本集简介
双语字幕
仅展示文本字幕，不包含中文音频；想边听边看，请使用 Bayt 播客 App。
回到2023年，我在纽约石溪参加一个关于暗物质的会议。
So back in 2023, I was at a conference, a conference on dark matter, in Stony Brook, New York.
有一天晚上，和会议上的几个人一起吃晚饭，我们一群人热烈地讨论起了AI在物理学中的作用。
And over dinner one night with a group of the people at the conference, a bunch of us got into a pretty energized conversation about the role of AI in physics.
那时候，差不多四年前吧，我的大多数同事都对AI会在我们的研究中变得非常重要持怀疑态度，至少在中短期内是这样。
At that time, almost four years ago or so, most of my colleagues were pretty skeptical that AI would be very important in our research, at least in the kind of short to medium term.
但那时我已经有了不同的看法。
But I already had a somewhat different view by that point.
我刚写完一本关于粒子宇宙学和天体物理学的研究生教材。
I had just finished writing a graduate level textbook on particle cosmology and astrophysics.
出于好奇，我把那本书里的一些作业题输入到了ChatGeBT里，那时候它还算是个新工具。
And I had out of curiosity fed a bunch of the homework problems from that book into ChatGeBT, which was still kind of a new tool then.
你知道，它表现得不算特别好，但很多问题它都答对了。
And, you know, it didn't do great, but it got a lot of the questions right.
考虑到当时大型语言模型进步的速度，我觉得用不了多久，它就能在这些任务上至少和人类做得一样好。
And considering how fast, you know, large language models were improving at that time, I thought it was just a short amount of time before it would do at least as well as human beings on these sort of tasks.
当时我的大多数同事都持反对意见，包括当晚一起吃饭的那几位。其中一位，至少在一定程度上，是我的朋友兼同事约尼·卡恩。
Most of my colleagues at that time disagreed, including those at the dinner, and one of them, at least to an extent, was my friend and colleague, Yoni Kahn.
约尼现在是多伦多大学的物理学教授。到了2023年那晚结束的时候，我们俩已经打了一个赌注不小的赌。
Yoni is now a professor of physics at the University of Toronto, And by the time that night was over back in 2023, the two of us had made a fairly high stakes wager.
如果到2026年的某个时候，一个公开可用的AI能通过约尼的研究生级别量子场论期末考试，我就赢；否则他赢。
I would win if a publicly available AI could pass Yoni's graduate level quantum field theory final exam by sometime in 2026, and he would win if it could not.
顺便提一句，约尼其实是个很厉害的圆号演奏家，而我吉他弹得也不错。
As an aside, Yoni turns out to be an accomplished French horn player, and I'm a pretty good guitarist.
我们俩都清楚这一点，所以赌注定得挺狠：输的人得学会弹赢家的乐器，还得在赢家选的地方公开表演。
We both knew this, so we made the stakes of the bet that the loser would have to learn to play the winner's musical instrument and publicly perform it at a venue of the winner's choosing.
考虑到要练成那样得花多少个小时，这赌注可真不小。
Considering this is a lot of hours of work it would have taken to do this, it was a really significant wager.
当时看来这赌注可能挺傻的，但你也知道，那是喝了几杯酒之后说的话。
It was probably a dumb wager to make at the time, but you know, it was a few drinks at night.
完全可以理解。
So understandable.
然后我们等了三年，今年早些时候我们揭晓了赌约的结果，是我赢了。
And then we wait three years and earlier this year, we found out who won the bet and it was me.
我觉得差距还挺大的。
I think it wasn't even close.
现在，任何主流的大语言模型都能轻松通过我们能拿出来的任何物理考试，不管难度如何。
Today, any of the major large language models can pass essentially any physics exam we can throw at it, at any level.
几年前，它连本科入门级的考试都勉强及格。
A few years ago, it could barely pass an undergrad intro exam.
如今，它能在最顶尖的学科里拿满分。
Today, it can ace the most advanced subject matter.
今天在这档节目《为何是这个宇宙》里，我要跟我的朋友约尼·卡恩聊聊，AI 在物理学领域现在能做到什么，以及我们觉得它未来很可能具备哪些能力。
Today on Why This Universe, I'm gonna talk with my friend, Yoni Kahn, about what AI can do in physics today and what it's likely or what we think it's likely to be able to do in the future.
我们会探讨这能让我们在对宇宙的认知上实现哪些突破，也会聊聊在科学这个宏大事业中，机器和人类各自大概会扮演什么样的角色。
We'll talk about what this might enable us to do when it comes to learning things about the universe, and we'll, you know, talk about what roles machines and human beings are likely to play in the grand endeavor we call science.
你正在收听的是《为何是这个宇宙》，一档致力于拆解物理学核心概念的播客节目。
You're listening to why this universe, a podcast where we break down the biggest ideas in physics.
我是丹·胡珀。
My name is Dan Hooper.
欢迎你，约尼·卡恩，来到《为什么是这个宇宙》。
Welcome, Yoni Kahn, to Why This Universe.
非常感谢。
Thanks so much.
做这个节目真的很有趣。
It's a lot of fun to do this.
所以，对于那些不认识约尼的听众来说，他是一位高能理论物理学家，跟我有点像，他也研究宇宙学和其他类似的问题。
So for those of you who don't know, Yoni is a high energy theorist, kind of like me, and he works in cosmology and other problems like me.
但最近，他非常投入于AI在物理学中的应用。
But lately, he's been really engaged in physics applications of AI.
而且他是个很有意思的人，知道很多有趣的事情，对这些话题也有独到的见解。
And he's just an interesting guy and knows a lot of interesting things and has interesting perspectives on these sorts of things.
所以我觉得他是今天聊这个话题的绝佳嘉宾。
So I thought he'd be a good guest to talk about this today.
尤尼，我们回到过去吧。
Let's go back in time, Yoni.
回到2023年左右，那时候大语言模型还算新鲜事物，你可能有过类似的对话。
Let's go back to that conversation or conversations you might have had like it back in around 2023 when LLMs were still a pretty new thing.
你当时对人工智能及其在物理学中的作用有什么想法？尽量回忆一下。
What were your thoughts, the best you can recall, about AI and its role in physics at that time?
我觉得很难不把现在的想法带进去，但我会尽量不搞修正主义历史。
So I think it's hard to do that without imprinting on my answer what I currently think about it, but I will do my best to not make a revisionist history about it.
我记得那时候，我在哥伦比亚大学做过一个讲座，有人给我看了弗雷德里克·德·内夫——他是哥伦比亚大学的教授——和ChatGPT关于量子场论的聊天记录。
I think around that time, I had given a talk at Columbia and somebody had shown me some chat transcripts that Frederic De Neff, who's a professor at Columbia, had had with ChatGPT about quantum field theory.
那让人不安，因为它似乎在鹦鹉学舌地说一些物理相关的东西，但你越看越发现那些内容不太对劲。
And it was disconcerting because it seemed to parrot physics ish things where the more you looked, the more you realized that wasn't quite right.
就像有人学外语，大概能发出所有正确的音，但顺序不对，听起来就不通。
It was sort of like somebody trying to learn a foreign language by kind of making roughly all the right sounds even though the sounds didn't quite make sense in order.
这很让人困扰，因为如果一个人不理解那些符号串，是不可能把它们按正确顺序排列出来的。
And it's disturbing because no human being would be able to put those strings of symbols together in anything resembling the correct order they didn't already understand.
所以我觉得，对我来说，AI很明显在语法知识层面有一定水平，但这些知识未必能转化成我们在物理学中那种更抽象的计算和交流方式。
So I think what was clear to me about AI was that there was obviously a level of, like, syntactical knowledge that wasn't necessarily being translated into the somewhat more abstract way that we calculate and communicate in in physics.
说实话，当时我真的不太确定它的能力到底有多强。
And so to be honest, I I wasn't really sure what the capabilities were at that point.
但三年前就很明显了，如果你真的想解决一个难题，一个作为专业物理学家都不容易搞明白的问题，那时候的ChatGPT还帮不上什么忙。
But it was clear three years ago, like, if you really wanted to know the answer to a hard problem, a problem that you couldn't, get to the bottom of easily as a professional physicist, you wouldn't find much utility in ChatGPedia yet.
我觉得没错。
I think that's correct.
我觉得即使在那时候，我们也不会用它来做开放式的头脑风暴，因为它太不可靠了。
I think even back then, weren't really using them for open ended brainstorming yet because it was sort of too unreliable.
那现在呢？
And what about today?
现在已经是2026年了，快到年底了。
Now we're 2026, you know, near the end of this year.
你现在怎么看这些工具？
How do you view these tools now?
我觉得我和AI的关系挺复杂的。
I think I have a complicated relationship with AI.
确实，我很多研究上的问题，只要提示词给得对，AI都能回答。
I think it's surely true that many research questions that I might have could be answered by AI with the right series of prompts.
就像Anthropic今天刚发布的一个例子，算九圈振幅，有时候你只需要让它继续、再努力点就行。
And as many examples show, one that was just released today from Anthropic about calculating some nine loop amplitude, Sometimes the only prompts you need are, keep going, try harder.
我先去睡了。
I'm going to sleep.
六个小时后回来找我。
Come back to me in six hours.
所以我觉得，这些工具能做到这一点，并不一定让我对这个问题有更好的理解。
And so the fact that these tools can do that, I feel like doesn't lead me necessarily to a better understanding of the problem.
所以我可以告诉你，我为了让自己心里过得去而做出的个人妥协是：我基本上只用AI来研究AI。
So I can tell you that the personal compromise that I have made that can reconcile this in my head is that I basically exclusively use AI to do research on AI.
换句话说，这些模型号称最擅长的事情，就是让自己变得更强。
So in other words, the thing that these models supposedly are the best at is in making themselves better.
所以我一直是一个社区的一员，这个社区认为可以用物理学的思想和工具来理解AI在做什么。
And so I have been part of a community that thinks that you can use ideas and tools from physics to try to understand what AI is doing.
比如，经常用模拟的物理数据作为很好的基准数据集。
Often, for example, using simulated physics data as a good benchmark dataset.
还有一系列项目，就是在物理数据集上训练一堆AI模型，以提取它们行为的一些特征。
And there's a series of projects that just involve training a bunch of AI models on physics datasets in order to extract something about how they're behaving.
这些项目的技术复杂性非常烦人，所以我非常高兴AI模型能一次性搞定我在集群上跨四个GPU训练扩散模型神经网络的事情，而且不会给我报编译错误。
And the technical complications of that are so annoying that I am totally happy that an AI model can just one shot how I train diffusion model neural network across four different GPUs on my cluster, whatever, without ever giving me compiler errors.
这太棒了。
That's wonderful.
好的。
Alright.
让我们……
Let...
让我们停一下。
Let's stop for a second.
所以你的意思是，现在大多数物理学家都在用AI来帮助他们了解物理，而你感兴趣的是用物理来帮助你了解AI。
So what you're saying, most physicists today are using AI to help them learn about physics, whereas you are interested in using physics to help you learn about AI.
对。
That's right.
而且用AI作为工具来帮助你了解它，这很有意思，三年前似乎还不太可能。
And using AI as a tool to help teach you about it is an interesting thing that didn't necessarily seem possible three years ago.
我得说，你知道，我……
I will say that, you know, my...
我成为物理学家的动机当然是为了了解宇宙。
The motivation for becoming a physicist is, of course, to learn about the universe.
我觉得AI有很多东西可以教我们关于宇宙的，但如果我不理解工具的工作原理，我有点不愿意把它们当作神谕或黑箱来使用。
And I think AI has a lot to teach us about the universe, But I am kind of allergic to using tools as oracles or black boxes if I don't understand how they work.
所以我认为更好地理解AI的工作原理，是使用AI来了解宇宙的一个垫脚石。
And so I consider understanding how AI works better as a stepping stone to using AI to learn about the universe.
因为我觉得你问的那个问题的答案，有一部分可能是：如果我想知道怎么做一个计算，我就让ChatGBC来做，它就会给我答案。
Because I think part of an answer to the question that you asked could have been, well, if I want to know how to do a calculation I just ask ChatGBC to do it and it gives me the answer.
但我不确定我是否真的能学到东西。我可以让它解释，不过我发现现在的模型即使你让它解释某个点，它给出的回答往往啰嗦得没必要。
But I'm not sure that I would have learned anything I could ask it to explain it to me but often what I found even with the models today if you ask a model to explain something, it can give you an answer in an unnecessarily verbose way.
如果你让它详细解释它回答里的某一个小细节，它会给出一个极其冗长的回复。
That if you ask it to explain one little tiny element of its explanation, it'll give you an extremely verbose answer to that.
我发现它还没能做到高效地提炼知识并传达出来。
And I found that it hasn't really been able to distill knowledge in a way that it communicates it most efficiently.
我跟AI互动学习物理时，有点像在做自由联想。偶尔它会提到某个概念或说出些我没想过的内容。
When I interact with AI to learn things about physics, I kind of treat it like a free association session Occasionally it will just mention some concept or say something that I hadn't thought of.
我会自己去深入研究这个问题。
I'll go up and learn about that on my own.
通过这种方式，我觉得自己的知识会不断加深。
And in that way I think that my knowledge tends to deepen.
不过我得承认，这种使用工具的方法确实挺独特的。
But I will admit that this is a fairly idiosyncratic way of using these tools.
好的。
All right.
我们很多听众可能都有……
A bunch of our listeners probably have...
或者他们可能不是物理学家，或者很多人不是物理学家，而且他们肯定没用过AI工具来做物理。
Or maybe they're not physicists or a lot of them aren't physicists, and they certainly haven't used AI tools to do physics.
所以我们就大概描述一下这些工具现在能做什么。
So let's just, like, kinda sketch the landscape of what these tools can do right now.
所以如果我给Claude或ChatGPT一些手写的，或者不是手写的，但比如不是基于网络的本科物理考试题，我敢打赌它能把每道题都做得完全正确，99%的情况下都不会出错。
So if I gave Claude or Cachi BT some handwritten or not handwritten, but like non internet based undergrad physics exam, I would wager it would get every problem perfectly correct 99% of the time without any errors.
你同意吗？
Do you agree?
我唯一可能有点异议的就是那个99%的情况。
So the only quibble that I might have with that is the 99% of the time.
这些模型本质上就是随机的。
These models are fundamentally stochastic.
它们的工作原理是预测一系列文本中的下一个标记或单词。
They work by predicting the next token or word in a series of text.
这并不意味着它们不能创造，也不意味着它们不能推理，因为显然它们可以。
That doesn't mean that they can't be created or it doesn't mean that they can't do reasoning because obviously they can.
但归根结底，模型给出的东西本质上是概率性的。
But at the end of the day the thing that the model says is fundamentally probabilistic.
所以我们并不清楚这些模型答对问题的概率到底是多少。
So we don't have a good calibration on the rate at which these models get things right.
确实，它……
It is true that it is...
我是说，几乎可以肯定它高于50%。
I mean, it is almost certainly true that it's above 50%.
比如，你第一次给模型输入提示，它很可能会给出正确答案。
Like, the first time you put a prompt into the model, it will probably give you the right answer.
如果你做一千次，会不会有一千分之一答错？
If you did it a thousand times, would one out of a thousand get the wrong answer?
我不知道。
I don't know.
所以，我是说，就我个人经验来看，我知道这远远超过50%。
So, I mean, I I just from personal experience, I know it's wildly above 50%.
但是，我……
But I...
而且那……
And that...
这并不是在说它怎么运作，或者它是不是在思考之类的。
That's not a statement about how it works or if it's thinking or anything.
我觉得那些是让人兴奋的元问题。
I think those are meta questions that are exciting.
但是，我是说，如果我让它解决一个我随便想出来的问题，不是那种前沿研究什么的，我，我……
But, I mean, if I ask it to solve some problem that I can come up with the top of my head that isn't, you know, cutting edge research or something, I I...
至少有90%以上的概率。
There's a at least 90 something percent chance.
它会给你正确的答案，对。
It will give you the right Yeah.
我不确定我会给出一个具体的数字，但，是的，我同意它高于50%，而且不是1%，也不是100%。
I'm not sure I would commit to a particular number, but but, yes, I would agree it's higher than 50% and it's not one or it's not a 100%.
你知道，我不确定那个界限在哪里。
You know, I'm not sure where that boundary is.
所以，今年早些时候我们把你那份量子场论考试卷子喂进去的时候，你还记得我们最后给它的分数是多少吗？或者我们的同事兼评委戈登·克尼亚克给它的分数？
So when when we fed in your quantum field theory exam earlier this year, do you remember what the percent of the ultimately the score we gave it was or our colleague and judge Gordon Kerniak gave it?
我记得大概是80分左右。
So I think it was about 80 About
80%。
80%.
说实话，我当时那个赌注下得挺蠢的。
And so speaking frankly, I was probably pretty dumb with the bet.
我……
I...
赌注的条件是及格，而至少在我当时所在的学校，及格线大约是60%。
The the conditions of the bet were pass, which at least at my institution at the time was defined as being about 60%.
我可能应该说90%才对。
I probably should have said 90%.
但我愿意说它能及格，这反映了我当时对这类工具发展速度的看法。
But I think fact that I was willing to say pass reflected what I thought of the pace of development of these tools at the time.
我看了那份考试里的很多内容，我知道很多芝加哥大学的高级研究生，还有现在麦迪逊大学的学生，都不一定能稳定地答出那些题。
I mean, looked at a lot of the in that exam and I knew that, I've seen many, many advanced grad students at the University of Chicago and now at the University of Madison who wouldn't be able to get those answers consistently.
这些题很难。
These were hard.
这是一门很难的课，选课的都是理论粒子物理最顶尖的学生。
This is a hard class taken by the top students in theoretical particle physics.
总之，它当时能做到那样，让我印象深刻。
Anyway, I was impressed by what it was able to do.
而且想想它在2023年肯定做不到的事情，也很有意思。
And it's also interesting to think about the things that it for sure was not able to do in 2023.
其中一道题是计算某个东西，然后画出答案的草图。
One of the questions was calculate this thing and then sketch the answer.
这在2023年是完全不可能的。
That was 100% not possible in 2023.
而且那可能真的是我们在2026年评估模型之前的最后一块拼图。
And that might have actually been the last piece of the puzzle before we evaluated the the models in 2026.
比如，现在它能调用代码了。
Like, now it can call code.
它能画图了。
It can make a plot.
它还能通过某种方式，比如PNG图像之类的，把图变成草图。
It can turn a plot into a sketch via some, you know, like, PNG graphic or something like that.
这些都是三年前根本不存在的、令人惊叹的工具使用例子。
Like, these are all kind of amazing, like, tool use examples that did not exist three years ago.
另一个在AI和物理学应用中让人特别兴奋的方面，是关于大数据应用的。
The other thing that's been really exciting to see in terms of uses of AI and physics has to do with, like, kind of, like, big data applications.
有很多例子，实验中收集了大量事件，每个事件都有细微不同的特征，你希望能以不同方式对它们进行分类，或者测量某些潜在的性质。
There are all sorts of examples where experiments collect huge numbers of events, all with slightly different characteristics, and you'd like to be able to categorize them in different ways or measure certain underlying properties.
比如，在我参与的冰立方中微子天文台，就有不同类型的事件。
Like, for example, with the, Ice Cube Neutrino Observatory that I'm part of, there are different kinds of events.
有轨迹事件，是由高能缪子穿过望远镜产生的。
There are there are track events that are created by energetic muons passing through the telescope.
还有我们所说的级联事件，它们产生于电磁或强子簇射。
There are what we call cascade events, which are produced in either electromagnetic or hadronic showers.
我们编写算法来查看数据，并尝试将每个事件分类为轨迹、电磁级联或强子级联，但这并不完美。
And we write algorithms to look at the data and try to categorize each event as either a track or electromagnetic cascade or a hadronic cascade, but it's imperfect.
效果不是很好。
It's not that good.
我们还尝试提取这些粒子有多少能量、来自哪个方向等等。
And we also try to extract how much energy did those particles have and what direction did they come in and whatever.
但结果同样不理想。
But again, with imperfect results.
我们现在已经在大量模拟和真实数据上训练了机器学习算法。
We have trained machine learning algorithms now on a bunch of simulated and real data.
有些时候，我们甚至训练出了连自己都搞不懂的算法。
And in some cases, we've produced algorithms that we don't understand.
它们就像个黑盒子，但能比我们自己写的算法更准确地给出粒子的方向和能量。
They're a big black box, but they tell us the direction and energy of those particles with much more accuracy than we can with our human made algorithms.
比如，IceCube发现来自银河系盘面方向的 neutrino，我觉得如果没有机器学习，这结果恐怕很难实现。
And for example, when IceCube discovered neutrinos coming from the plane of the Milky Way, from the disk of the Milky Way, was something that I don't think could have honestly been done without machine learning.
我跟大型强子对撞机那边的人聊过，他们现在能做出很多以前做不了的高精度测量，全靠机器学习帮忙。
I talked to people working at the Large Hadron Collider, and there are all sorts of measurements they're able to make with higher precision because of machine learning that they couldn't have made without those tools.
有没有什么理由让我们对这些应用不感到兴奋，还是说它们会越来越好？
Is there any reason to not be excited about these applications, or are they going to continue to get better?
我们在使用这些黑箱来获取测量数据时，会不会有失去某些东西的风险？
Are we in danger of losing something by using these black boxes to get at these measurements?
嗯，我想我先要反驳一下前提，就是这些东西永远都是黑箱。
Well, so I I guess let me first take issue with the premise that these things are always going to be black boxes.
我认为理解人工智能的目标正是要打破这种黑箱状态。
I think that's the goal of trying to understand AI.
咱们先区分一下机器学习和人工智能吧。每个人都有自己的定义，但我觉得一个合理的说法是，人工智能不仅能做机器学习，还能用英语或者其他人类语言交流。
And maybe let's distinguish between machine learning and artificial Everyone has their own definitions, but I think a reasonable one is that artificial intelligence is able to do machine learning but communicate in English or in in in some human language.
机器学习工具就是一大堆参数，目的是处理数据，把它变成某种函数。而当数据是语言的时候，你就是在处理语言，创造新的语言。
That a machine learning tool is a huge number of parameters whose goal is to do something that processes data and turns it into some function of Whereas the when that data is language then you're processing language and creating new language.
但语言能承载意义，还能进行计算。
But that language can carry meaning and it can do calculations.
然后突然间，它就能做所有我们以为只有人类才能做的事情了。
And then all of a sudden it can do all the kinds of things that we thought were just the providence of humans.
我觉得最大的区别就是，人工智能是语言上的那种体现。
I think that's the big difference is that that AI is sort of that manifestation in language.
所以对我来说，更有趣的问题是，你愿意把开发定制机器学习算法来分析冰立方数据的任务交给一个大型语言模型吗？然后说，好。
And so I think for me, the more interesting question is would you feel comfortable turning over the development of your bespoke machine learning algorithms to analyze IceCube data to a large language model and say, great.
你现在来写这个分析代码。
You now write the analysis code for this.
不管你写出什么，都行。
Whatever you come up with, great.
我会信服的。
I'll believe it.
我还没到那一步呢。
I'm not quite there yet.
我可能会让他去写，然后做大量的测试来验证结果。
I might ask him to do it and then and then, like, do a lot of extensive testing of it and see it.
而且你知道的
And, you know
没错。
That's right.
但我不会盲目使用它。
But I wouldn't use it blindly.
我认为这正是物理学相比其他大多数大数据领域的一个真正优势，因为我们对数据非常了解。
And I think this is where physics has a real advantage over almost every other big data field in that we know an enormous amount about our data.
我举个例子，跟我正在做的一些研究有关，那就是自牛顿以来，我们所有的实验都告诉我们能量是守恒的。
So I'll give one example that's related to some research that I'm doing, which is that every experiment we've ever done since Newton has told us that energy is conserved.
不管过程是怎样的，如果你把所有不同的能量来源都算进去，能量既不能被创造，也不能被消灭。
Whatever the process, if you account for all the different sources of energy, can't create or destroy energy.
它可以转化为质量。
It can turn into mass.
它可以变成其他形式的东西，但它永远不会消失。
It can turn into other kinds of things, but it never goes away.
好的。
Okay.
所以，如果你用一些数据训练一个机器学习模型，它永远不会精确地学到能量守恒。
So if you train a machine learning model on some data, it will never learn energy conservation exactly.
它根本做不到，因为这就像说我想让一个模型画一个完美的圆。
It simply can't because it's like saying I want a model to draw a perfect circle.
好的。
Okay.
画一个圆。
Draws a circle.
不是。
No.
不是。
No.
不是。
No.
不是。
No.
如果那条线有宽度，那就不是完美的圆。
If that line had a finite width, that's not a perfect circle.
那是个圆盘。
That's a disc.
你明白吗？
You know?
显然，完美的线是数学上的理想化，厚度为零。
Obviously, a perfect line is a mathematical idealization that has zero thickness.
但像能量守恒这样的东西，在物理学中已经是我们能最接近数学理想化的某种完美约束系统了。
But things like energy conservation are as close as we get in physics to the mathematical idealization of some sort of perfectly constrained system.
所以我觉得在物理学中，我们真正的优势在于，我们可以找到方法来施加这些约束，同时不影响模型的表现。
So I think in physics the real advantage for us is that we can find ways to impose these constraints in such a way that we don't affect the performance of the model.
然后我们的模型从构造上就更容易解释。
And then our model is sort of by construction more interpretable.
所以我几个月前写的一篇论文，就是尝试在模拟的高能对撞机数据上做这件事，我们找到了一种方法……
So a paper that I wrote a couple months ago was trying to do this on simulated high energy collider data that we found a way to...
在模拟数据中，你总能确保能量和动量是严格守恒的。
In simulated data you can always ensure that energy is exactly conserved and momentum is exactly conserved.
所以我们找到了一种方法，让模型学习数据的分布，这样无论它学到什么，即使没有完美学到分布，它总能严格守恒能量和动量。
And so we found a way to have a model learn that distribution of data in such a way that whatever it learned even if it didn't learn the distribution perfectly it always exactly conserved energy and momentum.
所以它就有这种，像是一个先验，能量需要
So it has this, like, a prior that energy needs
被守恒。
to be conserved.
就是说，它一开始就知道这一点。
Like, it knew that going into it.
对。
Exactly.
对。
Exactly.
这是一个硬性的物理先验，是以硬性方式强制的，而不是所谓的软性方式，软性方式就是如果模型在能量守恒上出错，你就惩罚它。
It was a hard physics prior that was enforced in a in a hard way rather than the so called soft way, which is like you you penalize the model if it gets energy conservation wrong.
你永远无法通过惩罚让它完全避免
You can never penalize it enough to never make it
犯错。
do it.
而且
And
所以我觉得，这种思考方式是找出利用AI帮助我们做物理学的最佳途径的好方法。
so I think I think this way of thinking about it is a good way of figuring out the best way of leveraging AI to help us do physics.
我的意思是……
I mean...
而且，你知道吗，有个有趣的事实是，我们不需要对语言做这个。
And, you know, an interesting fact is, like, we didn't need to do this for language.
我们从来不需要硬编码语法规则。
We never needed to hard code in the rules of grammar.
当然。
Sure.
但现在ChatGPT基本上永远不会犯语法错误。
But ChatGPT will basically never make a grammatical mistake now.
它只是读了足够多的语法例子就知道了。
It's just read enough examples of grammar to know.
对。
Exactly.
你知道吗？
You know?
但是另一方面，
But on the
训练数据里也有网上那些用不正确语法的人的例子，不管那是什么意思。
other hand, like, are examples in the training data on the Internet of people using not correct grammar, whatever that means.
如果你让ChatGPT用那种，你知道的，'was up LOL'那种短信语法来回答，它会给你一个用那种风格写的回复……
If you prompt chat GBT to use, you know, was up LOL kind of texting grammar, it will give you a response in...
你知道，就是根据那种文本风格来调整，如果它是在写正式的法律文件，语法就不会是那样。
You know, conditioned on that text where the grammar would not be correct if it was writing a formal legal document.
所以它知道怎么根据上下文来调整自己的回应。
So it knows how to adjust its kind of context or adjust its response based on its context.
你知道，这……
You know, that...
这挺有意思的。
That's interesting.
但是在物理方面，我们不需要让模型重新发现我们已经知道的那些物理定律。
But but in physics, like, we don't have to make the model rediscover all the laws of physics that we already know.
我觉得这是一个很大的优势，也是我对AI在物理学大数据应用中感到兴奋的地方。一旦我们能够准确且只放入我们真正需要的先验条件，模型就能发挥它的魔力，发现其余的部分。
And I think that's a big advantage and that's where I'm excited about the uses of AI in in kind of big data in physics is once we are able to put in exactly and only the priors that we really need to be there, then the model could do its magic and discover the rest of it.
然后我们就可以开始问这个问题了，好的。
And then we can start to ask the question, okay.
如果我训练一个机器学习模型来模拟宇宙的演化，那么在这个模型中，它是在哪里学到引力的呢？
If I if I trained a machine learning model to simulate the evolution of the universe, where in the model has it learned gravity?
如果它能完美地重现宇宙的演化，那它肯定学到了引力，但我们应该能够理解这个模型在哪里体现了引力。
It must if it's able to perfectly reproduce the evolution of the universe, but we should be able to understand where that model is representing gravity.
我觉得这些问题是可以解决的，这也是我现在花很多时间在研究的事情。
And I think these are solvable problems, and it's what I'm, you know, spending a lot of my time working on now.
所以在我看来，物理学中AI和机器学习影响最大的地方是在这些大数据领域，那里我觉得抽象程度不高。
So as I see it, like, the places in physics where AI and machine learning has been the most impactful have been in these big data arenas, where I would say there's not a lot of abstraction.
那里只是有大量的不同事物之间的相关性，机器学习能够识别这些相关性，并以人类难以做到的方式加以有效利用。
There's just, like, a a huge quantity of correlations between different things, and the the machine learning is able to identify those correlations and put them to powerful use in a way that human beings have struggled to.
这有点像看狗和猫的图片，然后分辨出哪个是哪个。
It's kind of like looking at pictures of dogs and cats and figuring out which are which.
这种事情，人类要写个算法来做真的很难，但人脑和机器学习算法都能做得不错，至少比人类写的算法要好。
And something that's really hard for a human being to write an algorithm to do that, but both the human brain and machine learning algorithms do it pretty well, or at least better than human written algorithms.
但在其他一些抽象程度很高的领域，AI也表现得非常好，特别是在数学和基础数学问题的研究上。
But there are other domains where there's a lot of abstraction, where AI is doing really, really well, in particular in the domains of mathematics and research of fundamental math problems.
你怎么看待机器学习在数学这类领域的应用？
How do you view those sorts of applications of machine learning and AI?
你觉得这些应用和我们在物理学里尝试做的事情有本质上的不同吗？
And do you think of them as fundamentally different than what we're trying to do in physics?
你能总结一下AI对数学的影响目前处于什么状态吗？
And could you summarize the state of AI's impact on math?
好的。
Yes.
我会尽力而为。
I will do my best.
也许我该先说，我们其实还不知道真正的边界或极限在哪里。
Maybe I'll just start by saying that we we don't know where the frontier where the limit really is.
目前，数学领域最令人惊讶、最突破性的成果似乎都来自那些未公开发布的AI模型，比如我们无法用它们来评估我们的赌注，因为你和我都没法使用。一旦这些模型向公众发布，肯定会有更多内部模型能做更强大的事情，而我们对此一无所知。
That right now the most surprising and boundary pushing results in math seem to come from AI models that are not publicly released that, for example, we would not be able to evaluate on for our bet because you and I can't And use once those models get released to the public, there will surely be more internal models that can do even more powerful things that we're not aware of.
但我认为，目前数学和计算机科学与物理学的区别在于，数学，除了像哥德尔不完备定理那样，说有些命题永远无法被证明或证伪，除了这类奇怪的边缘情况，数学和计算机科学都是形式上可验证的领域。
But I think what what distinguishes math and computer science from physics for the time being is that math, with the exception of things like the Godel incompleteness theorem, which says that some statements can never be proved or disproved, with with the exception of sort of weird edge cases like that, math and computer science are are formally verifiable domains.
你可以验证一个证明是否从公理推导出来，也可以验证代码是否能编译。
You can verify that a proof follows from its axioms, and you can verify that a code compiles.
而物理学，我们一度喜欢认为，它的优势在于有点模糊。
Physics, we've liked to think for a while, has an advantage in being a little bit squishy.
所以我跟学生们说，在……
So what I tell students in...
你知道，当我从物理学的角度教机器学习时，我会说，看。
You know, when I teach machine learning from a physics perspective, say, look.
我们在严谨性上放弃的，在灵活性上得到了补偿。
What we give up in rigor, we gain in flexibility.
比如，著名的量子场论从未被证明在三维空间中存在。停一下。
Like, famously, quantum field theory has never been proven to exist in three spatial three space Stop for a second
因为很多人听到这个会愣住，然后问：你这话是什么意思？
because a of people are gonna choke on that and say, what do you mean by that?
意思是，我们目前没有一个能让数学家满意的、正式的数学定义来描述粒子物理的标准模型。
It means we do not have a formal mathematical definition of the standard model of particle physics that would satisfy a mathematician.
我们有一堆计算和启发式方法，但显然我们做对了一些事情，因为我们现在能把电子的磁矩预测到一万亿分之一那么精确。
We have a bunch of calculations and heuristics, but we are clearly doing something right because we are able to predict things like the magnetic moment of the electron to one part in 10,000,000,000,000 right now.
这是人类历史上理论与实验之间最精确的对比。
It is the most precise comparison between theory and experiment that humankind has ever made.
所以很明显它是有效的，但我们无法证明它是正确的，这或许让人希望它能对AI更有抵抗力，不过我现在非常不愿意做任何预测。
So clearly it's working, but we can't prove that it's correct, which one might hope makes it more resistant to AI, though I am now very loath to make any predictions.
我同意你说的，而且我……
Agree with what you're saying and I I...
但我想扮演一个可能不懂你在说什么的人。
But I wanna, like, play the role of somebody who might not know what you're saying here.
好吗？
Okay?
那我想问的是，你说我们还没证明它有效，是什么意思？
So the question I would ask is, what do you mean we haven't proven it works?
实验和预测在十万亿分之一这个精度上吻合，这本身就是证明。
The fact that the experiment agrees with the prediction at one part in 10,000,000,000,000 is the proof.
答案是什么？
What's the answer?
那你怎么知道哪些实验该信，哪些不该信呢？
Well, how do you know what experiments you should trust and what you shouldn't?
实验嘛，大家都知道，有时候结果会互相矛盾。
So experiments, famously, sometimes give contradictory results.
实际上，我们没法把这个数字再提高一两位小数，原因在于电子磁矩的测量——基本上就是它在某种量子力学意义上转得多快——依赖于另一个参数，叫精细结构常数，它决定了电磁相互作用的强度。
And in fact, the reason that we can't push that number one or two decimal places higher is because the measurement of the electron magnetic moment, which is basically a measure of sort of how fast it's spinning in some quantum mechanical sense, depends on another parameter called the fine structure constant that governs the strength of electromagnetic interactions.
而测量这个常数的两个不同实验，得到了两个不同的结果。
And two different experiments that are measuring that same constant get two different answers.
好的。
Okay.
好的。
Okay.
但是，但是实验方面的问题是……
But but the experimental questions are are...
让我……
Let...
我只是想问，怎么……
I'm merely asking how does...
比如说，我说的是理论方面。
Like, I'm talking about the theory side.
所以我们有一个叫量子场论的东西，它告诉我们如何计算在各种实验中你会看到什么。
So we have a a thing we call quantum field theory, which tells us how to calculate what you will see in various experiments.
你告诉我那东西的基础很不牢靠。
You're telling me that that is on a flimsy foundation.
嗯，它在形式上的基础确实不牢靠，因为如果你想验证一个数学证明是否正确，你只需要一步步检查，看它是否从你开始的公理推导出来。
Well, it's on a flimsy formal foundation in that that if you want to verify that a mathematical proof is correct, you just go through and you check line by line, does it follow from the axioms you start with?
如果你想对物理预测做同样的事情，你会说，做个计算。
If you're trying to do the same thing for a physics prediction, you'd say, do a calculation.
那个计算和我的实验匹配吗？
Does that calculation match my experiment?
但接着你只能包括你信任的实验结果，而不是那些噪声更大、矛盾或令人困惑的结果。
But then you have to only include the experimental results that you trust and not the ones that are noisier or contradictory or confusing.
我想我试图表达的观点是，事后看来，我们可以说，哦，显然，我们知道粒子物理标准模型的理论。
And I think the point that I was trying to make is that with hindsight, we can say, oh, obviously, we know the theory of the Standard Model of particle physics.
它有这些粒子之类的。
It has these particles or whatever.
但如果你回到比如20世纪70年代的物理文献，有很多关于实验的论文，那些实验试图解释的实验结果是错误的，或者我们现在知道是错误的，因为实验当时认为看到的东西随着更多数据而消失了，或者由于某种原因变得模糊不清。
But if you go back to the physics literature from, say, the 1970s, there were a lot of papers about experiments that were trying to explain experiments that were just wrong or that we know now were wrong because the thing the experiment thought that it saw then went away with more data or was ambiguous for some reason.
而让我们得到粒子物理模型的那个推理过程，是关于丢弃那些与计算不符的实验，但这些实验本来也应该被丢弃，因为实际上它们在别的方面也是错误的。
And the process of reasoning that got us to our model of particle physics was about discarding the experiments that didn't fit the calculation but also should have been discarded because actually they were wrong in some other way.
这是一件危险的事情，因为你不想仅仅因为你不喜欢答案就抛弃一个实验。
And it's a dangerous thing to do because you don't wanna throw out an experiment just because you don't you don't like the answer.
这其实也是历史上经常出现的一种失败模式。
That also has been a failure mode historically.
我再补充一点，大多数物理学家和大多数数学家之间，有一种根本性的、性格上的差异。
And I'll just add that there is a fundamental, like, dispositional distinction between most physicists and most mathematicians.
如果我有某种公式或程序，数学上的程序，用来做预测，但其中某一步从未被数学上证明是有效的。
If I had some sort of formula or procedure, mathematical procedure to make predictions, and there was a step in it that had never been proven to work mathematically.
但我知道，只要我按照这个程序走，在很多不同的例子和很多实验里，长期下来，我都能得到正确的答案，或者和实验一致的结果，那我根本不在乎。
But I knew that when I followed this procedure, I got the right answer or the same answer the experiment gives me over a lot of different examples and over many experiments over a long period of time, I wouldn't care.
我会说，这就是正确的理论，或者至少是一个在经验上足够正确的理论，管它呢。
I'd say this is the right theory, or at least a sufficiently empirically correct theory, whatever.
但数学家不会这么看。
And a mathematician would not view it that way.
他们会坚持那种形式上的严谨，而作为物理学家，我根本不会去操心这个。
They would they would, insist on that kind of formal rigor that as a physicist, don't concern myself with.
作为一个经验主义者，我根本不会去操心这个。
As an empiricist, I don't concern myself with.
我觉得没错。
I think that's right.
而且我想说的是，这些AI模型的训练方式让它们在那些可以形式化验证的领域里进步飞快。
And and I think the point that I was making is that the way these AI models are trained is they are able to make progress incredibly fast in domains that are formally verifiable.
这不一定是个根本性的限制。
That is not necessarily a fundamental limitation.
只是它们目前的状态而已。
It is just where they are right now.
说说看
Describe some
AI和数学方面最大的成就有哪些。
of the biggest achievements in terms of AI and math.
好的。
Okay.
最近，OpenAI声称他们解决了所谓的纳维-斯托克斯问题。
So recently, there was a claim from OpenAI that they had resolved the so called Navier Stokes problem.
这是克莱数学千年大奖问题之一。
This is one of the clay millennium prize problems.
如果有人能证明，在这种情况下，流体流动的数学模型能产生无穷大，而计算一开始并没有无穷大，就能获得一百万美元的奖金。
There's a bounty of a million dollars for anybody that could provide a proof that in this case, a mathematical model of fluid flow could give you infinity when you never had infinity in the calculation to begin with.
所谓的奇点，在物理学中意味着你的理论失效了，你应该换一种方式来计算。
The so called singularity that are things that in physics that tells us that, like, your theory is breaking down and you should be calculating something different.
但纯粹从数学角度看，问题就是：你能从有限的数出发，最终得到无穷大吗？
But as just a purely mathematical question, can you basically start with small numbers and end up with infinity?
这个问题很难解决。
That was a hard question to resolve.
答案是肯定的。
And the answer was yes.
而且网上已经有一个证明，但据我所知，没有哪个人类能完全理解它。
And there is a proof that exists on the Internet that to my knowledge no human fully understands.
好的。
Okay.
那我们是不是……
So are we...
那些……
Are the...
数学界相信这个证明是对的吗？
Is the mathematics community convinced it's right?
我上周刚问了一个来我们这儿做讲座的学者这个问题。
So I just asked a speaker, a seminar speaker here that question last week.
答案基本上是肯定的，因为他们提供了一份所谓的 Lean 证书。
And the answer was largely yes because what they provided with it is a so called lean certificate.
Lean 是一种数学编程语言，能够自动形式化证明。
So lean is a mathematical programming language that is able to automatically formalize a proof.
它用代码定义公理，然后检查推导的每一步是否都遵循这些公理。
It defines axioms sort of in code, and then it just checks that every line of the derivation follows from those axioms.
所以只要用自然数学语言写的证明被正确翻译成了 Lean 代码。
And so as long as the proof that was written sort of in natural mathematical language was correctly translated into Lean.
而且只要所谓的挑战文件里对问题的陈述，是从真实问题的陈述正确翻译成这种编程语言的，那么基本上就可以通过代码编译来验证。
And as long as the statement of the problem in the so called challenge file was correctly translated from from the statement of the of the real problem into this this programming language, then it sort of follows by the fact that the code compiles.
只要没有bug，结果就一定是正确的。
As long as there aren't any bugs, the the result must be must be correct.
而且你知道，对于这种级别、重要性或影响力的证明，肯定有很多专家在逐行仔细审查。
And for a, you know, you know, a a proof of this kind of stature or or importance or consequence, I'm sure there are a lot of experts pouring over every line and and scrutinizing.
所以我们现在有的任何疑问，可能很快都会得到解决。
So any doubts that we have now will will eventually get resolved probably in the near near near term.
你同意吗？
Would you agree?
我数学水平不够，不好说。
I'm not enough of a mathematician to to say.
我觉得，无论是AI生成的用自然数学语言写的证明，比如PDF形式的，还是用Lean这种形式语言写的，都很难读。
I'd say that it's hard to read both the AI generated proofs in kind of natural mathematical language, like written as a PDF.
而且用Lean这种形式语言写的证明也很难读。
And it's also hard to read them in this formal language of Lean.
所以，研讨会主讲人指出的一个问题是，这些模型生成的代码看起来不像人类在思考如何写证明时会写的那种代码。
So one of the things the seminar speaker was pointing out is that these models produce code that doesn't look like the kind of code that humans would write if they were thinking about how to write a proof.
这又让我想起我在2023年初说过的话，它们能说出这些表面上看起来像是按正确顺序排列的东西，但实际上是陌生的。
And this is again very reminiscent of thing that I said in the beginning and kind of in 2023 that they're able to say these things that on the surface sort of look like they're putting things together in right order, but but it's sort of alien Mhmm.
对人类来说，看起来很奇怪。
To to a human to look at.
所以我一直在开的那个玩笑，我觉得最近在某个博客上看到的，就是：如果一条证明在森林里倒下，没人理解它，那它算发出声音了吗？
And so the the joke that I've been making that I think I saw on some blog recently is like, you know, if a proof falls in the forest and nobody understands it, does it make a sound?
就像，如果我们有一个我们知道是正确的结论，但我们没能从中学到任何东西，因为它的写法太晦涩难懂了。
Like, you know, if we have a result that we know is correct, but we we haven't been able to learn anything from it because the way that it's written is so inscrutable.
好的。
Okay.
所以，是的，让我给你另一个补充的视角。
So, yeah, let me give you another complimentary perspective.
很多人喜欢尝试为物理学研究做贡献。
So lots of people like to try to contribute to physics research.
互联网上有很多东西，都是人们声称发现了某种理论或某个被忽视的新事物，说物理学界忽略了这项惊人的进展。
There are a lot of things that exist on the internet of people claiming to have discovered some theory or some new thing that was overlooked that the physics community is ignoring this amazing development.
而让这种说法很难被反驳的最佳方法之一，就是写一份长达三百页的宣言。
And one of the best ways to make it very, very hard to refute that statement is to create a 300 page long manifesto
当然。
Sure.
里面全是些难以理解的东西，还有各种奇怪的……
Filled with inscrutable things and weird get all
经常出现在我的收件箱里。
the time into my inbox.
经常。
All the time.
没错。
That's right.
现在我们就有这么一份，还正式认证过是正确的。
We now have one of those that has a formal certification that is correct.
展开剩余字幕（还有 362 条）
这种情况处理起来非常棘手。
That's a very uncomfortable situation to deal with.
因为这不代表它是错的，但也不代表它更容易理解。
Because it doesn't mean it's wrong but it doesn't mean that it's any more comprehensible.
实际上，我觉得数学界有些人最好的办法是让另一个大语言模型来消化这个证明，把步骤变得更易懂。
And in fact, the best tool that I think some folks in the math community are using is to ask another LLM to simply digest the proof and make the steps more understandable.
也许我们就能通过这种提炼的过程达到目的。
And we might just get there via process of distillation.
但你知道，在这个过程中，人又处在什么位置呢？
But, you know, where in that process does the human sit?
我觉得这就是我在纠结的问题。
I think that's the the question that I'm wrestling with.
嗯，你知道，我好奇的是，比如说，为了争论的方便，我们假设一个人要……
Well, you know, here's what I wonder, like, just for the kind of sake of argument, let's imagine that for a human being to...
经过大量学习和努力，才能真正理解爱因斯坦的广义相对论，他们需要智商达到130。
With a lot of study and whatever to ever really understand Einstein's general theory relativity, they need an IQ of, 130.
好的。
Okay.
就比如……
Just for...
我们先停在这里。
Let's pause at that.
假设有一天AI提出了一个物理学理论，你不需要130的智商，但需要140、150、160。
And let's say AI comes up with a theory of physics someday that you don't need a 130, but you need a 140, a 150, a 160.
最终你会达到一个门槛，没有任何人类能理解它。
And eventually you reach a threshold where no human being understands it.
那它还算是一个理论吗？
Is it a theory?
我觉得它算，但它是一个超出我们生物大脑理解范围的理论。
I mean, I guess it is, but it's a theory that's beyond the grasp of our biological minds.
对，没错。
And That's right.
我觉得，这实际上就是一个预言机。
I mean, I think it's it's it's functionally an oracle.
它就是一个预言机，甚至可以直接查询。
It's an oracle even query.
没错。
That's right.
而且，你知道，它……
And, you know, it...
它保证能给你正确的答案。
It's guaranteed to give you the right answer.
我担心的是，如果我们真的到了那一步，基本上就回到了古希腊时代。
And I worry that if we get there, then we are basically back in ancient Greece.
但可能预测事物真相的能力会更强。
But with probably a greater ability to predict the truth of things.
但我想，我不愿意这么快就放弃人类的智慧。
But but I guess I'm not willing to give I'm not willing to give up on human intellect so quickly.
宇宙能被我们理解，这真是天大的运气。
The fact that the universe has been understandable to us is a fantastic stroke of luck.
本来不一定是这样的。
It didn't need to be the case.
比如，六十年代有一群人觉得量子场论永远行不通。
Like, there was a group of people in the nineteen sixties that thought that quantum field theory would never work.
根本没法理解粒子物理反应内部到底发生了什么。
There's no way of ever understanding what happens in the guts of a particle physics reaction.
你最多只能知道进去的是什么，出来的是什么。
All you could ever know is what went in and what came out.
结果呢，大自然跟我们开了个玩笑，实际上你花的钱越多，事情反而变得越简单。
And then it turned out that nature is playing a trick on us that actually things got simpler the more money you spent.
你花的钱越多，就能造出能量更高的对撞机。
So the more money you spent, the higher energy collider you could build.
所谓的耦合常数越小，预测起来就越容易。
The smaller the so called coupling constant went, the easier it became to predict things.
所以，简单来说……
So to put it in...
用我自己的话来说，在低能量下，很多物理计算都非常困难。
Put that same thing in my own words, at low energies, a lot of the physics was really hard to calculate.
而且……
And...
但是当你进入更高能量时，一些相互作用会变弱，变得更容易应用所谓的微扰理论。
But as you went to higher energies, some of the interactions got weaker and and became, more easy to apply something called perturbation theory to.
突然之间，你就能做出非常非常精确的预测，并用粒子对撞机来验证，这在低能量下是做不到的。
And suddenly, you can make very, very accurate predictions and confirm them with particle colliders in ways you couldn't at lower energies.
完全正确。
That's exactly right.
而且我觉得这是一个很好的类比，因为我们现在认为我们已经知道了强核力的基本理论。
And and I think this is a good analogy to to keep with because we think now that we know the fundamental theory of the strong nuclear force.
你可以把它写在黑板的一行上。
You could write it down on a single line of a blackboard.
它里面只有几个参数。
It only has a few parameters in it.
不过，还是有一些东西我们算不出来。
However, there are still some things that we cannot compute.
我们只能靠测量，因为我们已经没有那个好用的、可以展开的小参数了。
We can only measure because we no longer have this nice handle of a small number that we can do some expansion in.
有一些办法可以绕过这个问题。
There are some ways around it.
但我一直在和同事们讨论的是，一个AI模型，不管是纯粹靠推理，还是给它一大堆大型强子对撞机的数据，它能不能学会预测那些我们已经不知道该怎么计算的区域里发生的事情？
But one conversation I've been having with my colleagues is could an AI model, either just reasoning just from pure reason or given a whole bunch of data from the large Hadron Collider, could it figure out how to predict things that happen in this region where you don't know how to calculate anymore?
其实我也不确定这个问题的答案。
I'm actually not sure of the answer to that question.
在我们继续之前，我想先把这段数学讨论收个尾，做个总结。
So before we go on, I wanna wrap up a little bit, put a bow on this maths discussion.
你提到了纳维-斯托克斯方程，但还有哪些……
So you mentioned the Navier Stokes equation, but what are...
比如说，给我列一个清单，说说AI和机器学习在数学领域还有哪些其他重大成就。
Like, give me a laundry list of the other really big accomplishments in the in in the math domain of of AI and ML.
再说一次，我不是数学家，所以对这个圈子关注得没那么密切。
Again, with the caveat that I'm not a mathematician, so I don't follow this community as closely.
据我了解，很多数学家都在用AI助手来写证明。
My understanding is that a lot of mathematicians are using AI assistants to write proofs.
也就是说，你可以把大语言模型当作一种数学助手，帮你头脑风暴，帮你补充步骤和论证。
So in other words, you can use large language models as a sort of a mathematical assistant to to help you brainstorm, to fill in steps and arguments.
我觉得这是真的。
I think this thing is true.
想出一个证明。
Come up with a proof.
给我看一个证明文件，证明它一定是真的。
Show me a lien certificate to prove that it must be true.
太好了。
Great.
现在我就不用担心这个了。
Now I don't have to worry about that.
但是，有哪些关键的猜想或定理已经被证明了？
But what are the key, like, conjectures or theorems that have been proven?
这个，我不太清楚。
That, I don't know.
我是说，有一些例子，比如……
I mean, there's some examples of...
我记得有个叫“单位距离猜想”的东西，但我对它了解不够深，没法给你详细解释。
I think there's this thing called the the unit distance conjecture that I don't know well enough in order to fill it out for you.
你知道，关于其他已解决的问题，有一些传闻。
There are, you know, there are various rumors about other solved problems.
OpenAI 曾表示，数学领域有100个开放问题已经被解决了。
There's a statement from OpenAI that there's 100 open problems in math that have now been solved.
现在这些人正在想办法，如何负责任地把这些成果发布出去。
And now these folks are trying to figure out the best way to responsibly release those out into the wild.
但关于这个，我没有更多的信息了。
But I don't have any more information on that.
所以我觉得，是的，我们大致知道前沿在哪里。
So I think, yeah, we quite know where the frontier is.
但我认为这在数学界引发了一场严肃的讨论，关于证明在理解中扮演什么角色。
But I think it's causing a serious discussion in the math community as to what is the role of proof in understanding.
这是个很好的问题，丹。
That's an excellent question, Dan.
因为如果你还记得，那本来是打赌的原始条件。
Because if you will recall, that was gonna be the original terms of the bet.
哦，我不记得了。
Oh, I don't remember that.
真的吗？
Really?
我原本提议，五年后赌一下黎曼猜想是否会被解决。
I had proposed that we bet in five years on whether the Riemann hypothesis would be solved or not.
然后你说，不行。
And you said, no.
不行。
No.
不行。
No.
我完全无法想象五年后会发生什么。
I can't possibly imagine what will happen in five years.
我只赌三年内的事情。
I only bet on three year time horizons.
我们想一个三年内可能发生的事吧。
Let's come up with a thing that might happen in three years.
我仍然坚持认为，我们的能力正在呈指数级增长。
And I still maintain that we are on an exponential of increasing capabilities.
而且当你处于指数增长阶段时，如果时间窗口上差一点点，性能上可能会差出十万八千里。
And when you're on an exponential, if you're off by a tiny bit in the time window, you could be off enormously in the performance.
你能简单向我们的听众解释一下黎曼猜想是什么吗？
Can you briefly explain to our listeners what the Riemann hypothesis is?
好的。
Yes.
有一个数学函数叫黎曼ζ函数，它和质数分布的一些事实有关。
So there is a certain mathematical function called the Riemann zeta function that is related to facts about the distribution of prime numbers.
有一个猜想认为，涉及这个函数的某个方程的所有解都具有某种性质。
And it is a conjecture that the solutions to an equation involving that function all have a certain property.
也就是说，这些解都是所谓的复数，而且所有这些复数的虚部都是特定的。
That they're all so called complex numbers and all of the complex numbers have a specific imaginary part.
而且有大量的数值证据表明这是真的。
And there is unbelievable numerical evidence that this is true.
很多例子，已知的例子。
Lots of examples, known examples.
很多例子。
Lots of examples.
从来没有人找到过反例。
No one has ever found a counterexample.
但是，在数学的这个领域里，有时候要找到反例，你得去数那些数字，数到高得离谱，任何一个人或者所有人类加起来都不可能数完。
However, this is in a domain of mathematics where sometimes to find a counterexample you would need to go count numbers so high that no human or collection of humans could ever possibly count them.
你唯一能做的就是证明，在某个极其巨大的数字之后，这个假设就不再成立了。
All you can do is prove that there must be some unbelievably large number at which point this thing no longer is is true.
我觉得大多数数学家都相信这个假设一定是真的。
I would say most mathematicians, I think, believe that this hypothesis must be true.
但我不想评判这是不是数学里最重要的未解问题。
But I I I I don't wanna judge on whether this is the most important unsolved problem in in in math.
但肯定是很有名的一个。
But Certainly one of the famous.
它肯定是最有名的之一。
It's certainly one of the most famous.
它是这些千禧年难题中最古老的一个。
It is one of the oldest of these clay problems.
黎曼是十九世纪的数学家。
Riemann was a nineteenth century mathematician.
我非常好奇，AI能不能解决这个问题。
And I would be very curious as to the answer whether AI is able to solve it.
如果它解决了，是通过找到反例解决的，那这大概是最糟糕的情况了。
And if it did, whether it did it by finding a counterexample, which would sort of be the dumbest possible world
那确实。
to live.
那会非常令人兴奋。
That would be very exciting.
或者
Or
它给出一个建设性的证明，不仅被验证是正确的，还能教给我们一些新东西。
to to have a constructive proof and and and have one that that not only was able to be certified as correct but that was able to teach us something.
因为你知道，这些非常非常难的问题往往会把那些你从未想过有联系的领域连接在一起。
Because, you know, these really, really hard problems have a propensity for bringing together different fields that you never thought were connected.
众所周知，纽结理论，就是研究一根绳子有多少种不同的打结方式的理论，跟弦理论有关联，这里借用一下双关语，抱歉。
Famously, knot theory, the theory of, you know, how many different ways can you tie a string together is related to, with all apologies, pun intended, string theory.
当然。
Sure.
不是那种显而易见的关联，而是那种完全意想不到的关联。
In not the obvious way but in like a totally not obvious way.
如果人工智能能对数学中的一个未解难题给出证明，并且能像那样把不同领域联系起来，那么我认为我们真的可以说它推进了人类知识的边界。
If if an AI generated proof of an outstanding problem in math were able to bring those fields together in a similar way, then I think we could really say that it had advanced the, you know, the frontier of of human knowledge.
好的。
Okay.
那我们把话题拉回现实一点。
So let's kinda bring this conversation a little closer to Earth for a little bit.
我们来聊聊人工智能、机器学习和物理学是如何改变我们实际的工作方式的。
Let's talk about how AI and machine learning and physics has changed the way we actually go about our jobs.
比如，它如何改变了科研经费的分配格局？
Like, how has it changed the funding landscape?
它已经改变了，或者也许应该改变，我们在不同层次上雇佣来做物理研究的人员吗？
How has it, or maybe should it, changed the personnel that we employ to do physics research at different levels?
它已经改变了科学期刊或预印本档案的出版形式吗？
How has it changed what the publication scientific journals look like or the preprint archive?
它是否或者应该改变我们培养研究生和博士后的方式？
Does it or should it change how we train grad students and postdocs?
你有什么想法？
Any thoughts?
我对所有这些都有想法，回答其中任何一个问题本身就能做一个小时的播客。
I have thoughts on all of those, and answering any one of those questions would be an hour long podcast on its own.
所以让我试着总结一下。
So let me try to summarize.
我认为由于AI的发展，这个领域变化太快了，我们无法谈论它已经做了什么。
I think the field is changing so fast due to the developments in AI that we can't talk about what it has done.
我们还必须谈论它在短期和中期内能够或将会做什么。
We have to talk also about what it can or will do in the short and medium term.
我说的短期，就是指接下来的三个月。
And by short term, I mean like next three months.
中期的话，我指的是五年，也就是一个人完成博士学位所需的时间。
And by medium term, I mean five years, which is the time it takes for somebody to complete a D.
我觉得一个物理学博士，
I think that a Ph.
博士学位，
D.
现在开始读和五年后毕业，整个经历会完全不一样。
In physics will look completely different between somebody starting one right now and finishing it five years later.
我不知道具体会变成什么样，但肯定不会是现在这个样子。
I don't know what that will look like, but it can't look the same.
所以我觉得，你问的那个目标问题……
So I think that, you know, asking about, you know, what what is the goal...
作为在职物理学家，我们有两个目标。
As as working physicists, we have two goals.
我们做研究，也教书。
We perform research and we also teach.
我觉得我们需要重新强调教学的作用，因为再说，如果一个证明落在森林里没人理解，那它算发出声音了吗？
I think that we need to reemphasize the role of teaching because, again, if like a proof falls in the forest and nobody understands it, does it make a sound?
如果一个物理理论被创造出来，但没人能理解，那我们作为人类知识的探索者，又得到了什么呢？
If like a physics theory is created but nobody can understand it then what have we as inquisitors of human knowledge?
我们得到了什么？
What have we gained?
所以我教量子场论的时候，很大程度上几乎是像在教历史课。
So when I teach quantum field theory I teach it largely almost like a history class.
比如，这是一个令人惊叹的智力大厦。
Like, here is this amazing intellectual edifice.
这可能是人类智慧的巅峰成就。
Like, probably the crowning achievement of the intellect of humanity.
我是物理学家，所以我有资格说，物理学的成就是人类做过的最好的事情，这有点自我吹嘘。
I'm a physicist so I'm allowed to be a little self serving when I say that a physics accomplishment is best thing that humanity has done.
确实是这样。
I mean it is.
这显然是的，Yoni。
It obviously is, Yoni.
太棒了。
It's awesome.
但量子场论中的大多数结果现在已经有几十年历史了。
But most of the results in quantum field theory are decades old now.
虽然不是全部，但描述我们在地球上做过的每一个实验的标准粒子物理模型的主要组成部分，在1980年代就已经完成了。
Not all of them but the main ingredients of the standard model of particle physics that describe every experiment we've ever done on earth were worked out by the 1980s.
所以我觉得人们能理解这一点真是太神奇了。
And so I think it's just amazing for people to be able to understand that.
所以我想让每个上我课的人都能达到那种感觉，就像把量子场论下载到自己脑子里一样。
So I want to get everybody who takes my class to the point where they feel like they have quantum field theory downloaded into their own brain.
而且我们越能在其他学科上做到这一点，就越容易与AI互动，并尝试推动知识的边界向前发展。
And the more we can do that for other subjects, the easier it will be to interact with AI and try to push that boundary of knowledge forward.
我觉得这有点像很久以前文艺复兴时期大学运作的方式。
And I think that's the ways that universities kinda used to function a long time ago kind of in the Renaissance.
那时候大学就是训练人们把知识记在脑子里，这样就不用总是跑去图书馆，那个藏了所有书的图书馆。
It was it was a way of training people to carry knowledge around in their own heads so they didn't always have to, like, go to the library, like, one library that had all the books.
所以有一件事我觉得不太一样。
So one thing that I I think is different.
你之前说过我们应该重新强调教学。
So you said before that we should reemphasize teaching.
我觉得我们其实一直都应该更重视教学。
I think we should have been emphasizing teaching more all along.
但我觉得AI和机器学习带来的变化可能是，我们应该教的内容变了。
But I think what AI and ML do is maybe shift what we should be teaching.
所以也许不是教多教少的问题，而是教的东西不一样了。
So maybe not more or less, just different.
我记得我读研究生学粒子物理的时候，花了很多时间练习计算截面，比如那些迹定理什么的。
I think about when I was studying particle physics as a grad student, I spent a lot of time drilling how to do mechanics of cross section calculations, things like trace theorems and whatever.
这些计算又长又繁琐。
And these were long and tedious.
一道题就要写七页纸，我右手都因为写太多得了腱鞘炎。
One problem would take seven pages of paper, and I I literally got tendinitis in my right wrist from writing so much.
真的特别折磨人。
Like, it it was just brutal.
我不会让我的学生……
I would not ask my student...
我现在正好也在教同样的课程，我不会让学生花太多时间在这上面，因为现在用任何大语言模型查一下答案太容易了。
I'm teaching the same course right now by coincidence, and I I would not ask my students to spend much time on that now because it is so trivial to to just, you know, look up the answer, you know, essentially using any LLM of your choice.
你只需要看一次它是怎么运作的。
You should see how it works once.
你需要理解基本原理，但就像长除法或者大数相乘一样。
You should understand the underpinnings, but kind of like a long division or multiplying large numbers together or something.
我们现在有计算器了。
We have calculators now.
你可能不需要花太多时间在这些练习上。
You probably don't need to spend a whole lot of time practicing that stuff.
我觉得我作为物理学家学到的很多东西都属于这一类。
I I view a lot of the stuff that I learned as a physicist falls into that category.
所以我同意，这也是为什么我在量子场论课上花了大量时间来定义什么是粒子。
So I agree which is why I spend an enormous amount of time in my quantum field theory course trying to define what is a particle.
量子场论里这部分内容一直让我觉得有点别扭。
This is the part about quantum field theory that kind of always bothered me.
确实，在真正开始计算之前，往往得花上一到两个月来做那些必要的铺垫和形式化推导。
That that there is by necessity like, one to two months of kind of throat clearing formalism before you get to, like, actually calculating something.
没错。
That's right.
所以你知道的，在我这门课的第一次作业里，我安排了一个简化的计算例子，它其实跟你能实际测量的东西有关联。
So, you know, on the first homework for my class, I'm able to give a kind of a toy calculation which actually relates to a measurement you can do.
那就是电子在磁场中的跃迁速率。
It's the transition rate of an electron in a magnetic field.
这有点像量子场论的入门版，你得凭空造出一个光子来。
This kind of a baby version of quantum field theory where you have to make a photon out of nothing.
所以你得做一个计算，涉及制造一个原本不存在的粒子。
So you have to do a calculation that involves making a particle that wasn't there.
这其实就是量子场论的核心所在。
That's kind of the crux of quantum field theory.
然后接下来两个月都在讨论，我们说的粒子到底是什么意思？
And then it's like two months of like what do we mean by a particle?
比如，什么定义了粒子？
Like what defines a particle?
它怎么和狭义相对论以及量子力学保持一致？
How is that consistent with special relativity in quantum mechanics?
然后两个月之后，你才能做那些计算。
And then after two months now you could do those calculations.
所以我觉得……
And so I think...
而且我并不是故意把课程设计成防AI的。
And it's not like I designed my course like that to be AI proof.
但有很多工作只是验证你希望这个数学对象成立的性质确实成立。
But there's a lot of just the verifying that the thing that you want to be true about this mathematical object is in fact true.
就连这些步骤现在可能也能自动化了。
That even those steps could probably be automated now.
但即使知道该问哪些问题，这本身就很关键，对吧。
But even knowing that those are the right questions to ask Right.
我觉得只有少数几本教科书以让我满意的方式提出了这些问题。
Only a few textbooks I think asked those questions in a way that I found satisfying.
而且这些书大多是老教材。
And by and large they're older textbooks.
所以我觉得你说得对。
And so I think you're right.
我认为强调机械计算并不好。
I think emphasizing mechanical calculation is not good.
不过我觉得计算是帮助加深理解的一种工具。
But I think calculation is a tool to further understanding.
就像如果你知道怎么进行某项计算，那就说明你清楚所有符号的含义。
Like if you know that you can do a calculation, it means that you know what all the symbols mean.
而弄懂这些符号的意思，在我们所涉及的这类物理学习中，就已经成功了一大半。
And knowing what all the symbols mean is more than half the battle in the kind of physics that we
要是连基本的计算都做不到，哪怕只是理论上也不行，那很可能说明你并没有真正理解那些概念。
If you can't do the calculations even in principle, then you probably don't understand the concepts.
好的。
Alright.
那我问你一个问题。
So let me pose this question to you.
比如，我们回想一下，我先把范围缩小……
Like, if we think back, I'm limiting it...
目前只讨论理论物理方面的成就，不包括实验和观测。
Limiting the discussion right now to theoretical achievements in physics, not experimental, not observational.
但是，你知道，我们可以看到，偶尔会有那种巨大的范式转变的理论发展，比如狭义相对论和广义相对论，量子场论的某些部分，标准模型，再往前追溯到麦克斯韦方程组或者牛顿什么的，诸如此类。
But, you know, we can see, like, the occasional huge paradigm shifting theory development, you know, special and general relativity, certain parts of quantum field theory, you know, standard model, go back to Maxwell's equations or or Newton or whatever, things like this.
这些在历史上比比皆是。
They're they're littered across history.
要过多久，才会有某种人工智能或机器学习代理之类的东西，在思想史上制造出这样的事件呢？
How long will it be before some sort of AI or machine learning agent, whatever, produces one of those events in intellectual history.
你的预测是多少？我想要一个具体的年数。
What you're over under, I want a number of years.
我想要一个具体的年份。不，这个不行。
I want a year this No.
不。
No.
不。
No.
我不能再赌了。
I can't bet anymore.
我做不到这个。
I can't do this.
好吧。
Okay.
那让我尽量多说一些，不设定具体的时间范围。
So so let me try to say as many things I can without putting a time scale on it.
胆小鬼。
Coward.
所以……
So...
你知道吗？
Well, you know what?
行吧。
Fine.
我愿意接受这一点。
I'm willing to accept that.
我……
I...
你知道吗？
You know?
嗯，等我安排好吉他表演的时候，你就可以判断我不愿意再打这种赌是不是有道理了。
Well, by the time that I schedule my guitar performance, you know, you can you can judge whether my my reluctance to to enter another one of these wagers is is merited or not.
好的。
Okay.
所以物理学中有一些成就纯粹是计算上的，是领域内顶尖的成就，并不依赖于任何实验结果。
So there are achievements in physics that are purely calculational, that are towering achievements of the field that are not tied to any experimental result.
而我想到的那个……
And the one that I've...
比如，我立刻想到的就是弦理论中所谓的反常消除。
Like, comes to mind immediately is the so called anomaly cancellation in string theory.
当然。
Sure.
有个著名的故事，说剑桥的物理学家迈克尔·格林和其他几个人，在雷雨交加的夜晚拼命计算，连续好几个星期，就是为了证明弦理论在数学上有可能自洽。
So there's a famous story that Michael Green, a physicist from from Cambridge, and some other folks, like, furiously calculated into the night with thunderstorms, all kinds of stuff for weeks and weeks and weeks to show that string theory had a chance of being mathematically consistent.
一大堆数字，最后必须全部加起来等于零。
An enormous number of numbers just all had to just add up to zero.
这本来不一定是这样的。
It just did not have to be that way.
对。
Yep.
这种计算，今天我可以百分之百肯定，人工智能完全能做。
That kind of calculation today, I am a 100% sure that AI could do.
是的。
Yes.
所以，哪怕你只是有一点……
And so if you even had a...
一点预感，或者猜测这是真的，你就可以直接让人工智能帮你验证一下。
An inkling, a guess that this were true, you could just ask AI to to to verify it for you.
但你首先得猜到这个可能是真的。
But you had to kinda guess that this was gonna be true in the first place.
所以这就是有趣的地方。
And so this is kinda where you get the interesting stuff.
就是说，你得先提出一个好的猜想，然后再去验证它是否正确。
It's like, first you need to come up with a good conjecture and then check if that conjecture is true.
好的。
Okay.
物理学里就有这样的例子。
There is an example of that in physics.
OpenAI的一个模型对某个量子场论的振幅提出了猜想，并且成功证明了它。
A model from OpenAI made a conjecture about some certain quantum field theory amplitude and then was able to prove it.
但那种情况可能存在于量子场论中某个比较奇怪的正式角落。
But that sort of exists in kind of maybe a weird formal corner of quantum field theory.
我觉得更大的问题是，物理学中那些依赖于实验输入的重大成果。
I think the bigger question is like for the towering results in physics that did depend on experimental input.
比如，这个粒子到底存不存在？
Like, is this particle there or not?
它有什么性质？
What is its properties?
就像我和你一样，大家都想知道暗物质的质量是多少。
Like, you and I would both love to know what's the mass of dark matter.
所以我想问问你？
So maybe can I ask you?
我可以反过来问吗？
Can I turn this around?
AI 对这个问题会怎么回答呢？
What would an AI answer to that question even look like?
如果明天 ChatGPT 告诉你，暗物质的质量是 51.2 GeV。
If if chat GBT told you tomorrow, the mass of dark matter is 51.2 GeV.
我会问它为什么。
I would ask it why.
太好了。
Great.
那它得给出什么样的答案才能让你信服呢？
What kind of answer could it give you that would convince you?
我觉得，那几乎肯定得是经验性的。
I mean, it would almost certainly have to be empirical.
嗯，也不一定……
Well, not...
好吧。
Okay.
它可能从两个不同的方向来。
It could come from two different avenues.
比如，它可能说，我看了大量数据，做了分析，而之前看这些数据的人理解得不对。
It could be, I've looked at a bunch of data and I've analyzed it and the people who'd looked at it previously didn't understand it correctly.
如果理解正确的话，数据表明有一个暗物质粒子正在和我们的探测器相互作用。
If you understand it correctly, the data says there's a dark matter particle interacting with our detector.
它有这个质量。
It has this mass.
这是一种方式。
That would be one way.
另一种方式是，他们可以说，你知道，我研究了整个弦论模型的多重宇宙，或者其他类似的模型。
Another way is they could say, you know, I've looked through the whole string multiverse of all string theory models or some other model like that.
结果发现，只有那些与我们宇宙观测性质一致的模型才包含一个51 GeV的暗物质粒子。
And it turns out the only ones that are consistent with the observed properties of our universe contain a 51 GeV dark matter particle.
然后那个证明会有很多步骤我想去检查，但我可以想象，不是今天，而是十年、二十年或三十年后，类似的事情会发生。
And then there would be all sorts of steps of that proof I'd wanna check, but I could imagine, not today, but I could imagine ten or twenty or thirty years from now, something like that happening.
我……
I...
我不确定那个解释会不会让我满意，因为我觉得，也许下一步就是，好吧。
I'm not sure if I would find that explanation satisfying because I think I think maybe then the the next step is, okay.
去做个预测吧。
Go and make a prediction.
对。
That's right.
好。
Good.
所以它确实如此。
So then it does.
它说，你应该去做这个实验。
It says, you should go do this experiment.
如果那个实验要花一百亿美元呢？
What if that experiment costs $10,000,000,000?
你愿意一直坚持……
Are you willing to stay...
我的意思是，我都不知道明天的资金情况会怎样，更别说十年后了。
I mean, I don't know what the funding landscape's look like tomorrow, let alone in ten years.
但我觉得这就是我卡住的地方，作为物理学家，我们唯一感到满意的就是能通过实验来验证的预测。
But I think this is where I kinda get stuck is like, as physicists, the only way that we are satisfied is by, you know, a prediction that can be checked with an experiment.
那如果它说，是的。
And what if it says, yeah.
我知道答案，但要验证它得花掉全球的GDP，你知道吗？
I know the answer, but it would cost, you know, the world GDP to check it?
当然。
Sure.
所以，我的意思是，我们举个更接地气的例子，但这是个反事实的例子。
So, I mean, let's let's put it in a slightly more grounded example, but a a counter historical example.
当广义相对论刚提出时，几乎没什么办法能测试它。
When general relativity was proposed, there was very little that one could do to test it.
水星的轨道在一定程度上证实了它。
We had the orbit of Mercury confirmed it to a limited extent.
我们也能观察太阳周围的光线偏折，但说实话，那并不是很有说服力。
We could look at deflection around the sun, but frankly, that wasn't all that compelling.
直到六十年代发现并观测了双脉冲星，才真正清楚广义相对论是对时空和引力的准确描述。
It wasn't until the discovery of binary pulsars and the observations of them in the sixties that it really became clear that general relativity was an accurate description of space time and gravity.
那么假设这些望远镜的成本不是我们刚才说的那些相对适中的数字，而是100亿、1000亿甚至一万亿美元。
So let's say instead of costing whatever those telescopes cost, which was relatively modest as it turns out, but let's say they cost 10,000,000,000, a $100,000,000,000, a trillion dollars.
那样的话，我们还会去验证广义相对论吗？
Would we have tested general relativity in that case?
这是个好问题。
That's a good question.
我觉得吧……
I mean, I guess...
其实你可以说，LIGO探测器在五十年研发过程中的累计成本，就是一场这样的豪赌。
Well, you could argue that the LIGO detector, the cumulative cost of the LIGO detector over, like, the fifty years of its development was exactly that bet.
不是我们刚才说的那些数字。
Not the numbers we were just saying.
抱歉。
Well, sorry.
如果把从1960年到LIGO启动期间所有博士生的花费，算上通货膨胀，全部加起来的话，对。
If you add up literally the cost of every PhD student accounting for inflation from 1960 through when LIGO turned on, okay.
我不太确定。
I'm not sure.
可能是在十亿级别吧，是的。
Probably in the billion range, yeah.
十亿美元。
1,000,000,000.
但我觉得很多人对广义相对论不会感到满意。
But I think there are a lot of people that would have felt unsatisfied with general relativity.
所以让我说说我可能想到的那种答案。
So let me tell you the kind of answer that I might imagine.
我认识另一位物理学家，我们聊过这个，他担心以我们现有的数据和我们处理所有天体物理观测的方式，我们永远无法把暗物质粒子的质量确定到比十亿分之一更好的精度。
So there's another physicist that I've talked to about this who is worried that with the data that we have, with the way that we have of processing all of our astrophysical observations, we could never nail down the mass of the dark matter particle to, like, better than a factor of a billion.
比如，有……
Like, there...
有，有十亿……
There's there's a billion...
或者说，有十亿种可能的暗物质质量范围，都和我们现在掌握的数据一致。
Or, you know, a range of a billion possible dark matter masses that are all consistent with the data that we have right now.
就目前
As far
我们所知。
as we know.
就目前所知。
As far as we know.
不过，也许那是因为我们人类渺小的脑袋，只擅长看那些能画在纸上的图表。
However, maybe that's because our stupid puny human brains are only good at looking at plots that we can draw on a piece of paper.
你知道的，就是一个变量对另一个变量。
You know, one variable versus another variable.
在那堆庞大的数据里，藏着我们需要的相关性，用来确定暗物质粒子的质量。
Buried somewhere in all of that giant data are the correlations that we would need to nail down the mass of the dark matter particle.
我自己就曾在公开数据集中发现过那些创建数据集的人没发现的东西。
I mean, I've personally found things in public datasets that the people who created those datasets didn't find.
我完全相信，合适的AI算法也能做到同样的事情。
I have every confidence that the right AI algorithm could do the same thing.
所以我的猜测是，当涉及数据分析时，会是AI来编写机器学习算法。
So my bet would be when it involves data analysis, it will be AI writing a machine learning algorithm.
当然。
Sure.
好的。
Okay.
那么你可以说，太好了。
So then you could say, great.
所以，你知道，也可能是一个聪明的人类，足够聪明，写出了那个机器学习算法。
So, you know, it could have been a smart human that was clever enough to write that machine learning algorithm.
现在，如果我们更好地开发工具来理解机器学习的内部运作，这也是AI的基础，但也许我们还没有完全打通这个环节，我们就能更好地解释，到底是那些相关性中的什么导致了那个结果。
Now if we better develop our tools to understand the inner workings of the machine learning, which is also what underpins the AI, but maybe we haven't closed that circle yet, we will be better able to interpret what it is exactly about those correlations that led it to that.
然后我们可能就会拍着自己的脑袋说，当然，这太明显了，我怎么没想到。
And then we could just smack ourselves in the head and say, of course, it's so obvious why didn't I think of that.
最后再补充一点，这种对系统运作机制的理解，目前能帮我们让AI变得更安全、更可靠。
And maybe one more plug that I wanna make is that this same kind of understanding of what is it that makes these systems tick is the kind of thing that right now can help us make AI safer and more reliable.
我们现在能够构建出那些能做出惊人成就的系统，却并不清楚它们具体是怎么做到的，这在人类历史上是前所未有的。
The fact that we are able to build systems where we can make them do amazing things we don't know how they're doing it is pretty unprecedented in the history of humanity.
这就好比回到物理学领域，费曼晚年曾在黑板上写下那句名言：‘我无法创造的，我就不理解。’
So, like, to bring it back to physics, so Feynman on his blackboard at the end of his life had this famous quote, what I cannot create, I do not understand.
而我们现在的情况正好相反。
So we now have the negation of that statement.
所以这在逻辑上跟找到这个结论是独立的。
So it's actually logically independent from finding this statement.
但我们创造了一个我们并不理解的东西，明白。
But we have created a thing that we do not understand Mhmm.
这就需要我们作为一个整体，尤其是物理学家群体，去承担这项任务，我觉得我们在这方面非常有优势。
Which sort of demands that we as a community and physicists I think are are excellently well positioned to do this.
我们需要更深入地了解这些工具内部的运作机制，一方面是为了让它们在融入社会时能更平稳过渡，另一方面也是为了让它们能帮我们做出更好的科学研究。
We need to understand more about what is going on inside these tools both to try to help us, you know, you know, smooth the landing path for how these enter society but also so they can help us do better science.
尤尼，这期节目已经比平时长很多了。
Yoni, this is already turning out to be a much longer than normal episode.
所以我想尽快收尾，但我有一个问题想作为结尾。
So I think I wanna wrap up pretty quick, but I have one question I kinda wanna close on.
这更像是一个私人问题，你会认为自己或形容自己是AI乐观主义者，还是更偏向AI悲观主义者？
It's, like, it's more of a personal question, but would you consider yourself or describe yourself as an AI optimist or more of an AI pessimist?
我觉得我处于一种类似薛定谔的猫的叠加态，同时包含这两种状态。
I think I exist in a shorting your cat type superposition of those two of those two states.
如果所有AI发展现在就停止，我们只花时间去了解这些工具能做什么，我会完全没问题。
I would be perfectly fine if all AI development froze right here and we just spent time learning what these tools can do.
我觉得这可能吗？
I think Is that possible?
我们有没有……
Do we have the...
我的意思是，我们怎么才能做到这一点？
I mean, how would we make that happen?
我们到底能做什么来阻止它继续发展呢？
What would what would we possibly do to stop it from advancing?
哦，抱歉。
Oh, sorry.
嗯，这就像……
Well, that's like...
我以为你会问我们怎么能理解它。
I thought you're gonna ask how could we possibly understand it.
不。
No.
不。
No.
不。
No.
我们怎么能阻止它？
How could we stop it?
我……
I...
不是。
No.
不对。
No.
不知道。
Don't know.
我只是在想象一个假设的宇宙，在那里一切瞬间都静止了。
I'm just imagining a hypothetical universe where where everything froze right now.
就像那样，这就足够了。
Like, that would be enough.
我……
I...
你知道，我可能会感到悲观，这种悲观源于一种失落感：我进入学术界是为了能慢慢、仔细地思考问题，而我现在再也没有那种从容思考的奢侈了。
You know, pessimism that I might feel comes from the sinking feeling that I entered academia in order to think slowly and carefully about problems, and I no longer have the luxury of thinking slowly
还有仔细。
and carefully.
我来回答我自己的问题。
I'm gonna answer my own question.
在……方面，我是个非常乐观的AI支持者。
I'm very much an AI optimist in the...
如果你问我，我觉得AI会变得有多强大、速度有多快，我认为它会变得比几乎所有人目前想象的要强大得多，而且实现的速度也会比几乎所有人的准备都要快。
If you're asking me how powerful do I think AI is going to become and how quickly, I think it's gonna become much more powerful than almost anybody is currently imagining, and I think it's gonna happen faster than almost anyone's prepared for.
说到这背后的后果，我几乎可以肯定我是个AI悲观主义者。
And I am almost certainly an AI pessimist when it comes to what the consequences of that are going to be.
我想象着在不久的将来，会出现一种便宜、高效且普及的AI，它或许结合了机器人技术，能比人类更高效、更经济地完成几乎所有人类任务。
I imagine a world in the pretty near future where there are cheap, efficient, widespread AI that can do almost any human task, maybe combined with robotics, better and more cost effectively than we can.
顺便说一句，理论物理将是首批受到这种影响的领域之一。
And by the way, theoretical physics will be one of the first things that that will be true for.
一旦人类在经济上失去了价值，就很难理解为什么AI还要把我们留下来，或者是那些拥有AI的人为什么要留着我们，反正都是个问题。
And once human beings have no economic value, it's hard to understand why the AI would keep us around, or maybe the people who own the AI, either way.
我不知道我们会不会全被杀掉之类的，但我肯定觉得我们不会有多少自主权、权力或者对自己世界的掌控力。
I don't know if we're gonna be all killed or something, but I certainly don't think we'll have much autonomy or power or control over our world.
然后从一个更诗意的角度来说，作为一个物理学家，我更愿意生活在一个我们不知道所有答案、但能享受寻找答案的过程的世界，而不是一个所有答案都直接给到我们、我们甚至没有能力完全理解它们的世界。
And then on a more poetic notion, I'll just say that as a physicist, I would much rather live in a world where we don't know all our answers, but we get to enjoy the act of searching for them than to live in a world where all the answers are handed to us and we don't really even have the capacity to fully understand them.
我们只是知道答案是什么。
We just know what the answers are.
所以我要说，我同意你所有的观点，而且你说得再好不过了。
So I'm gonna say I agree with you on all those points and I and I couldn't have said it better.
我觉得，你知道……
I think, you know...
嗯，你之前说“我无法想象五年后的未来”时，给我的印象特别深刻。
Well, you you had incredible impressions when you said, I can't possibly speculate five years in the future.
因为确实，我们打赌的五年后就是两年后，而差不多到那个时候，我就无法想象事情会变成什么样了。
Because indeed, five years from our bet is two years from now and that is about where I stop being able to imagine what things are gonna look like.
我觉得世界会变得完全不一样。
I think the world will look completely different.
对。
And yeah.
所以再说一次，我想在你说的基础上加一点乐观的看法：我觉得像物理学家和数学家这样的人，他们习惯处理非常非常复杂的系统，但这些系统的基础相互作用却很简单。他们正是那种可以在接下来一到两年里运用自己的技能，尽一切努力去理解这些工具到底在做什么的人，也许还能在它彻底改变世界之前，窥见上帝的面容。
And so again, the one shade of optimism that I might put on top of what you said is that I think people like physicists and mathematicians who are used to dealing with very, very complicated systems that nonetheless have kind of simple fundamental interactions are the kind of people who can now put their skills to use in the next one to two years and doing everything we can to understand what it is these tools are doing to, I don't know, see the face of God before it completely changes the world.
带着这个乐观的结尾，这里是Yoni Kahn，感谢大家收听《为什么是这个宇宙》。
On that optimistic note, this has been Yoni Kahn, and, thanks all for listening to Why This Universe.
谢谢。
Thanks.
《为什么是这个宇宙》由我和Shalma Wegsman制作，由我们的好朋友Basal Distortion编辑。
Why This Universe is produced by myself and Shalma Wegsman and is edited by our good friend Basal Distortion.
我是Dan Hooper，威斯康星大学麦迪逊分校的物理学教授，也是威斯康星冰立方粒子天体物理中心的主任。
My name is Dan Hooper, and I'm a professor of physics at University of Wisconsin in Madison, and I'm the director of the Wisconsin Ice Cube Particle Astrophysics Center.
如果你想支持我们，可以考虑订阅我们的Patreon，每月只需3美元。
If you want to support us, consider subscribing to our Patreon for as little as $3 a month.
作为回报，你将获得无广告的节目，并且可以直接向我们提问，参与我们的“随便问”环节。
And in return for your support, you'll get ad free episodes and the ability to ask us questions directly for our ask us anything segments.
你可以在 patreon.com/whythisuniverse 找到我们。
You can find that at patreon.com/whythisuniverse.
感谢收听。
Thanks for listening.
关于 Bayt 播客
Bayt 提供中文+原文双语音频和字幕，帮助你打破语言障碍，轻松听懂全球优质播客。