所以可能性数量，我们现在可以搜索的东西，是巨大的。你问锂钴电池是不是最优电池——嗯，那 220 万种材料里有很多种，其中一种可能要好得多。我的意思是，锂实际上最优的概率是……

Wow, okay, wow, that's massive, right? And we're talking about graphene—there are 52,000 single-chain layered materials in that set. So the number of possibilities, the number of things that we can now search over, is immense. And you ask the question whether the lithium-cobalt batteries are the optimal batteries—well, there are so many things in that 2.2 million set that one of them could be much, much better. I mean, the chance that lithium is actually optimal is...

它叫什么？GNoME。是的。它代表什么？

What's it called? GNoME. Yes. What does it stand for?

图神经网络材料探索。

Graph Neural Networks for Material Exploration.

它是如何工作的？你如何决定尚未发现的新材料结构？

And how does it work? How are you deciding new material structures that haven't yet been discovered?

所以它本质上是从现有的结构和成分——我们已知的 5 万种——开始，然后说：‘好吧，让我改变其中一些，然后我学习一个模型来判断哪些变化、哪些修改后的材料是稳定的或不稳定的。’它能够非常高效地进行这些计算。这就是机器学习模型的用武之地——它能够以更准确的方式预测这些新成分和新晶体结构的稳定性和自由能。

So it essentially tries to start with existing structures and compositions—the 50,000 we know—and says, 'Okay, let me change some of those, and then I'll learn a model to say which of those changes, which of those modified materials, are stable or not stable.' And it is able to do those calculations very efficiently. That's where the machine learning model comes in—it's able to make predictions about the stability and the free energy of these new compositions and new crystal structures in a much more accurate way.

所以从某种意义上说，它是拿这 5 万种材料进行某种原子洗牌，如果你愿意这么说的话，只是尝试不同的组合。但巧妙之处在于，正如你描述的，你可以计算——而无需实际制造这种通过原子洗牌构建的幻想材料——就能判断它在零开尔文下是否稳定。

So in one sense, it's taking the 50,000 and doing some kind of atomic shuffling, if you like, just trying different combinations. But then the clever bit, as you're describing it, is that you can calculate—without actually having ever made this sort of fantasy material which has been constructed by atomic shuffling—you can tell whether or not it will be stable at zero degrees Kelvin.

是的，如果你制造了它，理论上你可以做到，但你需要进行大量非常复杂的计算。这个模型能够做的基本上是近似这些计算，并且计算成本低得多。

Yes, if you made it, you could do that theoretically, but you would need to do a lot of very complex calculations. What this model is able to do is basically approximate those calculations and do it much more computationally inexpensively.

在很多方面，我能看到这和 AlphaFold 的相似之处。你谈论的是原子层面的结构，直到氨基酸和原子，然后预测更大的结构特性。所以在这个特定案例中，你得到一个新晶体结构或新成分，然后被问：它会不会稳定？模型能够回答。

In many ways, I can see the similarities between this and AlphaFold. You're talking about the atomic structure of something right down at the level of amino acids and atoms, and then predicting larger structural properties as a result. So in this particular case, you are given a new crystal structure or a new composition and you are told: is it going to be stable or not? And the model is able to say that.

但你如何验证它呢？

How do you validate it, though?

我们通过两种方式验证了这些。一种验证方式是进行计算。当我说这些是 220 万种稳定结构时，我是什么意思？我如何说这些是稳定的？量子化学中有一些理论，它们提供了方法来判断某物是否稳定。这些计算非常困难，所以我们可以在 GNoME 的预测上运行这些计算，看看理论是否说这些是稳定的。这是一种方式。然后我们还取了其中一部分——最稳定的预测——并在实验室中通过实验进行验证。

We have validated these in two ways. One form of validation is by doing calculations. When I said these are 2.2 million stable structures, what do I mean? How do I say that these are stable? There are theories in quantum chemistry which give us approaches to figure out whether something is going to be stable. These are very difficult calculations, so we can run those calculations on the predictions that GNoME has made and see whether the theory says that those will be stable. That's one way. And then we have also taken a subset of those—the most stable predictions—and experimentally tried to validate them in the lab.

哦，像制造它？是的，哇，好吧。但等等——如果它给你一些幻想的原子结构，它会告诉你如何制造它吗？

Oh, like build it? Yeah, wow, okay. But then hold on—if it's giving you some fantasy atomic structure, does it tell you how to make it?

不，它不会。所以我们必须想办法制造它。

No, it doesn't. So then we have to figure out how to make it.

你已经用其中一小部分做到了这一点，对吗？它们成立吗？它们稳定吗？

And you have managed to do this with a small number of those, yes? And do they hold up? Are they stable?

是的，是的。而且其中很大一部分是稳定的。

Yes, yes. And a large proportion of them are stable.

给我数字。

Give me numbers.

我认为这不断在变化，但成功率相当高——我们在实验室尝试的那些中，70%到 90%最终是稳定的。

I think this is constantly changing, but the success rates are quite high—between 70% to 90% of those that we try in the lab end up being stable.

现在，一旦我们有了这些有效的材料，我们还必须考虑这些材料的性质。其中哪些更适合电池，或者对光伏有用，或者对超导性有好处，等等。那么它也能预测性质吗？

And now, once we have those valid materials, we also have to think about what are the properties of those materials. Which of them would be better for batteries, or would be useful in photovoltaics, or would be good in superconductivity, and so on. So can it also predict the properties then?

不，我们当前这一代不行。它只预测这些结构的稳定性。但我们正在为这些其他类型的问题开发新模型。

No, not our current generation. It just makes predictions about the stability of those structures. But we are working on a new model for these other types of problems.

你希望得到什么样的材料？我的意思是，电池材料是一个例子。你还希望 GNoME 帮助开发哪些其他东西？

What kind of materials are you hoping for? I mean, battery materials is one example. What other kind of things are you hoping that GNoME will help develop?

人们一直在思考的关键事情之一是能够表现出超导性的材料，这意味着它们表现出零电阻。为什么这很重要？因为如果你有这样一种材料，你可以产生非常强的磁场，你可以用这些材料储存大量能量。为什么磁场很重要？它们对从 MRI 扫描仪到制造聚变反应堆的一切都很重要。聚变反应堆中使用的磁铁需要非常高的磁场。所以超导性是一个极其重要的性质。目前，你必须将东西超冷才能达到那些高水平，否则它就会变得太热。

One of the key things that people have been thinking about is materials that can exhibit superconductivity, which means they exhibit zero resistance. Why is that important? Because if you have such a material, you can create very strong magnetic fields, you can store a lot of energy using those materials. Why are magnetic fields important? They're important for everything from MRI scanners to creating fusion reactors. The magnets used in fusion reactors require very high magnetic fields. So superconductivity is an extremely important property. At the moment, you have to supercool things to get up to those high levels, because otherwise it just gets too hot.

没错。所以材料科学中的圣杯是发现室温超导体。已经有很多尝试，一些错误的开始，还有一些相当厚颜无耻的假装。

Exactly. And so what is the Holy Grail in materials science is to discover a room-temperature superconductor. And there have been many attempts at it, and some false starts, and some quite cheeky pretending has happened.

是的，也有这种情况。但我认为如果有一天我们发现了它，那将是变革性的。

Yes, that too. But I think if one day we discover it, it will be transformational.

你认为 AI 会参与那个发现吗？

And do you think AI will be involved in that discovery?

我绝对相信 AI 将有助于寻找超导体，至少会加速这个过程。

I'm absolutely sure that AI will aid in the search for superconductors, at least accelerate the process.

那么好吧，不过制造它的那部分——我的意思是，你描述的方式是，‘是的，我们必须想办法制造它。’感觉我们跳过了过程中相当棘手的一部分。有没有办法也能自动化那部分？

So okay, that part of making it, though—I mean, the way that you described it was like, 'Yeah, we've got to figure out how to make it.' It feels like we're skipping over quite a bit of a tricky part of the process. Is there any way that you can automate that bit too?

一旦你知道材料是什么，那就是材料科学工作的另一个完整分支，即如何廉价地制造某物。而且有很多……

Once you know what the material is, then it's a whole other stream of materials science work, which is how to make something cheaply. And there is a lot of...

好，另一个开放性问题。你认为第一个由 AI 开发或借助 AI 辅助开发的材料，要多久才能产生重大影响？

Okay, here's another open question. How long do you think it will be before the first AI-developed material, or material developed with the assistance of AI, makes a big difference?

我认为与生物学不同，材料科学目前更加依赖实验。药物发现仍在使用计算方法，但材料科学更偏实验，所以需要时间。但我认为在未来 5 到 10 年内，我们会看到计算机模拟材料设计开始产生重大影响。

I think unlike biology, material science is much more experimental at the moment. Drug discovery still uses computational methods, but material science is much more experimental, so it will take time. But I think in the next 5 to 10 years, we will see in silico material design starting to make a big impact.

我想再稍微拉远一点视角。你坐在这里非常自信地跟我们谈论材料科学，还有生物学，而你是个计算机科学家。你是怎么进入这个领域的？

I want to zoom out a little bit further. You're sitting here talking to us very confidently about material science, but also about biology, and you're a computer scientist. So how come you got into this?

我上一次学习科学还是在中学。我来 DeepMind 是希望研究通用智能系统，让它们安全可靠。但我一直着迷于智能系统如何产生现实世界的影响。我曾和 Demis 讨论过这个问题。有一天他来到我的办公室说：‘Push，我觉得有个好角色适合你。’我以为是某个产品领域，终于可以用上机器学习。他说：‘我们想成立一个科学团队，你来领导。’我说：‘为什么是我？我没有科学背景。’他说：‘你热衷于跨学科研究，想理解问题并产生现实影响。还有什么比在自然科学中推动知识边界更好的地方呢？’于是我同意了。接下来的几个月，我才意识到自己卷入了什么。如果有人要设计一个让人体验冒名顶替综合征的工作，就是它了。你和诺贝尔奖得主一起工作，试图理解你一无所知的问题。你从维基百科开始，速成学习。但科学界很棒——人们愿意坐下来教你，或者和你一起踏上这段旅程。

The last time I studied science was in school. I came to DeepMind hoping to work on general intelligence systems and make them safe and reliable. But I was always fascinated about how intelligent systems can have real-world impact. I used to have discussions with Demis about this. One day he came to my office and said, 'Push, I think we have a good role for you.' I assumed it was some product area where I'd finally use machine learning. He said, 'We want to start a science team and you should lead it.' I said, 'Why me? I have no background in science.' He said, 'You are into multidisciplinary research, you want to understand problems and have real-world impact. What better place to have impact than pushing the boundaries of knowledge forward in the natural sciences?' So I agreed. Over the next couple of months, I found out what I'd gotten myself into. If someone was designing a job to experience imposter syndrome, this is it. You work with Nobel Prize winners, trying to understand problems you have no clue about. You start from Wikipedia, get a crash course. But the scientific community is amazing—people are willing to sit down and teach you or go on that journey with you.

真正关键的是，你真正理解这些算法的核心，理解它们擅长什么类型的系统，在哪里能大放异彩。从某种意义上说，你谈论很多不同领域——气象学、材料科学、生物学——我能看到你设计的系统背后有技术上的相似性。所以在很多方面，做这类工作需要通才。

The really key point is that you really understand the heart of these algorithms, the type of systems they perform well in, where they can excel. In some ways, the fact that you're talking about lots of different areas—meteorology, material science, biology—I can see behind the scenes technical similarities between the systems you're designing. So in many ways, being a generalist requires a generalist to do this kind of stuff.

绝对如此。如果你思考智能是什么，如何创造智能？智能不会凭空产生。一个能获得直觉并做出新发现的系统是通过经验实现的。这种经验有多丰富？这对人类也成立。我们生活中看到的东西给了我们不同的视角。对于这些模型，它们看到的数据的丰富性——无论是科学家在实验室收集的，还是模拟生成的——如果足够丰富，模型就会被迫使学习通用概念来解释这些数据。另一个要素是好的评估指标。机器学习方法非常擅长作弊；它们极其擅长记忆。它们就像能记住任何东西的学生。如果你给它们大量数据，它们会记住，如果你问任何接近训练集的问题，它们会给出答案，但实际上并不理解发生了什么。

Absolutely. If you think about intelligence, what is it? How do you create intelligence? Intelligence doesn't happen in a vacuum. A system that gets intuitions and makes new discoveries does so through experience. How rich is that experience? This holds true for humans too. The things we see in life give us a different perspective. With these models, the richness of the data they see—whether collected by a scientist in a lab or generated by simulations—if it's rich enough, the model is forced to learn general concepts to explain that data. The other element is good evaluation metrics. Machine learning methods are very good at cheating; they are amazing at memorization. They are like the student that can memorize anything. If you give them a lot of data, they will memorize it, and if you ask any question close to the training set, they'll give the answer, but they don't actually understand what's going on.

我猜这在蛋白质研究中肯定发生过不少次——它们偷看了课本后面的答案。

I guess that must have happened quite a lot with proteins at some point—they've peaked at the back of the textbook.

没错。所以我们必须要确保测试极其困难，让人无法作弊。一个好的机器学习系统的真正考验是它泛化到未见示例的能力，对从未见过的事物做出预测。理解不同科学学科的正确数据和正确评估指标是问题的一部分。另一部分在于这些 AI 模型的设计：你希望从数据中学习，但也要尽可能将关于问题的任何信息融入模型本身的设计。这就是领域知识非常有意义的地方。如果没必要，你不想从零开始。在我们所有的项目中，我们都从非常跨学科的角度出发，请来专家，问他们什么有效、什么无效，并尝试将所有已知信息注入模型的设计中。

Exactly. So what we had to do was make sure that test was extremely hard, made it impossible to cheat. The true test of a good machine learning system is its ability to generalize to unseen examples, to make predictions on things it has never seen before. Understanding what is the right data and the right evaluation metric for a different scientific discipline is one part of the problem. The other part is in the design of these AI models: you want to learn from data, but also bake in any information you can about the problem into the design of the model itself. That's where domain knowledge makes a lot of sense. You don't want to start from scratch if you don't have to. In all our projects, we started with a very multidisciplinary viewpoint, got experts in, asked them what was working and what was not, and tried to inject everything currently known into the design of the model.

值得注意的是，你描述的所有内容都有一个共同主题：存在一个成功的版本——你寻找的正确答案，比如蛋白质的真正底层结构、它是否真正稳定、你是否真的在预测天气。这对这些项目来说绝对关键吗？

It's noticeable that the common theme of everything you've described is where there is a version of success—a right answer you are looking for, like the true underlying structure of the protein, whether it's truly stable or not, whether you really are predicting the weather. Is that absolutely crucial for any of these projects?

是的，绝对如此。你真的需要清楚成功是什么样子。

Yeah, absolutely. You really need to have a good sense of what success looks like.

说到选择项目，因为我的意思是，如果你像你说的那样，是一个通才，能够利用特定领域的知识将这些算法应用到几乎任何事情上，你如何决定哪些事情值得你花时间？

When it comes to choosing projects, because I mean if you are, as you say, the generalist who can, with domain-specific knowledge, adapt these algorithms to sort of anything, how do you decide what's worth your time?

是的，这是一个非常好的问题。我们考虑多个维度。一个维度是：是否有数据？如果没有数据，没有经验，机器学习模型无法自己学习，对吧？所以这是一个重要的考虑因素。第二点，这也是我们从一开始就一直在考虑的，就是关注根节点问题——那些非常基础的问题，一旦解决，会对许多不同的应用产生影响。例如，蛋白质折叠，正如我所说，对药物发现、设计用于塑料分解的新酶等都有影响。材料也是如此。如果你能以变革性的方式理解材料发现，那会有很多不同的应用。这比仅仅研究某种特定药物（比如阿尔茨海默症药物）要有效得多。

So yeah, that's a very good question. And there are a number of sort of dimensions that we look at. One dimension is: is there data? If there's no data and there's no experience, the machine learning model will not learn by itself, right? So that's one of the important considerations. The second thing, and this is something that we have also always considered from the very start, is a focus on root node problems—problems that are so fundamental that once you unlock them, they have implications for a number of different applications. For instance, protein folding, as I said, had implications for drug discovery, to design new enzymes for plastic decomposition, and so on. And the same thing is true for materials. If you can understand material discovery in a transformational way, that has so many different applications. It's much more effective than only looking at, say, one particular drug for Alzheimer's or whatever it might be.

那么你们是否有一个清单？这栋楼里某个地方有没有一块白板，上面列着潜在的项目，然后你们根据这些标准来评估——哪些有好的数据，哪些是更重要的问题？

So do you sort of have a list? Is there a whiteboard somewhere in this building where you have a list of potential projects and then you're evaluating them on that basis—which ones have good data, which ones are more important problems than others?

是的，绝对如此。我们一直在这样做。而且我们必须处理很多不确定性。所以我们以科学的方式处理问题：我们进行实验，做探索性研究，看看我们是否在取得进展，我们的一些假设是否正确。这意味着我们现在可以做出重大承诺，致力于解决那个问题。我们运作方式的另一个不同点是，我提到的大多数课题，我们都有非常专注的团队来攻克。他们不是以 6 个月或 12 个月为周期运作，而是以多年为周期。这些团队中的研究人员和工程师将整个职业生涯都奉献给了那个课题。所以我们非常认真地对待选择正确问题的责任。如果我们认为一个问题可以在牛津、麻省理工、哈佛、伯克利或帝国理工解决，我们就不想做了，因为我们想解决那些真正需要我们的规模和跨学科团队才能解决的问题。

Yeah, absolutely. We are constantly doing that. And there is a lot of uncertainty that we have to deal with. So we approach the problem in a scientific manner: we conduct experiments, we do exploration studies, and see if we are making progress, if some of our assumptions were correct. That means we can now make that big conviction and commit to that problem. And the other difference in how we are operating is that most of the topics I've mentioned to you, we have very focused teams that pursue these problems. And they're not operating at the level of 6 months or 12 months; they are operating at the level of many years. They are dedicating researchers and engineers in those teams—they are dedicating their whole careers to that topic. So we take that responsibility very seriously as to which are the right problems that we should be working on. If we believe that a problem can be done at Oxford or MIT or Harvard or Berkeley or Imperial, we wouldn't want to work on it, because we want to work on the problems that really require the scale and the type of multidisciplinary team that we have to come together to solve them.

好的，那我们换个话题吧。因为我认为过去几年让世界非常兴奋的大事是生成式 AI，特别是大型语言模型。你们是否已经开始将大型语言模型整合到你们的科学研究中？

Okay, well let's pick another topic then, because I think the big thing that the world's got very excited about in the last couple of years is generative AI, and in particular large language models. Have you started to incorporate large language models into your research for science?

是的，我们正在全面研究这个问题。我们正在探索两个主要方向。一个是，到目前为止，我们在科学领域的大部分工作都使用结构化数据——比如基因组数据或蛋白质结构预测数据，其中序列和结构直接相关。但是，科学出版物中蕴含着大量以自由文本形式存在的科学直觉和知识。如何从这些经验中学习，从过去几个世纪在该领域工作的关键科学家的日记中学习？大型语言模型使我们能够吸收所有这些数据并从中学习。所以这是一个关键想法。另一种方式基本上是使用大型语言模型在某些领域生成答案。其中一个例子是我们的算法发现项目，恰如其分地命名为 FunSearch，代表函数搜索。叫 FunSearch 更好——我更喜欢这个名字。我认为我们的团队对这个名字感到非常自豪。

Yeah, so we are looking at it across the board. There are two main themes that we are exploring. One is that, till now, most of the work we have been doing in scientific areas was using structured data—data like genomic data or protein structure prediction data, where you have sequences and structures directly connected to each other. But there's a lot of scientific intuition and knowledge embedded in scientific publications in free-form text. How do you learn from that experience, from the diary entries of key scientists that have worked in that area over the last many centuries? So large language models give us the ability to now ingest all that data and learn from all of it. So that's one key idea. The other way is basically where you use the large language model to generate answers in certain domains. One example of that is our project on algorithmic discovery, aptly named FunSearch, which stands for function search. It's much better named FunSearch—I much prefer that. I think our team was very proud of the name.

那么 Fun 团队是做什么的？

So what does the Fun team do?

没错，Fun 团队研究 FunSearch。他们试图为计算机科学中的重要问题发现新算法。模型被问到：这里有一个计算机科学中的重要问题，无论是像装箱问题这样的概念性问题——给定一组盒子和一些物品，如何紧凑地将这些物品装入盒子？你可能会想，我在家也这么做，这有什么意义？这个概念性问题无处不在。它出现在快递公司如何将杂货送到你家，或者云提供商如何在不同计算机上调度计算任务。所以这是一个非常重要的现实世界问题。这个模型会尝试提出新算法。其中许多算法可能是它之前见过的已知算法，我们告诉它，这很好，但尝试改进这个特定部分并优化它。它继续尝试优化。有时它会犯错，我们告诉它，你犯了错误。它继续这样做，试图得到正确的解决方案，我们反馈给它。在这个过程中，有时它会发现一些全新的、以前未知的东西，最终提高性能，或者提出一种新的启发式方法或新算法，以显著不同且更高效的方式解决问题。

Exactly, the Fun team working on FunSearch. They are trying to discover new algorithms for important problems in computer science. The model is asked: here's an important problem in computer science, whether it's a conceptual problem like the bin packing problem—how do you, given a set of boxes and some items, compactly pack those items in those boxes? You might think, well, I do it at home, what's the relevance? That conceptual problem is everywhere. It's in how delivery companies deliver your groceries to your homes, or how cloud providers schedule computational jobs on different computers. So it's a very important problem in the real world. So what this model does is it then tries to propose new algorithms. Many of those algorithms are maybe well-known algorithms that it had seen before, and we say, well that's fine, but try to improve this particular part and try to refine it. And it goes on trying to refine it. Sometimes it makes a mistake, and we tell it, here's the mistake you have made. And it keeps doing that, trying to get the right solutions, and we feed it back. In that process, sometimes it discovers something completely new that was not known before, and ends up improving the performance, or coming up with a new heuristic or a new algorithm which solves the problem in a remarkably different and much more efficient manner.

所以它能够做到这一点，是因为知识之间存在更深层的联系网络，而这些联系不一定为我们所见。这就是为什么有时大型语言模型会以某种看似合理的方式产生幻觉吗？我不知道，也许有人会说玛丽·居里发明了青霉素，当然她没有，但她确实在同一时期做出了非常重要的发现。所以它似乎可以交换信息片段，因为它看到了我们看不到的更深层联系。

So the reason it's able to do that is because there's this deeper network of connections between knowledge that isn't necessarily visible to us. Is that why sometimes large language models will hallucinate in a way that sort of makes sense? I don't know, maybe one would say Marie Curie invented penicillin, and of course she didn't, but she did come up with a discovery that was really important around the same time. So it sort of can swap over pieces of information as it were, because it's seeing deeper connections that aren't visible to us.

绝对如此。这基本上是因为它在那个潜在空间中运作。它感觉到某些事物是相互关联的。在玛丽·居里的例子中，那是一个错误，如果某人不知道，那是一个有问题的错误。这是不正确的信息。但在我们的案例中，这是一个可以接受的错误，因为我们有一个评估函数，它与 FunSearch 耦合，而 FunSearch 又与大型模型耦合，它可以调用类似……

Absolutely. It is basically because it's operating in that latent space. It feels that there are certain things which are related to each other. In the case of the Marie Curie example, that was a mistake, and that's a problematic mistake if somebody did not know about it. It's incorrect information. But in our case, that's a fine mistake to have, because what we have is an evaluation function which is coupled with FunSearch, which is coupled with the large model, which can call like a like a...

真相检测器，没错，它能很快地指出‘这说不通’。所以重要的是创造力。如果它提出完全有创意的内容，我们就采纳那部分说‘太好了’，而它说的糟糕内容我们很容易就能过滤掉。但时不时地，它说的有创意的东西揭示了更多关于底层知识网络的信息，对吧？这基本上就像你把幻觉用在了好的方面。

Truth detector, yeah exactly, which can sort of call out and say, 'Oh, this doesn't make sense,' very quickly. So then what matters is creativity. So if it comes up with something completely creative, we pick that part and say, 'Oh yeah, great,' and the bad stuff that it says we are able to filter it out very easily. But every now and then, the creative stuff it says is revealing something more about the underlying network of knowledge, right? It's basically like you've harnessed hallucinations for good.

是的，在这个特定案例中，幻觉是好的，如果你能以某种方式利用创造力，保留好的部分并过滤掉所有无效的部分。所以我的意思是，别让 Fun 团队知道，他们现在可以自称‘为善幻觉’团队了。可能会得意忘形。

Yes, hallucinations are good in this particular case if you can somehow leverage creativity and keep the good part and filter out all the invalid part. So I mean, don't let the Fun team know that, they can call themselves 'The Hallucinations for Good' team now. Might get ahead of themselves.

那么它是在发现计算机科学家或数学家不知道的东西吗？

So is it finding stuff that wasn't known to computer scientists or mathematicians?

是的，绝对如此。事实上，用这种方法我们在计算机科学中得到了一个新结果。有一个非常有趣的问题叫做 capset 问题，它是在一个结构化的图中寻找独立集。这就像一个图论问题，你要找到具有某些属性的特定节点和边。这个问题已经被研究了很长时间，是计算机科学中一个非常有趣的问题。而 FunSearch 能够做到的是，首次产生了一个没人能得出的结果。它不仅得出了那个结果，而且它生成的用于寻找该结果的算法程序具有极其有趣的子结构和直觉，以至于与我们合作的数学家看到后，觉得这个程序提取了问题中的一个新对称性。

Yes, absolutely. In fact, with this method we were able to get a new result in computer science. There is a very interesting problem called the capset problem, which is trying to find independent sets in a structured graph. It's like a problem where you have a graph and you're trying to find certain nodes and edges with certain properties. It has been studied for a long time, a very interesting problem in computer science. And what FunSearch was able to do was, for the first time, produce a result that nobody had been able to produce. Not only was it able to produce that result, but the algorithm, the program that it had generated to find that result, had extremely interesting substructure and intuitions that when mathematicians working with us saw it, they felt that the program had extracted a new symmetry in the problem.

哦，所以在这种幻觉中，它真的偶然发现了一些未被探索但被证明是真实的东西，并且它利用了问题中一些我们没告诉它的非常有趣的特性。

Oh, so in this hallucination, it had really stumbled upon something that had been unexplored but turned out to be true, and it was leveraging some very interesting property about the problem that we had not told it about.

是的。

Yeah.

哇。我能看出那可能很有用。好的，给我讲讲奥林匹克竞赛吧。

Wow. I can see how that might be useful. Okay, tell me about the Olympiad.

有国际数学奥林匹克竞赛，这是一个学生参加的竞赛。来自世界各地的一些最优秀的学生参加这个比赛。它包含非常难的数学题。如果表现好，可以获得铜牌、银牌或金牌。国际数学奥林匹克竞赛的问题难度极高。例如，当前一代的 AI 系统无法解决这些问题。它们需要极高水平的横向思维、逻辑和对数学概念的深刻理解，尽管这些问题是面向学龄儿童的。

There is the International Math Olympiad, a competition that students take part in. Some of the best students from all around the world come to this competition. It has very hard math problems. If you perform well, you can get a bronze, silver, or gold medal. The level of problems asked at the International Math Olympiad are extremely hard. For example, the current generation of AI systems are not able to tackle those problems. They require an incredible level of lateral thinking, logic, and deep understanding of mathematical concepts, despite the fact that they are aimed at school children.

是的，绝对。那些不是普通的游戏，对吧？它们是世界的难题，你懂我的意思吗？没错。我不确定 Demis 是否参加过，但确实，它们很特别。对于计算系统来说，能够解决其中任何一个问题都是一个长期的挑战，因为这些是极其困难的数学问题。数学，不像国际象棋或围棋，是一个开放式的环境，没有特定的步数；步骤是无限的，且可以向任何方向。所以你要推理的空间极其巨大。因此最复杂的 AI 系统也无法解决这些问题。所以我们有一个叫 AlphaGeometry 的系统，它首次展示了 AI 系统可以解决国际数学奥林匹克级别的几何问题。

Yeah, absolutely. Those are not your normal games, yes. They are the Demises of the world, you know what I mean? Exactly. I'm sure I don't know whether Demis participated in one, but yeah, they are exceptional. And it has been a long-standing challenge for a computational system to be able to solve any of these problems because these are extremely hard mathematics problems. Mathematics, unlike the game of chess or Go, is an open-ended environment in the sense that it does not have a specific number of moves; the moves are infinite and in any direction. So it's an extremely large space that you are reasoning over. And so the most sophisticated AI systems are not able to tackle any of these problems. So we had a system called AlphaGeometry, which for the first time showed that an AI system can solve geometry problems that are at the International Math Olympiad level.

那我们怎么描述它们呢？你会得到一张有圆、三角形、正方形等图形的图片，然后被问到一些，比如，如何仅凭你对圆、正方形和三角形规则的理解，从这张图中计算出一些看似不可能的东西。类似那样。真的很难。

So how can we describe them? You'll get a picture of some circles and triangles and squares and things, and then you'll be asked something about, I don't know, how to calculate some seemingly impossible thing from this image based only on what you understand about the rules of circles and squares and triangles. Something like that. It's really hard.

是的，相当难。你不仅要非常理解几何，还需要能够提前规划，思考什么样的解决方案有意义并能引导你得到最终答案。

Yeah, it's quite hard. You really have to understand not only geometry very well, but you also need to be able to plan ahead and think about what kind of solution can make sense and lead you to the final answer.

那么 AlphaGeometry 是如何工作的呢？

So how does AlphaGeometry work then?

AlphaGeometry 的工作原理是将问题（有时以文本形式给出，有时以图像形式）转换成一种形式语言，即它自己的领域特定语言，在这种语言中它可以推理问题。然后它尝试用那种语言解决问题。AlphaGeometry 非常聪明地生成了大量该语言的合成问题以及相应的解决方案。这样，他们可以在数十万个这样的问题上训练机器学习模型，这使它变得极其有效。给定一个那种形式的新问题，它就能解决；它只是从自己的经验中知道。

AlphaGeometry works by transforming the problem, which is given sometimes in text, sometimes in image, into a formal language, its own domain-specific language, in which it can reason about the problem. Then it tries to solve the problem in that language. What AlphaGeometry did very smartly was generate a very large number of problems synthetically in that language, along with the corresponding solutions. With this, they could train the machine learning model on hundreds of thousands of these problems, and that made it extremely effective. Given a new problem of that form, it could solve it; it just knows from its own experience.

是的，太棒了。实际上，我再次看到了那和材料科学之间的相似之处。就像，‘好吧，我们就用组合数学，掷骰子，一开始想出大量随机的东西，然后不断缩小范围，直到最终它能看着一个单一的问题说，‘我知道怎么解决它。’

Yes, amazing. And actually, I can see the similarities there again between that and the material science thing. It's like, 'Okay, let's just use combinatorics, let's just roll the dice, come up with loads of random things in the beginning, and then narrow it down and narrow it down until eventually it can look at a single problem and say, 'I know how to solve it.'

是的，没错。

Yeah, exactly.

我挺嫉妒你的工作。好了，我们聊了一大堆话题，但你接下来希望解决什么？

I'm quite jealous of your job. Okay, we have gone on a wild list of topics here, but what are you hoping to tackle next?

我认为在我们工作的任何领域，无论是理解蛋白质、理解基因组、理解天气还是理解材料科学，都有许多问题尚未解决。还有很多工作要做。我们刚才谈到材料；我们只是对这些材料的稳定性进行预测，但如何将其扩展到预测这些材料的性质，然后合成它们呢？

I think there are many problems that remain unsolved in any area we're working on, whether it's understanding proteins, understanding the genome, understanding the weather, or understanding material science. There is just so much work that still remains to be done. We were talking about materials; we are only making predictions about stability of these materials, but how do you extend that to making predictions about properties of these materials and then synthesizing them?

但你有最喜欢的吗？有没有一个让你觉得‘那是我真正想要的’？

Have you got a favorite though? Is there one that you're like, 'That's the one I really want'?

不，你不能问我那个问题。不，我觉得所有……不是所有孩子都平等？得选一个最喜欢的。我是计算机科学家，我喜欢计算机科学的工作，但所有项目都有不同的元素，让你在某天感叹‘哇，这太棒了’。比如当你试图学习一个控制聚变反应堆磁体的策略时，那天……

No, you can't ask me that question. No, I think all... not all your babies are equal? Got to take a favorite. I'm a computer scientist, I love the computer science work, but all of them have different elements which on the day make you go, 'Wow, this is so amazing.' Like when you are trying to learn a policy that is going to control the magnets of a fusion reactor, the day that...

那个实验无论其他地方发生什么都会进行。你最关心那个，没错。但你觉得哪一个会产生最大的影响？

That experiment happens regardless of what's happening anywhere. I'm, I'm you care most about that, yeah exactly. Is there one that you think though will have the biggest impact?

我认为我们在理解生物学、化学和材料方面所做的工作。这些是如此基础、如此根源性的问题，以至于很难限制其影响。它们的效果难以预测。

I think the work that we are doing in understanding biology and in understanding chemistry and materials. I think there it's so fundamental, such root node problems, that it's difficult to even sort of limit what the impact would be. Their effects are difficult to predict.

是的，完全同意。我非常期待再次回来和你聊聊，看看还有哪些重大进展。Pushmeet，非常感谢你接受我的采访。

Yeah, absolutely. I'm very much looking forward to coming back and talking to you again to see what other massive things have been happening. Pushmeet, thank you very much for joining me.

谢谢。

Thank you.

与 Pushmeet 会面后，我印象深刻的是，科学的发展通常非常缓慢，但偶尔，也许一代人一次，会出现一次地震般的转变，一次巨大的进步，从根本上改变我们的理解。但这次利用人工智能进行科学的转变，不仅仅会影响物理学或宇宙学中的某个方程，而是将以最根本的方式改变所有科学。好吧，我知道我听起来像是被炒作冲昏了头脑，但不要只相信我的话。去问问你身边友好的科学家，他们对此有何看法，因为真正理解这对幕后世界意味着什么的人，他们无法用言语形容这次转变有多么巨大。您正在收听的是 Google DeepMind 播客，我是 Hannah Fry 教授。如果您喜欢这一集，请订阅我们的 YouTube 频道，您也可以在您喜欢的播客平台上找到我们。接下来，我们还有更多内容要探索，包括深入探讨 AI 如何增强教育，以及 AI 助手是否应该具备人类特质。这些内容即将在 Google DeepMind 播客中呈现。

What I'm struck by after meeting Pushmeet is that the thing about science, it moves really slowly, and then every now and then, maybe once in a generation, you get this seismic shift, you get a big step forward that fundamentally changes our understanding. But the thing about this shift of using artificial intelligence for science is that this isn't just going to make a difference to physics or an equation in cosmology, it's that this is going to transform all of science in the most fundamental way. And okay, I know that I sound like I've just swallowed the hype here, but don't just take my word for this. Go and ask your friendly neighborhood scientists what they think of what is happening here, because the people who really understand what this means for the world behind the scenes, they don't have words for how big this shift is or is going to be. You have been listening to Google DeepMind the Podcast with me, Professor Hannah Fry. If you enjoyed that episode, do subscribe to our YouTube channel, and you can also find us on your podcast platform of choice. Now we have lots more to explore on this podcast, including a deep dive into how AI could enhance education, plus should AI assistants be given human traits? That is coming up soon on Google DeepMind the Podcast.