AI：与Pushmeet Kohli一同加速科学探索

但我想，从2万种跃升到220万种，我们面临的完全是新的局面了。

So, but I think the fact that there is this new place that we find ourselves in in that from 20,000 to 2,200,000, it's a different ballgame.

确实如此。

Absolutely.

好的。

Okay.

这里还有一个开放性问题。

Here's another open question.

是的。

Yeah.

你认为距离首个由AI开发或AI辅助开发的材料产生重大影响还需要多久？

How long do you think it will be before the first AI developed or or material developed with the assistance of AI makes a big difference?

没错。

Yeah.

所以我认为与生物学不同，目前材料科学更偏向实验性，对吧？

So I I think unlike biology, material science is much more experimental at the moment, right?

药物研发和生物学领域已经运用了计算方法，但材料科学仍以实验为主。

Drug discovery is still and biology still has used computational method, but material science is much more experimental.

因此这需要时间。

So it will take time.

但我认为未来五到十年内，正如我所说，我们将看到计算机模拟材料设计开始产生重大影响。

But I think in the next five to ten years we will see really, as I say, in silico material design starting to make a big impact.

我想把视角再放宽些。

So I wanna zoom out a bit further, right?

因为你看，你现在正信心十足地和我们谈论材料科学，同时也涉及生物学。

Because, okay, so you're sitting here talking to us very confidently about material science, but also about biology.

而你其实是计算机科学家，对吧？

And you're a computer scientist, right?

你是怎么进入这个领域的？

So how come you got into this?

我最后一次学习科学是在学校时期，加入DeepMind是希望能研究通用智能系统并确保其安全可靠。

I the last time I studied science was in school, and I came to DeepMind hoping to work on general intelligence systems and make them safe and reliable.

但我一直着迷于智能系统如何能在现实世界中产生影响。

But I was always fascinated about how can intelligence systems have real world impact.

这正是我过去常与德米斯讨论的话题。

And that's what I used to keep on sort of having discussions Demis with.

有一天他来到我办公室说，哦，普什米特，我觉得有个职位很适合你。

And one day he comes to my office and says, oh, Pushmit, I think we have a good role for you.

我问，好吧，是什么职位？

And I said, okay, what's the role?

我当时猜想可能是某个产品领域，终于能让我应用机器学习了。

And I was assuming some sort of product area that I'm going to finally use machine learning for.

他说，我们想组建一个科学团队，希望由你来领导。

And he said, we want to start a science team and you should lead it.

我问，为什么是我？

And I said, why me?

我对科学领域完全没有背景。

Like science, I have no background.

我是说，在学校学过科学课程。

I mean, studied science in school.

上学时我很喜欢科学，但后来成了计算机科学家，就一直从事这个领域。

I loved it at school, but then I became a computer scientist and then that's what I did.

但他说，你看，你擅长跨学科研究。

But he said, well, see, you are into multidisciplinary research.

你想要理解问题并产生实际影响，还有比推动自然科学知识边界更好的地方吗？

You want to understand the problems and you want to have real world impact and what better place to have impact in pushing the boundaries of knowledge forward in the natural sciences?

于是我说，好吧。

So I said, Okay.

我同意了。

I agreed.

在接下来的几个月里，我逐渐明白自己陷入了怎样的境地——如果有人专门设计一个会让人产生冒名顶替综合征的职位，那这个职位绝对当之无愧。

And then over the next sort of couple of months, I found what I've got myself into because if somebody was designing a job to experience impostor syndrome, this is the job.

因为你将和诺贝尔奖得主共事，试图理解那些难题，而你却毫无头绪。你试图有所作为，却只能从维基百科开始摸索。

Because you will be working with Nobel Prize winners and trying to understand those problems and you don't have no clue and you're trying to make a difference and you'll start from Wikipedia.

从维基百科开始。

Start from Wikipedia.

没错，正是这样。

Yeah, exactly.

你从维基百科起步，然后尝试快速掌握该领域的入门知识。

You start from Wikipedia and then, and try to get a crash course in in the area.

但我想说，科学界真的太棒了。

But I mean, the scientific community is amazing.

总有人愿意坐下来指导你，陪伴你共同探索这段旅程。

People are there, who are willing to sit down with you, are willing to sort of teach you, go on that journey with you.

不过我认为最关键的是，你必须极其透彻地理解这些算法的核心本质。

I guess the really key point though is that you really, really, really understand the heart of these algorithms.

你要真正明白它们擅长处理的系统类型，以及它们能够大放异彩的应用场景。

You really understand the type of systems that they can perform well in and and where they can really excel.

从某些方面来说，你谈到的气象学、材料科学、生物学等不同领域，我确实能看出这些系统设计背后存在技术层面的共性。

And so in some ways, I mean, the fact that you're you're talking about lots of different areas, know, meteorology on the one hand, material science on another, then biology, actually, I can see how there are these behind the scenes, these actual technical similarities between the systems that you're designing.

所以在很多情况下，要做这类工作，就必须具备通才的能力。

So in in many ways, being a generalist, it it requires a generalist to be able to do this kind of stuff.

是的。

Yeah.

完全正确。

Absolutely.

说到智能，究竟什么是智能？

If you think about intelligence, what what is intelligence?

对吧？

Right?

如何创造智能，对吧？

How do you create intelligence, right?

智能不会凭空产生。

Intelligence does not happen in a vacuum.

突然间，你就有了一个能获得直觉并做出新发现的系统。

Suddenly, like, you have a system which gets the intuitions and is able to make these new discoveries.

它是通过经验产生的。

It happens through experience.

那么经验的形式是什么？你的经验有多丰富？

And what form of experience, how rich is your experience?

这对人类也同样适用，对吧？

And like this holds true for humans as well, right?

我们一生中所见的事物数量，给了我们不同的视角。

The amount of things that we see in our life, it gives us a different perspective.

因此对这些模型而言，它们所接触数据的丰富程度至关重要——无论这些数据是科学家在实验室收集的，还是通过模拟生成的。

And so with these models, the richness of the data that they see, whether that data is collected by a scientist in lab or whether that data is generated by simulations.

如果数据足够丰富，你希望数据丰富到迫使模型必须学习通用概念来解释这些数据。

If it is rich enough, you want the data to be rich enough so that the model is forced to learn general concepts to explain that data.

另一个关键要素是要有良好的评估指标。

And then the other sort of element is to have good evaluation metrics.

机器学习方法非常擅长作弊。

Machine learning methods are very good at cheating.

它们有着惊人的记忆力。

They are amazing at memorization.

它们就像那种能记住任何东西的学生，对吧？

They are like the student that can memorize anything, right?

所以如果你给他们大量数据，他们就会记住这些内容，当你提问接近训练集的问题时，他们就会回答'哦对，这就是答案'。

So if you just give them a lot of data, they will memorize it and you ask them any question from the which is close to the training set and they will say, Oh yeah, is the answer.

但他们实际上并没有

But they actually don't

真正理解发生了什么

like Understand what's going

。

on.

没错。

Exactly.

我猜这种情况在蛋白质研究中肯定经常发生。

I guess that must have happened quite a lot with proteins at some point or another.

他们只是看到了文本末尾的峰值数据

They've just seen they've peaked at the back of the text

正是如此。

Exactly.

因此我们必须确保这个测试极其困难，杜绝任何作弊可能。

And so what we had to do is make sure that we made that test extremely hard, made it impossible to cheat.

优秀机器学习系统的真正考验在于其泛化能力——对未见过的样本做出准确预测。

And the true test of a a good machine learning system is its ability to generalize to unseen examples, to make those predictions and things which it had never seen before.

所以理解不同科学领域需要什么样的数据和评估标准，就是问题的一部分，也是我需要与团队共同解决的课题。

And so understanding what is the right data and what is the right evaluation metric for a different scientific discipline is one part of the problem, was one of the things that I needed to work with our teams.

解决方案的另一部分在于机器学习模型的设计——既要从数据中学习，也要将问题相关的先验知识融入模型架构。

And then the other part of the solution is in the design of these machine learning models and these AI models, you want to learn from data, but you also want to bake in any information that you can about the problem into the design of the model itself.

这正是领域知识发挥价值的地方。

And that is where domain knowledge makes a lot of sense.

在非必要情况下，我们没必要从零开始。

You don't want to start from scratch if you don't have to.

因此，在我们所有的项目中，我们最初都采用了多学科的视角。

So in all our projects, we started with a very multidisciplinary viewpoint.

我们邀请了专家，询问他们哪些方法有效、哪些无效，并尝试将所有已知信息融入模型的设计中。

We got in experts, we asked them what was working, what was not working, and tried to inject everything that was currently known into the design of the model itself.

但值得注意的是，你所描述的一切都有一个共同主题——即存在某种成功版本，对吧？

But it's noticeable that the common theme of everything that you've described there is where there is a version of success, right?

你寻找的是一个正确答案，比如蛋白质的真实底层结构、它是否真正稳定，或者你是否真的在预测天气。

There's a right answer that 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.

那么这些项目的绝对关键点是什么？

Is absolutely crucial then for any of these projects?

是的，完全正确。

Yeah, absolutely.

你确实需要清楚了解成功的标准是什么。

You really need to get a have a good sense of what does success look like.

那么在挑选项目时...是的。

So then when it comes to choosing projects Yeah.

因为我的意思是，如果你像自己说的那样是个通才——能凭借领域知识让算法适应各种情况，你如何判断什么值得投入时间？

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?

这是个非常好的问题。

So, yeah, that's a very good question.

我们会从多个维度来考量。

And there are a number of sort of dimensions that we look at.

其中一个维度是数据。

One dimension is the data.

如果没有数据和经验，机器学习模型就无法自主学习——这是重要的考量因素之一。

If there's no data and there's no experience, intelligence would not, the machine learning model will not learn by itself, So that's one of the sort of important considerations.

第二点是我们从一开始就始终考虑的：聚焦根节点问题——那些基础到一旦解决就能影响下游众多应用的问题。

The second thing, and this is something that we have also always sort of 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 on stream.

正如我所说，蛋白质折叠对药物研发有重要影响，比如设计新型酶用于塑料分解等领域。

Protein folding, as I said, had implications for drug discovery, to design new sort of enzymes for plastic decomposition and so on.

材料领域也是如此，对吧？

And the same thing is true for materials, right?

如果能以变革性的方式理解材料发现，这将带来多种多样的应用。

If you can understand material discovery in a transformational way, that has so many different applications.

这比只关注某种特定药物（比如阿尔茨海默症药物）要高效得多。确实如此。

It's much more effective than only looking at, say, one particular drug for Alzheimer's or whatever Exactly.

可能是吧。

It might Yeah.

所以...是的。

So Yeah.

那么你们是否...你们有项目清单吗？

Then do you sort of Do you have a list?

这栋楼里有没有一块白板，上面列着潜在项目清单，然后你们根据这个来评估？

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?

哪些项目有可靠数据？

Which ones have got good data?

哪些问题比其他问题更重要？

Which ones are more important problems than others?

是的，完全正确。

Yeah, absolutely.

我们一直在这样做，同时需要应对很多不确定性。

We are constantly doing that and there is a lot of uncertainty that we have to deal with.

因此我们以科学的方式处理问题。

So we approach the problem in a scientific manner.

我们会进行各种实验。

We sort of conduct experiments.

我们进行探索性研究后发现，是的，我们正在取得进展。

We do exploration studies and see, oh yes, we are making progress.

我们的一些假设是正确的。

Some of our assumptions were correct.

这意味着我们现在可以全力投入解决那个问题。

So that means that we can now make that big commitment to that problem.

我们运作方式的另一个不同点在于我刚才提到的大部分课题。

And the other difference in how we are sort of operating is most of the topics that I've mentioned to you.

我们有非常专注的团队来解决这些问题，他们的工作周期不是六个月或十二个月。

We have very focused teams that pursue these problems, and they're not operating at the level of six months or twelve months.

他们的工作周期是很多年。

They are operating at the level of many years.

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这些课题。

They are dedicating researchers and engineers in those teams are dedicating their whole careers to that topic.

因此我们非常认真地对待选择正确研究课题的责任。

So we take that responsibility very seriously as to which are the right problems that we should be working on.

如果一个问题我们认为牛津、MIT、哈佛、伯克利或帝国理工学院就能解决，我们就不想研究它，因为我们想研究那些真正需要规模化和多学科团队协作才能解决的问题。

Like if a problem can 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 a scale and the type of multidisciplinary team that we have to come together to solve them.

好的。

Okay.

那么让我们换个话题吧，我认为过去几年让世界非常兴奋的是生成式AI，特别是大语言模型。

Well, let's 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.

你们是否已经开始将大语言模型融入科学研究？

Have you started to incorporate large language models into your research for science?

是的。

Yeah.

我们正在全面研究这个方向。

So we are looking at it across the board.

我们主要探索两个主题方向。

I mean, are two main sort of themes that we are exploring.

其一是迄今为止，我们在科学领域的大部分工作都使用结构化数据。

One is that till now, most of the work that we have been doing in the scientific areas, we were using structured data.

这些数据无论是基因组数据还是蛋白质结构预测数据，都包含序列和结构信息。

So this is data, whether it's like genomic data or protein structure prediction data, where you have sequences and structures.

它们是直接相互关联的

They're directly connected to each

没错，它们都是表格形式的数据对吧？

Exactly, they're tables, right?

但大量科学直觉和知识都蕴含在科学出版物和自由格式文本中。

But there's a lot of scientific intuition and knowledge which is embedded in scientific publications, in free form text.

而且

And

如何从这些经验中学习？如何从过去几个世纪该领域关键科学家的研究日志中汲取智慧？

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?

大语言模型让我们能够消化所有这些数据并从中学习。

So large learning models give us the ability to now ingest all that data and learn from all of it.

这就是第一个核心理念。

So that's one key idea.

另一个方向是利用大语言模型在特定领域生成答案。

The other way is basically where you use the large language model to generate answers in certain sort of domains.

这方面的典型案例是我们的算法发现项目，它被恰当地命名为FunSearch（函数搜索）。

And one example of that is our project on algorithmic discovery, aptly named sort of FunSearch, which stands for function search.

FunSearch这个名字确实更胜一筹。

It's much better named FunSearch.

我也更喜欢这个名字。

Much prefer that as well.

是的。

Yeah.

我想我们的团队一定为此感到非常自豪。

I think that the our team was very proud of I bet.

对这个名字。

Of the name.

所以呢，欢乐团队。

So what what The fun team.

没错。

Yeah.

正是如此。

Exactly.

欢乐团队正在研究FunSearch。

The fun team working on FunSearch.

这个模型的功能是尝试为计算机科学中的重要问题发现新算法。

So what this model does is they are trying to discover new algorithms for important problems in computer science.

模型被要求解决计算机科学中的重要问题，比如装箱问题这类概念性问题——给定一组箱子和若干物品，如何高效地将这些物品装入箱子？

And the model is asked, here's the problem, an important problem in computer science, whether it's conceptual problem like the bin packing problem, where how do you, given a set of boxes and some items, how do you compactly pack those items in those boxes?

你可能会想，我在家也会做这种事。

And you might think, well, I do it at home.

这有什么实际意义呢？

What's the relevance of it?

概念性问题无处不在。

Conceptual problem is everywhere.

它体现在快递公司如何将杂货配送到你家，或是云服务提供商如何在他们拥有的不同计算机上调度计算任务。

It's in how delivery companies go and deliver your groceries to your homes or how cloud providers like schedule computational jobs on different computers that they have.

因此这在实际应用中是个极其重要的问题。

So it's it's a very, very important problem in the real world.

这个模型的作用是尝试提出新的算法。

And so what this model does, it then tries to propose new algorithms.

其中许多算法可能是它之前见过的知名算法。

And many of those algorithms are maybe well known algorithms that it had seen before.

我们告诉它：这样没问题，但要试着改进这个特定部分并不断优化。

And we tell it, well, that's fine, but try to improve this particular part and try to refine it.

它会持续优化过程，有时会犯错，我们就指出：这里你犯了个错误。

And it goes on, sort of trying to refine it, and sometimes it makes a mistake and we tell it, oh, here's the mistake that you have made.

然后它继续这个循环。

And it keeps on doing that.

我们努力获取正确解决方案并反馈给它。

We try to get the right solutions and feed it back.

在这个过程中，有时它会发现前所未有的新事物，最终提升性能或提出全新启发式方法，以截然不同且更高效的方式解决问题。

And in that process, sometimes it discovers something completely new that was not known before and ends up improving the performance or coming up with new heuristic or a new algorithm which solves the problem in a remarkably different and much more efficient manner.

所以它能做到这点，是因为知识间存在我们未必能察觉的深层关联网络。

So the reason it's able to do that then is because there's this this deeper network of of connections between knowledge that that isn't necessarily visible to us.

这就是为什么大语言模型会产生看似合理的幻觉吗？

Is that why sometimes large language models will hallucinate in a way that sort of makes sense?

比如有人可能会说居里夫人发明了青霉素，

So I don't know, maybe one would say Marie Curie invented penicillin.

而

And,

然

course,

她并没有，但她的确在同一时期做出了重大发现。

she didn't, but she did come up with a discovery that was really important around the same time.

所以它能够交换信息片段，因为它看到了我们无法察觉的深层关联。

So it sort of it it can swap over pieces of information as it were because it's seeing deeper connections that aren't visible to us.

确实如此。

Absolutely.

这主要是因为它在潜在空间中运作，能感知到某些事物之间存在关联。

It is basically because it's operating in that latent space, it feels that there are certain things which are related to each other.

以Marikuri的例子来说，那是个错误，对吧。

And in the case of the Marikuri example, that was a mistake, right.

如果人们不知道这个错误，那就会造成问题，对吧。

And that's a problematic mistake if somebody did not know about it, right.

这是错误信息。

It's incorrect information.

但在我们的案例中，这个错误无伤大雅，因为我们有评估函数——它与基金搜索系统结合，又与能识别胡言乱语的大语言模型联动。

But in our case, that's a fine mistake to have because what we have is an evaluation function which is coupled with fund search, which is coupled with the large language model, which can call bullshit.

就像个真相检测器。

Like a truth detector.

没错，正是如此。

Yeah, exactly.

它能快速识别出不合逻辑的内容并指出问题。

Which can sort of call bullshit and say, oh, doesn't make sense, right, very, very quickly.

所以关键在于创造力。

So then what matters is creativity.

如果它产生了极具创意的内容，我们就会采纳这部分并予以肯定。

So if it comes up with something completely creative, we sort of pick that part and say, oh, yeah.

很棒。

Great.

至于它产生的糟糕内容，我们能轻松过滤掉。

And the the the bad stuff that it says, we we are able to filter it out very easily.

但偶尔它提出的创意内容，会揭示出底层知识网络中更深层次的东西。

But the every now and then, the the creative stuff it says is is revealing something more in more about the underlying network of knowledge.

确实如此。

Exactly.

没错。

Right.

这基本上就像是你把幻觉用于了好的一面。

It's basically like you've harnessed hallucinations for good.

是的。

Yes.

在这个特定案例中，幻觉是有益的。

Hallucinations are good in this particular case.

如果你能设法利用创造力，保留有益部分并过滤掉所有无效内容。

If you can somehow leverage creativity and sort of keep the good part and filter out all invalid part.

所以，我是说，别让创意团队知道他们可以自称'有益幻觉团队'。

So, I mean, don't let the fun team know that they can call themselves the the hallucinations for good team.

是啊。

Yeah.

现在可能会得意忘形。

Now might get ahead of themselves.

那么它发现的是计算机科学家或数学家们之前不知道的东西吗？

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

是的。

Yes.

完全正确。

Absolutely.

事实上，用这种方法，我们能在计算机科学领域获得新成果。

So so in fact, this this method, with this method, we were able to get a new result in computer science.

有个非常有趣的问题叫CAPSET问题，它试图在结构化图中寻找独立集合。

So there is a very interesting problem called the CAPSET problem, which is trying to find independent sets in a structured graph.

这就像一个计算问题，你有一个图，试图找到具有特定属性的某些节点和边。

It's like a computational problem where you have a graph and you're trying to sort of find certain nodes and edges with certain properties.

而且这个问题已经被研究很长时间了。

And it has been studied for a long time.

这是计算机科学中一个非常有趣的问题。

It's a very interesting problem in computer science.

而FunSearch首次实现了一个前所未有的成果。

And, what FunSearch was able to do was for the first time produce a result which nobody had been able to produce.

它不仅得出了那个结果，其生成的求解程序还包含极其有趣的子结构和直觉——当与我们合作的数学家看到时，他们认为程序提取出了问题中新的对称性。

And not only was it able to produce that result, but the program that had written, that it had generated to find that result had extremely interesting substructure and intuitions that when mathematicians who were working with us saw it, they felt that the program had extracted a new symmetry in the problem.

哦，所以它在这种幻觉中...嗯。

Oh, so it it in this hallucination Yeah.

它确实偶然发现了未被探索的领域...是的。

It had really stumbled upon something that that had been unexplored Yeah.

但结果证明是正确的。

But turned out to be true.

是的。

Yeah.

而且它利用了问题中某个非常有趣的特性——这个特性我们从未告诉过它。

And it and and it it was leveraging some very interesting property about the problem that we had not told it about.

哇。

Wow.

是啊。

Yeah.

我能理解这为什么会有用。

I can see how that might be useful.

好的。

Okay.

讲讲奥林匹克竞赛的事。

Tell about the Olympiad.

嗯。

Mhmm.

国际数学奥林匹克竞赛是一项学生参与的数学竞赛。

So there is the International Maths Olympiad, which is competition that students take part in.

全球最优秀的学生都会来参加这个比赛。

Some of the best students all around the world come to this competition.

题目难度非常高。

It has very sort of hard maths problems.

如果表现优异，可以获得铜牌、银牌或金牌。

And if you perform well, then you can get a sort of a bronze, silver or gold sort of metal.

国际数学奥林匹克的题目难度极大。

The level of problems that are asked at the International Math Olympiad are extremely hard.

比如当前的人工智能系统，甚至都无法解决这类问题。

The current generation of AI systems, for example, even they are not able to sort of tackle those problems.

尽管面向中小学生，但这些题目需要极强的横向思维能力、逻辑能力以及对数学概念的深刻理解。

They require an incredible level of sort of lateral thinking, of logic, of deep understanding of mathematical concepts, despite the fact that they aimed at school children.

没错，完全正确。

Yeah, absolutely.

这不是普通的游戏。

They not your normal games.

不是普通的游戏。

Are not your normal games.

它们就像是数学界的巅峰对决，

They They're are like the Demises of the world, do

正是如此。

know what Exactly.

确实如此。

Exactly.

我不确定Demise是否参与了，但他们确实出类拔萃。

I'm I don't know whether Demise participated in But, they are exceptional.

对计算系统而言，解决这类问题一直是长期挑战，因为它们属于极其困难的数学难题。

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.

与象棋或围棋不同，数学是一个开放式领域，因为它没有固定的步骤限制。

And 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.

可能性是无限的。

The moves are infinite.

而且方向任意。

And in any direction.

方向也是任意的。

And in any direction.

因此你需要在一个极其庞大的推理空间里思考。

So it's an extremely large sort of space that you are reasoning over.

即便最先进的人工智能系统也无法处理这类问题。

And so the most sophisticated AI systems are not able to tackle any of these sort of problems.

所以我们开发了Alpha Geometry系统，首次证明AI能解决国际奥赛级别的几何题。

So we had a system called Alpha Geometry, which for the first time showed that an AI system can solve geometry problems that are at the International Mass Olympiad level.

这该怎么形容呢？

So this is like, I mean, how describe can them?

你会看到一些由圆形、三角形和正方形等组成的图形。

So you sort of will get a picture of some circles and triangles and and squares and things.

对。

Yeah.

然后题目会要求你仅凭对几何规则的理解，从图中计算出某些看似不可能的结果。

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.

是的。

Yes.

差不多是这样。

Something like that.

对。

Yeah.

这真的非常非常难。

It's really it's really hard.

没错。

Yeah.

这确实相当困难，你不仅要非常精通几何，还需要能够提前规划，思考什么样的解法能带你找到最终答案。

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

那么Alpha Geometry是如何工作的呢？

So how does alpha geometry work then?

Alpha Geometry的工作原理是，它将这个问题——有时以文本形式给出，有时以图像形式——转换成一种形式化语言，即它自己的领域特定语言，以便进行推理。

So alpha geometry, how it works, it transforms this problem, which is given in sometimes in text, sometimes in image, into a formal language, into its own domain specific language in which it can reason about the problem.

然后它尝试用这种语言来解决问题。

And then it, tries to solve the problem in that language.

Alpha Geometry的做法非常聪明。

Now, what the Alpha Geometry did was very smart.

他们生成了数量极其庞大的合成问题及其对应解法。

They generated a very, very large number of, problems, synthetically generated problems in that language and the the corresponding solutions.

通过这种方式，他们可以用数十万个这样的问题来训练机器学习模型。

So with this, they they could train the machine learning model on hundreds of thousands of these problems.

对吧？

Right?

这使得它在面对同类新问题时，能够极其高效地解决。

And then that made it extremely effective in given a new problem of that form, it could solve it.

它只是根据自己的经验来认知。

It just knows from its own experience.

是的。

Yes.

太棒了。

Amazing.

实际上，可以再次看到它与材料科学之间的相似之处，就像，好吧，我们就用组合数学来...对。

And actually, can see the similarities there again between that and the and the material science thing is like, okay, let's just use combinatorics to Yeah.

正是如此。

Exactly.

我们就像掷骰子一样，一开始提出大量随机想法，然后不断筛选、缩小范围，直到最终能针对单个问题说：我知道怎么解决。

Let's just like roll the dice, come up with loads of random things in the beginning and then narrow it down and narrow it down and narrow it down until eventually it can look at a single problem and say, I know how to solve.

没错。

Yeah.

完全正确。

Exactly.

我真的很羡慕你的工作。

I'm quite jealous of your job.

好吧，我们已经跑题跑得够远了，你接下来希望解决什么问题？

Okay, we have gone on a wild list of topics here, but what are you hoping to tackle next?

是的，我认为还有很多未解决的问题。

Yeah, so I think there are many problems that remain unsolved.

在我们研究的任何领域，无论是理解蛋白质、基因组、天气还是材料科学，都有大量工作亟待完成。

In any area that we're working on, whether it's understanding proteins, whether it's understanding the genome, whether it's understanding the weather, whether it's understanding material science, there is just so much work that still remains to be done.

我们刚才在讨论材料。

We were talking about materials.

我们现在只能预测这些材料的稳定性，但如何进一步预测它们的特性并实现合成呢？

We are only making predictions about stability of these But how do you extend that to making the prediction about properties of these materials and then synthesizing them?

就像完成这段旅程还有大量工作要做。

So it's like there is so much work that remains to be done to complete the journey.

不过你有特别喜欢的吗？

Have you got a favorite though?

有没有哪一个让你觉得‘就是它了’？

Is there one that you're like, that's the

我真正想要的那个？

one I really want?

没有。

No.

你不能问我那个

You can't ask me that

问题。

question.

我不能。

I can't.

我觉得全都没有。

I think all No.

并非所有‘孩子’都平等。

Not all not all your babies are equal.

你心里肯定有个最偏爱的。

You've got a you've got a picker favorite.

作为计算机科学家，我热爱计算机科学工作，但每个项目都有不同的闪光点，在某些时刻会让我觉得‘哇，这太神奇了’。

Like, I'm a computer scientist, I I love the computer science work, but, like, all of them have different elements which are on the day, they they are like, wow, this is so amazing.

比如当你试图学习控制聚变反应堆磁体的策略时，实验成功那天，无论其他事情如何，我都会

Like, if when you are trying to learn a policy that is going to control the magnets of a fusion reactor, the day that that experiment happens, regardless of what's happening anywhere, I'm You

最关注那个。

care most about that.

是的，是

Yeah, Is

你认为哪一个会产生最大的影响？

there one that you think though will have the biggest impact?

我认为我们在理解生物学、化学和材料学方面所做的工作，这些基础研究至关重要。

I think the work that we are doing in understanding biology and in understanding chemistry and materials, think it's so fundamental.

这类根本性问题。

Such root node problems.

是的，这些都是根本性问题，甚至很难限定其影响范围，对吧？

Yeah, there are such root node problems that it's difficult to even sort of limit what the impact would be, right?

它们的影响难以预测。

Their effects are difficult to predict.

是的，完全同意。

Yeah, absolutely.

我非常期待下次再与你交流，看看又有什么重大突破。

I'm very much looking forward to coming back and talking to you again to see what other massive things have been happening.

Pushmeat，非常感谢你的参与。

Pushmeat, thank you very much for joining me.

谢谢。

Thank you.

与Pushmeat会面后让我感触最深的是：科学发展看似缓慢，

I'm struck by after meeting Pushmeat is that the thing about science, it moves really slowly.

但偶尔——可能一代人才能遇到一次——会出现颠覆性突破，

And then every now and then, maybe once in a generation, you get this seismic shift.

这种重大飞跃会彻底改变我们的认知。

You get a big step forward that fundamentally changes our understanding.

而人工智能应用于科学研究的特殊之处在于，它不会仅影响物理学或宇宙学中的某个方程，

But the thing about this shift of using artificial intelligence for science is that this isn't just gonna make a difference to physics or or one equation in cosmology.

正是这件事将以最根本的方式改变整个科学界。

It's that this is going to transform all of science and in the most fundamental way.

好吧，我知道我现在听起来像是被炒作冲昏了头脑，但别只听我的一面之词。

And, okay, I know that I sound like I've just swallowed the hype here, but don't just take my word for this.

去问问你身边友好的科学家们，看看他们对正在发生的事情有何看法。

Go and ask your friendly neighborhood scientists what they think of what is happening here.

因为真正理解这件事对幕后世界意味着的人们，根本无法用语言形容这场变革的规模——无论是已经发生的还是即将发生的。

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.

您正在收听的是由我——汉娜·弗莱教授主持的《谷歌DeepMind》播客。

You have been listening to Google DeepMind the podcast with me, professor Hannah Fry.

如果您喜欢本期节目，请订阅我们的YouTube频道，您也可以在任意播客平台找到我们。

If you enjoyed that episode, do subscribe to our YouTube channel, and you can also find us on your podcast platform of choice.

本播客还有更多精彩内容待您探索，包括关于人工智能如何提升教育的深度探讨。

Now we have lots more to explore on this podcast, including a deep dive into how AI could enhance education.

此外，我们还将讨论：是否应该赋予AI助手人类特征？

Plus, should AI assistance be given human traits?

以上内容即将在《谷歌DeepMind》播客中呈现。

That is coming up soon on Google DeepMind, the podcast.