数字速度下的科学：AI 对发现的影响

Demis，我们从你开始吧。你是 Google DeepMind 的 CEO 兼联合创始人，也是 Isomorphic Labs 的 CEO 兼联合创始人。你用过一句话，说我们正在见证的是「数字速度的科学」。那么，或许先简单介绍一下背景，让我们快速了解目前的情况——AI 对科学做了什么？

Demis, if we could start with you. You are CEO and co-founder of Google DeepMind. You're also CEO and co-founder of Isomorphic Labs. And you have used this phrase that what we're witnessing is science at digital speed. So perhaps just set the scene to bring us up to speed very quickly with what we've seen so far. What AI has done to science so far?

是的，「数字速度的科学」这个说法是我在 2020 到 2021 年看到 AlphaFold 之后提出的。我们花了大约一年时间，折叠了科学界已知的所有 2 亿种蛋白质。这让我意识到，我们正在用技术和工程领域的方法，但现在应用到了科学课题上。我指的是三个方面。第一，解决方案本身的速度。AlphaFold 不仅非常准确，足以让生物学家使用，而且能在几秒钟内折叠一个普通蛋白质。第二，解决方案的传播速度。通过 AlphaFold，我们与欧洲生物信息学研究所合作，很快就把结构数据放到了数据库里。现在有来自 190 个国家的超过 300 万研究人员在使用它。第三，AI 本身正在加速科学发现。我认为我们现在正处于这个阶段的初期。上周我引用了「奇点山麓」这个说法，引起了一些轰动，但我是认真的。十年后回顾，我们会觉得这个时期具有里程碑意义。AI 是我对科学的贡献，是一种用于发现的元工具。我认为 AGI 就是这样的工具，而我们正在开始看到它。

Yeah, that phrase 'science at digital speed' I sort of coined after seeing AlphaFold in 2020, 2021. We spent about a year folding all 200 million proteins known to science. That struck me that we were using techniques from technology and engineering, but now applied to a scientific subject. I mean it in three ways. First, the speed of the solution itself. AlphaFold was not only very accurate, accurate enough for biologists, but also able to fold an average protein in a few seconds. Second, the dissemination of the solution. With AlphaFold, we collaborated with the European Bioinformatics Institute and put the structures on a database quickly. Now over 3 million researchers from 190 countries have used it. Third, AI itself is accelerating scientific discovery. I think we're in the beginnings of that now. I quoted 'foothills of the singularity' last week, which caused a stir, but I really mean it. In 10 years, we'll look back and see this period as monumental. AI is my contribution to science as a meta-tool for discovery. I think AGI is that, and we're starting to see it.

你选择用「奇点」这个词，「奇点山麓」，很有意思。

The choice of the word 'singularity', 'foothills of the singularity' is an interesting one.

是的。我用这个词是因为它关乎技术周围更广泛的环境。我们把这项技术称为 AGI，这个词是我的联合创始人 Shane Legg 创造的。但这只是技术本身。我认为它会影响一切，包括科学、经济等方方面面。那个时代就是我们即将经历的。

It is. I used that phrase because it's about the wider milieu around the technology. We call the technology AGI, coined by my co-founder Shane Legg. But that's just the technology. What's going to happen to society, including science, economics, everything, I think it's going to affect everything. That era is what we're about to live through now.

当你谈到 AI 对发现过程的贡献时，比如上周《自然》杂志上发表的两篇论文，其中一篇来自 Google DeepMind，AI 进行了实验并提出了假设。科学发现渗透到普遍使用和接受的速度，是科学的一个优势，因为需要时间来评估事物。你觉得人们还有那个时间吗？还是事情发展得太快，变得有些困难了？

When you talk about AI contributing to the discovery process, for instance, two papers published in Nature last week, one from Google DeepMind, where AI conducted experiments and made hypotheses. The pace of scientific discoveries filtering into general use and acceptance is a strength of science that it takes time to assess things. Do you feel people have that time anymore, or are things moving so fast that it becomes difficult?

我希望我们有更多时间，这样就能更严格地应用科学方法。如果我们能理解我们构建的系统，而不仅仅是黑箱，了解它们运作的深层细节，那会更好。人们正在研究这个，但它落后于性能进步的速度。现在它已经成为一种商业技术，聊天机器人和大语言模型更是火上浇油。我希望科学能赶上工程。我既是一名科学家，也是一名工程师。但市场力量在推动这一点。这对进步来说是惊人的，但尤其是在应用于科学领域时，如果能更慎重一些会更好。然而，在 AGI 到来之前，我们没有太多时间了。我认为现在只有几年时间了。社会已经得到了预警，所以现在是我们所有人更认真对待这件事的时候了，不仅仅是那些构建技术的人。我指的是经济学家、社会科学家，以及社会科学家。AlphaFold 在 2020 年完成，这在 AI 领域已经是古老的历史了。距离 AlphaGo 已经过去 10 年了。

I'd like us to have more time so we could apply the scientific method more rigorously. It would be better if we understood the systems we're building, not just as black boxes, but the deep details of how they work. People are working on that, but it's lagging behind the pace of performance progress. Now it's become a commercial technology with chatbots and LLMs, adding fuel to the fire. I'd love to see the science catch up with the engineering. I'm both a scientist and an engineer. But market forces push that. It's amazing for progress, but especially when applied to scientific areas, it would be better to be more considered. However, we don't have a lot of time before AGI arrives. I think we're only a few years out now. Society has had advance warning, so it's time for all of us to take this more seriously, not just those building the technologies. I mean economists, social scientists, and scientists of society. AlphaFold was finished in 2020, which is ancient in AI terms. It's 10 years since AlphaGo.

我刚从韩国回来，见了总统，庆祝那场比赛十周年。回头看，那确实标志着现代 AI 时代的开端。所以我们是有时间的——对关注的人来说并不意外。我不知道我们作为社会是否明智地度过了那五六年，但确实有一记警钟。很多人以为那是异常或离群值。它确实超前于时代，因为之后五年没有发生其他重大事件。但这些通用系统现在变得非常强大，我们应该认真对待。

I just came back from Korea, meeting the president and celebrating the 10-year anniversary of the match. Looking back, it really marked the beginning of the modern era of AI. So we've had time—it's not a surprise for those paying attention. I don't know if we wisely spent those five or six years as a society, but there was a warning shot. Many people thought it was an anomaly or outlier. And it was ahead of its time, given nothing else significant happened in the following five years. But these general systems are getting really good now, and we should take that very seriously.

谢谢。Alison，你想插话吗？

Thank you. Alison, do you want to jump in?

一个挑战是每个人都想争第一。我们不奖励放慢脚步，尤其是在学术界。如果你不是第一，就发不了论文。我们需要考虑奖励机制，鼓励人们去探索和理解这些方法。这对未来发展很重要。

One challenge is that everyone wants to be first. We don't reward slowing down, especially in academia. You won't get a paper published if you're not first. We need to think about reward systems that encourage people to explore and understand these methodologies. That's important for going forward.

谢谢，好观点。在我们继续之前，先喘口气。你们俩都是皇家学会院士和诺贝尔奖得主。你们坐在一起很合适，而且你们有很长的交情。你们认识很久了。

Thank you. Nice point. Before we continue, let's take a breather. You two are both Fellows of the Royal Society and Nobel laureates. It's fitting you're sitting together, and you have a long history. You've known each other a long time.

是的。我们甚至穿得一样。

We do. We even dress the same.

没错。

Exactly.

我上来时在想，他是国际象棋大师，我们这里有黑子和白子。不过，Demis 可以提醒我具体时间，但我想我们第一次交谈是在你还在剑桥读本科或刚毕业时。我们讨论过读博，但你想创办一家游戏公司。

I was thinking as we came up, he was a chess master and we have the black and white pieces. But anyway. Demis can remind me exactly, but I think we first spoke when you were still an undergraduate at Cambridge or just after. We talked about a PhD, but you wanted to set up a gaming company.

是的，她当时不赞成，我想。

Yeah, which she disapproved of, I think.

嗯，我觉得我应付不来。所以，这差不多 30 年，或者超过 30 年了。

Well, I thought I can't handle that. I can't handle it. So, this is nearly 30 years or maybe over 30 years.

是的，我把 Paul 视为我在生物学上的导师。我们讨论虚拟细胞之类的东西已经 30 年了，Paul。

Yes, and I consider Paul my mentor in biology. We've talked about things like virtual cells for 30 years now, Paul.

是的。我们没取得太大进展，但……

We have. We haven't got very far, but...

它说我们很快就会做到。很快就会。是的，没错。但当时 Demis 显然非常有趣。

It says we will do soon. We will do soon. Yeah, that's true. But it was clear Demis was very interesting at the time.

你能看出来。

You can spot them.

但不是在实验室里。

But not that they're in the lab.

当然。好，我们继续讨论正在发生或将要发生的事情。什么样的问题适合你们的方法？我想听听大家的意见，比如 Isomorphic Labs，一个非常雄心勃勃的重新发明药物发现的项目。你是顾问，Paul，所以你的想法会很好。

For sure. Okay, let's continue with what's happening or what's going to happen. What sort of problem is amenable to your approach? I'd like everyone's input, for instance on Isomorphic Labs, a very ambitious project to reinvent drug discovery. You're an advisor, Paul, so your thoughts would be good.

嗯，我大致解释一下。在 AlphaGo 之后，我意识到我们开发的方法适用于具有巨大组合空间的问题——比如围棋的走法数量，比宇宙中的原子还多，或者蛋白质构象，估计有 10^300 种。你不能用暴力破解。第二，你需要一个明确的目标函数，比如赢棋或最小化自由能。第三，你需要数据——真实数据或精确的模拟器。理想情况下两者都有。对于 AlphaFold，PDB 中的 15 万个蛋白质不够；我们创建了一个早期版本来生成一百万个蛋白质，选出我们确信准确的 20-30 万个，然后反馈回去。使用合成数据时必须小心，确保分布匹配。有了这三样东西，你可以用深度学习模型学习该领域的 世界模型，引导搜索过程，使棘手的组合空间变得可处理。这就是 AlphaGo 和 AlphaFold 所做的——大海捞针。药物发现和 Isomorphic 是这一思路的延续。大约有 10^50 种可能的类药化合物，其中一种或多种可能符合疾病特征。问题在于找到并合成它。Isomorphic 正在开发六种以上类似 AlphaFold 的系统——AlphaFold 是一个组件，蛋白质结构，但只是药物发现的一小部分。我们正在研究生物化学和化学，以预测毒性、ADME 性质等。

Well, I'll explain in general. After AlphaGo, it dawned on me that the methods we developed are useful for problems with a huge combinatorial space—like the number of moves in Go, more than atoms in the universe, or protein conformations, estimated at 10^300. You can't brute force. Second, you need a clear objective function, like winning a game or minimizing free energy. Third, you need data—real data or an accurate simulator. Ideally both. With AlphaFold, the 150,000 proteins in the PDB weren't enough; we created an early version to generate a million proteins, selected 200-300k we were confident in, and fed them back. You have to be careful with synthetic data that the distribution matches. With those three things, you can use a deep learning model to learn a world model of the domain, guiding a search process to make the intractable combinatorial space tractable. That's what AlphaGo and AlphaFold did—finding the needle in the haystack. Drug discovery and Isomorphic are a continuation. There are 10^50 possible drug-like compounds, and one or more may fit the disease profile. The question is finding and synthesizing it. Isomorphic is developing half a dozen more AlphaFold-like systems—AlphaFold is one component, protein structure, but only a small part of drug discovery. We're looking at biochemistry and chemistry to predict toxicity, ADME properties, etc.

我想稍微回溯一下。我是个湿实验科学家，不像我的两位同事那样是 AI 专家。所以，我们可以问，它为我做了什么？

I wanted to go back a bit. I'm a wet scientist, so to speak, and not an AI specialist like my two colleagues. So, what has it done for me, we could ask.

这对我的实验室来说，我们从相当简单的事情开始，机器学习。我们不应该忽视这类事情。我的意思是，我们现在可以在几分钟内分析图像，而这过去需要几天。这节省了大量时间。我们可以在几分钟内完成媒体调查。坦率地说，大型语言模型返回的结果相当无聊，但它能让你起步。所以，这实际上也有帮助。现在，我认为这就是 Demis 所指的，我们如何超越那些相当简单的事情，我不想贬低它们，它们非常重要。顺便说一句，我要说 AlphaFold 非常了不起。我们做实验，尝试想象可能存在的机制，通过 AlphaFold 运行，看看什么可能与其他东西相互作用。它可能成功也可能不成功，但它立即产生想法和假设。这一切都很棒。那么下一步呢？我认为人们正在思考的下一步是如何应用一种连接的工作循环，在实验室中，AI 可以在每个阶段提供帮助。所以，这不仅仅是一种特定的技术，而是融入到我们实际工作的过程中。现在，世界各地都在思考这个问题并尝试实施。但其中存在一些困难，部分在于你需要分析的数据类型。AlphaFold 使用的是非常有用的数据，大量以相同方式收集的数据，而且是在公共数据库中。对于那些质疑为什么我们要做公共资助的科学的人，没有这些数据，AlphaFold 根本不可能实现。

And what it's done for my lab, we start with actually rather simple things, machine learning. We shouldn't just disregard this type of thing. I mean, we now can analyze images in minutes, which used to take us days. It saves an awful lot of time to actually be able to do it. We can do a media survey in a minute or two. It's pretty boring what you get back with the large language models, if I can be blunt with you, but it gets you started. So, that sort of actually helps as well. Where now, and this is what Demis is referring to, I think, is how can we now go beyond those rather simple things, which I don't want to denigrate. They're very important. And by the way, I should say AlphaFold has been amazing, you know. We do experiments. We can try and imagine what mechanism there might be. We can put it through AlphaFold. We can see what might touch something else. It may or may not do that, but it immediately creates ideas and hypotheses. All of that is great. So, what about the next steps? And the next steps I think people are thinking about is how can we apply a sort of connected loop of ways of working in a laboratory where AI can help at every stage as we go around. So, it's not just simply a particular technique, but it's integrated in the process by which we can actually work. Now, various places in the world are thinking about this and thinking about doing it. And I mean, there's difficulties in part of it. Part of it is the type of data you have to analyze. AlphaFold worked with extremely useful data. I mean, immense amounts of data collected in the same way, in a public database, by the way. Should I mean, for those who wonder why do we do public funded science, AlphaFold would have been totally impossible without that. And we

他们后来将其开源了。

They open-sourced it afterwards.

当然。而且，他们做得非常得体。但这是得益于能够访问以完全相同方式整理的公共数据。我认为很多生物数据，虽然 AI 会很有用，但实际上并不是以那种方式收集的。所以我们需要获取并分析这些数据，这可能需要不同的方法。我们拥有的往往是深度，而不是大量相似的数据，就像在不同地方、不同深度的井。我们得想办法把这些东西连接起来。目前，我认为这有点棘手，我们需要找到方法。但我认为关键在于如何连接数据、假设、假设检验、产生想法、再产生更多数据，循环往复，以及 AI 如何帮助我们。在我看来，最终目标是理解生命。我们从细胞开始。我们如何理解细胞？

Absolutely. And by the way, they did it very decently, if I can say that. But it was a consequence of that access to public data curated in exactly the same sort of way. Much of the biological data, where I think this is going to be useful, isn't actually collected in that way. So, we need to get and then this may have views about it, ways of analyzing it, which is of a different type. What we tend to have is, if you like, depth rather than lots of similar data like this, we have sort of wells in different places and different lengths. And somehow we've got to connect all this sort of thing together. And at this moment, I think that's a bit tricky. And we need to get the methods there. But I think the idea is how we can connect all the stages of data, hypothesis, hypothesis testing, generating an idea, back producing more data, and going round that cycle, and how AI can actually help us. With in my view, the ultimate objective of understanding life. And we start there with the cell. How do we understand the cell?

谢谢。我们稍后会谈到构建虚拟细胞，但 Alice，你想就这一点补充吗？

Thank you. We'll come on to building a virtual cell in a minute, but Alice, did you want to chip in on this point or

嗯，我想在我的领域，情况非常相似。如果你拿起一个超声探头进行扫描，机器人做不到，你需要人类的专业知识。所以，现在人们正在思考，既然我们已经解决了这些单独的小任务，如何将它们组合起来并整合每一步。拿起探头、移动、施加一定的压力，有些步骤你可以理解该怎么做，有些则是人类的本能。正是这些细粒度的信息目前很难建模。所以，最终目标是让机器人能够执行所有这些任务并整合在一起。这就是人们正在努力的方向。我认为五年前，我们会说这是一个梦想，现在人们认为这是可能的。

Well, I guess in my space, it's very very similar. If you pick up an ultrasound probe and you want to scan, you can't do that with a robot. You actually require the human expertise. So, now people are very much thinking about having solved these individual little tasks, which is what we start with, is how can we now put them together and integrate every step. So, pick up, you move the probe, you apply certain pressure, and how do we... Some of it you can understand what to do. Some of it humans just do. And it's that fine grain information that is very difficult to be able to model at the moment. So, the ultimate would be able to get a robot to be able to perform all these tasks and integrate them together. And that's what people are working. I think 5 years ago, we would say that's a dream. Now people think that that's possible.

事情发展如此之快，因为现在我们谈论的是 AI 和人类完美协作，整合工作流程，互相帮助。但人们已经开始担心接下来会发生什么。年轻科学家在职业生涯起步阶段，越来越多地问：“当我试图建立自己的实验室时，情况会怎样？地位如何？创造性科学家的角色是什么？”这是一个完全不同的话题，但我想探讨一下。

Things are moving so fast because now we're talking about AI and humans working together beautifully integrating workflows, it helping each other. But already people are worrying about what will happen next. Young scientists coming up through the starting out career paths, I think are increasingly asking, "Where will it be when I'm trying to set up my own lab? What will the status be? And what will the role of the creative scientist be?" And this is a whole another topic, but I'd like to explore it.

首先要知道，在湿实验室做研究非常无聊。真的，很无聊。我们大部分时间都在将少量液体从一个试管转移到另一个试管。这根本无法激发任何人，更不用说他们的智力了。所以，我们可以研究如何让机器人来做这些事。在我工作的克里克研究所，我们有很好的技术核心，但还没有以我们想要的方式整合。但我们有资源和思维方式来实现。我说它无聊，是因为我认为实验室里的大多数同事都不会介意不必一直做这些事。然而，我们需要保持和发展的是创造性思维。所以，我认为重点应该是：“我们如何将这些工具转化为帮助人类在思维上更有创造力？”大型语言模型在思维上并不具有创造性。它们可能会把你推向某些方向。但我认为，与 AI 过程合作，你作为人类思维参与其中不同部分，实际上可能提供一种操作方式，以这种方式思考实验，可能会非常富有成效。因为什么能激发我们？不是将少量液体从一个试管转移到另一个试管，而是创造性的想法，当然，大多数想法都是错的，但你可以证明它们是错的，然后继续前进。所以，这就是让我兴奋的地方，与机器合作，如果我可以这么说的话，来产生创造性思维。

Well, the first thing to realize is doing research in a wet lab is enormously boring. I mean, it really is boring. We spend most of our time transferring little volumes of liquid from one tube to another. And there is no way to excite anybody, let alone their intellect, okay? So, we could work on how we can get robots to do much of that. And in the Crick Institute, where I work, we have very good technical cores, which are not yet doing this in a way that is integrated in what we would like. But we have the resources and the way of thinking that would allow it. The reason I said it's boring is I think most of my colleagues in the lab wouldn't mind if they didn't have to do all of that all the time. However, what we do need to keep and to actually develop is creative thinking. So, I think the focus here should be, "How can we turn these tools into assist the human being to be more creative in thought?" Now, large language models are not creative in thought. They might sort of push you in certain directions. But working, I think, with AI processes, where you as your human mind is involved in different parts of this, may actually provide a way of operating in this sort of way of thinking about experiments, which could be very productive. Because what switches us on? It isn't transferring little volumes of liquid from one tube to another. It is a creative idea, most of which are wrong, of course, and which you can then prove are wrong and move on. So, that's what excites me, working with the machine, if I can put it this way, to generate creative thought.

但是，是的。虽然确实，很多时候你在做移液、观察弯月面时，那是思考实验的好时机。

But, yes. Although it's true that a lot of time when you're doing pipetting and just watching the meniscus and that's a good time to be thinking about your experiment.

回到 70 年代，我甚至不用做那些。我只是观察事物，有很多时间思考。但有一个问题：当我们使用复杂技术时，你所有时间都花在让那该死的技术工作上，你停止思考你试图研究的生物过程。如果技术简单，你确实有很多时间思考。但是，如果机器人来做这些事，也许我们也有足够的时间思考。

Back in the '70s, I mean, when I could did you know, I didn't even do that. I was just looking at things and I had a lot of time to think. There is a problem, though. When we use complex technology, you spend all your time getting the damn technology to work. I mean, you and you stop thinking about the process, the biological process you're trying to study. If the technology is simple, it has to be said, you do have a lot of time for thinking. But, if the robot does it, maybe we have enough time to think, too.

我同意保罗说的。这也是为什么当年你没法把我吸引进湿实验室，保罗。我理解这一点，所以我选择留在数字世界。但你看，我认为如果使用得当，这些技术应该能让我们有更多时间进行创造性思考——就是伟大科学家那种品味：提出假设、选择正确的问题，甚至提出正确的问题。我一直觉得提出正确的问题比解决问题难得多。什么才是真正正确的问题？你如何用非常具体的方式表述它，并围绕它设计控制条件？目前的 AI 系统还完全做不到这一点。

I agree with what Paul said. It's one of the reasons you couldn't attract me into the wet lab back in the day, Paul. I kind of understood that. So I would stay in the digital world. But look, I think if used correctly, these technologies should enable us to have more time for creative thinking—the kind of taste that great scientists have: setting the hypothesis, choosing the right problem, even asking the right question. I always think asking the right question is much harder than solving it. What is actually the right question? And how can you state it in a really concrete way, with the controls around it? Current AI systems have no way of doing that today.

我想支持这一点。就是提出问题，提出正确问题的能动性。实际上，你是否在问正确的问题，那些无穷无尽的问题，有时就这么冒出来。我不认为我们目前的方式能实现这一点。

Want to support that. It's asking the question, the agency of the right question. And actually, whether you're even asking the right question, asking endless questions, which sometimes just come to you from over here. And I don't think the ways we do it can deliver that.

不，目前还不行。我的意思是，总有一天 AI 系统能做到——我不认为这不可能——但肯定不是未来 5 到 10 年内。接下来我们能看到的可能是它会加速迭代。想想看，在你的想法空间里，也许我们可以更像今天的工程学：快速构建原型、测试、然后继续。这将是一种有趣的新方式，把这种方法引入更多科学学科。我认为它还会解放博士生和博士后，让他们从事更高层次的工作。例如，不用再像 20 年前我们做 fMRI 分析那样费力地编写支持向量机，AI 系统会完成那些分析，我们可以用这些时间去思考实验方面——下一个好问题或假设是什么。单个学生能完成的工作量将是惊人的。

No, not at the moment. I mean, one day I think AI systems could do that—I don't think it's impossible—but certainly not in the next 5 or 10 years. Probably the next bit we can see is that it will enable faster iteration. If you think about it, through your idea space, maybe we can be a little more like how engineering is today: you quickly prototype something, test it, and move on. This will be an interesting new way to bring that approach into more scientific disciplines. I think it will also free up PhD students and postdocs to do higher-level work. For example, instead of painstakingly writing support vector machines like we did 20 years ago for fMRI analysis, AI systems will do that analysis, and we can use that time to think about the experimental side—what would be a good next question or hypothesis. The amount a single student will be able to do is incredible.

在科学之外，甚至对今天的年轻人来说，木已成舟——AI 不会回到盒子里。但你能做的，就像我这一代在 90 年代随着家用电脑和互联网成长起来那样，是拥抱它，变得非常精通这些技术，包括理解它们的工作原理。所以我仍然推荐 STEM 学科，因为如果你理解这些工具，你就能更好地使用它们。几年后一个博士生能产出的成果，可能相当于以前整个实验室的产出。对于有进取心、聪明、有干劲、有动力的学生来说，这将是巨大的赋能。而且这些工具遍布全球——就像手机和应用程序一样——所以你可以从任何国家访问一些最先进的技术。这令人兴奋：来自任何地方的最有天赋的学生都能做前沿科学，而今天如果不搬到顶尖中心，这是不可能的。

In general outside of science, even for the youth of today, the cat is out of the bag—AI is not going back into the box. But what you can do, like I did in my generation in the '90s growing up with home computers and then the internet, is lean into that and become incredibly au fait with the technologies, including understanding how they work. So STEM subjects—I still recommend that to students today, because you'll be able to use the coding tools way better if you understand them. The output of what a PhD student might be able to do in a few years is like what a whole lab would have needed before. For enterprising, smart, go-getting, motivated students, that could be hugely empowering. And these tools are distributed everywhere—like phones and apps—so you can access some of the most advanced technology from any country. That could be exciting: the most talented students from anywhere could do cutting-edge science that would not be possible today without relocating to a top center.

至于把裁剪论文的时间用来思考，我认为我们可以用别的东西替代——去乡下散个长步。但我确实认为这一点很重要。我想用我们的 Gemini 聊天机器人——我在 Google 也负责的大语言模型这边——做的一件事是，当这些成为真正能干的助手时（可能还需要几年），我希望我们能少用技术，而不是多用。因为今天我们被那些试图劫持我们注意力的不智能系统轰炸，无论是社交媒体还是别的。我们不得不跳进那股洪流，因为我们需要信息，但如果你有一个为你工作的 AI 系统，你可以委托它说：‘听着，我打算从 10 点到 6 点进行深度思考。别打扰我。在一天结束时总结一下世界上发生了什么。’而且你可以真正信赖它。所以我认为这能保护我们的心智空间和注意力空间。我的梦想是：度过这个阶段，然后拥有智能、个性化、知道你需要什么的技术，这样你就可以把信息收集委托出去，从而有更多时间思考。

In terms of using the paper-cutting time for thinking, I think we can replace that with something else—go for a long walk in the country. But I do think this is an important point. One thing I want to do with our Gemini chatbot, the LLM side of the house that I also run at Google, is that when those become really capable assistants—which we're maybe a couple of years away from—I would love it that we can then use technology less, not more. Because today we're bombarded by unintelligent systems trying to hack our attention, whether social media or whatever. We have to dip into that torrent because we need the information, but if you had an AI system that worked on your behalf, you could delegate to it and say, 'Look, I'm going to do some deep thinking between 10 and 6. Don't interrupt me. Summarize what's happening in the world at the end of the day.' And you could really rely on it. So I think that could protect our mind space and attention space. That's my dream: getting through this period and then having technology that's smart, personalized, and knows what you need, so you can delegate information gathering and have more time to think.

这很好……我只是在想象我的研究生想要离开的样子：‘我现在要去进行深度思考了。’

It's a nice... I'm just trying to imagine my graduate student wanting to leave. 'I'm just off now to do my deep thinking.'

对，就是这样。

Yes, that's exactly it.

Demis 说了很多。你想评论一下工程问题吗，还是作为生物工程教授？

There was a lot in what Demis said. Do you want to comment on the engineering question, or was it as a professor of biological engineering?

我正想说，我觉得他说得很好。在我自己的团队里，我已经看到计算方向的博士后们现在惊讶于他们编程和获得软件辅助的方式。他们都说：他们有更多时间思考，更多时间创造，并提出更先进的解决方案，而几年前你可能还需要几个硕士生才能做到。

I was going to say I think he's said it well. In my own group, what I'm already seeing is postdocs—computational postdocs—who now find it amazing how they can program and get assistance with their software. And that's what they say: they have more time to think, more time to be creative, and come up with much more advanced solutions where normally a few years ago you would have needed a few master students as well.

我们这里需要小心，因为我们确实把这一切都视为训练的一部分。所以，我们不能忘记还需要考虑教育。但我确实认为这些工具非常易用。所以，正如你所说，我们会看到更多由博士生完成的工作。团队的概念——虽然我之前提到过——团队中的人员构成，你仍然需要跨学科来理解如何构建合适的模型。但也许是小团队，而不是大团队。

We need to be careful here because we do treat that all as part of training. So, we mustn't forget that we do need to think about the education as well. But, I do think that these tools are very accessible. So, we are just going to see as you say more done by a PhD student. And the notion of a team is although I mentioned it earlier, the notion of who will be in a team, you still need interdisciplinarity to understand how to build the appropriate models. But, maybe they're small teams and not the large teams.

所以，从你所说的我们可以延伸出很多方向。有一点要提的是，也许就是信息的大量涌现。你提到，这些系统在正确的人手中应该能让人们更有创造力。但我感兴趣的是，你总体上是否认为，目前 AI 在科学领域的推广方式导致了更多创造性的科学，还是说人们做创造性的事情和因为能做、因为资助机构和期刊希望看到新工作方式而做的事情混杂在一起？然后回到你之前提到的出版系统，一个完全不同的问题：如果事情如此之多而资源仍然紧张，我们如何有效地传播科学？所以内容很多。抱歉。谁想回答？

So, there were many directions we could go from you said. One thing to pick up on is perhaps that just the deluge of information that's coming out. Now, you mentioned that these systems should in the right hands be making people more creative. But, I'd be interested in your opinions of whether in general the way that AI is rolling out at the moment in science is leading to lots more creative science or a great mix of people doing creative things and people doing things because you can and because grant awarding bodies and journals are wanting to see new ways of working. And also then coming back to the point you made earlier about the publication system, a whole different question of how we can communicate science effectively if there's just so much more going on and resources are still stretched. So there's a lot there. Sorry. Who wants to?

你。

You.

哦，是的。我想说，我认为这项技术在某些方面已经危险地被诱惑了。如果我现在看看一些高知名度的期刊，他们会发明某种看待问题的方法，并积累大量数据。这很聪明，也很昂贵，所以只有少数人能做到。但它并没有真正得出任何有趣问题的明确结论。事实上，我感觉他们甚至对得出重大结论不感兴趣，因为他们对自己开发了这项技术来做某事感到非常自满。这实际上开始让我担心。有一本叫《细胞》的期刊，曾经是一本伟大的期刊。我现在再也读不下去了，因为里面全是这种东西，而且没有结论，因为他们没有思考这一切意味着什么。我认为我们必须重新引入思考和提问，因为技术的诱惑是一个问题，它不会通过算力解决，因为这实际上是一个人类问题，他们需要思考。这一切的目标是理解，而不仅仅是收集数据。

Oh, yeah. Say something about I think the technology has got dangerously seduced in some ways. If I look at some of our high-profile journals now, they will invent some way of looking at a problem and accumulate a lot of data. And it's clever. It's expensive. So only a few people can actually do it. But it doesn't actually come to very firm conclusions about anything of interest. In fact, I sense they're not even interested in coming to conclusions of great interest because they're so full of themselves of having developed this technology to do something. It's actually something I'm beginning to worry me. There's a journal called Cell which used to be a great journal. I can't bear to read it anymore because it's just full of this stuff and they just have no conclusion because they are not thinking about what it all means. And I think we've got to get thinking and asking the questions back into this because I think the seduction of the technologies is a problem and it won't be solved by the compute because it's actually a human problem that they need to think. The objective of all of this is to understand and not just collect data.

是的，我同意这一点，而且有些科学领域我本来不想点名，但你提到了《细胞》，所以也许我不该说。我认为这是一种练习，就是尽可能多地获取数据，而没有真正思考它如何告诉我们关于功能和其他事情的信息。你知道，我认为早期的连接组学可能就是那样，当我在麻省理工学院做博士后时研究它。我当时想，这到底能告诉我们关于大脑功能的什么？但这显然是一种非常诱人的数据收集方法。我认为有很多这样的领域，我们需要思考科学问题。我以不同的方式看待 AI。我们已经被大数据这个概念淹没了，对吧？如果你还记得，那是 21 世纪初的热词。是大数据。现在的热词显然是 AI，但我曾经把我们在 DeepMind 做的事情描述为：大数据是问题，AI 是解决方案。这是我试图向早期风险投资人描述我们工作的一种方式。他们仍然对我们试图做的事情感到困惑，显然，比如解决智能。这到底是什么意思？但我认为大数据是问题，而且越来越严重，如果你把它看作我们社会创造的信息洪流，显然在科学中，我们有这些大机器、大型强子对撞机等等，它们只是产生大量数据，其中可能有一些洞见，也可能没有，或者一些结构或模式，但可能没有足够多的科学家真正在思考它。所以，首先，我们需要更多科学家来做这件事，也许甚至在他们生成数据之前，但当然，这种复杂性超出了任何人类心智，甚至是我们最聪明的头脑可能理解的水平。而且它们通常处于那些不容易转化为优雅数学方程的领域。所以，这就是为什么我常说 AI 是生物学的完美描述语言，就像数学之于物理学一样。因为生物学，我认为，很多人试图——我不认为我们会得到细胞的牛顿三定律。它太涌现了，太动态了。这并不意味着没有规则，没有东西可以发现和理解，但我认为它更像是一个模拟，这就是为什么我们在考虑虚拟细胞。所以，我认为如果你看得更广，经济学，这样一个涌现的动态系统。显然，以人类为原子单位，而人类已经难以置信地复杂。我们如何真正理解这些类型的系统？如何在这些系统上做科学？我有各种各样的想法，但都涉及使用 AI 来理解数据并找到结构，也许在此基础上构建一个模拟器，就像我们对天气系统所做的那样。这绝对可能吗？我们已经在很多方向上证明了。显然，通过 AlphaFold，但最近还有预测飓风及其方向。英国气象局正在使用它。那非常复杂。过去你需要用超级计算机进行两周的流体动力学计算。现在，我们可以构建一个模型，同样准确，甚至更准确，并在几小时内给出结果，如果你想警告一个国家飓风路径，这是一个巨大的差异，对吧？就像我们去年为飓风梅丽莎所做的那样。所以，我认为我们只是刚刚触及真正可能实现的事情的表面。至于回答你的第一个问题，我认为科学需要时间来适应如何最好地使用这些工具。我认为在很多情况下，这些工具没有被以非常有用的方式使用，或者是以相当平凡的方式使用。仍然有用。再次强调，我不是对此表示怀疑。

Yeah, I agree with that and there's some areas of science which I wasn't going to name them, but then you named cells, so maybe I shouldn't have. I think it is an exercise in let's just get as much data as possible without really thought about how it tells us anything about functionality and other things. You know, I think the early days of connectomics maybe was like that when I was looking at that as a postdoc at MIT. I was like, well, what is this actually going to tell us functionally about the brain? And but it's obviously a very seductive data gathering approach. And I think there's many areas like that where we need to think about the scientific question. It's sort of I see this in a different ways as AI. I we already were being deluged by this idea of big data, right? That was the buzzword if you remember in the early 2000s. It was big data. Now the buzzword is obviously AI, but I used to pitch what we were doing at DeepMind as big data is the problem, AI is the solution. That was one way I used to try and describe what we were doing to early venture capitalists. They were still a bit confused about what we were trying to do, obviously, like solve intelligence. What does this mean? But I think big data is the problem and it's increasingly, if you view it as the deluge of information we're creating as a society, obviously in science, we have these big machines, large hadron colliders, all these things, and it's just producing reams of data where maybe there's some insight in it, maybe not or some structure or some pattern, but it's beyond maybe not enough scientists are actually thinking about it. So, first of all, we need more scientists to do that, perhaps even before they generate that data, but certainly it's sort of the complexity is beyond the level any human mind, even the smartest minds we have can probably comprehend. And they're usually in areas which are not amenable to making into kind of elegant mathematical equations. Right, so that's why I always used to say that AI is the perfect description language for biology, but just like math was for physics. Because biology, I think, a lot of people try to I don't think we're going to get Newton's three laws of motion for a cell. It's just too emergent. It's too dynamic. Doesn't mean there aren't rules and there aren't things to discover and understand about it, but I think it's going to be more like a simulation, which is why we're thinking about virtual cells. And so I think if you think about wider than that, economics, such an emergent dynamic system. Obviously, with humans as the atomic unit, which are already unbelievably complicated. How are we going to actually understand those types of systems? How does one do science on those types of systems? And I have all sorts of ideas about that, but it involves using AI to try and make sense of data and find the structure, maybe build a simulator on top of that like we've done with weather systems. Is this definitely possible? We've shown it in many directions. Obviously, with AlphaFold, but also more recently with predicting hurricanes and which direction they go. The Met Office is using that. That's super complicated. You used to have to have supercomputers doing fluid dynamics calculations for 2 weeks. Now, we can build a model that is as accurate, maybe more accurate, and it'll give you the result back in a few hours, which is a huge difference if you want to warn a country about the hurricane path, right? Like we did last year actually for Hurricane Melissa. And so I think this is just we're just scratching the surface of what really is going to be possible. And just to answer your first question about that, I think there's going to take time for science to adapt to how to best use these tools. I think in a lot of cases, the tools are not being used in very useful ways, or it's pretty mundane ways. Still useful. Again, not to cast any doubt on that.

就像它对处理我们每天必须做的文书工作或解读图像非常有用，但这只是第一步。所以，我们现在就应该做这件事。这显然很有用，但理解如何将这些工具应用于你心中的问题本身就是一个创造性过程。我认为这对商业和科学都适用。现在，过去几年里，CEO 们从没听说过 AI，到认为他们需要 AI。他们经营一家保险公司或其他什么，然后觉得只需要把 AI 应用到某件事上。我经常对他们说，你看，你不需要 AI；你只需要用统计学。如果你试图使用最新的系统之一，它实际上会妨碍那个问题。它根本不是适合所提方法的问题。我认为这本身就是一个品味问题。你不仅能提出正确的问题，还能以正确的方式使用工具来服务于那个问题吗？这取决于科学家。你必须非常了解你的领域，同时也要很好地理解机器学习，或者让理解它的学生来做，然后为那个新世界正确地塑造它。我认为这是所需的新品味的一部分。

It's like it's very useful for dealing with paperwork that we have to do every day or just interpreting images, but that's really step one. So, we should do that right now. That's clearly useful, but it's in itself a creative process to understand how to apply these tools to the question that you have in mind. And I think that holds for business as well as science. Now, CEOs for the last couple of years have gone from not hearing about AI to thinking they need something on AI. And they're running an insurance company or whatever, and they think they just need to apply AI to something. I often used to say to them, look, you don't need AI for that; you just need to use stats. If you try to use one of the latest systems, it would actually get in the way. It's just not the right problem for the proposed method. And I think that in itself is a taste question. Not only can you come up with the right question, but can you use the tools in the right way in service of that question? That's on the part of the scientist. You've got to understand your domain really well and also understand machine learning well, or get students who understand it, and then shape that correctly for that new world. I think that's part of this new taste that is required.

谢谢。我们几次提到虚拟细胞，我认为我们应该提高目标，谈谈虚拟细胞的模拟。这是你们共同的梦想。

Thank you. We've mentioned the virtual cell a few times and I think we should step up the ambition level and talk about the simulation of the virtual cell. A dream that you share.

嗯，我先开始。这很复杂。人们确实尝试过。他们会拿一个简化的细胞，比如细菌细胞，把它减少到 4500 个基因，这样问题就比 4000 个好一点。然后他们试图识别所有反应，建立 400 个微分方程，把它们放在一起，预测一些东西。我的评论是，如果你有 400 个微分方程却无法预测任何东西，那你微分方程学得不好。所以，我认为这从一开始就是无稽之谈。我不想说你不应该做，因为我们应该让人们尝试，我可能错了，但我认为它极不可能成功，尤其是考虑到所有这些动力学几乎肯定是在体外测定的，在体内几乎不会那样工作。我举这个例子只是想说明我们离解决这个问题还差得远。现在，想想我们有什么。我们周围是无序。在细胞内，我们有序，物理学家对此有点担心，因为热力学第二定律，但一旦解释清楚它是孤立的并且你输入能量，他们就失去兴趣了，因为他们看到这并不违反第二定律。但它不仅仅是创造有序，它是在创造有目的的有序。有目的的有序是为了实体——细胞的生存和繁殖，以便自然选择可以发挥作用。这极其困难。我对此很挣扎，因为我看不太清怎么做。我想到的最接近的方法——虽然我不认为这是个好主意，但这是我最好的想法——是看看我们是否能把细胞分成几个域，比如一些黑箱，我们不在乎里面发生了什么。我们只需要知道关键输入和输出，看看这种简化是否能产生一些东西。现在，我的实验室做了一个实验，我认为它可能告诉我们一些东西。我们测量了细胞中蛋白质合成的速率。如果我们有一个培养物并测量它，平均值会完全相同。但如果我们观察群体，变异性非常高。会有很多细胞是正常水平的 50%或高出 50%。现在，我们生物学家认为一切都被严格调控。实际上，并非如此。它非常松散或马虎。此外，如果你有一个钟形曲线，你在外面，10 分钟后你会回到平均值。这很奇怪，因为你会认为蛋白质合成需要数百个反应；它们会平均化并变得紧密，但事实并非如此。它非常松散。我认为这可能告诉我们一些东西，因为我们生物学家过于认为一切都被高度控制。但如果它被高度控制，如果你进入控制空间的一个奇怪部分，你可能永远无法逃脱。你看，我们的电脑多久会出一次奇怪的问题？我们怎么做？我们关掉它再重新打开。我们不知道它为什么出错了。细胞不能那样做。但如果它们松散而马虎，也许它们永远不会卡在错误的地方。这可能全是胡扯，但像这样思考，借助 Demis 正在做的那种东西，是我会尝试解决这个问题的方式。

Yeah. Well, I'll start. This is complicated. People do have a go at it. They'll take a simplified cell, a bacterial cell, reduce it to 4,500 genes, so the problem is better than 4,000. And then they'll try to identify all the reactions, build 400 differential equations, put it all together, and predict something. My comment there is if you have 400 differential equations and you can't predict whatever you like, you're not very good at differential equations. So, I actually think it's nonsense to start with. I don't want to say you shouldn't do it because we should let people try, as I could be wrong, but I think it's exceedingly unlikely to work, especially given that all these kinetics have been determined almost certainly in vitro and almost certainly won't work like that in vivo. I give that example simply to say we are nowhere near dealing with this problem. Now, let's think what we have. We have disorder around us. Within a cell, we have order, and physicists worry about that a bit because of the second law of thermodynamics, but as soon as it's explained that it's isolated and you put energy in, they lose interest because they can see it's not contradicting the second law. But it isn't just creating order; it's creating order with purpose. Order with purpose for the entity, the cell to survive and also to reproduce in a way upon which natural selection can work. This is ferociously difficult. And I struggle with it because I can't quite see how to do it. The nearest I've got, which I don't think is a good idea, but it's the best I could, is to see whether we can divide up the cell into domains, like a number of black boxes where we don't really care what's going on inside. We just need to know the critical inputs and outputs and see whether that simplification might produce something. Now, my lab has done an experiment which I think could be telling us something. We measured the rate of protein synthesis in the cell. If we have one culture and measure it, the mean will be absolutely identical. But if we look in the population, the variability is incredibly high. There will be a lot of cells that are 50% of the normal level and 50% more. Now, us biologists think everything is tightly regulated. Actually, it isn't. It's very floppy or sloppy. Furthermore, if you have a bell curve and you're out here, 10 minutes later you go back to the mean. This is weird because you'd think hundreds of reactions are required in protein synthesis; it would all average itself out and be tight, but it isn't. It's very floppy. I think this may be telling us something because we biologists too much think everything is highly controlled. But if it's highly controlled, if you get into a funny part of control space, you may never escape. Look, how often does our computer do something weird? What do we do? We switch it off and switch it back on. We don't know why it went wrong. Cells can't do that. But if they're floppy and sloppy, maybe they never get stuck in the wrong place. This could be all hooey, but thinking like this with the assistance of the sort of stuff that Demis is doing is the way I would try to approach this problem.

这就是关键。带来这样的思考，或者像你当初做最终获得诺贝尔奖的工作时的思考，你把人类基因撒在不复制的酵母细胞上，这有点疯狂。不，我认为任何认真思考的人都不会那样做，但它成功了。

This is the key. Bringing thinking like this, or thinking like you did when you did your work that was eventually awarded the Nobel Prize, where you just scattered human genes on yeast cells that weren't replicating, which is kind of crazy. No, I don't think anyone thinking about it very hard would have done that, but it worked.

嗯，你的意思是必须有点疯狂。也许你需要从你的模型中获得输入，创造这种疯狂，把疯狂调高。

Well, what you're saying is you have to be a bit crazy. And maybe you need to have an input from your models that creates the craziness, dial up the craziness.

那正是恰到好处的疯狂。

That's just the right amount of crazy.

是的。所以，我认为保罗是对的，当我们第一次讨论这个时，可能 30 年前，我会尝试用启发式方法来做，比如一堆微分方程或其他规则。那是当时构建 AI 系统的方式，比如像深蓝这样的国际象棋计算机。但在我本科期间，我就非常清楚，这种方法无法实现通用智能，甚至无法理解语言或任何那些东西。

Yeah. So, I think Paul's right in that when we first talked about this maybe 30 years ago, I would have tried to do it in a heuristic way like a bunch of differential equations or some other rules. That's how AI systems were built back then, like chess computers like Deep Blue. But it was very obvious to me actually during my undergrad that this was not going to work to get to general intelligence or even understand language or any of those things.

你知道，我们在剑桥学的是，一阶逻辑基本上就是语言作为一阶逻辑规则。这显然是错的，因为我们一半时间说话都不符合语法，但我们仍然能理解彼此，即使不遵守逻辑系统。所以我们显然在做别的事情，至少人脑是这样。后来我在 DeepMind 早期关于学习系统的工作中，以及最终在 AlphaGo 上，完全证实了我的怀疑。你无法直接为围棋系统编写能达世界冠军水平的启发式规则。这不可能，因为围棋太深奥、太直觉、太依赖模式。没有像物质价值那样容易编写启发式规则的东西，因为围棋每个棋子价值相同。所以所有在国际象棋中有效的方法在围棋中都不管用。它必须直接从经验和数据中学习，系统通过自我对弈，找出自己的直觉启发式规则，关于游戏和哪些策略有用。它 famously 发现了人类棋手从未发现的新策略，尽管我们已经下了两千多年围棋。这非常了不起。它不仅能够建模原本难以处理的东西，还有可能超越我们人类目前所知。这说得通，因为用已知方程编程的国际象棋计算机很难超越其创造者的知识。这是让我们非常兴奋的事情之一，现在每个人都意识到了这些现代系统的特点。它们学习，它们是通用的，所以它们有潜力在人类专家的帮助下超越我们当前的知识水平，这对于我们作为物种建造的任何其他工具来说都是前所未有的。你的车不会突然飞起来。当你设计一辆车时，它可能不像你想象的那样工作，但它不会突然做你没有设计它做的事情。所以这是一种非常不寻常的机器，独一无二，这就是为什么我从小就对它着迷。

You can't, you know, we were taught at Cambridge that first-order logic is basically language as first-order logic rules. It's obviously wrong because half the time we don't speak grammatically, and we still understand each other even if it doesn't obey the logic system. So we're obviously doing something else, at least the human brain was. And then I sort of realized and completely confirmed my suspicions with our early work at DeepMind on learning systems, and eventually AlphaGo. You can't program heuristics directly for a Go system that would be world champion level. It's impossible because it's too esoteric, too intuitive, too about patterns. There are no easy things to write heuristics about, like material value, because every piece in Go is worth the same. So all the things that worked in chess do not work in Go. It had to be learned directly from experience and data, with the system playing against itself, figuring out its own intuitive heuristics about the game and what strategies were useful. And it famously discovered new strategies that had never been discovered by human players, even though we've played Go for over 2,000 years. That was extraordinary. Not only can it model something otherwise intractable, it can potentially go beyond what we currently know as humans. That makes sense because a chess computer programmed with equations we already know can hardly go beyond the knowledge of its creators. That's one of the very exciting things we got excited about, and everyone is realizing now about these modern systems. They learn, they are general, so they have the potential, with the help of human experts, to go beyond our current level of knowledge, which was never true of any other tools we've built as a species. Your car isn't suddenly going to fly spontaneously. When you design a car, it might not work as imagined, but it won't suddenly do something you didn't design for. So this is a very unusual type of machine, unique, which is why I've been fascinated since I was a kid.

关于模拟，我们多年来一直在争论的问题是：什么是好的限定系统？这归结为两点。除非你要模拟整个星球直到量子层面，也许有一天我们会做到，否则你必须描述某个孤立的、自包含的系统，并以某种方式近似其外部的一切。第二个问题是：你需要以什么样的粒度来建模系统内部，才能使你的预测有价值？化学是一个很好的例子。我很惊讶在物理学和生物学之间还有化学的空间。我一些最好的朋友是剑桥的化学家，我们经常讨论这个。怎么会有化学这样一个学科，你可以抽象掉物理学和量子物理学，而不像生命涌现系统那样复杂，并且可以独立研究？你可以把它分离或上下近似。仔细想想，这挺神奇的。也许科学其他部分也有类似的方法。我目前的想法，我们还没时间讨论，也许我们可以做类似细胞核的东西。也许这是我和 Zico 正在研究的一个好的起点。但可能仍然太复杂。

With simulations, the issue we've wrestled with over the years is: what is a good circumscribed system? It comes down to two things. Unless you're going to simulate the entire planet down to the quantum level, which maybe one day we will, you have to describe some siloed, self-contained system that you can approximate everything outside of it in some way. The second question is: what granularity do you need to model the system inside for it to be valuable for your predictions? Chemistry is a great example. It's amazing to me that there's room between physics and biology for chemistry. Some of my best friends are chemists from Cambridge, and we talked about this a lot. How can there be a subject like chemistry, where you can abstract away physics and quantum physics, and not get as complicated as the emergent systems of life, and it's studyable on its own? You can separate it or approximate it up and down. It's kind of amazing if you stop to think about it. Maybe there are other ways to do that with other parts of science. My current thinking, which we haven't had time to discuss yet, is maybe we can do something like the cell nucleus. Perhaps that's a good starting point that Zico and I are working on. But maybe that's still too complex.

非常复杂。

Very complex.

是的，可能仍然太复杂。

Yeah, probably still too complex.

这么多方向，时间却这么少。Alison，Demis 提到了一项潜力，即世界各地的人们都能从这些可分布的技术中受益，但你也提到了算力获取的障碍。我们如何确保世界其他地方能够参与，而不仅仅是随波逐流？

So many places to go, so little time. Alison, one thing that Demis mentioned was the potential for everybody around the world to benefit from these technologies that can be distributed, but you also mentioned the barrier to access of compute power. How do we ensure that the rest of the world can participate rather than just be carried along?

人们正在研究这个问题。例如，如果工作依赖于高性能计算机，那么如何以不同的方式进行计算，使其不依赖于大型算力？还有提供访问，以及使用像联邦分析这样的技术，你可以共享算力并汇总结果。但这其中一些取决于你研究的问题，以及构建模型所需的保真度。我认为人们正在通过各种方式推进，这只是技术解决方案的演进。

People are working on this. For example, if the work depends on high-performance computers, then how do you do computation differently so it doesn't depend on large computes? Also providing access, and using techniques like federated analysis where you can share the compute and bring results back. But some of this depends on the question you're looking at, the fidelity needed to build a model. I think there are a variety of ways people are moving forward, and that's just evolution of technological solutions.

是的，我认为我们必须解决计算问题。在这个国家，这也是一个能源成本问题。我们必须获得能源，现在基本上意味着通过芯片和数据中心获得智能。这将是直接相关的。我们拥有世界上最昂贵的能源之一，这对扩大我们所需规模来说很棘手。话虽如此，我认为这是一个创造力和想象力的问题。是的，你需要数百亿美元的算力来构建最新的前沿模型。但如果你在学术界或研究所，你绝对不应该考虑做这个。这毫无意义。你无法跟上那个规模和所需的人才。而且，这不是该做的事。已经有足够多的公司在做——美国有五家，中国有三家。

Yeah, I think we've got to sort out the computing problem. It's also an energy cost problem in this country. We've got to get energy, which basically means intelligence now, via chips and data centers. It's going to be a one-to-one correlation. We have some of the most expensive energy in the world, which is tricky for scaling up what we need. Having said that, I do think this is a creativity and imagination problem. Yes, you need tens of billions of dollars of compute to build the latest frontier model. But if you're in academia or an institute, there's no way you should be thinking about doing that. It's pointless. You can't keep up with that and all the people you need. Also, it's not the thing to be doing. There are enough companies doing that—five in the US, three in China.

就像我们一直在讨论黑箱之类的问题。我跟很多院系都说过这话，效果不一。他们总问我：‘作为顶尖大学的计算机教授，我们该研究什么？’如果我在学术界，我会研究黑箱的分析和理解，压力测试它们的能力和极限，以及如何部署监控工具。有很多事情需要做，比如不需要大量算力的基准测试。因为我会用开源模型，比如 Gemma 4，或者中国模型，它们都很棒。它们只比前沿落后六个月到一年。实际上，去年这个时候它们还具备前沿能力。它们体积很小——比如我们的 Gemma 4 模型，甚至可以在单台笔记本上运行，更不用说小型大学集群了。你可以轻松运行很多实例，中国的小模型也一样。你可以用它们做很多关于 AI 的科学研究，也可以将其应用于其他领域。一年前你根本做不到这些。所以，这有点像错失恐惧症，觉得错过了什么激动人心的东西。但在我看来，最优秀的科学家，如果你相信自己在做的事，就会屏蔽噪音。就像我们在 2010 年那样，当时没人研究 AI。真的没人。所有人都觉得我们疯了。甚至在学术界，这也被认为是死胡同——‘90 年代我们试过 AI，行不通。’但如果你真的相信并热爱你所做的事，就应该能屏蔽噪音，走自己认为有独特优势能做出贡献的路。我认为多学科交叉就是这样的领域之一。我相信这是新时代的伟大之处之一。现在做真正的多学科工作比以往任何时候都容易，因为我们可以用这些工具。你可以用它们在你非专长的领域快速达到合理水平，比以往快得多。所以，我认为跨学科研究有很多事可做，AI 本身的科学也是如此。我只是觉得，人们有点缺乏创造性想象力。

Like what it is, we've been talking about black boxes and other things. I've said this to many departments with varying degrees of success. They always ask me, 'What should we work on here as professors of computer science at whichever top university?' If I were in academia, I would look at the analysis and understanding of the black boxes, stress-testing what they can do, their limits, and how we can put monitoring tools in place. There are so many things that need to be done, benchmarks for that, which don't require a lot of compute. Because what I would do is use open-source models like Gemma 4, or Chinese models, which are very good. They're only six months to a year off the frontier. Literally, this time last year, they would have been frontier capability. They're tiny in size—our model Gemma 4, for example, is built to run on a single laptop, let alone a small university cluster. You can easily run many instances of these things, same with the smaller Chinese models. You could do plenty of great science on AI and also use it for plenty of great science, applying that kind of AI to another field. You literally wouldn't have had that a year ago. So, it's a bit of FOMO, missing out on something exciting over here. But the best scientists, in my opinion, if you believe in what you're doing, you block out the noise. Just like we did back in 2010, when no one was working on AI. Literally nobody. Everyone thought we were mad. Even in academia, it was considered a dead end—'We tried AI in the '90s, it doesn't work.' But if you really believe in what you're doing and you're passionate about it, you should be able to block out that noise and go in your lane where you think you have unique advantages to contribute. I think multidisciplinary is one of those areas. I believe that's one of the great things of the new era. It's never going to be an easier time to do proper multidisciplinary work because we can use these tools. You can use them to get up to reasonable speed on a number of other areas that are not your domain, far faster than you ever could before. So, I think there's plenty to be done interdisciplinary, but also the science of AI itself. I just think there's a kind of lack of creative imagination, I would say.

谢谢。我来问几个问题。很多人对 AGI 感兴趣，想知道系统是否会拥有意识，这意味着什么。这是个非常复杂的话题。我知道你的目标是创造 AGI 并用它解决世界问题。这是你的方向。也许这个可以改天再谈，但现在，有什么让你担心的吗？我们谈了一些小担忧，但什么让你真正担忧？

Thank you. I will turn to some of these questions. There were a lot of people interested in AGI and whether systems will become conscious and what that means. It's a very complicated thing to talk about. I know that you set out to create AGI and use it to solve world problems. That's your direction. Maybe that could be talked about another time, but now, is there anything that worries you about any of this? We've talked about small worries, but what worries you?

嗯，有很多大担忧。显然，我们谈过机遇——我整个职业生涯都在构建这一切，就是为了用 AI 推动科学，尤其是医学。这从 AlphaFold 和 Isomorphic 的工作中就能看出。但有几个担忧。我也可以谈谈意识问题，但两个主要担忧是：第一，恶意行为者将通用技术用于有害目的。这些恶意行为者可能是个人、生物恐怖分子，甚至流氓国家。这是其一。第二是技术性的 AGI 风险。随着系统变得更加自主——这也是我们迈向智能体时代的原因，这只是第一步——它们会变得更自主、更强大。我们能确保设置的护栏足够强大，将系统行为约束在我们最初意图之内吗？这也是一个超级难题。所以这是两个问题。我把它们归为技术问题。除此之外，即使我们解决了这些，还有经济问题：如何尽可能广泛地分享利益，让尽可能多的人和国家受益。最后，如果解决了那个，还会有关于意义和目的的哲学问题。所以有很多要解决的。我仍然非常乐观，因为我坚信人类的创造力。只要我们用心去做，尤其是在压力巨大的时刻，我认为人类总能挺身而出。但首先，我们需要认识到挑战。例如，我很惊讶没有更多经济学家朋友认真对待这个问题，研究后 AGI 时代的经济体系会是什么样子。

Well, there are huge worries. Obviously, we talked about the opportunity—the reason I spent my whole career building this is to advance science, especially medicine, with AI. That's clear from my direct work with AlphaFold and Isomorphic. But there are a couple of worries. I can touch on the consciousness question as well, but the two main worries are: first, bad actors repurposing general-purpose technologies for harmful ends. Those bad actors could be individual rogue people, bioterrorists, up to rogue states. That's one. Second is the technical AGI risk. As systems become more autonomous—and that's why we're moving towards an agentic era, this is the first step—they'll become more autonomous and more powerful. Can we make sure the guardrails we set are strong enough to constrain the behavior of those systems to what we originally intended? That's a super hard problem too. So those are two problems. I classify those as technical problems. Beyond that, even if we solve those, there's the economics problem of how to share the benefits as widely as possible to benefit as many people and countries as possible. And finally, if we solve that, there will be the philosophical question about meaning and purpose. So there's a lot to solve. I'm still very optimistic because I'm a huge believer in human ingenuity. If we put our minds to it, especially in acute times of stress, I think humanity has always stepped up. But first, we need to recognize the challenges. I'm surprised, for example, that there aren't more of my economist friends taking this seriously and working on what a post-AGI economic system would look like.

你认为我们能构建一个护栏根本无法被移除的系统吗？你觉得可能吗？

Do you think we can ever build a system where the guardrails simply cannot be removed? Do you think it is possible?

这是个重大问题。假设一个系统变得比我们更智能——AGI 就会如此——你如何给这样的东西设置护栏？这是一个极其困难但也非常有趣的研究问题。学术界或公民社会可以做的研究之一——实际上，可能比大科技公司自己给自己打分更好——就是这个确切的问题。我猜测，从我们当前系统的可控性来看，我对此有些乐观。但这是个未解问题。它被称为对齐问题。肯定还没解决。任何认为这就像‘开关在哪’那么简单的人，应该去读点东西。不是这样的。这正是阿西莫夫在他的机器人故事里所写和警告的。整个故事就是一个警告：机器人三定律行不通。加上零定律也行不通，因为在某些语境下会被误解。

That's the big question. Assuming a system gets more intelligent than us—which AGI would be—how do you keep guardrails on something like that? That's an extremely hard but also very interesting research problem. As one example of the sort of research that could be done in academia or civil society—in fact, it might be better done there rather than big tech companies marking their own homework—is that exact question. I suspect, just from looking at how steerable our current systems are, I'm somewhat optimistic about that. But it's an unsolved problem. It's called the alignment problem. It's for sure not solved. Anyone who thinks it's as simple as 'Where's the off switch?' should do some reading. It's not. This is what Asimov wrote and warned about with all his robot stories. The whole point was a warning: the three laws of robotics don't work. Adding a zero law also doesn't work because it gets misinterpreted in certain contexts.

但是，在一切发展如此之快的情况下，人们真的有精力去做这项工作吗？是否需要放缓？如果需要放缓，会是什么样子？

But, is there the bandwidth for people to actually do this work while everything's developing so fast? Is a slowdown needed? And if there was a slowdown needed, what would it look like?

这不是我 20 年前想象的技术发展方式，但它一直是我希望 AlphaFold 之类的东西以及推动医学进步所能做到的。但在我看来，如果我们能像 CERN 那样以更科学的方式进行基础研究，那会好得多。然后你仍然可以非常快速地推进应用。比如 AlphaFold，或者你可以在 AGI 到来之前尝试治愈癌症。但直到我们作为一个社会做好准备之前，我们不必处理存在性问题。只是技术并没有那样发展，因为语言机器人、聊天机器人变得可能且非常商业化，这改变了一切。所以现在覆水难收了。我们必须想出其他办法，如何在标准或认证流程方面开展国际合作。而且必须是国际性的，这在当前地缘政治和国际机构碎片化的情况下非常困难。所以这是一个糟糕的汇合点，就在我们需要强大的国际机构和合作的时候，我们却处于那个阶段的低谷。所以我有一些想法。我认为这将是一个需要巧妙处理的棘手问题。

This isn't the way I imagined it 20 years ago in terms of the technologies where it's at, but it's what I always hoped it would do in things like AlphaFold and trying to advance medicine. But it would be much better, in my opinion, if we'd been doing the basic research in a more scientific way, like a CERN-like international collaboration. And then you could still move very fast with the applications. So things like AlphaFold, or you could try to cure cancer before AGI arrived. But we wouldn't have to deal with the existential question until we were ready as a society. It's just that the technology hasn't gone like that because the language bots, the chatbots becoming possible and also very commercial, changed all that. So that's not going back in the box now. So we have to think of some other ways of how we can get international collaboration around maybe standards or certification processes. And it has to be international, which is really hard right now with the geopolitics and the fragmentation of our international institutes. So it's a sort of bad confluence that just at the moment we need strong international institutions and collaboration, we're sort of at the nadir of that phase. So I have some ideas. I think it's going to be a tricky needle to thread here.

抱歉，Demis，我不想一直问你，但既然你提到了意识，人们确实喜欢听这个。就简单问一下，你认为人脑能做的一切都是可计算的吗？

Sorry, Demis, I don't want to keep coming back to you, but since you mentioned you talk about consciousness and people do like to hear about it. Just very quickly, do you think that everything that the human brain can do is computable?

嗯，我一直以来的英雄是图灵。所以我会说，我在大学里很喜欢研究他的图灵机等等。而且我认为证据——我和 Roger Penrose 这样的人谈过很多，显然他会不同意，他认为大脑中有某种量子效应，还有 Stuart Hameroff。我认为到目前为止没有任何证据。我觉得这有点像‘这里有两个我们不完全理解的奇怪事物，让我们把它们放在一起。’但无论如何，我不想——Roger 是个了不起的人，所以他可能是对的。但在我所有的神经科学工作中，我没有看到任何证据。所以我的假设是，大脑中的大多数事物——一切在极限上都是可计算的。这并不意味着——但我认为我们正在进行的这场冒险，我想构建 AI 的原因之一是为了帮助将其用作神经科学的工具，但也可能作为比较器，帮助我们作为一个对照，看看心智有什么特别之处，如果有的话。所以我们之后会谈到这个。

Well, my all-time hero is Turing. So I would say, and I loved studying his Turing machines and all that at college. And I think the evidence—I've talked a lot with people like Roger Penrose and obviously he would disagree and he thinks there's something quantum in the brain, and Stuart Hameroff. I don't think there's any evidence of that so far. I feel like it's sort of 'here are two strange things we don't fully understand, let's put them together.' But anyway, I don't want to—Roger is an amazing guy, so maybe he's right. But I haven't seen any evidence in all my neuroscience work. So my assumption would be that most things in the brain—everything is computable in the limit. That doesn't mean that—but I think this adventure we're on, one of the reasons I wanted to build AI was to help use it as a tool to help neuroscience, but also maybe as a comparator to help us as a control really to see what was special about the mind, if anything. And so we'll come to that.

抱歉，这里的一个问题是我们并不真正知道意识是什么。我的意思是，我们甚至不太清楚我们在——

Sorry, one problem here is that we don't really know what consciousness is. I mean, so we don't even know quite what we're—

这不是一个定义明确的问题。

It's not a well-defined problem.

是的。

Yeah.

但我认为我们有一些方面可能是必需的，但必要不充分，对吧？比如自我意识、身份感这类东西。我的观点一直是 AGI。我们创造 AGI 这个术语而不是用旧术语强 AI——以前叫强 AI 和弱 AI——我不喜欢强 AI 因为它暗示了意识。所以它被混淆了。智能就像一种通用智能与意识。我认为这些是可分离的属性，这是我的猜测。事实上，如果你看动物，我们的宠物狗和猫，它们感觉——它们有很多东西感觉很有意识，但它们不如人类聪明。所以感觉有一个梯度，它可能是可分离的。而且我不觉得至少我们正在构建的工具在任何意义上是有意识的，无论我们用什么定义。我不觉得它们有那种表象，但我认为它们可以。所以我的建议，如果我能挥动魔杖，就是让我们先跨过一条卢比孔河，那就是 AGI 和这些非常聪明、超级能力的工具。然后利用那给我们作为社会争取时间，并利用这些工具来更好地提出意识是什么的问题，用它们做神经科学。我认为有了这些工具，10 年内我们就能做到。然后那可能是社会的下一个问题：我们是否想跨过第二条卢比孔河，真正尝试创造有意识的实体？在我看来，我认为我们不想——那些本身就是足够大的挑战。我们真的想混淆两者并同时跨越它们吗？

But I think we have some aspects of it that are probably required, but necessary but not sufficient, right? Like self-awareness, a sense of identity, these types of things. And my view is and has always been AGI. The reason we coined the term AGI rather than use the old term strong AI—it used to be called strong and weak AI—and I didn't like strong AI because it implied consciousness. So it was conflated. Intelligence like a sort of general intelligence with consciousness. I think those are dissociable properties, would be my guess. In fact, if you look at animals, our pet dogs and cats, they feel—there's a lot of things about them that feel pretty conscious, but they're not as smart as humans. So it feels like there's a gradation and it could be separable. And I don't feel like at least the tools we're building are in any way conscious yet, by whatever we mean by that definition. I don't feel they have a semblance of that, but I think they could. And so my recommendation, if I could wave a magic wand, is let's cross one Rubicon first, which is AGI and these really smart and super capable tools. And then use that to give ourselves time as a society and use those tools as well to maybe better pose this question of what is consciousness, do some neuroscience with that. I think given 10 years with those kinds of tools we would be able to. And then maybe that'll be the next question for societies: do we want to cross the second Rubicon of actually trying to create conscious entities? And in my view, I don't think we want to—those are big enough challenges in themselves. Do we really want to conflate that and cross both of them at the same time?

Allison，我想所有科学家有时都会参与这些关于意识的对话。你想插句话吗？

Allison, I guess that all scientists get involved in these conversations about consciousness sometime. Do you want to chip in here?

嗯，我想我的工作实际上试图避开那个，因为我在应用领域，但当然它会在对话中出现，而且回到科学的背景下，你会如何使用那种能力，它实际上如何帮助你？我认为这也是一个完整的循环。

Well, I guess my work I try and steer away from that actually, being in the applied area, but yes, of course it comes up in the conversation and also how would you use that capability going back in the context of science is how would that actually help you as well? I think it's a full circle one as well.

是的，确实如此。从观众那里得到问题的一个好处是，它们会带你去你未必会去的地方。你已经提到了哲学，但有一个很好的问题是：科学哲学在科学的未来中扮演什么角色？很多科学家之间关于科学走向的对话都没有哲学家参与。所以，有人想——

Yes, indeed. One of the advantages of getting questions from the audience is they take you in places that you wouldn't necessarily have gone. And you mentioned philosophy already, but a nice question is: what is the role of philosophy of science in the future of science? And a lot of conversations between scientists about where science is going happen without philosophers being involved. So, does anyone want to—

嗯，我认为哲学在思考科学方面非常重要，我不得不说。以及什么是知识，什么可以测试，什么不能测试。所以我觉得，正如我已经说过的，波普尔在这里很有用，因为你并没有真正证明某事是正确的，你只是证明了它是错误的。过了一段时间，你有点接受它可能是正确的，因为演绎与归纳等问题。所以我认为你最终会陷入一点纠结。但我们没有正确教授的是思考科学以及如何做科学，这确实有哲学基础。现在，哲学家们有点纠结于事物。我的意思是，我认为我们应该阅读他们，但不要完全被他们驱动。但我们没有行动——很难说我们不知道科学方法是什么。我的意思是，我们可以说一些一般性的东西，比如我们应该收集数据，数据应该可靠、可重复，我们应该挑战自己的想法。

Well, I think philosophy is quite important in thinking about science, I have to say. And what knowledge is and what you can test and what you can't test. So I find, as I've already said, Popper useful here in the sense that you don't really prove something's right, you only prove that it's wrong. And after a while you sort of accept that it might be right because of the problem of deduction versus induction and so on. So I think you do end up in a bit of a tangle. But what we don't teach properly is thinking about science and thinking how you do science, which does have a philosophical basis. Now, the philosophers get a bit tangled up with things. I mean, I think we should read them, but not actually be totally driven by them. But we don't act—it's difficult to quite say we don't know what the scientific method is. I mean, we can say general things, that we should collect data, it should be reliable, it should be reproducible, we should challenge our own ideas.

我们可以列出一系列特征。确实可以。但我认为我们需要教授这一点，因为它有助于我们实践。不过，我强烈反对存在某种黄金标准的科学方法，因为不同领域的方法各不相同。物理学的研究方式与生物学或气候学不同。一些最伟大的物理学家在思考气候问题时束手无策，因为他们认为必须对事物有从 A 到 Z 的完全理解。因此，他们批评气候学，因为你没有做到这一点。这是因为他们试图将自己的科学方法移植到其他领域，罗杰在一定程度上也是如此。所以，这是一个复杂的问题。我们可以应对它，只需要认识并描述它。

We can do a list of attributes. I mean, for sure. But, I think that and I think actually we need to teach that because it helps us in doing it. But, I really do push back that there is some sort of gold standard scientific method of doing things because it differs from area to area. How you do science in physics is different from how you do it in biology or climatology. Some of the greatest physicists in the world were hopeless in thinking about climate because they simply thought they had to have complete understanding from A to Z about something. And so, they criticize it because you weren't producing that. But that is because they tried to translate how they did their science and that's what Roger does to some extent to how people do science in other areas. So, it's a complicated thing. It is something that we can deal with. We just have to recognize it and describe it.

继续。我想说的是，我认为哲学家的时代已经到来。如果我现在是哲学家，这将是史上最激动人心的时刻，因为你不仅需要一种新的科学哲学或更新版的科学哲学——我们触及了所有需要更新的东西。比如，我们应该如何使用模拟或黑箱，如何逆向工程它们？这需要被纳入科学方法。是的，还有，现在什么是理解？因为这些黑箱也可以被还原成数学方程，如果我们想这么做的话。所以也许这是一个两步过程。因此，我认为那里有一种新哲学。更广泛地说，我认为现在正是需要一些伟大新哲学家的时候，比如康德、维特根斯坦或斯宾诺莎，这些我喜爱的古代哲学家。比如，好吧，那么什么是新的伦理哲学、美德和目的，以及我们刚刚讨论的所有这些东西？如果我们把其他技术问题都解决了，这些将是最重要的问题。我认为我们需要一些关于人类处境的新哲学。

Let's keep going. I just want to say on that I think the philosopher's time has come. I would say if I were a philosopher right now, this would be the most exciting time ever because you not only need a new philosophy of science or an updated one because we touched on all the things that need updating. Like, how are we supposed to use simulations, or black boxes, reverse engineering them? That needs to be included as part of the scientific method. Yes, and what is understanding now? Because also these black boxes can be potentially backed out into mathematical equations if that's what we wanted to do. So maybe it's a two-step process. So I think there's a new philosophy there. And more broadly, I think we need the time is now for some great new philosophers like a Kant or Wittgenstein or Spinoza, some of my favorites from the ancient times. Like, okay, so what is going to be new ethical philosophy and virtue and purpose and all these things that we just discussed? If we get all the other technical things right, those are going to be the most important questions. And I think we're going to need some new philosophies of the human condition.

如果是维特根斯坦，我们就一直沉默不语了。

If it's Wittgenstein, we'll just be silent all the time.

好的，最后两件事。这里还有一个问题，是给艾莉森和保罗的。可能来自一位年轻科学家：AI 是否有可能帮助我们改善研究文化，尤其是针对年轻科学家的骚扰？

Okay, two last things. One more question from here, and this one is for Alison and Paul. Possibly from a young scientist, is there a possibility that AI will help us improve research culture and especially the kind of harassment of young scientists?

你说的研究文化是什么意思？

What do you mean by research culture?

我指的是世界各地实验室中的研究文化。你能看到 AI 让年轻研究者更舒心的方式吗？

I mean the culture of research that happens in laboratories around the world. And can you see a way that AI can just make it nicer to be a young researcher?

你知道，这并没有人们说的那么可怕。

You know, it's not so horrible as people make it out.

我们已经热情洋溢地谈了一个半小时，我只是觉得你确实听到……

We've talked about it all in glowing terms for an hour and a half, and I just feel you do hear...

嗯，你看，科学家和其他人一样也会犯错。所以有时他们会行为不当等等。我只是认为我们不应该对此过于纠结，以至于……我确实认为媒体喜欢捏造事实，说他们都在编造等等。这种情况很少见。很少见。所以我们的文化并不太糟。它总是可以改进的。我不想听起来好像……当然可以改进，但我想说，我们不要为此过于自责。

Well, look, you know, scientists are bad just like everybody else. So sometimes they misbehave and so on. I just don't think we should get so tangled up with it that we'll absolutely... I do think the media love making things up, you know, that they're all making stuff up and so on. It's rare. It's rare. And so our culture isn't too bad. It's always improvable. I don't want to sound as if... Of course it is, but let's not beat ourselves up too much over it, I would say.

爱丽丝？

Alice?

嗯，如果他们的意思是压力很大。现在要出成果，他们担心自己的论文可能已经在 arXiv 上发表了。你可以看到当前年轻科学家承受着很大压力。但我想我们也讨论过这一点，他们也许应该停下来反思，思考什么是科学家以及科学家应该做什么。但问题是，如果出版物和简历上的内容一直很重要，那么资深人士就应该站出来说：‘不，我希望你成为最好的科学家，无论发表了多少论文。’并且拿出几篇杰出的论文，而不是考虑工作量。

Well, if they mean in terms of there seems to be lots of pressure. And being to produce results now and they're worrying because on arXiv their work might have been published. You can see that there is a lot of stress on the current young scientists. But I think we've also talked about this to now, you know, they should maybe pause and reflect and think about what being a scientist is and what scientists should be doing. But that is attention that if publications matter and what's on their CV all the time, then it's up to more senior people to stand up and say, 'No, I want you to be the best scientist regardless of the number of publications.' And to come out with a few outstanding papers rather than think about the volume of work.

数量根本不重要。

Volume doesn't matter at all.

正是。

Exactly.

重要的是你一直产出的质量，而不是数量。这太明显了。人们竟然没有意识到，真是令人惊讶。

What matters is quality of what you produce all the time. Not quantity. And it is so obvious. It just is astonishing that people don't realize it, really.

非常感谢。我想重申，你们提交的问题确实为这次讨论提供了信息。所以我希望我至少代表了你们的一些想法。最后，Demis，当 AlphaGo 在 2016 年击败李世石时，他退役了，说‘这不再是同样的游戏了。’我只是想……

Thank you very much indeed. I'd like to reiterate that the questions that you submitted have really helped inform this discussion. And so I hope I've represented at least some of your thinking in this conversation. Just to finish, Demis, when AlphaGo beat Lee Sedol in 2016, he retired from the game saying, 'It's not the same game anymore.' And I just wanted to...

嗯，现在更舒服了。

Well, it's been more comfortable.

正是。但好吧，也许最后一个问题是一个构建的前提。但有了 AI 之后，科学还是同样的游戏吗？有什么改变吗？我希望你们都谈谈这个。

Exactly. But okay, maybe a constructed premise for a last question. But is science the same game with AI attached? Has something changed? And I'd like you all to just talk about that.

是的，我认为会有所改变。你看，我最近去韩国时见到了李世石。和他叙旧真是太好了。他过得很好。他在谈论……我是说，在某种程度上，他对 AI 侵入他的领域有着切身的体验，我认为全世界现在都可以从中学习。他当时也接近职业生涯的尾声了，这只是……我是说，他曾经是……我曾把他描述为围棋界的罗杰·费德勒，他确实是。他是 18 次世界冠军，但他是前十年最强的棋手。仍然处于巅峰，但正接近职业生涯的终点。所以这些事情有点混淆了。但即使对于 AlphaGo 的比赛，我也感到一些苦乐参半和悲伤，尤其是作为一个棋类游戏玩家。我年轻时曾非常专业地学习国际象棋，实际上在比赛前夕我们聊过，我们都开玩笑，他喜欢科技和电脑。如果不是因为生活中的一些不同事件，我们也许可能站在棋盘的两边，对吧？因为我曾一度要成为职业国际象棋棋手。所以，我非常尊重那种艺术、那种技巧和那种精通。毫无疑问，当 AI 系统出现时，它改变了这一点，就像在国际象棋中一样，它改变了关系。

Yeah, I think something will change. Look, I saw Lee Sedol recently when I went out to Korea. It was really nice to catch up with him. He's doing great. He's talking about... I mean, in a way he has a visceral experience of what it's like for AI to encroach on his area and I think the whole world can probably learn from that now. He was near the end of his career anyway, that's just something to... I mean, he was... I used to describe him as the Roger Federer of Go, which is what he was. He was an 18-time world champion, but he was the strongest player of the previous 10 years. Still at the top of his game, but he was getting towards the end of his career. So those things got a little bit conflated. But there was some bittersweetness and sadness for me even with the AlphaGo match, especially as a games player myself. I used to play chess very professionally when I was young and we had a chat actually on the eve of the match where we both joked and he likes technology and computers. We could have been, but for some different events in our lives, maybe we could have been on the opposite sides of the board, right? Because I at one point was going to be a professional chess player. But so, I really respect that art and that skill and that mastery. And there's no doubt that when an AI system comes along, it changes that, like it did with chess and it changes the relationship.

但话虽如此，现在国际象棋比以往任何时候都更受欢迎，人们喜欢观看，但人们并不关心看象棋计算机互相对弈，就像我们仍然关心百米短跑运动员尤塞恩·博尔特一样，尽管我们有汽车等能比人类跑得更快的东西，但我们并不在乎那些。我们想看看我们的人类同胞能做什么。所以，我认为在这些领域，这是其中的一部分。

But having said that, chess is more popular than ever now, you know, to watch and people don't care about watching chess computers play each other, just like we don't care about, you know, we we still care about 100 m uh sprinter Usain Bolt, even though we have cars and things that can go faster than a human can, but who you know, we don't care about that. We want to see what what our fellow human beings can do. So, there's I think in those areas there's there's part of that.

我唯一想说的另一件事，可能与科学相关，就是围棋选手过去常描述——我的意思是，他们现在仍然这样描述——围棋在亚洲被认为不仅仅是一种游戏。它是一种神秘的艺术，宇宙的一些奥秘被认为体现在游戏中，因为它是一种非常美丽的阴阳游戏。当然，最好的围棋选手曾告诉我——我有很多朋友是各种运动的职业选手，也包括围棋——他们想知道游戏的奥秘。但他们真的想知道吗？这是个问题，对吧？所以，我认为科学也是如此。

The only other thing I would say, which is probably relevant to science, is Go players used to describe I mean, they still I think describe it and Go is regarded as more than just a game in Asia. It's a sort of mystical art and um some of the mysteries of the universe are thought to be embodied in the game cuz it's it's just such a beautiful kind of yin-yang game. And um and and and of course, they the best Go players used to tell me and I have a lot of friends who are professional gamers in various sports go as well, that they want to know the mystery of the game. But do they really? Is the question, right? So, and I think that's the same with science.

我有一种永不满足的好奇心，这就是驱动力。这可能是病态的。它驱使我一生，而人工智能——以及我们在座的大多数科学家——都有这种好奇心，否则我们就不会成为科学家。所以，我认为我们努力理解自然法则和宇宙法则，在我看来，就像我个人对现实本质的看法，这些大问题，我们很想知道答案。人工智能肯定会对此有所帮助，但会有代价。

I have an insatiable curiosity, that's what's driven me. It's maybe pathological. It's like driven me my whole life and AI is and most of us in the room who are scientists have that, otherwise we wouldn't be scientists. And so we we're striving I think to understand the laws of nature and the laws of the universe and in my my opinion like my my personal nature of reality, um these big questions, we would love to understand those answers. Um and AI will definitely help with that. But it will come at a cost.

太棒了。非常感谢你，丹尼斯。

Awesome. Thank you very much, Dennis.

我本来想说类似的话，从某种意义上说，人工智能绝对是一个工具，可以帮助你以不同的方式做科学。但这是积极的，不是消极的。

I was going to say something similar in the sense AI is definitely a tool that can help you will be helping you do science in different ways. But that's a positive, it's not a negative.

我认为科学家的驱动力是理解我们以前不理解的东西。我们的动机是对未知的好奇。坦率地说，我们用来支持刚才所说的工具很重要，但我认为它们不会改变科学这一基本方面的性质，即理解世界。

The driver in science I think for um the you know, a scientist is to understand something that we didn't understand before. Our motivation is curiosity about the unknown. Frankly, the tools we use to really back up what's just been said are important, but I don't think they change the character of that fundamental aspect of science, it's understanding the world.

也许我可以以乐观的语气结束。我的意思是，我非常确定在接下来的十年里，我们将进入一个我称之为新文艺复兴的时期。这将是一个科学发现的新黄金时代。我不知道之后会发生什么，也许会有别的事情，一个新的时代。但我认为至少我们将在未来 10 到 20 年经历一个黄金时代，希望当我们回顾时，AlphaFold 只是我们在人工智能帮助下能够破解的一个例子。

Maybe I could just finish on a optimistic note. I mean, I I am for sure very sure about in the next 10 years we are going to be entering what I feel like is a new renaissance. It'll be a new golden age of scientific discovery. I don't know what happened beyond that, maybe something else will happen, there'll be a new sort of era. But I think at least we're going to I think we're going to sort of live through a kind of golden age the next 10 20 years and hopefully when we look back AlphaFold will just be one example of the things that we were able to crack with the help of AI.

非常感谢。谢谢大家。听你们讲话是莫大的荣幸，也很高兴能和这里的观众在一起。我只想说，非常感谢大家。

Thank you very much indeed. Thank you all of you. It's been an enormous pleasure listening to you and it's been a pleasure being here with this audience. I'd just like to say thank you all very much indeed.