吴恩达：从五岁编程到教育数百万人

以下是与安德鲁的对话，他是人工智能和技术领域最具影响力的教育家、研究员、创新者和领导者之一。他共同创立了 Coursera 和 Google Brain，创办了 DeepLearning.AI、Landing AI 和 AI Fund，并曾担任百度首席科学家。作为斯坦福大学教授，并通过 Coursera 和 DeepLearning.AI，他帮助教育和激励了数百万学生，包括我。这是人工智能播客。如果你喜欢，请在 YouTube 上订阅，在 Apple Podcasts 上打五星，在 Patreon 上支持，或者直接在 Twitter 上联系我@LexFriedman。像往常一样，我现在会做一两分钟的广告，中间绝不会插播广告打断对话。希望这不会影响你的收听体验。本期节目由 Cash App 呈现，它是 App Store 中排名第一的金融应用。下载后使用代码'LexPodcast'，你将获得 10 美元，Cash App 还会向我最喜欢的组织之一捐赠 10 美元，该组织致力于帮助全球年轻人推进机器人和 STEM 教育。现在，开始我与安德鲁的对话。

The following is a conversation with Andrew, one of the most impactful educators, researchers, innovators, and leaders in artificial intelligence and technology. He co-founded Coursera and Google Brain, launched DeepLearning.AI, Landing AI, and the AI Fund, and was the chief scientist at Baidu. As a Stanford professor and with Coursera and DeepLearning.AI, he has helped educate and inspire millions of students, including me. This is the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, give it five stars on Apple Podcasts, support on Patreon, or simply connect with me on Twitter at @LexFriedman. As usual, I'll do one or two minutes of ads now and never any ads in the middle that can break the flow of the conversation. I hope that works for you and doesn't hurt the listening experience. This show is presented by Cash App, the number one finance app in the App Store. When you get it, use code 'LexPodcast'. Cash App lets you send money to friends, buy Bitcoin, and invest in the stock market with as little as one dollar. Brokerage services are provided by Cash App Investing, a subsidiary of Square, and member SIPC. Since Cash App allows you to buy Bitcoin, let me mention that cryptocurrency in the context of the history of money is fascinating. I recommend 'The Ascent of Money' as a great book on this history. Debits and credits on ledgers started over 30,000 years ago. The US dollar was created over 200 years ago, and Bitcoin, the first decentralized cryptocurrency, was released just over ten years ago. So given that history, cryptocurrency is still very much in its early days of development, but it's still aiming to and just might redefine the nature of money. So again, if you get Cash App from the App Store or Google Play and use the code 'LexPodcast', you'll get $10, and Cash App will also donate $10 to one of my favorite organizations that is helping to advance robotics and STEM education for young people around the world. And now, here's my conversation with Andrew Ng.

你在斯坦福教授机器学习课程，后来又在 Coursera 上授课，教育和激励了数百万人。所以我想问你：在你年轻的时候，是什么人或想法激励你进入计算机科学和机器学习领域？你是什么时候第一次爱上这个领域的？

The courses you taught on machine learning at Stanford and later on Coursera have educated and inspired millions of people. So let me ask you: what people or ideas inspired you to get into computer science and machine learning when you were young? When did you first fall in love with the field?

我在香港和新加坡长大，五六岁时开始学习编程。那时我在学习 BASIC 编程语言。他们会拿这些书，告诉你把程序输入电脑。于是我把那些程序输入电脑，经过一番打字，我就能玩一些非常简单的射击游戏，这些游戏是我自己实现的。在我的旧电脑上，我觉得很神奇，作为一个小孩，我竟然能写出代码——其实只是把书上的代码抄到电脑里——然后玩这些酷炫的视频游戏。另一个时刻是在我十几岁的时候，我父亲是医生，他在读关于专家系统和神经网络的书。所以他让我也读了一些这些书，我觉得能写出展现智能的电脑真的很酷。我记得高中时在新加坡实习，做了很多复印工作。我是办公室助理，工作中最精彩的部分就是能用碎纸机。所以十几岁的我在想：‘天哪，复印这么多。要是我们能写软件、造个机器人来自动化这个，也许我就能做点别的事了。’所以我认为从那以后我的很多工作都围绕着自动化这个主题。甚至今天我思考机器学习的方式——我们非常擅长编写学习算法，可以自动化人们能做的事情。或者推出第一批 MOOC（大规模开放在线课程），后来促成了 Coursera——我也在尝试自动化我在校园教学中可以自动化的部分。教育过程——我试图自动化其中的一部分，使其更具可扩展性，让单个教师、单个教育者产生更大的影响。

Growing up in Hong Kong and Singapore, I started learning to code when I was five or six years old. At that time, I was learning the BASIC programming language. They would take these books and tell you to type this program into your computer. So I typed those programs into my computer, and as a result of all that typing, I would get to play these very simple shoot-'em-up games that I had implemented on my own. On my old computer, I thought it was fascinating as a young kid that I could write this code—really just copying code from a book into my computer—to then play these cool video games. Another moment for me was when I was a teenager, and my father, because he was a doctor, was reading about expert systems and about neural networks. So he got me to read some of these books, and I thought it was really cool that you could write a computer that started to exhibit intelligence. Then I remember doing an internship in high school in Singapore, where I remember doing a lot of photocopying. I was an office assistant, and the highlight of my job was when I got to use the shredder. So the teenager me was thinking, 'Boy, this is a lot of photocopying. If only we could write software, build a robot, something to automate this, maybe I could do something else.' So I think a lot of my work since then has centered on the theme of automation. Even the way I think about machine learning today—we're very good at writing learning algorithms that can automate things that people can do. Or even launching the first MOOCs (Massive Open Online Courses) that later led to Coursera—I was trying to also automate what could be automatable in how I was teaching on campus. The process of education—I tried to automate parts of that to make it more scalable, to have more impact from a single teacher, a single educator.

你在斯坦福早期亲自教授数千人，后来又在线上教授数百万人，有没有特别喜欢的回忆？

Do you have favorite memories from your early days at Stanford teaching thousands of people in person and then millions of people online?

你知道吗，在线教学——很多人不知道的是，很多视频是在晚上 10 点到凌晨 3 点之间拍摄的。很多时候，我们看着第一批 MOOC，那是我们为预告课程拍摄的，但已经有十万人注册了。我们刚开始写代码，实际上还没有拍摄视频。所以压力很大——十万人等着我们制作内容。所以很多周五和周六，我会出去和朋友吃饭，然后我在想：‘好吧，我现在是回家，还是去办公室拍视频？’想到可以帮助十万人学习机器学习，这让我决定：‘好吧，我去办公室，去我的临时录音室。’我会调整我的罗技网络摄像头，调整我的 Wacom 数位板，确保我的领夹麦克风开着，然后经常录制到凌晨 2 点或 3 点。我觉得幸运的是，看起来不像是在深夜录制的，但这真的很鼓舞人心——想到我们可以制作内容来帮助这么多人学习机器学习。

You know, teaching online—what not many people know was that a lot of those videos were shot between the hours of 10 p.m. and 3 a.m. A lot of times, we were watching the first MOOCs that we filmed with our announcement course, but a hundred thousand people had signed up. We just started to write the code, and we had not yet actually filmed the video. So there was a lot of pressure—a hundred thousand people waiting for us to produce the content. So many Fridays and Saturdays, I would go out, have dinner with my friends, and then I was thinking, 'Okay, do I want to go home now, or do I want to go to the office to film videos?' And the thought that I could help a hundred thousand people potentially learn machine learning unfortunately made me think, 'Okay, I'm going to go to my office, go to my makeshift recording studio.' I would adjust my Logitech webcam, adjust my Wacom tablet, make sure my lapel mic was on, and then I was often recording until 2 a.m. or 3 a.m. I think I'm fortunate that it doesn't show that it was recorded that late at night, but it was really inspiring—the thought that we could create content to help so many people learn about machine learning.

那是什么感觉？你很可能独自一人，也许和一两个朋友一起，用罗技网络摄像头录制，然后在凌晨一两点独自回家？

How does that feel? The fact that you're probably somewhat alone, maybe with a couple of friends, recording with a Logitech webcam and going home alone at 1 a.m. and 2 a.m.?

晚上知道这将会触及成千上万、最终数百万的人，那是什么感觉？是不是有一种坚持到底的满足感？

At night and knowing that that's going to reach thousands of people, eventually millions of people, what's that feeling like? Is there a feeling of just satisfaction of pushing through?

我觉得这很令人谦卑。我当时并没有考虑自己的曝光度。有一件事我很自豪地说我们从早期就做对了：我告诉我的整个团队，首要任务是做对学习者最有利的事，最好地服务学生。所以当我走进录音室时，我脑子里想的只有：我该说什么？我该如何设计幻灯片，让这些概念对学习者来说尽可能清晰？我有时看到讲师会受到诱惑：'嘿，我们来谈谈我的工作吧。如果我教我的研究，也许有人会多引用我的论文几次。'我认为我们在推出最初几门 MOOC 以及后来建立 Coursera 时做对的一件事，就是确立了这条基本原则：只做对学习者最有利的事，然后忘记其他一切。事实证明，这条指导原则对于这场运动的发展非常重要。

I think it's humbling. I wasn't thinking about what I was viewing. One thing I'm proud to say we got right from the early days was I told my whole team back then that the number one priority is to do what's best for learners, to best serve students. So when I went into a recording studio, the only thing on my mind was: what can I say? How can I design my slides to make these concepts as clear as possible for learners? I've seen sometimes instructors are tempted: 'Hey, let's talk about my work. Maybe if I teach about my research, someone will cite my papers a couple more times.' I think one thing we got right launching the first few MOOCs and later building Coursera was putting in place that bedrock principle: let's just do what's best for learners, then forget about everything else. That guiding principle turns out to be really important to the rise of the movement.

你想象中的学习者尽可能广泛，尽可能全球化，所以真正试图触及尽可能多对机器学习和 AI 感兴趣的人。我真的很想帮助任何对机器学习有兴趣的人进入这个领域。有时人们问我：'嘿，你为什么花那么多时间解释梯度下降？'我的回答是：如果我看学习者会从什么中受益，我觉得对基础有很好的理解，回归基本，会让他们在建立长期职业生涯时处于更有利的位置。所以你一直试图根据这个原则做决策。

The kind of learner you imagined in your mind is as broad as possible, as global as possible, so really try to reach as many people interested in machine learning and AI as possible. I really want to help anyone that had an interest in machine learning to break into the field. Sometimes people ask me: 'Hey, why do you spend so much time explaining gradient descent?' My answer was: if I look at what I think learners would benefit from, I felt that having a good understanding of the foundations, coming back to the basics, would put them in a better stead to build a long-term career. So you've tried to consistently make decisions on that principle.

你当时向狭义的 AI 社区以及全世界揭示的一件事是，实际上对 AI 感兴趣的人数比我们想象的要大得多。通过你教授这门课以及它变得如此受欢迎，这表明哇，这不只是去 NeurIPS 的一小群人。它要大得多。有开发者，有来自世界各地的人。我是俄罗斯人，所以俄罗斯每个人都真的很感兴趣。有大量程序员对机器学习感兴趣：印度、中国、南美，到处都有。只是有数百万人对机器学习感兴趣。从你的角度来看，你觉得这个数字有多大？

One of the things you revealed to the narrow AI community at the time, and to the world, is that the amount of people who are actually interested in AI is much larger than we imagined. By you teaching the class and how popular it became, it showed that wow, this isn't just a small community of people who go to NeurIPS. It's much bigger. It's developers, it's people from all over the world. I'm Russian, so everybody in Russia is really interested. This is a huge number of programmers who are interested in machine learning: India, China, South America, everywhere. There are just millions of people who are interested in machine learning. How big a sense do you get that this number of people is interested from your perspective?

我认为这个数字随着时间的推移而增长。这有点像那些看似一夜成名的事情，但作为内部建设者，它花了多年时间。所有那些一夜成功都花了多年才达到。我第一次涉足这类教育是在我们拍摄我的斯坦福课程并把视频放到 YouTube 上，然后还有一些其他事情，比如上传整个课程等等。基本上，我们放到 YouTube 上的那个一小时十五分钟的视频，然后我们还有四五个其他版本的网站，你们大多数人从未听说过，因为它们只触及了小众受众。但这让我能够迭代，让我的团队和我能够创新，学习哪些想法有效，哪些无效。例如，有一个我非常兴奋、非常自豪的功能：建立一个网站，允许多人同时登录。今天，如果你访问一个网站，如果你登录了，然后我想登录，你需要退出，在同一个浏览器、同一台电脑上。但我想：如果两个人，比如你和我，一起在电脑前看视频呢？如果网站可以让你输入你的名字和密码，然后我输入我的密码，现在电脑知道我们俩一起在看，并且给我们俩都记上我们作为一组所做的任何事情呢？我们在旧金山的一所高中推出了这个功能，Sacred Heart Cathedral Prep，老师很棒。我们有大约二十多个用户。你猜怎么着？零人使用这个功能。原来在线学习的人想自己看视频，这样你可以按自己的速度播放、暂停，而不是分组。所以这是一个从众多中学到的小教训的例子，它让我们能够磨练出正确的功能集。

I think the numbers have grown over time. It's one of those things that maybe feels like it came out of nowhere, but as an insider building it, it took years. All those overnight successes took years to get there. My first foray into this type of education was when we were filming my Stanford class and sticking the videos on YouTube, and then some other things with uploading the whole course and so on. Basically, the one hour fifteen minute video that we put on YouTube, and then we had four or five other versions of websites that most of you would never have heard of because they reached small audiences. But that allowed me to iterate, allowed my team and me to innovate, to learn what ideas work and what doesn't. For example, one of the features I was really excited about and really proud of was building this website where multiple people could be logged into the website at the same time. Today, if you go to a website, if you're logged in and then I want to log in, you need to log out on the same browser, same computer. But I thought: what if two people, say you and me, were watching a video together in front of the computer? What if a website could have you type your name and password, then me type in my password, and now the computer knows both of us are watching together, and it gives both of us credit for anything we do as a group? We rolled out this feature in a high school in San Francisco, Sacred Heart Cathedral Prep, great teacher. We had about 20-something users. And guess what? Zero people used the feature. It turns out people studying online want to watch the videos by themselves, so you can playback, pause at your own speed, rather than in groups. So that was one example of a tiny lesson learned out of many that allowed us to hone in on the set of features.

这听起来像是一个绝妙的功能，所以我想从中得到的教训是，有些东西在纸面上看起来很棒，但没人使用，它实际上并没有你想象的那么大的影响力。所以你确实经历了很多不同的功能和很多想法，才最终到达了 Coursera 的最终形态，那个强大的东西向世界展示了 MOOC 可以教育数百万人。而且我认为机器学习运动也是如此，它不是凭空出现的。相反，随着越来越多的人学习机器学习，他们会告诉他们的朋友，他们的朋友会看到它如何应用于他们的工作，社区不断增长。我认为我们仍在增长。我不知道未来有多少比例的开发者会成为 AI 开发者。我很容易看到超过 50%，对吧？因为广义上的 AI 开发者——不仅仅是做机器学习建模的人，还有基础设施、数据管道、所有围绕机器学习模型的软件——可能甚至更大。我觉得今天几乎每个软件工程师都对云有一定了解。但也许这个微控制器开发者不需要做云。但我觉得今天绝大多数软件工程师都欣赏云。我认为在未来，也许我们将接近百分之百的开发者以某种方式成为 AI 开发者，或者至少对机器学习有所了解。我的希望是，有一种效应：那些对成为程序员或从事软件工程并不真正感兴趣的人——比如生物学家、化学家、物理学家，甚至机械工程师——所有这些学科现在越来越多地拥有大型数据集。他们原本不认为自己会对编程感兴趣，直到他们有了这个数据集，并意识到有一套机器学习工具可以让他们利用这个数据集。所以他们实际上学会了编程，成为了新的程序员。所以不仅仅是更大比例的开发者成为机器学习人员，而且成为开发者的人群种类也在显著增长。

It sounds like a brilliant feature, so I guess the lesson to take from that is something that looks amazing on paper and then nobody uses it doesn't actually have the impact that you think it might have. So you really went through a lot of different features and a lot of ideas to arrive at the final Coursera, its final kind of powerful thing that showed the world that MOOCs can educate millions. And I think with the machine learning movement as well, it didn't come out of nowhere. Instead, what happened was as more people learned about machine learning, they would tell their friends, and their friends would see how it applies to their work, and the community kept on growing. I think we're still growing. I don't know in the future what percentage of developers would be AI developers. I could easily see it being more than 50%, right? Because so many AI developers broadly construed—not just people doing the machine learning modeling, but the infrastructure, data pipelines, all the software surrounding the machine learning model—maybe is even bigger. I feel like today almost every software engineer has some understanding of the cloud. But maybe this microcontroller developer doesn't need to do the cloud. But I feel like the vast majority of software engineers today have an appreciation of the cloud. I think in the future, maybe we'll approach nearly a hundred percent of all developers being in some way an AI developer, or at least having an appreciation of machine learning. And my hope is that there's this kind of effect: people who are not really interested in being a programmer or being into software engineering—like biologists, chemists, physicists, even mechanical engineers—all these disciplines that are now more and more sitting on large data sets. They didn't think they were interested in programming until they have this data set and they realize there's a set of machine learning tools that allow them to use the data set. So they actually learn to program and become new programmers. So not just a larger percentage of developers become machine learning people, but the kinds of people who are becoming developers is also growing significantly.

人类中有一部分是识字的，能读会写。也许你曾认为不是每个人都需要学会读写——你只需去听几个僧侣给你读就够了。或者我们只需要少数几个作者来写畅销书，其他人都不需要写。但我们发现，让尽可能多的人——在某些国家几乎是所有人——具备读写能力，极大地增强了人与人之间的交流。我们现在可以为一个人写作，比如我给你发邮件，你给我回邮件。我认为在计算领域，我们仍然处于这样一个阶段：会编程的人很少，以至于代码大多必须面向相对广泛的受众。但如果每个人——或者说大多数人——都在某种程度上成为开发者，就像发达经济体中的大多数人都有一定读写能力一样，我很乐意看到夫妻店的老板用很少的代码来定制他们本周特价商品的电视展示。我认为这增强了人机交流，这在今天变得越来越重要。

Part of humanity was literate; you could read and write. And maybe you thought not everyone needs to learn to read and write—you just go listen to a few monks read to you, and that was enough. Or maybe we just need a few handful of authors to write the bestsellers, and no one else needs to write. But what we found was that by giving as many people—in some countries almost everyone—literacy, it dramatically enhanced human-to-human communications. And we can now write for an audience of one, such as if I send you an email, you send me an email. I think in computing, we're still in that phase where so few people know how to code that the code mostly has to be for relatively large audiences. But if everyone—well, most people—became developers at some level, similar to how most people in developed economies are somewhat literate, I would love to see the owners of a mom-and-pop store, with very little code, customize the TV display for their special this week. And I think it enhances human-to-computer communications, which is becoming more and more important today as well.

所以你认为机器学习有可能变得类似于读写能力，就像你说的——夫妻店的老板，基本上各行各业的人，都会有一定程度的编程能力？我可以看到社会朝着那个方向发展。

So you think it's possible that machine learning becomes kind of similar to literacy, where—like you said—the owners of a mom-and-pop shop, basically everybody in all walks of life, would have some degree of programming capability? I could see society getting there.

还有一件有趣的事。如果我去和夫妻店老板聊，或者和许多从事日常职业的人聊，我以前很难解释为什么他们应该学习编程。我们给了他们一些理由，但随着机器学习和数据科学的兴起，我发现，在日常生活和工作中对数据科学有具体用途的人数，可能甚至超过了对软件工程有具体用途的人数。例如，如果你真的经营一家小夫妻店，我认为如果你能分析关于销售和客户的数据，那就有真正的价值——甚至可能比传统软件工程更有价值。所以我发现，对于我许多从事不同职业的朋友——招聘人员、会计师，或者我在工厂打交道越来越多的人——我觉得如果他们能在某种程度上成为数据科学家，他们就能立即在工作中运用这些技能。所以我认为，对很多人来说，数据科学和机器学习可能比软件工程更容易进入开发者世界。

There's one other interesting thing. If I go talk to the mom-and-pop store, if I talk to a lot of people in their daily professions, I previously didn't have a good story for why they should learn to code. We gave them some reasons, but what I found with the rise of machine learning and data science is that I think the number of people with a concrete use for data science in their daily lives and their jobs may be even larger than the number of people with a concrete use for software engineering. For example, if you actually run a small mom-and-pop store, I think if you can analyze the data about your sales, your customers, there's real value there—maybe even more than traditional software engineering. So I find that for a lot of my friends in various professions—being recruiters, accountants, or people that work in factories, which I deal with more and more these days—I feel if they were data scientists at some level, they could immediately use that in their work. So I think that data science and machine learning may be an even easier entree into the developer world for a lot of people than software engineering.

这很有趣。我同意，但你说得很好。我们生活在一个大多数课程和演讲都有幻灯片、PowerPoint、Keynote 的世界里，但你却以经常使用马克笔和白板而闻名。这种简单性很有说服力，至少对我来说，看起来很有趣。所以让我问一下：为什么即使在最大的舞台上，你也喜欢用马克笔和白板？

That's interesting. And I agree, but that's beautifully put. We live in a world where most courses and talks have slides, PowerPoint, Keynote, and yet you famously often still use a marker and a whiteboard. The simplicity of that is compelling, and for me at least, fun to watch. So let me ask: why do you like using a marker and whiteboard even on the biggest of stages?

我认为这取决于你想要解释的概念。对于数学概念，最好一次构建一个方程，而白板马克笔或触控笔是一种非常容易的方式，可以在你讲解的同时逐步构建方程、逐步构建复杂概念。有时这能增强可理解性。书写的缺点是慢，所以如果你想要一个长句子，很难写出来。所以我认为各有利弊。有时我用幻灯片，有时我用白板或触控笔。白板的慢也是它的优点：它迫使你把一切简化到基础。你的一些演讲用了白板——我的意思是，真的没有别的——但你进行得很慢，真正专注于最简单的原则。这强化了一种思想的极简主义，我认为这令人惊讶地——至少对我来说——非常适合教育。我认为一个好的演讲不是内容很多的演讲；一个好的演讲是只清晰地表达几个简单想法的演讲。我认为白板在某种程度上强化了这一点。

I think it depends on the concepts you want to explain. For mathematical concepts, it's nice to build the equation one piece at a time, and the whiteboard marker or the pen stylus is a very easy way to build up the equation, build up a complex concept one piece at a time while you're talking about it. And sometimes that enhances understandability. The downside of writing is that it's slow, so if you want a long sentence, it's very hard to write that. So I think there are pros and cons. Sometimes I use slides, and sometimes I use a whiteboard or a stylus. The slowness of a whiteboard is also its upside: it forces you to reduce everything to the basics. Some of your talks involve the whiteboard—I mean, there's really none—but you go very slowly and you really focus on the most simple principles. And that enforces a kind of minimalism of ideas that I think is surprisingly—at least for me—great for education. A great talk, I think, is not one that has a lot of content; a great talk is one that just clearly says a few simple ideas. And I think the whiteboard somehow enforces that.

Peter Abbeel，现在是世界上顶尖的机器人学家和强化学习专家之一，是你的第一个博士生。我提起他，是因为我想象这一定是你生命中一段有趣的时光。你有什么关于和你的第一个学生 Peter 一起工作的美好回忆吗？尤其是在深度学习真正爆发之前那段不确定的时期？

Peter Abbeel, who's now one of the top roboticists and reinforcement learning experts in the world, was your first PhD student. So I bring him up just because I kind of imagine this must have been an interesting time in your life. Do you have any favorite memories of working with Peter, your first student, in those uncertain times especially before deep learning really sort of blew up? Any favorite memories from those times?

我很幸运能有 Peter Abbeel 作为我的第一个博士生。我认为我长期的职业成功也建立在早期的基础或早期工作上，而 Peter 对此至关重要。所以我非常感激他和我一起工作。没有多少人知道研究有多难。Peter 的博士论文是用强化学习来驾驶直升机。实际上，即使在今天，网站 heli.stanford.edu——我不知道它是否还在——你可以看到我们用强化学习让直升机倒飞、做筋斗、翻滚的视频。这很酷。有史以来最令人难以置信的机器人视频之一。所以你怎么还能看到它？哦，是的，谢谢。那是 2000 年，还是 2008、2007、2006 年？真的吗？我爸爸说，是的，所以已经超过十年了。那对很多人来说真的很鼓舞人心。但很少有人看到它有多难。Peter、Adam Coates 和 Morgan Quigley 和我一起研究了直升机的各种版本，很多事情都不成功。例如，我们遇到的最难的问题之一是：当直升机倒飞做特技时，你如何确定位置？如何定位直升机？我们尝试了各种方法。只有一个 GPS 单元不行，因为你在倒飞；GPS 单元朝下，所以看不到卫星。所以我们尝试了两个 GPS 单元——一个朝上，一个朝下——但如果你翻转，那也不行，因为如果你快速翻转，朝下的那个无法同步。Morgan Quigley 正在探索一种疯狂复杂的专用硬件配置来解释 GPS 信号，研究 FPGA——完全疯狂。花了大约一年时间研究这个；没有成功。我记得 Peter——一个很棒的人——他和我坐在我的办公室里，看着我们尝试的最新失败的东西，说：“你知道吗？现在怎么办？”因为我们尝试了太多东西，但就是不行。最后，我们做的——Adam Coates 对此至关重要——是把摄像机放在地面上，用地面上的摄像机来定位直升机。这解决了定位问题，这样我们就可以专注于强化学习和逆强化学习技术，这些技术最终让直升机飞了起来。

I was really fortunate to have had Peter Abbeel as my first PhD student. And I think even my long-term professional success builds on early foundations or early work that Peter was so critical to. So I was really grateful to him for working with me. Not a lot of people know just how hard research was. Peter's PhD thesis was using reinforcement learning to fly helicopters. Actually, even today, the website heli.stanford.edu—I don't know if it's still up—you can watch videos of us using reinforcement learning to make the helicopter fly upside down, do loops, rolls. This is cool. One of the most incredible robotics videos ever. So how do you still watch it? Oh yeah, thanks. That's from like 2000, eight or seven or six? That really? My dad's like yeah, so it's over ten years old. That was really inspiring to a lot of people. But not many people see how hard it was. Peter and Adam Coates and Morgan Quigley and I worked on various versions of the helicopter, and a lot of things did not work. For example, one of the hardest problems we had was: when the helicopter is flying around upside down doing stunts, how do you figure out the position? How do you localize a helicopter? We tried all sorts of things. Having one GPS unit doesn't work because you're flying upside down; the GPS unit is facing down, so you can't see the satellites. So we tried having two GPS units—one facing up, one facing down—but if you flip over, that didn't work because the downward-facing one couldn't synchronize if you're flipping quickly. Morgan Quigley was exploring this crazy complicated configuration of specialized hardware to interpret GPS signals, looking into FPGAs—completely insane. Spent about a year working on that; didn't work. I remember Peter—great guy—him and me sitting down in my office, looking at the latest things we had tried that didn't work, and saying, "You know what? What now?" Because we tried so many things, and it just didn't work. In the end, what we did—Adam Coates was crucial to this—was put cameras on the ground and used cameras on the ground to localize the helicopter. And that solved the localization problem so that we could focus on the reinforcement learning and inverse reinforcement learning techniques that actually enabled the helicopter to fly.

你在斯坦福做这项工作时，当时有很多强化学习的理论论文，但实际应用不多。所以自主直升机的工作是当时少数几个强化学习的实际应用之一，这使它变得相当知名。我觉得我们今天几乎又回到了原点，强化学习有那么多炒作和兴奋，但我们又在寻找更多应用。Delica 提出的所有这些好想法，面对大多数人都在做理论工作的事实，是什么驱使你在不确定性和挑战中让直升机完成应用工作，让实际系统运转起来？

Doing this work at Stanford around that time, there was a lot of reinforcement learning theoretical papers but not a lot of practical applications. So the autonomous helicopter work for fine helicopters was one of the few practical applications of reinforcement learning at the time, which caused it to become pretty well known. I feel like we might have almost come full circle with today, there's so much hype and excitement about reinforcement learning, but again we're hunting for more applications. And all of these great ideas that Delica has come up with, what was the drive, sort of in the face of the fact that most people doing theoretical work, what motivated you in the uncertainty and the challenges to get the helicopter to do the applied work, to get the actual system to work?

是的，面对恐惧、不确定性和挫折，我喜欢能用的东西。我了解物理世界。就像回到碎纸机一样。我喜欢理论，但当我自己做理论工作时——这是个人品味，我不是说别人应该做我所做的——但当我做理论时，如果我觉得我的工作会影响人们、产生积极影响、帮助别人，我个人会更享受。我记得很多年前和一位数学教授交谈，他说：‘嘿，你是做什么的？’然后他回答时眼里闪着星星。这位数学家，不是斯坦福的，是另一所大学，他说：‘我做我所做的事，因为它帮助我发现宇宙中的真理和美。’他眼里闪着星星。我想，那很好。我不想那样做。我觉得有人那样做很好，我完全支持，我非常尊重那样做的人。但我更有动力的是，当我能看到我和我的团队所做的工作如何帮助人们。世界需要各种各样的人。我只是其中一种，不希望每个人都和我做同样的事。但当我深入理论或实践时，如果我个人确信有一条帮助人们的途径，我会觉得更满足。要有那种信念，那就是你的道路。

Yeah, in the face of fear, uncertainty, and setbacks, I like stuff that works. I know the physical world. It's like this back to the shredder. I like theory, but when I work on theory myself—and this is personal taste, I'm not saying anyone else should do what I do—but when I work on theory, I personally enjoy it more if I feel that the work I do will influence people, have positive impact, will help someone. I remember many years ago speaking with a mathematics professor, and he said, 'Hey, what do you do?' And then he had stars in his eyes when he answered. This mathematician, not from Stanford, a different university, he said, 'I do what I do because it helps me to discover truth and beauty in the universe.' He had stars in his eyes. And I thought, that's great. I don't want to do that. I think it's great that someone does that, fully supportive, people that do that, I have a lot of respect for them. But I am more motivated when I can see a line to how the work that my teams and I are doing helps people. The world needs all sorts of people. I'm just one type, not hoping everyone should do things the same way as I do. But when I delve into either theory or practice, if I personally have conviction that here's a pathway to help people, I find that more satisfying. To have that conviction that that's your path.

在深度学习被广泛接受之前，你就是它的支持者。你在这个领域看到了什么让你有信心？在第一个十年，也就是 2000 年代，你的思考过程是怎样的？

You were a proponent of deep learning before it gained widespread acceptance. What did you see in this field that gave you confidence? What was your thinking process like in that first decade, the 2000s?

我可以告诉你我们做错的事和做对的事。我们真正做错的是对无监督学习重要性的认识。谷歌大脑早期，我们在无监督学习上投入了大量精力，而不是监督学习。大约 2005 年，在一个叫 NIPS（现在叫 NeurIPS）的会议上，杰夫·辛顿和我坐在会场外的自助餐厅里吃午饭聊天。杰夫拿出一张餐巾纸，开始在上面画一个论点。这个论点非常有说服力。他论证道：人脑大约有 100 万亿个，也就是 10^14 个突触连接。你大约会活 10^9 秒，也就是 30 年。实际上你活 2 到 3 乘以 10^9 秒。所以就算 10^9 秒。如果每个突触连接，也就是你大脑神经网络中的每个权重，只有一个比特的参数，那就是你一生中需要在最多 10^9 秒内学习的 10^14 比特。通过这个简单的论点——非常简化——你一生中每秒需要学习 10^5 比特。我有一个一岁的女儿。我没有每秒向她指出 10^5 比特的标签。而且我认为我是一个非常有爱心的父母，但我就是不会那样做。所以从这个粗略的、肯定有问题的论点来看，我们所知道的大部分知识不可能通过监督学习获得。你获得如此多视觉信息的方式是从图像、音频和世界中的那些经验中吸收。所以那个论点，以及许多其他论点，确实让我相信无监督学习有很大的力量。所以那是我们可能做错的部分。我仍然认为无监督学习非常重要，但在早期，10 年、15 年前，我们所有人都认为那是前进的道路。

I can tell you the thing we got wrong and the thing we got right. The thing we really got wrong was the importance of unsupervised learning. Early days of Google Brain, we put a lot of effort into unsupervised learning rather than supervised learning. There was an argument around 2005 at a conference called NIPS, now NeurIPS, and Geoff Hinton and I were sitting in the cafeteria outside the conference, having lunch, chatting. Geoff pulled out a napkin and started sketching an argument on a napkin. It was very compelling. He argued: the human brain has about a hundred trillion, so 10^14 synaptic connections. You will live about 10^9 seconds, that's 30 years. Actually you live for 2 to 3 times 10^9 seconds. So let's say 10^9. If each synaptic connection, each weight in your brain's neural network, has just a one-bit parameter, that's 10^14 bits you need to learn in up to 10^9 seconds of your life. So via this simple argument—which is very simplified—that's 10^5 bits per second you need to learn in your life. I have a one-year-old daughter. I am not pointing out 10^5 bits per second of labels to her. And I think I'm a very loving parent, but I'm just not going to do that. So from this crude, definitely problematic argument, there's just no way that most of what we know is through supervised learning. The way you get so much visual information is from sucking in images, audio, those experiences in the world. So that argument, and a lot of other arguments, really convinced me that there's a lot of power to unsupervised learning. So that was the part we maybe got wrong. I still think unsupervised learning is really important, but in the early days, 10, 15 years ago, all of us thought that was the path forward.

所以你是说，也许那在当时是错误的直觉。那是我们做错的部分。我们做对的部分是规模的重要性。亚当·科茨，另一位了不起的人，很幸运曾与他共事，他当时在斯坦福是我的团队成员。亚当在斯坦福做了这些实验，表明我们训练的学习算法越大，性能越好。正是基于他生成的那张图，x 轴和 y 轴向右上方倾斜：更大的模型让事情更好，性能准确度是纵轴。所以正是基于亚当生成的那张图，让我确信你可以将这些模型扩展到比我们在斯坦福的几台 CPU 上所能做到的更大，然后我们可以得到更好的结果。正是基于亚当生成的那一张图，让我有信心去找塞巴斯蒂安·特龙，提议在谷歌启动一个项目，后来成为谷歌大脑项目。直觉是：规模会带来系统性能的提升，所以我们应该追求越来越大的规模。我认为人们没有意识到这有多么具有开创性。这很简单，但更大的数据集会带来更好的性能，这是一个开创性的想法。这在当时是有争议的。我机器学习社区里一些好心的朋友——我不点名，但我们认识的人——过来试图给我友好的建议：‘嘿，安德鲁，你为什么这么做？这太疯狂了。这不是正确的架构。看看我们正在构建的这些架构。你只是追求规模？那是一个糟糕的职业举动。’所以我的好心朋友们试图劝我放弃。我发现，如果你想取得突破，有时你必须有信念，在它流行之前就去做，因为这样能让你产生更大的影响。

So you're saying that perhaps that was the wrong intuition for the time. That was the part we got wrong. The part we got right was the importance of scale. Adam Coates, another wonderful person, fortunate to have worked with him, he was in my group at Stanford at the time. Adam had run these experiments at Stanford showing that the bigger we train a learning algorithm, the better performance. It was based on that graph he had generated, where the x-axis and y-axis lines going up to the right: bigger models make the thing better, performance accuracy is the vertical axis. So it's really based on that chart that Adam generated that gave me the conviction that you could scale these models way bigger than what we could on the few CPUs we had at Stanford, and then we could get even better results. It was really based on that one figure that Adam generated that gave me the conviction to go to Sebastian Thrun and pitch starting a project at Google, which became the Google Brain project. And the intuition was: scale will bring performance for the system, so we should chase larger and larger scale. I think people don't realize how groundbreaking it is. It's simple, but it's a groundbreaking idea that bigger datasets will result in better performance. It was controversial at the time. Some of my well-meaning friends in the machine learning community—I won't name names, but people we know—came and tried to give me friendly advice: 'Hey Andrew, why are you doing this? This is crazy. It's not the right architecture. Look at these architectures we're building. You just go for scale? That's a bad career move.' So my well-meaning friends were trying to talk me out of it. I find that if you want to make a breakthrough, you sometimes have to have conviction and do something before it's popular, since that lets you have a bigger impact.

让我就这个话题问一个小问题。我发现自己和人们争论，他们说更大的规模，尤其是在主动学习的背景下——非常仔细地选择数据集——但扩大数据集的规模将导致深度学习的进一步突破。目前有人反对这个想法，认为更大的数据集不再那么有用。所以你想增加规模吗？

Let me ask you just a small tangent on that topic. I find myself arguing with people saying that greater scale, especially in the context of active learning—so very carefully selecting the dataset—but growing the scale of the dataset is going to lead to even further breakthroughs in deep learning. And there's currently pushback at that idea, that larger datasets are no longer that useful. So you want to increase the scale?

学习效率：你想要更好的学习机制，我个人认为，更大的数据集，即使使用我们现有的相同学习方法，也会带来更好的性能。你目前对此有什么直觉？我看到这种两面性：我们需要提出更好的学习架构，还是只需要更大、更好的数据集就能提高性能？

Efficiency of learning: you want to make better learning mechanisms, and I personally believe that bigger data sets will still, with the same learning methods we have now, result in better performance. What's your intuition at this time on those? I see this dual side: do we need to come up with better architectures for learning, or can we just get bigger, better data sets that will improve performance?

我认为两者都很重要，而且这也取决于具体问题。对于少数数据集，我们可能正在接近贝叶斯误差率，接近或超越人类水平的表现，然后我们会遇到那个理论天花板。但我认为还有很多问题，我们离人类水平的表现或贝叶斯误差率还很远，更大的数据集加上神经网络，即使没有进一步的创新，也足以让我们走得更远。但另一方面，如果我们看看最近使用 Transformer 网络在语言模型上的突破，那是新颖架构和规模共同作用的结果。如果我们看看 GPT-2 和 BERT 的情况，我认为规模是其中很大一部分原因。

I think both are important, and it's also problem-dependent. For a few data sets, we may be approaching the Bayes error rate, approaching or surpassing human-level performance, and then there's that theoretical ceiling that we will hit. But I think there are plenty of problems where we're still quite far from either human-level performance or from Bayes error rate, and bigger data sets with neural networks, but without further elaborate innovation, will be sufficient to take us further. But on the flip side, if we look at the recent breakthroughs using Transformer networks for language models, it was a combination of novel architecture but also scale has a lot to do with it. If we look at what happened with GPT-2 and BERT, I think scale was a large part of the story.

是的，这经常被忽视：训练数据集的大小和质量。因为有些数据，比如 Reddit 上高赞的帖子，已经带有某种弱监督，这是一个很大的数据集，但人们很少谈论这一点，对吧？

Yeah, that's not often talked about: the scale of the data set it was trained on and the quality of the data set. Because there's some, so it was like Reddit threads that had been upvoted highly, so there's already some weak supervision on a very large data set that people don't often talk about, right?

我发现如今我们在管理代码方面有成熟的过程，比如 Git、版本控制。我们花了很长时间才演化出好的流程。我记得我和朋友们曾经通过电子邮件互相发送 C++文件，但后来我们有了 CVS、Subversion、Git。未来可能还会有别的。但在管理数据方面，我们非常不成熟。想想我们如何创建和管理数据。我对混乱的数据问题非常热衷。我认为那里有很多创新空间。

I find it today we have maturing processes for managing code, things like Git, version control. It took us a long time to evolve good processes. I remember when my friends and I were emailing each other C++ files in email, you know, but then we had CVS, Subversion, Git. Maybe something else in the future. We're very immature in terms of managing data. And think about how we create and how we manage data. I'm very hot on messy data problems. I think there's a lot of innovation there to be had.

我仍然喜欢你们通过电子邮件进行版本控制的想法。我给你举个例子。当我们与制造公司合作时，经常会有多个标签列表相互矛盾的情况，对吧？所以我们在做视觉检测工作。我们会拿一个塑料零件给一个检查员看，检查员有时非常主观：‘哦，这明显是个缺陷，有划痕，明白吗？所以要检查这个零件。’把同一个零件给另一个检查员，不同的，非常主观：‘很明显划痕很小，没问题，别扔掉，你会让我们亏钱的。’然后有时你拿同一个塑料零件，下午给同一个检查员看，和上午看的结果完全不同。上午他说明显没问题，下午同样自信地说明显是缺陷。那么，如果一个人甚至在一天之内都不认同自己，AI 团队该怎么办？所以我认为这些是非常实际、非常混乱的数据问题，我的团队一直在与之斗争。在大型消费互联网公司，你有十亿用户，你有很多数据，你不用担心，你只需取平均值，它就能工作。但在其他行业环境中，我们没有大数据，只有小数据，非常小。用户可能只有 100 个有缺陷的零件或 100 个缺陷示例。如果你只有 100 个示例，这些小标签错误：如果你的 100 个标签中有 10 个是错误的，那实际上是 10%，影响很大。那么你如何清理这些数据？你应该怎么做？这是 Landing AI 团队一直在处理的问题的一个例子，处理小数据，一旦你离开消费互联网，这种情况就经常出现。

I still love the idea that you were versioning through email. I'll give you one example. When we work with manufacturing companies, it's not at all uncommon for there to be multiple label lists that disagree with each other, right? So we were doing work in visual inspection. We would take, say, a plastic part and show it to one inspector, and the inspector sometimes is very opinionated: 'Oh, clearly that's a defect, a scratch, understand? So gonna check this part.' Take the same part to a different inspector, different, very opinionated: 'Clearly the scratch is small, it's fine, don't throw it away, you're gonna make us lose money.' And then sometimes you take the same plastic part, show it to the same inspector in the afternoon versus in the morning, and very different. In the morning they say clearly it's okay, in the afternoon equally confident clearly this is a defect. So what is the AI team supposed to do if sometimes even one person doesn't agree with himself or herself in the span of a day? So I think these are the types of very practical, very messy data problems that my teams wrestle with. In the case of large consumer internet companies where you have a billion users, you have a lot of data, you don't worry about it, you just take the average, it kind of works. But in the case of other industry settings, we don't have big data, just small data, very small. The users maybe 100 defective parts or 100 examples of a defect. If you have only 100 examples, these little labeling errors: if 10 of your 100 labels are wrong, that's actually 10%, it has a big impact. So how do you clean this up? What are you supposed to do? This is an example of the types of things that my teams at Landing AI wrestle with, to deal with small data, which comes up all the time once you're outside consumer internet.

是的，这很有趣。那么你会投入更多精力和时间来思考实际的标注过程：标签是什么，分歧如何解决，所有这些务实的现实世界问题。这是一个迷人的领域。

Yeah, that's fascinating. So then you invest more effort and time in thinking about the actual labeling process: what are the labels, how are disagreements resolved, all those kinds of pragmatic real-world problems. That's a fascinating space.

是的，我发现实际上当我在斯坦福教书时，我越来越鼓励斯坦福的学生尝试为期末项目找到自己的项目，而不是仅仅下载别人整理好的干净数据集。如果你需要去定义自己的问题并找到自己的数据集，实际上要困难得多，而不是去那些有精心整理的数据集的网站，直接拿来用。

Yeah, I find it actually when I'm teaching at Stanford, I increasingly encourage students at Stanford to try to find their own project for the end-of-term project rather than just downloading someone else's nicely clean data set. It's actually much harder if you need to go and define your own problem and find your own dataset, rather than going to one of the several good websites with nicely curated datasets that you could just work on.

你现在在运营三个项目：AI Fund、Landing AI 和 DeepLearning.AI。正如你所说，AI Fund 专注于从零开始创建新公司，Landing AI 帮助已有公司实施 AI，而 DeepLearning.AI 则面向其他所有人，或者那些有兴趣进入该领域并取得卓越成就的个人。那么，我们也许可以谈谈这些领域。首先，DeepLearning.AI：一个基本问题，对深度学习感兴趣的人如何入门？

You're now running three efforts: the AI Fund, Landing AI, and DeepLearning.AI. As you've said, the AI Fund is involved in creating new companies from scratch, Landing AI is involved in helping already established companies do AI, and DeepLearning.AI is for education of everyone else, or of individuals interested in getting into the field and excelling. So let's perhaps talk about each of these areas. First, DeepLearning.AI: the basic question, how does a person interested in deep learning get started in the field?

DeepLearning.AI 致力于创建课程，帮助人们进入 AI 领域。所以我在斯坦福教授的那门机器学习课程至今仍是 Coursera 上最受欢迎的课程之一。它可能是那种课程，如果我问某人‘你是怎么进入机器学习领域的？’或‘你是怎么爱上机器学习的？’或‘是什么让你感兴趣的？’，他们总会追溯到那门课。你影响的人数之多简直不可思议。为此，我相信我代表很多人说一声非常感谢。

DeepLearning.AI is working to create courses to help people break into AI. So my machine learning course that I taught through Stanford is one of the most popular courses on Coursera to this day. It's probably one of those courses where if I ask somebody 'How did you get into machine learning?' or 'How did you fall in love with machine learning?' or 'What got you interested?', it always goes back to that course at some point. You won't find the amount of people you influence is ridiculous. So for that, I'm sure I speak for a lot of people saying big thank you.

不，谢谢你。你知道，我曾经读一篇新闻文章，我想是《技术评论》，我可能会记错数据，但我记得一篇文章说大约三分之一的程序员是自学的。我可能记错了数字，也许是三分之二。但我读了一篇文章，心想这说不通：每个人都是自学的，因为你是自己教自己。我不教别人。这不好。哈哈。那么，一个人如何开始深度学习，DeepLearning.AI 在其中扮演什么角色？

No, thank you. You know, I was once reading a news article, I think it was Tech Review, and I'm gonna mess up the statistic, but I remember reading an article that said something like one-third of all programmers are self-taught. I may have the number wrong, maybe it was two-thirds. But I read an article and I thought this doesn't make sense: everyone is self-taught because you teach yourself. I don't teach people. It's no good. Haha. So how does one get started in deep learning and where does DeepLearning.AI fit into that?

所以 DeepLearning.AI 提供的深度学习专项课程，我想它叫深度学习专项课程，可能现在还是。这是一个非常受欢迎的方式，人们通过这个专项课程学习从神经网络到如何调优网络的一切。比如什么是 CNN，什么是 RNN，或序列模型，或什么是注意力模型。深度学习专项课程逐步引导大家学习这些算法，让你深入理解并能从头实现和使用它们。

So the Deep Learning Specialization offered by DeepLearning.AI is, I think it was called the Deep Learning Specialization, it might still be. It's a very popular way for people to take that specialization to learn about everything from neural networks to how to tune your network. So what is a CNN, what is an RNN, or sequence model, or what is an attention model. And the Deep Learning Specialization steps everyone through those algorithms so you deeply understand them and can implement and use them for whatever, from the very beginning.

那么，对于想要学习深度学习专项课程的人来说，你认为需要什么先决条件，比如数学或编程背景？

So what would you say the prerequisites are for somebody to take the Deep Learning Specialization, in terms of maybe math or programming background?

你需要了解基本编程，因为课程中有 Python 编程练习。

You need to understand basic programming, since there are programming exercises in Python.

深度学习专业课程有哪些先修要求？

What are the prerequisites for the deep learning specialization?

先修要求相当基础。不需要微积分，当然如果你懂微积分会更好，能获得更直观的理解。但我们特意尝试在不要求微积分的情况下教授这门专业课程。我认为高中数学就足够了。如果你知道如何做两个矩阵的乘法，那基本上就是线性代数，甚至是非常基础的线性代数。还需要一些编程知识。学过机器学习专业课程的人会觉得深度学习专业课程稍微容易一些，但直接学习深度学习专业课程也是可以的。只是会稍微难一点，因为我们往往会更快地讲解一些概念，比如梯度下降是如何工作的，以及什么是目标函数，这些主要在机器学习课程中涵盖。

The prerequisites are quite basic. No calculus is needed, though if you know calculus, it's great and you get better intuitions. But we deliberately try to teach the specialization without requiring calculus. I think high school math would be sufficient. If you know how to multiply two matrices, that's basically linear algebra, even very basic linear algebra. And some programming. People who have done the machine learning specialization also find the deep learning specialization a bit easier, but it's also possible to jump into the deep learning specialization directly. It'll be a little bit harder since we tend to go over concepts faster, like how gradient descent works and what an objective function is, which is covered mostly in the machine learning course.

你能简要介绍一下学生在最初几个月或第一年内应该学习的深度学习关键概念吗？

Could you briefly mention some of the key concepts in deep learning that students should learn in the first few months or first year?

如果你学习深度学习专业课程，你会学到基础：什么是神经网络，如何从单个单元、堆叠层、不同的激活函数构建神经网络，以及如何训练神经网络。在这门专业课程中，我非常自豪的一点是，我们涵盖了大量关于如何实际让这些东西工作的实用知识。比如，不同优化算法之间的区别是什么？如果算法过拟合了，你该怎么办？如何判断算法是否过拟合？什么时候应该收集更多数据？什么时候不应该费心去收集更多数据？我发现，即使到今天，不幸的是，仍然有工程师会花六个月时间试图追求某个方向，比如收集更多数据，因为我们听说更多数据是有价值的。但有时你可以运行一些测试，并且可能在六个月前就已经发现，对于这个问题，收集更多数据是行不通的。所以不要花六个月收集更多数据；把你的时间花在修改架构或尝试其他方法上。我们还涵盖了大量实用知识。当你学习深度学习专业课程时，你会有技能非常高效地构建神经网络。所以直接上手摆弄网络，训练它，在特定数据集上进行推理，建立直觉，而不要把它搞得太大，以至于你花了六个月构建大项目，却没有对数据的一个小方面建立直觉，而这个小方面可能已经告诉了你关于该数据的一切你需要知道的东西。

If you take the deep learning specialization, you learn foundations: what is a neural network, how do you build up a neural network from a single unit, a stack of layers, different activation functions, and how to train neural networks. One thing I'm very proud of in that specialization is that we go through a lot of practical know-how of how to actually make these things work. So what are the differences between different optimization algorithms? What do you do if the algorithm overfits? How do you tell the algorithm is overfitting? When should you collect more data? When should you not bother to collect more data? I find that even today, unfortunately, there are engineers who will spend six months trying to pursue a particular direction, such as collecting more data because we heard more data is valuable. But sometimes you could run some tests and could have figured out six months earlier that for this problem, collecting more data isn't going to cut it. So just don't spend six months collecting more data; spend your time modifying the architecture or trying something else. We also go through a lot of the practical know-how. When you take the deep learning specialization, you have the skills to be very efficient in how you build a neural net. So dive right in to play with the network, to train it, to do inference on a particular dataset, to build an intuition about it without building it up too big, where you spend six months building your big project without building an intuition of a small aspect of the data that could already tell you everything you need to know about that data.

还有关于如何构建实用机器学习的系统性思维框架。打个比方，当我们学习编程时，我们必须学习某种编程语言的语法，对吧？无论是 Python 还是 C++ 或其他。但同样重要，甚至更重要的是理解如何将这些代码行串成连贯的东西。所以你知道什么时候应该把东西放在函数调用中，什么时候不应该，如何思考抽象。这些框架才是让程序员高效的原因，甚至比理解语法更重要。我记得我在卡内基梅隆大学读本科时，有一个朋友调试代码的方法是先尝试编译，然后那是 C++ 代码，每一行都有语法错误。他们想尽快修复语法错误。那么怎么做呢？他们会删除每一行有语法错误的代码。所以对于修复语法错误非常高效，但其他方面就很糟糕了。我想我们学会了如何调试。而在机器学习中，调试机器学习程序的方式与在传统软件工程中使用二分查找或调试器跟踪代码的方式非常不同。这是一个不断发展的学科，但我发现真正擅长调试机器学习算法的人，在让东西工作起来的速度上轻松达到 10 倍，甚至 100 倍。调试的基本过程是：这里的 bug 是为什么这个东西不学习？或者为什么它没有改进？它涉及过拟合以及所有这类问题。这就是在神经网络中调试发生的逻辑空间。常见的问题是：为什么它还不工作？好吧，我能期望它最终工作吗？我可以尝试哪些事情？改变架构？更多正则化？不同的优化算法？不同类型的数据？所以系统地回答这些问题，这样你就不会花六个月时间走进死胡同，然后有人过来说：‘你为什么花了六个月做这个？’

And also the systematic frameworks of thinking for how to go about building practical machine learning. To make an analogy, when we learn to code, we have to learn the syntax of some programming language, right? Be it Python or C++ or whatever. But equally important, maybe even more important, is to understand how to string together these lines of code into coherent things. So you know when should you put something in a function call and when should you not, how do you think about abstraction. Those frameworks are what make the programmer efficient, even more than understanding the syntax. I remember when I was an undergrad at Carnegie Mellon, one of my friends would debug their code by first trying to compile it, and then it was C++ code, and every line had a syntax error. They wanted to fix the syntax errors as quickly as possible. So how do you do that? Well, they would delete every single line of code with a syntax error. So really efficient for fixing syntax errors, but horrible otherwise. I think we learned how to debug. And in machine learning, the way you debug a machine learning program is very different than the way you do binary search or use the debugger to trace through the code in traditional software engineering. It's an evolving discipline, but I find that people who are really good at debugging machine learning algorithms are easily 10x, maybe 100x faster at getting something to work. The basic process of debugging is: the bug in this case is why isn't this thing learning? Or why isn't it improving? It goes into the questions of overfitting and all those kinds of things. That's the logical space that the debugging is happening in, in a neural network. The often question is: why doesn't it work yet? Well, can I expect it to eventually work? And what are the things I could try? Change the architecture? More regularization? Different optimization algorithm? Different types of data? So to answer those questions systematically, so that you don't spend six months heading down a blind alley before someone comes and says, 'Why did you spend six months doing this?'

你认为学生在深度学习中哪些概念最难掌握？或者说，他们克服那个让他们着迷、激励他们、真正理解的门槛的最大挑战是什么？

What concepts in deep learning do you think students struggle with the most? Or what is the biggest challenge for them to get over that hill that hooks them and inspires them and they really get it?

类似于学习数学，我认为深度学习的挑战之一是很多概念是相互叠加的。如果你问我数学难在哪里，我很难指出一件事。是加法？减法？进位？乘法？长除法？有很多东西。我认为学习数学和学习某些技术领域的挑战之一是很多概念相互依赖，如果你错过了一个概念，那么你就错过了后面某个东西的先决条件。所以在深度学习专业课程中，我们尝试分解概念，以最大化每个部分的可理解性。所以当你进入更高级的内容时，你会学到信心。希望你有足够的早期部分的直觉，从而理解为什么我们以某种方式构建信心，然后最终为什么我们以某种方式构建神经网络或注意力模型，这些都是建立在早期概念之上的。

Similar to learning mathematics, I think one of the challenges of deep learning is that there are a lot of concepts that build on top of each other. If you ask me what's hard about mathematics, I have a hard time pinpointing one thing. Is it addition? Subtraction? Is it carrying? Is it multiplication? Long division? There's a lot of stuff. I think one of the challenges of learning math and of learning certain technical fields is that a lot of concepts build on each other, and if you miss a concept, then you're kind of missing the prerequisite for something that comes later. So in the deep learning specialization, we try to break down the concepts to maximize the chance that each component is understandable. So when you move on to the more advanced thing, you learn confidence. Hopefully you have enough intuitions from the earlier sections to then understand why we structure confidence in a certain way, and then eventually why we build neural networks or attention models in a certain way, building on top of the earlier concepts.

我很好奇，你也做了很多教学。在你的教学中，有没有一个你最喜欢的‘这是最难的概念’的时刻？

I'm curious, you do a lot of teaching as well. Do you have a favorite 'this is the hard concept' moment in your teaching?

嗯，我不认为有人曾经反过来采访过我。我觉得这是个非常好的问题。真的很难捕捉到他们挣扎的时刻。我认为你说得非常贴切。我确实认为有些时刻像是‘啊哈’时刻，真正激励了人们。我想出于某种原因，强化学习，尤其是深度强化学习，是一种非常好的方式来真正激励人们，并理解神经网络能做什么。尽管神经网络只是深度强化学习框架的一部分，但它是一种很好的方式来描绘神经网络从零开始学习、一无所知、并探索的完整图景。

Well, I don't think anyone's ever turned the interview on me. I think that's a really good question. It's really hard to capture the moment when they struggle. I think you put it really eloquently. I do think there are moments that are like 'aha' moments that really inspire people. I think for some reason, reinforcement learning, especially deep reinforcement learning, is a really great way to really inspire people and get what the use of neural networks can do. Even though neural networks are really just a part of the deep RL framework, it's a really nice way to paint the entirety of the picture of a neural network being able to learn from scratch, knowing nothing, and explore.

我发现很多顿悟时刻发生在你用深度强化学习教人们神经网络的时候，这有点反直觉。我觉得很多灵感和热情来自强化学习领域。你觉得学习是教学过程中有用的一部分吗？

I find that a lot of the aha moments happen when you use deep RL to teach people about neural networks, which is counterintuitive. I find a lot of the inspired fire and people's passion comes from the RL world. Do you find learning to be a useful part of the teaching process or not?

我仍然在斯坦福的一门课和我的博士论文中教授强化学习。但如果我想教学生今天最有用的技术，我会减少讲强化学习的时间。这不是今天最有效的方法。我们的世界变化很快，也许几年后情况会完全不同。我认为强化学习还需要一些东西才能达到那个水平。我的一个团队正在研究强化学习用于机器人控制任务，所以我看到了应用。但如果你把它看作我们所做事情中影响的一部分，至少今天，除了玩电子游戏和一些游戏之外，范围是有限的。我们一群人曾站在一起说：‘嘿，你们有什么实际部署的强化学习应用的最佳例子？’在机器学习研究人员中，有一些新兴的例子，但并没有很多伟大的例子。

I still teach reinforcement learning in one of my Stanford classes and my PhD thesis. But if I'm trying to teach students the most useful techniques for them to use today, I end up shrinking the amount of time I talk about reinforcement learning. It's not what's working today. Our world changes so fast, maybe it will be totally different in a couple years. I think we need a couple more things for reinforcement learning to get there. One of my teams is looking at reinforcement learning for some robotic control tasks, so I see the applications. But if you look at it as a percentage of all the impact of the types of things we do, at least today, outside of playing video games and a few games, the scope is limited. A bunch of us were standing around saying, 'Hey, what's your best example of an actual deployed reinforcement learning application?' Among machine learning researchers, there are some emerging ones, but not that many great examples.

嗯，我完全同意。可悲的是，强化学习还没有一个大的、有影响力的现实世界应用。它最大的影响是在玩具领域、游戏领域和小型例子中。出于教育目的，它似乎是一个探索神经网络的有趣工具。但从你的角度来看，如果你想用一个简单的例子来教育人们，说明这实际上可以扩展到规模并产生现实影响，那么也许专注于监督学习的基础，在像 MNIST 这样的简单数据集上，是正确的道路。我看到人们在强化学习中获得的乐趣很多，但在现实世界的应用影响上却不多。所以这是一个权衡：你想要多少影响，你想要多少乐趣。

Well, I think you're absolutely right. The sad thing is there hasn't been a big, impactful, real-world application of reinforcement learning. Its biggest impact has been in the toy domain, in the game domain, in small examples. For educational purposes, it seems to be a fun thing to explore neural networks with. But from your perspective, if you're trying to educate with a simple example to illustrate how this can actually be grown to scale and have real-world impact, then perhaps focusing on the fundamentals of supervised learning in the context of a simple dataset, even like an MNIST dataset, is the right path to take. The amount of fun I've seen people have with reinforcement learning has been great, but not in the applied impact on the real-world setting. So it's a trade-off: how much impact you want to have versus how much fun you want to have.

是的，这真的很酷。我觉得世界实际上需要多样化的技能。即使在机器学习内部，深度学习非常令人兴奋，但 AI 团队不应该只使用深度学习。我的团队使用一套工具组合。也许这不是什么激动人心的话，但有些日子我们用神经网络，有些日子我们用 PCA。前几天我和团队坐下来查看 PCA 残差，试图弄清楚 PCA 在制造问题中的应用。有时我们使用概率图模型，有时我们使用知识图谱。其中一些东西有巨大的行业影响，但学术界关于知识图谱的讨论相比实际影响要少得多。所以我认为强化学习应该在这个组合中，问题在于平衡我们教授所有这些内容的时间。世界应该拥有多样化的技能。如果每个人都只学一个狭窄的东西，多样化的技能能帮助你发现适合工作的工具。

Yeah, that's really cool. I feel like the world actually needs diverse skills. Even within machine learning, deep learning is so exciting, but the AI team shouldn't just use deep learning. My team uses a portfolio of tools. Maybe that's not the exciting thing to say, but some days we use neural nets, some days we use PCA. The other day I was sitting down with my team looking at PCA residuals, trying to figure out what's going on with PCA applied to a manufacturing problem. Sometimes we use a probabilistic graphical model, sometimes we use a knowledge graph. Some of these things have tremendous industry impact, but the amount of chat about knowledge graphs in academia is really thin compared to the actual real-world impact. So I think reinforcement learning should be in that portfolio, and it's about balancing how much we teach all of these things. The world should have diverse skills. If everyone just learns one narrow thing, diverse skills help you discover the right tool for the job.

对你来说，深度学习中最美丽、最令人惊讶或最鼓舞人心的想法是什么？一些在规模上吸引你想象力的东西，比如通过规模实现的性能，或者其他想法。

What is the most beautiful, surprising, or inspiring idea in deep learning to you? Something that captivated your imagination at scale, like the performance you can achieve with scale, or other ideas.

如果我唯一的工作是一名学术研究员，拥有无限的预算，不必担心短期影响，我会把所有时间花在研究无监督学习上。我仍然认为无监督学习是一个美丽的概念。在 NeurIPS 和 ICML 上，我参加了关于自监督学习的研讨会，这是无监督学习的一个垂直分支。让我总结一下这个想法。例如，在自监督学习中，我们从互联网上获取大量未标记的图像。有了无限量的这类数据，我取每张图像，将其随机旋转 90 度的倍数，然后训练一个监督神经网络来预测原始方向：是旋转了 90、180、270 还是 0 度。这样你可以生成无限量的标记数据，因为你知道旋转后的标签。各种研究人员发现，通过使用未标记数据，创建标记数据集，并在这类任务上训练大型神经网络，你可以获取隐藏层表示，并将其非常强大地迁移到不同的任务上。学习词嵌入是另一个例子：我们取一个句子，去掉一个词，然后预测缺失的词，这是我们学习嵌入的一种方式。现在有一整套技术来生成这些人为任务。另一个叫做拼图：你取一张图像，将其切成 3x3 的网格，像九块拼图，打乱九块，然后让神经网络预测它来自九的阶乘种可能排列中的哪一种。许多团队，包括 OpenAI、Yann LeCun 等，都在做这方面的工作。Facebook、Google Brain、DeepMind 和 Aaron van den Oord 在 CPC 目标上也有出色的工作。所以很多团队都在做令人兴奋的工作。我认为这是一种生成无限数据的方法，我觉得这是无监督学习中非常令人兴奋的一部分。

If my only job was being an academic researcher with an unlimited budget and didn't have to worry about short-term impact, I would spend all my time doing research on unsupervised learning. I still think unsupervised learning is a beautiful idea. At both NeurIPS and ICML, I was attending workshops on self-supervised learning, which is one vertical segment of unsupervised learning. Let me summarize the idea. For example, in self-supervised learning, we grab a lot of unlabeled images off the internet. With infinite amounts of this type of data, I take each image and rotate it by a random multiple of 90 degrees, then train a supervised neural network to predict the original orientation: whether it was rotated 90, 180, 270, or 0 degrees. So you can generate an infinite amount of labeled data because you know the label from the rotation. Various researchers have found that by taking unlabeled data, making up labeled datasets, and training a large neural network on these tasks, you can then take the hidden layer representation and transfer it to a different task very powerfully. Learning word embeddings is another example: we take a sentence, leave out a word, and predict the missing word, which is one way we learn embeddings. There's now a portfolio of techniques for generating these made-up tasks. Another one is called jigsaw: you take an image, cut it into a 3x3 grid like nine puzzle pieces, shuffle the nine pieces, and have a neural network predict which of the nine factorial possible permutations it came from. Many groups, including OpenAI, Yann LeCun, and others, have been doing work on this. Facebook, Google Brain, DeepMind, and Aaron van den Oord have great work on the CPC objective. So many teams are doing exciting work. I think this is a way to generate infinite amounts of data, and I find this a very exciting piece of unsupervised learning.

你认为这种无监督学习会解锁机器学习系统的很多能力吗？

Do you think that's going to unlock a lot of power in machine learning systems, this kind of unsupervised learning?

我不认为这是全部。它只是其中一部分。这一部分，自监督学习，正在获得 traction。我们非常接近它变得有用。词嵌入非常非常有用。我认为我们越来越接近在现实世界中产生重大影响，也许在计算机视觉和视频领域。但我觉得这个概念，以及围绕它的其他概念，比如我研究过的其他无监督学习的东西，都让我兴奋。我曾经对稀疏编码和慢特征分析非常兴奋。我认为这些都是我们大约十年前研究过的想法，后来我们都因为监督学习的效果太好而分心了。

I don't think it's the whole enchilada. I think it's just a piece of it. This one piece, self-supervised learning, is starting to get traction. We're very close to it being useful. Word embeddings are really, really useful. I think we're getting closer and closer to having a significant real-world impact, maybe in computer vision and video. But I think this concept, and then other concepts around it, like other unsupervised learning things I've worked on, I've been excited about. I was really excited about sparse coding and slow feature analysis. I think all of these are ideas that various of us were working on about a decade ago, before we all got distracted by how well supervised learning was working.

如果你要给学习深度学习的人推荐一个每日计划，你会推荐什么？

If you were to recommend a daily schedule for people learning deep learning, what would you recommend?

我认为养成学习的习惯是关键，这意味着规律性。例如，我们每周三发送时事通讯《The Batch》，所以人们知道它来了。周三你可以花点时间通过《The Batch》了解最新动态。我自己养成了每周六和周日花时间阅读或学习的习惯。周六醒来时，我不需要决定今天是否想阅读或学习；这就是我做的事。习惯让事情变得更容易。如果有人能养成这个习惯，就像每天早上刷牙一样——你不需要思考。它不消耗认知负荷。这种一致性让学习感觉更轻松。在我自己的生活中，我每天弹吉他，哪怕只有五分钟。这时间短得离谱，但因为养成了习惯，一两年内你能取得的进步令人难以置信。通过每天做很短的时间，你可以在某些方面变得非常出色。时间会累积。这通常不是靠爆发式的持续努力和熬夜，因为你只能有限次数地那样做。关键是长期持续的努力。每周读两篇论文很好，但力量在于每周读两篇论文，坚持一年——那么你就读了一百篇论文，你真的学到了很多。所以规律性和让学习成为习惯是关键。

I think getting the habit of learning is key, and that means regularity. For example, we send out our weekly newsletter, The Batch, every Wednesday, so people know it's coming. On Wednesday, you can spend a little time catching up on the latest news. For myself, I've picked up a habit of spending some time every Saturday and Sunday reading or studying. I don't wake up on Saturday and have to decide whether I feel like reading or studying; it's just what I do. The fact that it's a habit makes it easier. If someone can get into that habit, it's like brushing your teeth every morning—you don't think about it. It takes no cognitive load. That consistency makes it feel easier. In my own life, I play guitar every day, even if just for five minutes. It's a ridiculously short period, but because I've gotten into that habit, it's incredible what you can accomplish in a year or two. You can become exceptionally good at certain aspects by just doing it every day for a very short time. It adds up over time. It's often not about bursts of sustained effort and all-nighters, because you can only do that a limited number of times. It's the sustained effort over a long time. Reading two research papers is nice, but the power is in reading two research papers a week for a year—then you've read a hundred papers and you actually learn a lot. So regularity and making learning a habit is key.

你还有其他关于深度学习的学习建议吗？

Do you have other study tips for deep learning?

当我试图深入学习某件事时，我仍然会做的一件事是手写笔记。我知道很多人利用通勤时间学习深度学习课程，那时可能不方便做笔记。但当我上 Coursera 课程时——我仍然偶尔会这么做；最近一次是临床实验课程——我会拿出我的小 Moleskine 笔记本，坐在书桌前，记下讲师所说的内容。做笔记，最好是手写笔记，能提高记忆力。当你观看视频时，暂停一下，把基本要点写在纸上。有研究表明，手写笔记因为书写速度较慢，会促使你更多地用自己的话重新编码知识，而这种重新编码的过程能促进长期记忆。相比之下，打字更快，所以人们倾向于逐字转录，这减少了重新编码，导致长期记忆效果较差。

One thing I still do when trying to study something deeply is take handwritten notes in theory. I know many people take deep learning courses during commutes, where it might be hard to take notes. But when I take courses on Coursera—I still do that every now and then; the most recent was a course on clinical trials—I get my little Moleskine notebook and sit at my desk, taking down notes on what the instructor is saying. That act of taking notes, preferably handwritten, increases retention. As you watch the video, pause and take the basic insights down on paper. There have been studies showing that handwritten notes, because handwriting is slower, cause you to recode the knowledge in your own words more, and that process of recoding promotes long-term retention. In contrast, typing is faster, so people tend to transcribe verbatim, which reduces recoding and results in less long-term retention.

这很有趣。你花了很多时间研究教学法。你是如何应用的？

That's fascinating. You've spent so much time studying pedagogy. How do you apply that?

我在制作视频或撰写《The Batch》时做的一件事是思考：在这里花一分钟是否比在其他地方花一分钟更高效？我们真的努力让它时间高效。每个人都很忙。在编辑时，我经常告诉我的团队，每个词都需要为自己的存在而战。如果删除它不会损失意义，我们就删掉。这样就不会浪费学习者的时间。

One thing I do when creating videos or writing The Batch is think: is one minute spent here going to be a more efficient learning experience than one minute spent anywhere else? We really try to make it time-efficient. Everyone is busy. When editing, I often tell my teams that every word needs to fight for its life. If we can delete it without losing meaning, we do. That way, we don't waste the learner's time.

一个人如何将对深度学习的兴趣转化为职业？给那些想将其作为终身事业的人一些建议。

How does one make a career out of an interest in deep learning? Give advice for people who want to make it a lifelong journey.

或者至少是一二十年的旅程。你是怎么做到的？

Or at least a journey of a decade or two. How did you do it?

最重要的事情是正确起步。我认为在职业生涯早期，像深度学习专项课程这样的课程学习是掌握这些材料非常高效的方式。像我或其他讲师，或者教 TensorFlow 专项课程的 Laurence Moroney，都努力让学习新概念更省时。课程学习对于初学者进入新领域确实非常高效。事实上，我在斯坦福看到一些博士生想马上投入研究，我通常会告诉他们，刚进入博士项目时，花时间上课打好基础很重要。头几年效率低一点没关系，长期来看你会更好。但到了一定阶段，有些材料太前沿，课程里没有。还有一些实践经验我们还不擅长在课程中教授。在高效的课程学习之后，大多数人需要去做项目，或者通过阅读博客文章和研究论文继续学习。实践非常重要。我认为从小处着手很重要。今天就做点什么。如果你读到深度学习，看到别人在做激动人心的事情，你可能会想：“我又不构建改变世界的神经网络，有什么意义？”其实，构建一个很小的神经网络——比如 MNIST 或升级到 Fashion-MNIST——做自己的趣味项目，正是你获得技能的方式，让你能逐步做更大的项目。我发现这对个人和组织都适用。公司要擅长机器学习，有时正确的做法不是处理大项目，而是做一个小项目让组织学习，然后逐步发展。对个人和公司来说，迈出第一步然后小步前进是关键。

The most important thing is to get started right. I think in the early part of a career, coursework like the Deep Learning Specialization is a very efficient way to master this material. Instructors like me or someone else, or Laurence Moroney who teaches our TensorFlow Specialization, spend effort to make it time-efficient for you to learn new concepts. Coursework is actually a very efficient way for people to learn concepts in the beginning and break into new fields. In fact, one thing I see at Stanford: some of my PhD students want to jump into research right away, and I tend to say, look, when you first come to the PhD program, spend time taking courses because it lays the foundation. It's fine if you're less productive in your first couple of years; you'll be better off in the long term. Beyond a certain point, there's material that doesn't exist in courses because it's too cutting edge for the courses we created. There's some practical experience that we're not yet good at teaching in a course. After exhausting the efficient coursework, most people need to go on to either work on projects or continue learning by reading blog posts and research papers. Doing practice is really important. I think it's important to start small. Just do something today. If you read about deep learning and see all these people doing exciting things, you might think, "I'm not building a neural network that will change the world, so what's the point?" Well, the point is sometimes building that tiny neural network—be it for MNIST or upgrading to Fashion-MNIST—doing your own fun hobby project is how you gain the skills to let you do bigger and bigger projects. I find this to be true at the individual level and also at the organizational level. For a company to become good at machine learning, sometimes the right thing to do is not to tackle the giant project but instead to do a small project that lets the organization learn and then build up from there. This is true for individuals and for companies: just taking the first step and then taking small steps is the key.

学生应该攻读博士学位吗？你认为没有博士学位也能做很多事情吗？你怎么看？人们应该读研究生吗？应该读博士吗？

Should students pursue a PhD? Do you think you can do so much without ever getting a PhD? What are your thoughts? Should people go to grad school? Should people get a PhD?

我认为有多种好的选择，博士学位是其中之一。如果被顶尖博士项目录取——麻省理工、斯坦福等顶尖学校——我认为那是很好的经历。或者如果在一流组织的一流 AI 团队找到工作，那也是很好的经历。有些事情仍然需要博士学位才能做。如果某人的志向是成为顶尖学术大学的教授，那就需要博士学位。但如果目标是创办公司、做伟大的技术工作，我认为博士学位是很好的经历，但我会看看不同的选择——哪里能找到工作，哪里能进入博士项目——然后权衡利弊。

I think there are multiple good options, of which doing a PhD could be one. If someone is admitted to a top PhD program—MIT, Stanford, top schools—I think that's a very good experience. Or if someone gets a job at a top organization on a top AI team, I think that's also a good experience. There are some things you still need a PhD to do. If someone's aspiration is to be a professor at a top academic university, you just need a PhD to do that. But if it's to start a company, build a company, do great technical work, I think a PhD is a good experience, but I would look at the different options available—where you can get a job, where you can get into a PhD program—and weigh the pros and cons.

你认为人们最终应该有什么梦想和目标？他们应该探索哪些选择？你可以在工业界工作，比如谷歌、Facebook、百度这样的大公司，这些大公司都有庞大的机器学习工程师团队。你也可以做工业界的研究组，比如谷歌研究、谷歌大脑。你也可以走学术界的教授路线。还有什么？你可以创办自己的公司，做初创企业。这些选择之间有什么特别突出的吗，还是说它们都是人们应该考虑的不同美好旅程？

What final dreams and goals do you think people should have? What options should they explore? You can work in industry for a large company like Google, Facebook, Baidu—all these large companies have huge teams of machine learning engineers. You can also do industry research groups like Google Research, Google Brain. You can also do the professor track in academia. And what else? You can build your own company, do a startup. Is there anything that stands out between those options, or are they all beautiful different journeys that people should consider?

我认为影响你体验的更多不是你在公司还是学术界或工业界。我认为影响你体验最大的是你每天和谁互动。即使在一些大公司，不同团队的个人体验也大不相同。最重要的不是你每天走进大楼时门上的标志。最重要的是你每天互动的 10 个人或 30 个人是谁。所以我通常建议人们：如果你从一家公司得到工作机会，也要问你的经理是谁，你的同事是谁，你实际上会和谁交流。我们都是社会性动物；我们往往会变得更像周围的人。如果你和优秀的人一起工作，你会学得更快。如果你在一家好公司或好大学找到工作，也许标志很响亮，但你实际上被困在一个做你不感兴趣的工作的团队里，那其实是很糟糕的经历。这对大学和大公司都是如此。对于小公司，你很快就能弄清楚你会和谁一起工作。我通常建议人们：如果一家公司拒绝告诉你你会和谁一起工作，而是说“加入我们吧，轮岗制度会解决的”，我认为这是一个令人担忧的回答，因为这意味着你可能最终不会和优秀的同事和优秀的人一起工作。

I think the thing that affects your experience more is less whether you are in a company versus academia or industry. I think the thing that affects your experience most is who are the people you're interacting with on a daily basis. Even if you look at some of the large companies, the experience of individuals in different teams is very different. What matters most is not the logo above the door when you walk into the giant building every day. What matters most is who are the 10 people or the 30 people you interact with every day. So I actually tend to advise people: if you get a job offer from a company, also ask who is your manager, who are your peers, who are you actually going to talk to. We're all social creatures; we tend to become more like the people around us. If you're working with great people, you will learn faster. If you get a job at a great company or a great university, maybe the logo is great, but you're actually stuck on some team doing work that doesn't excite you, and that's actually a really bad experience. This is true both for universities and large companies. For small companies, you can kind of figure out who you'd be working with quite quickly. I tend to advise people: if a company refuses to tell you who you'll work with and says something like "Oh, join us, the rotation system will figure it out," I think that's a worrying answer because it means you may not end up on a team with great peers and great people to work with.

这真的是非常深刻的建议，我们有时会忽略。我们没有严谨或仔细地考虑，你身边的人真的很重要。尤其是当你取得伟大成就时，似乎伟大的成就都是因为你身边的人。所以不是关于你学这个还是那个，或者像你说的头顶的标志——而是人。这很迷人，而且这是一个非常艰难的寻找过程，就像找到合适的朋友或结婚对象一样。这是一个非常艰难的寻找过程，一个人际搜索问题。

That's actually really profound advice that we sometimes sweep under the rug. We don't consider rigorously or carefully that the people around you are really important. Especially when you accomplish great things, it seems the great things are accomplished because of the people around you. So it's not about whether you learn this thing or that thing, or like you said the logo that hangs up top—it's the people. That's fascinating, and it's such a hard search process, just like finding the right friends or somebody to get married with. It's a very hard search process, a people search problem.

是的，但我认为当有人面试大学或大公司的研究实验室时，最好坚持问清楚有哪些人，我的经理是谁。如果他们拒绝告诉我，我会想，也许是因为你没有好的答案。可能不是我喜欢的。如果你没有特别合得来，如果感觉和那些人不对劲，那就不要坚持。这是一个非常重要的信号需要考虑。

Yeah, but I think when someone interviews at a university or a research lab at a large corporation, it's good to insist on asking who are the people, who is my manager. And if they refuse to tell me, I'm going to think, well, maybe that's because you don't have a good answer. It may not be someone I like. And if you don't particularly connect, if something feels off about the people, then don't stick with it. That's a really important signal to consider.

是的，在我的标准课程 CS230 中，我做过一次 AC 演讲——我想我做过一个小时的职业建议演讲，包括求职过程，其中一些是你的观点。

Yeah, and I am in my standard course CS230, I gave an AC talk—I think I gave like an hour-long talk on career advice, including on the job search process, and some of these are yours.

如果你能在其他平台找到那些视频，我会推荐给大家。太好了。那么 AI Fund 帮助 AI 初创公司起步，或者你可以详细说说它发展出的有趣之处。对于如何建立成功的 AI 初创公司，你有什么建议？

If you can find those videos on also and others, I'll point people to them. Beautiful. So the AI Fund helps AI startups get off the ground, or perhaps you can elaborate all the fun things it's evolved with. What's your advice on how to build a successful AI startup?

在硅谷，很多初创公司的失败是因为构建了没人想要的产品。所以当你拥有很酷的技术时，谁会用它呢？我认为我倾向于非常注重结果和客户至上。最终，我们无法投票决定成功或失败，只有客户才能给出赞或踩。长期来看，这才是真正重要的。因此，当你建立初创公司时，问题是：客户会对此点赞吗？我认为那些非常关注客户、深入了解客户并致力于服务客户的初创公司更有可能成功，但前提是我们所有人都应该只做那些创造社会效益、让世界更美好的事情。我个人不想仅仅因为有利可图就构建让人上瘾的数字产品。但如果我们能找到以有意义的方式服务人们的方法，我认为无论是在学术环境、企业环境还是真正的初创环境中，这些都是值得做的好事。

You know, in Silicon Valley, a lot of startup failures come from building products that no one wanted. So when you have cool technology, but who's gonna use it? I think I tend to be very outcome-driven and customer-obsessed. Ultimately, we don't get to vote if we succeed or fail; only the customer gives a thumbs up or thumbs down. In the long term, that's what really matters. So as you build a startup, the question is: will the customer give a thumbs up on this? I think startups that are very customer-focused, deeply understand the customer, and are oriented to serve the customer are more likely to succeed, with the provision that I think all of us should only do things that create social good and make the world better. I personally don't want to build addictive digital products just because they could be lucrative. But if we can find ways to serve people in meaningful ways, I think those can be great things to do, whether in an academic setting, corporate setting, or a real startup setting.

那么你能告诉我你为什么创办 AI Fund 吗？

So can you give me an idea of why you started the AI Fund?

我记得当我离开百度的 AI 团队时，我的工作有两部分。一部分是构建一个 AI 引擎来支持现有业务，这部分运行得很好。第二部分是尝试利用公司的 AI 能力系统性地开拓新业务线。所以自动驾驶汽车团队来自我的团队，语音助手团队类似于美国的亚马逊 Alexa，但我们比亚马逊更早宣布。所以我们不是跟随亚马逊，那是我团队出来的。我发现那是我工作中最有趣的部分。所以我想做的是将 AI Fund 打造成一个创业工作室，利用我们现在能用 AI 做的所有事情，系统性地创建新的初创公司。我认为建立新团队去追逐这个丰富的机遇空间，是一个非常重要的机制来完成这些将推动世界前进的项目。不幸的是，只有少数团队能产生有意义的积极影响，我觉得我们或许可以更系统、更可重复地做到这一点。所以创业工作室是一个相对较新的概念；现在大概有几十个创业工作室。但我感觉很多团队仍在试图弄清楚如何系统性地建立高成功率的公司。我认为，尽管我的风投朋友们似乎越来越多地在构建公司而不是投资公司，但我发现这是一件迷人的事情：找出机制，让我们能够系统性地在我们认为有意义的领域建立成功的团队和成功的企业。所以创业工作室是一个让初创公司从零到成功的场所和机制。我们试图制定一个蓝图；它实际上是我们从零开始构建初创公司的地方。我们经常引入创始人并与他们合作，或者甚至可能有现成的想法，我们将其与创始人匹配，然后希望这能启动并成为成功的公司。

I remember when I was leaving the AI group at Baidu, I had two parts to my job. One was to build an AI engine to support the existing businesses, and that was running well by itself. The second part was to try to systematically initiate new lines of businesses using the company's AI capabilities. So the self-driving car team came out of my group, the speaker team similar to Amazon's Alexa in the US, but we announced it before Amazon did. So we weren't following Amazon; that came out of my group. I found that to be the most fun part of my job. So what I wanted to do was to build AI Fund as a startup studio to systematically create new startup firms with all the things we can now do with AI. I think the ability to build new teams to go after this rich space of opportunities is a very important mechanism to get these projects done that will move the world forward. It's unfortunate that only a few teams have a meaningful positive impact, and I felt that we might be able to do this in a more systematic, repeatable way. So a startup studio is a relatively new concept; there are maybe dozens of startup studios right now. But I feel like many teams are still trying to figure out how to systematically build companies with a high success rate. I think even though my venture capital friends seem to be more and more building companies rather than investing in companies, I find it a fascinating thing to do: to figure out the mechanisms by which we could systematically build successful teams and successful businesses in areas that we find meaningful. So a startup studio is a place and a mechanism for startups to go from zero to success. We try to develop a blueprint; it's actually a place for us to build startups from scratch. We often bring in founders and work with them, or maybe even have existing ideas that we match founders with, and then this launches hopefully into successful companies.

那么你离找到一种方法来自动化从零开始并建立成功的 AI 初创公司的过程有多接近？

So how close are you to figuring out a way to automate the process of starting from scratch and building a successful AI startup?

是的，我认为我们一直在不断改进和迭代我们的流程。比如我们需要打多少客户电话才能获得客户验证，如何确保这项技术能够很好地构建。我们所有的业务都需要尖端的机器学习算法，所以即使它在研究论文中有效，但将其投入生产真的很难。让这些东西在现实生活中发挥作用有很多问题，你在学术界是看不到的。那么你如何验证它实际上是可行的？如何建立团队并获得教育、医疗或我们关注的任何领域的专业知识？我认为我们实际上在给企业家提供高成功率方面做得越来越好了，但我认为我们仍处于摸索的早期阶段。

Yeah, I think we've been constantly improving and iterating on our processes. Things like how many customer calls do we need to make to get customer validation, how do we make sure this technology can be built well. All of our businesses need cutting-edge machine learning algorithms, so even if it works in a research paper, it turns out taking it to production is really hard. There are a lot of issues for making these things work in real life that you don't see in academia. So how do you validate it's actually doable? How do you build a team and get specialized domain knowledge in education or healthcare or whatever sectors we're focusing on? I think we're actually getting much better at giving entrepreneurs a high success rate, but I think we're still in the early phases of figuring it out.

你认为那个过程的某些方面是否可以从一个初创公司转移到另一个？

Do you think there are some aspects of that process that are transferable from one startup to another?

非常如此。对大多数企业家来说，创办公司是一件非常孤独的事情。我见过很多企业家不知道如何做出某个决定，比如什么时候需要做 PDP 销售？如果你不知道，那真的很难。或者如何高效营销，而不是购买广告，那太贵了？有更高效的策略。从机器学习项目来看，基本决策可以改变机器学习产品是否成功的走向。企业家需要做出成百上千个决定，而在几个关键决定上犯错可能会对公司的命运产生巨大影响。所以我认为创业工作室提供了一个支持结构，让创办公司不再那么孤独。此外，在面临关键决策时，比如试图招聘你的第一位工程副总裁，好的选择标准是什么？我该不该雇佣这个人？通过围绕企业家和创始人建立一个生态系统来提供帮助，我认为我们在关键时刻帮助他们，并希望循环往复地让这个过程更愉快、成功率更高。在这些非常困难的决策点上，有人可以一起头脑风暴，并且帮助他们认识到他们甚至可能没有意识到这是一个关键决策点。这可能是最重要的部分。

Very much so. Starting a company, to most entrepreneurs, is a really lonely thing. I've seen so many entrepreneurs not know how to make a certain decision, like when do you need to do PDP sales? If you don't know that, it's really hard. Or how do you market efficiently other than buying ads, which is really expensive? There are more efficient tactics. From machine learning projects, basic decisions can change the course of whether a machine learning product works or not. There are so many hundreds of decisions that entrepreneurs need to make, and making a mistake in a couple of key decisions can have a huge impact on the fate of the company. So I think a startup studio provides a support structure that makes starting a company much less of a lonely experience. Also, when facing key decisions like trying to hire your first VP of Engineering, what's a good selection criteria? Should I hire this person or not? By having an ecosystem around the entrepreneurs and founders to help, I think we help them at the key moments and hopefully cyclically make it more enjoyable and with a higher success rate. There's somebody to brainstorm with in these very difficult decision points, and also to help them recognize what they may not even realize is a key decision point. That's probably the most important part.

在这些非常困难的决策点上，有人可以一起头脑风暴，并且帮助他们认识到他们甚至可能没有意识到这是一个关键决策点。对，这可能是最重要的部分。

There's somebody to brainstorm with in these very difficult decision points, and also to help them recognize what they may not even realize is a key decision point. Right, that's the first probably the most important part.

是的，我还可以说另一件事。我认为创办公司是一回事，但我感觉真正重要的是我们建立能够推动世界前进的公司。例如，Lavinia 的团队曾经有一个新公司的想法，如果成功的话，会导致人们在某个狭窄的垂直视频类型上观看更多视频。我们看了看；商业案例没问题，收入案例也没问题，但我说我不想做这个。我实际上不想让更多人看这种类型的视频；它没有教育意义。

Yeah, you can say one other thing. I think building companies is one thing, but I feel it's really important that we build companies that move the world forward. For example, Lavinia's team once had an idea for a new company that, if it had succeeded, would have resulted in people watching a lot more videos in a certain narrow vertical type of video. We looked at it; the business case was fine, the revenue case was fine, but I said I don't want to do this. I don't actually just want to have a lot more people watch this type of video; it wasn't educational.

所以这个想法的基础是，没想到它真的能帮助人们。无论是创办公司、为企业工作，还是做个人项目，我认为我们每个人都要想清楚自己想用 AI 在世界上创造什么不同。通过学习 AI，你帮助了已经成立的公司发展他们的 AI 和机器学习工作。一家大公司如何将机器学习整合到他们的工作中？

So the idea on the basis that didn't think it would actually help people, so whether building companies, working for enterprises, or doing personal projects, I think it's up to each of us to figure out what's the difference we want to make in the world. With learning AI, you've helped already established companies grow their AI and machine learning efforts. How does a large company integrate machine learning into their efforts?

AI 是一种通用技术，我认为它将改变每一个行业。我们的社区已经改变了逻辑中心软件互联网行业。除了最顶尖的五六家或三四家之外，大多数软件互联网公司已经具备合理的机器学习能力，或者正在接近。还有改进的空间。但当我看到软件互联网行业之外，从制造业、农业到医疗保健，它们才刚刚起步。有太多机会，但很少有人在做。所以我认为 AI 的下一波浪潮是改变所有这些其他行业。麦肯锡的一项研究估计，这将带来 13 万亿美元的全球经济增长。美国 GDP 是 19 万亿美元，所以 13 万亿美元是一个很大的数字。或者普华永道估计是 16 万亿美元。不管具体数字是多少，它都很大。但对我来说有趣的是，很多影响将来自软件互联网行业之外。所以我们需要更多的团队与这些公司合作，帮助他们采用 AI。我认为这是让我希望推动全球经济增长、让人类更强大的事情之一。正如你所说，影响是存在的。

AI is a general-purpose technology, and I think it will transform every industry. Our community has already transformed the logic center software internet sector. Most software internet companies outside the top right five or six or three or four already have reasonable machine learning capabilities or are getting there. There's still room for improvement. But when I look outside the software internet sector, everything from manufacturing, agriculture, healthcare, they're just getting started. There are so many opportunities that very few people are working on. So I think the next wave for AI is to transform all of those other industries. There was a McKinsey study estimating 13 trillion dollars of global economic growth. The US GDP is 19 trillion dollars, so 13 trillion is a big number. Or PwC estimated 16 trillion dollars. So whatever number, it's large. But the interesting thing to me was that a lot of that impact would be outside the software internet sector. So we need more teams to work with these companies to help them adopt AI. And I think this is one thing that makes me hope to drive global economic growth and make humanity more powerful. And like you said, the impact is there.

正如你所说，影响是存在的。那么，在软件科技行业之外，哪些是最好的行业、最大的行业，AI 可能在那里发挥作用？

And like you said, the impact is there. So what are the best industries, the biggest industries where AI can perhaps outside the software tech sector?

坦率地说，我认为是所有行业。我花了很多时间的一些行业是制造业、农业，还有医疗保健。例如，在制造业中，我们做了很多视觉检测的工作，目前有人工用肉眼检查塑料零件或智能手机是否有划痕或凹痕。我们可以用相机拍照，使用算法、深度学习等来检查是否有缺陷，从而提高工厂质量和产量。事实证明，我们遇到的实际问题与大多数研究论文中读到的问题非常不同。数据集非常小。工厂里的问题不断变化。环境在变化，所以它在你的测试集上表现良好，但猜怎么着，工厂里有些东西变了，灯开了或关了。最近，有一个工厂，一只鸟飞过工厂，在什么东西上拉了屎，所以情况变了。因此，提高我们算法对工厂中所有变化的鲁棒性，我发现我们遇到了很多在学术界没有被广泛讨论的实际问题。在很多人甚至意识到有问题存在之前，解决这些问题真的很有趣，处于前沿。

Frankly, I think it's all of them. Some of the ones I'm spending a lot of time on are manufacturing, agriculture, looking to healthcare. For example, in manufacturing, we do a lot of our work in visual inspection, where today there are people standing around using their eyes to check if a plastic part or a smartphone or something has a scratch or dent in it. We can use a camera to take a picture, use an algorithm, deep learning, and other things to check if it's defective or not, and thus improve factory quality and improve throughput. It turns out the practical problems we run into are very different than the ones you might read about in most research papers. The data sets are really small. The problems in the factories keep on changing. The environment changes, so it works well on your test set, but guess what, something changes in the factory, the lights go on or off. Recently, there was a factory in which a bird flew through the factory and pooped on something, so that changed stuff. So increasing our algorithm's robustness to all the changes that happen in the factory, I find that we run into a lot of practical problems that are not as widely discussed in academia. And it's really fun being on the cutting edge, solving these problems before maybe many people are even aware that there is a problem there.

那真是一个迷人的领域。你说得完全正确。但公司应该采取的第一步是什么？从人眼检测到数字化这个过程，拥有摄像头和算法，这是一个可怕的飞跃。第一步是什么？你推荐这些公司采取的早期旅程是什么？

And that's such a fascinating space. You're absolutely right. But what is the first step that a company should take? It's a scary leap into this new world of going from the human eye inspecting to digitizing that process, having a camera, having an algorithm. What's the first step? What's the early journey that you recommend that you see these companies taking?

我发布了一份名为《AI 转型手册》的文档，在网上可以找到。我在 Coursera 的《人人 AI》课程中也简要谈到了公司应该采取的长期旅程。但第一步实际上是从小处着手。我看到很多公司因为起步太大而失败，而不是因为起步太小。就拿谷歌来说，大多数人没有意识到早期有多困难、多有争议。在谷歌大脑成立之前，它是有争议的。人们认为深度学习，他们试过，但没成功，为什么要做深度学习？所以我在谷歌的第一个内部客户是谷歌语音团队，这不是谷歌最赚钱的项目，也不是最重要的，不是网页搜索或广告。但通过从小处着手，我的团队帮助语音团队构建了更准确的语音识别系统，这让他们的同行，其他团队，开始对深度学习更有信心。我的第二个内部客户是谷歌地图团队，我们使用计算机视觉从街景图像中读取门牌号，以更准确地定位谷歌地图中的房屋，从而提高质量。只有在这些成功之后，我才开始与谷歌广告团队进行最认真的对话。所以有一个涟漪效应：你展示了它在这些案例中有效，然后它就会在整个公司传播，表明这个东西对我们有很多价值。我认为早期的小规模项目有助于团队建立信心，也有助于团队了解这些技术的作用。我仍然记得我们的第一个 GPU 服务器是放在某个人桌子下的服务器，这教会了我们早期的重要教训：如何让多个用户共享一组 GPU，这在当时是非常不明显的。但这些早期教训很重要。我们从第一个 GPU 服务器中学到了很多，后来帮助团队思考如何在不进行太多大规模部署的情况下进行扩展。

I published a document called the AI Transformation Playbook that's online, and I talk briefly in my AI for Everyone course on Coursera about the long-term journey that companies should take. But the first step is actually to start small. I've seen a lot more companies fail by starting too big than by starting too small. Take even Google. Most people don't realize how hard it was and how controversial it was in the early days. When it was not the Google Brain, it was controversial. People thought deep learning, they tried it, it didn't work, why would you want to do deep learning? So my first internal customer role in Google was the Google Speech team, which is not the most lucrative project in Google, not the most important, it's not web search or advertising. But by starting small, my team helped the Speech team build a more accurate speech recognition system, and this caused their peers, other teams, to start having more faith in deep learning. My second internal customer was the Google Maps team, where we used computer vision to read house numbers from Street View images to more accurately locate houses within Google Maps, thus improving the quality later. And only after those two successes did I then start the most serious conversation with the Google Ads team. So there's a ripple effect: you show that it works in these cases, and then it just propagates through the entire company that this thing has a lot of value and use for us. I think the early small-scale projects help the teams gain faith, but also help the team learn what these technologies do. I still remember when our first GPU server was a server under some guy's desk, and that taught us early important lessons about how to have multiple users share a set of GPUs, which was really non-obvious at the time. But those early lessons were important. We learned a lot from that first GPU server, which later helped the teams think through how to scale without too many large deployments.

公司面临哪些具体的挑战，对它们来说解决这些挑战很重要？

Are there concrete challenges that companies face that are important for them to solve?

我认为构建和部署机器学习系统很难。在你的笔记本电脑上的 Jupyter 笔记本中能运行的东西，与在工厂或发电厂等生产部署环境中运行的东西之间，存在巨大的鸿沟。所以我看到很多人让东西在笔记本电脑上运行，说‘哇，看我做了什么’，这很好，这是非常重要的第一步。但所有团队都低估了剩下的步骤。例如，我听过很多机器学习人员和业务人员之间的对话。机器学习人员说：‘看，我的算法在测试集上表现很好，我创建的干净测试集。’业务人员说：‘非常感谢，但你的算法很烂，它不起作用。’机器学习人员说：‘不，等等，我在测试集上做得很好。’我认为在硬盘上的测试集上表现良好与在部署环境中表现良好之间存在鸿沟。一些常见问题：鲁棒性和泛化能力。例如，在工厂里，也许他们砍掉了工厂外面的一棵树，所以树不再遮挡窗户，光照不同了，所以测试集变了。在机器学习中，尤其是在学术界，我们通常假设测试集代表部署分布，但在实践中，分布会发生变化。

I think building and deploying machine learning systems is hard. There's a huge gulf between something that works in a Jupyter notebook on your laptop versus something that runs in a production deployment setting in a factory or a power plant or whatever. So I see a lot of people get something to work on their laptop, say 'Wow, look what I've done,' and that's great, that's a very important first step. But all teams underestimate the rest of the steps. For example, I've heard this exact same conversation between a lot of machine learning people and business people. The machine learning person says, 'Look, my algorithm does well on the test set, the clean test set I created.' And the business person says, 'Thank you very much, but your algorithm sucks, it doesn't work.' And the machine learning person says, 'No wait, I did well on the test set.' And I think there is a gulf between what it takes to do well on a test set on your hard drive versus what it takes to work well in a deployment setting. Some common problems: robustness and generalization. For example, in a factory, maybe they chopped down a tree outside the factory, so the tree no longer covers the window and the lighting is different, so the test set changes. In machine learning, especially in academia, we often assume the test set is representative of the deployment distribution, but in practice, the distribution shifts.

我们不知道如何处理与训练集分布截然不同的测试集分布。虽然有领域自适应、迁移学习等研究，但我们在这方面确实不擅长。那么如何真正让系统工作呢？因为测试集分布会变化。另外，如果你看一个软件系统的代码行数，机器学习模型可能只占整个系统的 5%甚至更少。那么如何完成所有工作并使其可靠和系统化？良好的软件工程是构建成功的机器学习系统的基础。

We don't know how to deal with test set distributions that are dramatically different than the training set distribution. There's research like domain adaptation, transfer learning, but we're really not good at this. So how do you actually get this to work because your test set distribution is going to change? Also, if you look at the number of lines of code in a software system, the machine learning model is maybe five percent or even fewer relative to the entire software system we need to build. So how to get all that work done and make it reliable and systematic? Good software engineering work is fundamental to building a successful machine learning system.

是的，软件系统需要与人们的工作流程对接。机器学习是自动化的增强版。如果我们把工厂里众多任务中的一项——比如视觉检查——自动化，那会非常有价值，但你可能需要围绕这项任务重新设计许多其他任务。例如，机器学习算法说这个有缺陷，那么接下来是扔掉它，还是让人复核、返工或修复？所以你需要重新设计许多围绕这项自动化的任务。因此，规划变更管理，确保你编写的软件与新工作流程一致，并花时间向人们解释发生了什么，我认为 Landing AI 已经变得擅长这一点——我们通过犯错和痛苦的经历学到了——那就是与合作伙伴一起思考机器学习模型之外的所有事情：不只是做一个 Jupyter 笔记本，而是构建整个系统，管理变更过程，并找出如何以产生实际影响的方式部署它。大型软件科技公司使用的部署流程在许多其他场景下并不适用。例如，当我领导大型语音团队时，如果语音识别系统宕机，警报会响起，然后我会说‘嘿，你们 20 个工程师，请修好这个系统’。但如果你在工厂里有一个系统在运行，那里没有 20 个机器学习工程师等着，你只能呼叫值班人员来修复。那么如何处理维护、DevOps、MLOps 或其他方面的问题？这些概念我认为 Landing AI 和其他一些前沿团队正在探索，但我们甚至还没有系统的术语来描述我们所做的一些事情，因为我觉得我们是在即兴发明。

你提到有些人感兴趣的是发现宇宙中的数学之美和真理，而你感兴趣的是对世界产生巨大的积极影响。那么我问你：这两者并不矛盾，对吧？它们是一体的。我只是半开玩笑，因为你可能两者都有一点兴趣。但让我问一个浪漫化的问题：你的工作和我们今天的讨论大多集中在应用 AI 上，甚至可以说是狭义 AI，目标是创建自动化特定流程并为世界创造巨大价值的系统。但还有另一个 AI 分支，从艾伦·图灵开始，梦想创造人类级或超人类级的智能。这也是你的梦想吗？你认为我们人类会建造出人类级或超人类级的智能系统吗？

So you mentioned some people are interested in discovering mathematical beauty and truth in the universe, and you're interested in having big positive impact in the world. So let me ask: the two are not inconsistent, no? They're all together. I'm only half joking, because you're probably interested a little bit in both. But let me ask a romanticized question: so much of your work and our discussion today has been on applied AI, maybe you can even call narrow AI, where the goal is to create systems that automate some specific process that adds a lot of value to the world. But there's another branch of AI, starting with Alan Turing, the kind of dreams of creating human-level or superhuman-level intelligence. Is this something you dream of as well? Do you think we human beings will ever build a human-level or superhuman-level intelligent system?

我很想实现 AGI，而且我认为人类会实现，但需要一百年、五百年还是五千年，我觉得很难估计。

I would love to get AGI, and I think humanity will, but whether it takes a hundred years or 500 or 5,000, I find hard to estimate.

有些人担心这条道路可能带来的不同轨迹，甚至 AGI 系统的存在性威胁。你有这样的担忧吗，无论是短期还是长期？

Do you have some folks have worries about the different trajectories that path would take, even existential threats of an AGI system? Do you have such concerns, whether in the short term or the long term?

我确实担心人类的长期命运。我也很好奇。我确实担心火星上的人口过剩，但不是今天。我认为将来有一天，也许火星会被污染，有孩子死去，有些人会回看这个视频说‘Andrew，你怎么这么无情？你不在乎火星上那些垂死的孩子。’我向未来的观众道歉：我确实在乎那些孩子，但我不知道今天如何有效地解决这个问题。

I do worry about the long-term fate of humanity. I do wonder as well. I do worry about overpopulation on the planet Mars, just not today. I think there will be a day when maybe someday in the future Mars will be polluted, there are children dying, and some will look back at this video and say 'Andrew, how can you be so heartless? You didn't care about all these children dying on the planet Mars.' And I apologize to the future viewer: I do care about the children, but I just don't know how to productively work on that today.

你的照片会出现在字典里，给那些对火星人口过剩无知的人看。好吧，是的，所以这是一个长期问题。短期来看，我们是否应该考虑如何让 AI 系统的价值观与人类的价值观对齐？Russell 和其他人正在思考：随着系统不断发展，我们希望确保它代表我们本性中更好的一面，代表我们社会的伦理和价值观。

Your picture will be in the dictionary for the people who are ignorant about the overpopulation on Mars. Okay, yes, so it's a long-term problem. Is there something in the short term we should be thinking about in terms of aligning the values of our AI systems with the values of us humans? Something that Russell and other folks are thinking about: as this system develops more and more, we want to make sure that it represents the better angels of our nature, the ethics, the values of our society.

以自动驾驶汽车为例，自动驾驶汽车最大的问题不是电车难题。你开车时有多少次面临这种道德困境：‘我该撞你还是不撞？’我认为自动驾驶汽车遇到这个问题的频率和我们开车时差不多。自动驾驶汽车最大的问题是当一辆白色大卡车横在路中间时，你应该刹车而不是撞上去，但自动驾驶汽车失败了，撞了上去。所以我认为我们需要先解决那个问题。我认为关于 AGI 对齐、回形针问题的一些讨论，极大地分散了我们对今天真正需要解决的更困难问题的注意力。一些困难问题：我认为偏见是一个大问题。我担心财富不平等。AI 和互联网正在加速权力集中，因为我们现在可以集中数据并加以处理。所以一个又一个行业，我们影响了每一个行业。互联网行业有很多赢家通吃的动态，但它已经感染了所有其他行业，所以它们也有赢家通吃的味道。看看 Uber 和 Lyft 对出租车行业做了什么。这创造了巨大的财富，但我们如何确保财富公平分配？我认为，还有如何帮助那些工作被取代的人？我认为教育是其中的一部分，但可能需要做的比教育更多。我认为偏见是一个严重的问题。AI 的恶意使用，比如深度伪造被用于各种邪恶目的。所以我担心一些团队可能无意中——我希望不是故意——对遥远未来的问题制造大量噪音，而不是专注于这些更困难的问题。

If you take self-driving cars, the biggest problem with self-driving cars is not that there's some trolley dilemma. How many times when you're driving your car did you face this moral dilemma: 'Should I crash into you or not?' I think self-driving cars run into that problem roughly as often as we do when we drive our cars. The biggest problem for self-driving cars is when there's a big white truck across the road, and what you should do is brake and not crash into it, but the self-driving car fails and crashes into it. So I think we need to solve that problem first. I think the problem with some of these discussions about AGI alignment, the paperclip problem, is that it is a huge distraction from the much harder problems that we actually need to address today. Some hard problems: I think bias is a huge issue. I worry about wealth inequality. AI and the internet are causing an acceleration of concentration of power because we can now centralize data and use it to process. So industry after industry, we've affected every industry. The internet industry has a lot of winner-take-most dynamics, but it's infected all these other industries, so they also have winner-take-most flavors. Look at what Uber and Lyft did to the taxi industry. So this is creating tremendous wealth, but how do we ensure that wealth is fairly shared? I think that, and then how do we help people whose jobs are displaced? I think education is part of it, but there may be even more that we need to do than education. I think bias is a serious issue. There are adverse uses of AI, like deepfakes being used for various nefarious purposes. So I worry about some teams maybe accidentally, and I hope not deliberately, making a lot of noise about problems in the distant future rather than focusing on these much harder problems.

我们今天已经面临的问题被掩盖了，它们异常棘手，就像你提到的那些，甚至包括一些看似琐碎但影响巨大的问题，比如工厂窗外光照的变化，这最终决定了你所说的 Jupyter Notebook 和真正能变革整个行业的东西之间的区别。而且，当监管机构找上门说你的产品搞砸了，修复它可能会影响收入，跟他们讨论你如何承诺不毁灭人类反而更有趣。我们面临的问题确实很棘手。你的人生从教学到研究再到创业，是一段伟大的旅程。有两个问题：第一，有没有遗憾的时刻，如果能重来你会做得不同？第二，有没有你特别自豪、真正让你快乐的时刻？

The overshadow the problems that we have already today they're exceptionally challenging like those you said and even the silly ones but the ones that have a huge impact which is the lighting variation outside of your factory window that that ultimately is what makes the difference between like you said the jupyter notebook and something that actually transforms an entire industry potentially yeah and I think and then just to some companies when a regulator comes to you and says no your product is messing things up fixing it may have a revenue impact was much more fun to talk to them about how you promise not to wipe out humanity in this interface they're actually really hard problems we face so your life has been a great journey from teaching to research to entrepreneurship two questions one are there regrets moments that if you went back you would do differently and two are there moments you're especially proud of moments that made you truly happy

你知道我犯过很多错误，感觉每次我发现新东西时，都会想为什么我五年前甚至十年前没想到呢？有时我读一本书，会想如果我十年前读了这本书，我的人生会多么不同。虽然最近也发生过这样的事，我就在想，如果我们在创业时读了这本书，Coursera 可能会好得多。但我发现那本书在我们创办 Coursera 时还没写出来。但我觉得，发现的过程就是不断找出事后看来显而易见的东西，但总是比我希望的要花更长时间。至于第二个问题，有没有你特别自豪或特别快乐的时刻？有两个答案。第一，尽管有各种变动，当然，无论我花多少时间，我都觉得陪她不够。祝贺你。谢谢。第二是帮助他人。对我来说，生命的意义就是帮助他人实现他们的梦想，同时努力通过让人类整体更强大来推动世界前进。所以，我最快乐、最自豪的时刻，是当我感到别人让我有幸在他们的追梦路上帮了一点小忙。我想没有比谈论幸福和生命意义更好的结尾了。能和你对话是我的巨大荣幸，我和数百万人感谢你所做的一切。非常感谢你接受采访。

You know I've made so many mistakes it feels like every time I discover something I go why didn't I think of this you know five years earlier or even ten years earlier and Reese's and then sometimes I read a book and I go I wish I read this book ten years ago my life we've been so different although that happened recently and then I was thinking if only I read this book when we're a start-up Coursera could have been so much better but I discovered that book had not yet been written we're starting Coursera so that means even but I find that the process of discovery we keep on finding out things that seem so obvious in hindsight but it always takes us so much longer than then I wish to figure it out so on the second question are there moments in your life that if you look back that you're especially proud of or especially happy the that fills you with happiness and fulfillment well two answers one despite all turnover yes of course you say no matter how much time I spend for I just can't spend enough time with her congratulations weather thank you and then second is helping other people I think to me I think the meaning of life is um helping others achieve whatever are their dreams and then also to try to move the world forward by making humanity more powerful as a whole so the times that I felt most happy most proud works when I felt um someone else allowed me the good fortune of helping them a little bit on the path to their dreams I think there's no better way to end it than talking about happiness and the meaning of life so enter it's a huge honor me and millions of people thank you for all the work you've done thank you for talking to thank you so much

感谢收听本期与吴恩达的对话。感谢我们的赞助商 Cash App，下载并使用代码“LexPodcast”，你将获得 10 美元，同时 10 美元将捐给“First”组织，该组织激励和教育年轻人成为未来的科技创新者。如果你喜欢这个播客，请在 YouTube 上订阅，在 Apple Podcast 上打五星，在 Patreon 上支持，或者直接在 Twitter 上联系我@LexFriedman。现在，让我用吴恩达的一句智慧箴言作为结尾：问问自己，如果你正在做的事情取得了超乎想象的巨大成功，你是否显著地帮助了他人？如果没有，那就继续寻找其他值得投入的事情，否则你并没有发挥出全部潜力。感谢收听，下次再见。

Thanks for listening to this conversation with Andrew Aang and thank you to our presenting sponsor cash app downloaded use coal export cast you'll get ten dollars and $10 will go to first an organization that inspires and educates young minds to become science and technology innovators of tomorrow if you enjoy this podcast subscribe on youtube give it five stars and Apple podcast supported on patreon or simply connect with me on Twitter at Lex Friedman and now let me leave you with some words of wisdom from NGO Aang ask yourself if what you're working on succeeds beyond your wildest dreams which you have significantly helped other people if not then keep searching for something else to work on otherwise you're not living up to your full potential thank you for listening and hope to see you next time