Scale AI CEO 王亚历山大：22 岁融资 1 亿美元

大家好，欢迎收听《本周初创企业》。我是主持人 Jason Calacanis，这档播客中，我们与创始人畅谈他们改变世界的愿景。今天节目里来了一位有趣的家伙，他叫 Alexander Wang，是 Scale AI 的 CEO 兼联合创始人。他拿下了 scale.com 这个域名，这大概值一百万美元。我们稍后会问他花了多少钱。我想最引人注目的是，你在最近几轮融资中筹集了超过一亿美元，这可不是小数目。这笔钱来自 Founders Fund。而且你才 22 岁。22 岁。年轻有为的烦恼就是每次采访都从你的年龄开始，有点烦人。我也经历过，但那时我是 23 岁的《Cyber Server》出版人，或者 25 岁的《Silicon Alley》出版人。我会问年龄为什么重要？等你到了 35、40 岁，他们就不再提了，因为会觉得你 40 岁就该做出点有意思的事或成功。但你从多大开始经营这家公司？19 岁。你是 Thiel Fellow 吗？你是怎么入行的？

Hey everybody, welcome to This Week in Startups. I'm your host Jason Calacanis, and this is the podcast where we talk to founders about their vision for how they want to change the world. Today I've got an interesting cat on the program. His name is Alexander Wang, he is the CEO and co-founder of Scale AI. He got the domain name scale.com, that's like a million dollar domain name. We'll find out what he paid for it later. And I guess the thing that most people would think is remarkable is candidly that you've raised over a hundred million dollars in your last rounds of funding. That's a lot of money. It's quite a bit. It's quite a bit of money from Founders Fund. And I think you're 22 years old. 22. So that's annoying to be young and successful because then every interview starts with your age a little bit. It's annoying. I had it happen to me but with like 23 year old publisher of Cyber Server or 25 year old publisher of Silicon Alley. And I say why does my age matter? Now I tell you when you hit about 35, 40 they don't mention it anymore because they're like wow you're 40 you should be doing interesting things or be successful in the world. But you've been running this company since you were how old? 19. Now were you a Thiel Fellow or something? How did you get into the game?

不，我有一段有趣的小历史。我在新墨西哥州洛斯阿拉莫斯长大。我的父母是物理学家，在洛斯阿拉莫斯国家实验室工作。那个实验室最初是制造原子弹的地方。曼哈顿计划就在洛斯阿拉莫斯启动，当时非常保密。高中时我做了很多编程，参加各种编程比赛，高中时就被招募了。所以高中毕业后我直接出来工作，在 Quora 这家公司干了几年。你是怎么得到那份工作的？你只是申请了，他们看了你的代码就说行？他们招募了我，因为我在编程比赛里是匿名参赛者。你可以直接参加编程比赛，没人知道你的年龄。然后你自学编程？其实是互联网教会了我编程，所以算是互联网教的。我只是上网查，你在 YouTube 上找课程？对，我不太记得了，大概就是谷歌搜索。总之，我在 Quora 做了几年工程和基础设施之类的工作。然后没上大学？不，之后我上了大学。我去了麻省理工，但一年后觉得无聊，就创办了 Scale。你退学了？对。所以你父母当时对你离开麻省理工很伤心？目前是。即使你筹集了一亿美元，他们现在仍然伤心？天哪，这些父母标准真高。是啊，这虽然是个梗，但确实如此。好吧，听好了爸妈，他会在麻省理工建一栋以你们名字命名的大楼，所以放过他吧。会没事的。他以后会成为名誉教授。嗯，但愿吧。什么？我不会是名誉教授，我是个退学生。你会是荣誉教授。对，那可能是安抚我父母的方式。因为你工作了几年，然后去了麻省理工。这样做不对，因为你坐在那里，一切都很慢，一切都是理论。对，确实如此。你从高速驾驶赛车，然后他们把你放进维修区和卡丁车，说给你些卡丁车。老实说，我觉得学校的慢节奏让我烦躁，最终促使我创办了公司。我想在另一个世界里，我可能会在 Quora 之后继续在公司工作很多年。

No, I have a fun little history. I grew up in Los Alamos, New Mexico. My parents are physicists and they worked at the National Lab in Los Alamos. It was the lab where the atomic bomb was originally built. The Manhattan Project started in Los Alamos, was very secretive at that time. And then in high school I did a bunch of programming, I did all these coding competitions and I was getting recruited inbounds in high school. So after high school I actually came out here to work. I worked at this company Quora for a couple of years. How'd you get that job? You just applied and they saw your code and they were like okay. They recruited inbounds because I was an anonymous person on these coding competitions. And then you could just go into a coding competition, nobody knows your age. And then you teach yourself how to code? Well the internet taught me how to code, so I guess the internet. I just looked it up, you found courses online on YouTube or? Yeah, I don't remember, I think I just googled around. Anyway, so I worked at Quora for a couple of years doing engineering, infrastructure, etc. And then no college? No, well then I went to college after that. I went to MIT and then got basically bored after a year and started Scale. You left? I left. So your parents were heartbroken about you leaving MIT at that time? For now. For now they're still heartbroken even if you raise 100 million? My god, these parents have high standards. Yeah, I mean it's a meme but it's true. Alright, listen mom and dad, he's going to build a building at MIT with your name on it, the family name on it, so give him a break. It'll be okay. He'll be a professor emeritus at some point. Yeah, well one can hope. Something? I wouldn't be professor emeritus, I would be somebody who left. You'd be like an honorary professor. Yeah, that would probably be the way I'd appease my parents. Because you went to work for a couple years, then you went to MIT. That is not the way to do it because you'd be sitting there and everything's going to be going so slow and everything's theoretical. Yes, that's exactly what happened. You went from like running fast, you're driving race cars, and then they like put you in the pit and the go-karts and they're like here's some go-karts. And if I'm being honest, I think the slower speed of school is sort of what got me agitated enough to eventually start the company. I think there's an alternate world where I continued working at companies after Quora for many many years.

那么 Scale 的愿景是什么？你是怎么想到这个主意的？什么时候想到的？

So what was the vision for Scale? How did you get the idea? When did you have the idea?

我们的使命是加速 AI 应用的发展。我们从根本上认为 AI 和机器学习是技术领域一代人一次的变革。甚至可能是物种级别的。是啊，取决于发展情况。我们拭目以待。这显然被大肆炒作，但我们坚信这一点。你觉得它和互联网一样大还是更大？互联网本身连接了数十亿人。对，比互联网更大。比硅芯片的发明还大？我认为它更接近计算的出现，而不是互联网的出现，因为它使许多以前必须由人类完成的事情现在可以由机器完成。明白了。那你来排个序：AI、计算机、互联网？或者可能是计算机、AI，我们看看 AI 和计算机之间会发生什么。我们会在未来几十年密切关注。计算机确实极大地改变了我们的日常生活。对，AI 也会，通过自动驾驶汽车和手机上的各种助手。而且我认为它会不断发展。还有很多应用。好的，这是背景。那么你对阻碍 AI 发展的因素有什么见解吗？

Our mission is to accelerate the development of AI applications. I think we fundamentally view AI and machine learning as kind of a once in a generation shift in technology. Might be once in a species by the way. Yeah, depending on how this goes. We'll see. It's obviously very hyped, but I think we hold that belief quite strongly. And do you think it's as big as the internet or bigger? The internet itself took billions of people and connected them for the first time. Yeah, bigger than the internet. And bigger than the silicon chip being created? I think it's more comparable to the advent of computing than it is to the advent of the internet, because it's an enabler of all these things that previously had to be done by humans and now can be done by machines. Got it. Okay, for you to rank them: AI, computer, internet? Or maybe computer, AI, and we'll see what happens between AI and computer. We'll watch intently over the coming decades. Computers did change our day-to-day lives pretty significantly. Yeah, well AI will as well, with autonomous vehicles and all these assistants on your phone. And I think it'll go on and on and on. There's a lot more applications. All right, so that's the backdrop. And then you have some insight on what was holding back AI or something?

我回学校的主要原因其实是学习机器学习。我在 Quora 时，那是一家非常依赖机器学习的公司，但我没有很强的学术背景。所以我回到麻省理工深入学习。然后我有很多想构建的产品想法，但有一个房间里的大象问题，那就是你该如何……

The big reason I went back to school actually was to study machine learning. So I was at Quora, it was a very machine learning driven company, but I didn't have that strong an academic backing. And so I went back to MIT to really study this more deeply. And then I had all these ideas of products I wanted to build, but there was sort of an elephant in the room problem, which was how were you supposed to...

给我举个例子，然后为不太熟悉机器学习这个术语的观众定义一下。人们常把机器学习和 AI 放在一起说，你怎么解释它们各自是什么？

Give me an example of that and then define for the audience that's not super familiar with the term machine learning. What is the difference between saying the word machine learning and AI? People hear them together. How would you explain what each one is?

机器学习是 AI 的一个子集，是一种特定类型的 AI。我们训练这些程序，让它们能够完成人类通常做的各种事情，也就是传统上需要人类判断的任务。它们通过大量数据来学习。所以我们选一个人类用大脑完成的任务——涉及逻辑、直觉，谁知道人类大脑到底怎么做决策，这方面争议很大。你把大量数据喂给机器学习算法，然后它给出一个它认为接近人类答案或最佳答案的结果。但它知道最佳答案的唯一途径就是通过人类创造的所有这些数据。

Machine learning is a subset of AI, a particular kind of AI. We're training these programs that are able to do various things that humans normally do, tasks that traditionally require human judgment. They do that by feeding them lots of data. So we pick a task that humans have done with their brains—some combination of logic, intuition, who knows exactly how human brains make decisions, there's a lot of debate about that. You feed a bunch of data to a machine learning algorithm, and then it gives you an answer that it thinks would approximate a human's answer or the best answer. But the only way it's going to know what the best answer is is through all this data that humans have created.

咱们举个最杰出的例子。你给 VC 讲什么例子时他们直接投钱？

Let's come up with the most illustrious example. What's an example that when you gave it to VCs they threw money at you?

真正吸引全世界注意力的例子是自动驾驶汽车。这个例子很有说服力，因为首先没人喜欢开车，而且开车非常不安全。开车风险很大，所以事关重大。这个引人入胜的机器学习模型能够接收车辆的所有摄像头数据和其他传感器数据，理解周围发生的一切——这对人眼来说很容易，但在机器学习出现之前对机器来说非常困难——然后确定最佳路径，并学会自己驾驶。

The one that has really captivated the world's attention is autonomous vehicles. It's a compelling example because first, nobody likes driving, but also driving is very unsafe. There's a lot of risk in driving, so a lot at stake. The captivating machine learning model is one that can take in all the camera data and other sensor data from the vehicle, understand everything going on around it—something very easy for your eye but currently, or at least before machine learning, was very difficult for machines—and then determine the best path to take and figure out how to drive on its own.

我们看到高速公路上的车道标记、双黄线、双白线。我们知道要让车尽量平稳地保持在两条线之间。我们看到有人偏离车道进入我们的车道，我们知道要减速，给他们留点空间——也许他们喝醉了。机器天生不知道这些；我们必须教它。

We see the lane markers, double yellow markers, double white markers on the highway. We know to keep the car between those two lines as smoothly as possible. We see somebody deviate from their lane into ours, we know to slow down, give them some room—maybe they're drunk. The machine doesn't know that inherently; we have to teach it that.

没错。

Exactly.

Scale.com 做了哪些特斯拉和 Waymo 没做的事？他们用你们的软件吗？他们需要吗？

What does Scale.com do that Tesla and Waymo don't already do? Do they use your software? Do they need to?

这个问题问得好。核心问题是，机器不知道要做什么，除非有数据告诉它们该做什么。所以机器学习的一大瓶颈就是数据——告诉这些算法和模型该做什么的数据。这就是 Scale 的用武之地。我们就像一个数据精炼厂。我们接收客户的大量原始数据，进行处理，然后告诉机器它应该做什么。例如，对于自动驾驶汽车拍摄的图像，我们会标出人在哪里、车在哪里、车道标记在哪里等等，这样随着时间的推移，这些算法就能学会这些东西。

This is a great question. The core problem is that machines don't know what to do unless they have data that tells them what they're supposed to be doing. So one of the huge bottlenecks for machine learning is data—data that tells these algorithms, these models, what they're supposed to be doing. That's where Scale comes in. We are a data refinery. We accept a bunch of raw data from our customers, process it, and tell the machine what it should be doing. For example, given an image taken by a self-driving car, we would outline where the people are, where the cars are, where the lane markings are, etc., so that over time these algorithms can learn those things.

你这里有视频可以展示。我看到你在标出汽车、标出行人，机器正在识别出那是道奇皮卡的大致形状，那是丰田普锐斯，这些看起来像人的轮廓。但现在是人类在告诉机器那是什么。

You have a video of that you can show right here. I see you are highlighting cars, you're highlighting people, and the machine is figuring out okay that's the approximate shape of a Dodge pickup truck, that's a Toyota Prius, and these look like the silhouettes of people. But that's human telling the machine what it is for now.

没错。我们流程的核心方式是，很多工作最初由机器和我们自己的 AI 模型在后台完成，然后人类基本上提供输入并纠正错误，以确保最终数据极其准确，因为这最终对这些系统的安全性和低偏差等至关重要。所有这些对机器学习来说都是必不可少的。

Exactly. The core way our pipeline works is that a lot of work is done behind the scenes by machines and our own AI models originally, and then humans basically give input and correct mistakes to make sure the end data is extremely accurate, because that's ultimately what's important for the safety of these systems and for low bias, etc. All these things are imperative for machine learning.

你会去找像 Waymo 或 Uber 这样的客户。他们都是客户，对吧？他们会给你他们汽车行驶的视频，然后你为他们标注，以某种方式把数据放入数据库。

You would go to a customer like Waymo or Uber. They are both customers, right? They would give you videos of their cars driving, and then you would annotate it for them, put that data into a database somehow.

完全正确。例如，如果他们给我们这样的视频，最初的第一步是人工画框。然后一个已经通过所有这些数据预处理过的机器学习模型会确定那辆车随时间变化的路径。然后我们确认所有这些都正确，再把数据发送给客户，他们在此基础上训练机器学习模型。

That's exactly right. For example, if they gave us a video like this, originally the first step was a human drawing a box. Then a machine learning model that's already pre-processed through all this data has determined the path of that vehicle over time. Then we confirm that all this is correct and send that data over to the customer, and they train machine learning models on top of that.

我猜这就是其中一辆车被愚弄的方式。有人在地上画了一个转弯箭头，然后一辆车就跟着箭头走了。我们人类也会这么做。但他们基本上画了一个转向信号，想看看能不能骗过自动驾驶汽车，结果当然骗过了。

This is how I guess one of the cars got fooled. Somebody drew on the ground like an arrow turning, and a car followed the arrow. Which our human would do too. But they basically drew a turning signal to see if it would fool a self-driving car, and of course it did.

我没看到这个新闻，但我相信会发生这种事。顺便说一句，人类也会这样。我觉得这是有史以来最愚蠢的恶作剧。他们就像在说，‘看，我们能骗过开车的机器，让它转错弯。’这就像你也能骗人类转错弯一样。就像把出口匝道的‘禁止驶入’标志拿下来放到入口匝道上。恭喜你，《搏击俱乐部》，你刚搞了个愚蠢的恶作剧。

I didn't see this news, but I would believe that is what would happen. And that's what would happen to a human by the way. I thought that was the stupidest prank ever. They're like, 'Look, we can fool a machine that's driving cars to make a wrong turn.' It's like you would also fool a human to make the same wrong turn. It's like taking that 'Do Not Enter' sign off the off-ramp and putting it on the on-ramp. Congratulations, Fight Club, you just did some stupid prank.

完全正确。我的意思是，在很多方面，它们会遇到和人类驾驶相同的一些挑战。

That's exactly right. I mean, in a lot of ways, they'll have some of the same challenges that humans have driving.

等我们回来，我想知道你是为一家公司存储所有这些数据并标注，还是有一个宏伟计划让数据跨多家公司共享，这样大家就不必重复造轮子了？

When we get back, I want to understand if you are storing all this data and annotating it for one company, or is this some sort of grand plan to have it go across multiple companies so everybody doesn't have to recreate the wheel?

你其实更快乐，所以我们和 Calm 合作，它是排名第一的睡眠应用。你会得到 Calm 的一系列节目，旨在帮助你获得大脑和身体所需的睡眠，比如音景和超过 100 个睡眠故事。我和孩子们一起听这些，他们很喜欢，我自己也听，效果很棒。这是我的同事 Presh，他因为老板太严厉而睡不好，他试了试，找了一些非虚构作品，正在看《尼罗河上的鲁阿·克鲁兹》，所以马修·麦康纳，看马修·麦康纳读睡眠故事，他看了 ASMR 画作《美女与野兽》，然后他看了看睡眠，他听了《星星摇篮曲》——啊，太放松了。所以这是你的行动号召：本周末，Startup 听众将获得 Calm 高级订阅 25% 的折扣。没错，calm.com/twist。C-A-L-M / T-W-I-S-T。四千万人下载了 Calm，它是苹果 2017 年度应用。在 calm.com/twist 上了解原因。再次感谢 Calm。我是该公司的投资者，我喜欢这家公司，我为团队在那里所做的工作感到自豪。它真是一个了不起的应用和一个很棒的故事。好了，让我们回到这期精彩的节目。

You're actually happier, so that's why we're partnering with Calm, the number one app for sleep. You're gonna get a library of programs from Calm that are designed to help you get the sleep your brain and body needs, like soundscapes and over 100 sleep stories. I do these with my kids and they love it, and I do it myself and it is amazing. And here's my associate Presh who has been having trouble sleeping because his boss is too intense, and he goes through it and he finds some nonfiction and he's looking at 'Rua Cruz on the Nile', so Matthew McConaughey, watch Matthew McConaughey reading some sleep stories, he looks at the ASMR painting Beauty and the Beast, and then he looks at sleep, he does 'Lullaby to the Stars' — ah, so relaxing. So here's your call to action: this weekend, Startup listeners will get 25% off a Calm premium subscription. That's right, calm.com/twist. C-A-L-M / T-W-I-S-T. Forty million people have downloaded Calm and it was Apple's 2017 App of the Year. Find out why at calm.com/twist. Thanks again to Calm. I'm an investor in the company and I love the company and I'm so proud of the work the team over there is doing. It's just such an amazing app and such a great story. Okay, let's get back to this amazing episode.

周五，Alexandr Wang 来了。 blah blah blah，二十多岁，谁在乎呢。他向你展示了一些更厉害的东西，但一切都会变老。你最终会变老，人们不会再叫你 22 岁什么的。他正在打造 scale.com，那是一个六位数或七位数的域名。不评论，不评论那个域名的价格。它不便宜。它在字典里，而且少于六个字母，所以不便宜。它不便宜。但一个好域名确实有助于品牌建设，不是吗？嗯，我想我们以后会看到的，那会在未来几年发生。我们拭目以待。每个人都说，‘哦，你的邮箱是什么？比如 Alexandr@scale.com，只要少写一个 e，准备好少写一个 e。’你怎么说？等等，你的名字是 Alexandr，但你在 d 和 r 之间少了一个 e。是的，这实际上是对的。这是你出生证明上的拼写错误吗？甚至是有意为之，非常有意。所以是你父母或程序开的玩笑？我父母希望我的名字有八个字母，因为他们是中国人，那代表好运。八，八，八，八是非常吉利的。是的，非常吉利。是的，他们真的为了好运这么做了，而且奏效了。是的，奏效了。我的意思是，我有一个朋友，我和他打扑克，他非常迷信。在中文里，是的。打扑克时，如果他打得不好，如果他输了，他会买入八万八千美元。那就是不看他的牌，在德州扑克中盲推筹码。然后我们两三个人如果有 A 之类的就会跟注，然后他每次都赢。哦，哇。八万八千美元，我见过他连续这样做五次。他是洛杉矶的传奇。是的，我只能说这么多。他听起来真的很厉害。是的，我喜欢打扑克。这太神奇了，因为我们认为扑克中最差的牌通常是 8-2 不同花之类的，所以你甚至有机会。但是的，这真是难得一见。

Friday, Alexandr Wang is here. Blah blah blah, twenty-something year old who cares. He's shown you some more dory, it's all get old eventually. You'll get old eventually and people will stop saying that you're 22 and whatever. And he is building scale.com, that's like a six or seven figure domain name. No comment, no comment on the price of that domain. It's not cheap. It's in the dictionary and it's a less than six letter, so it's not cheap. It was not, it was not inexpensive. But a good domain name does help the branding, does it not? Well, I guess we'll see though, that will happen in the out years. We'll see what happens. Everybody, it's like 'Oh, what's your email? Like Alexandr at scale.com, just leave out an e, pick the ready to leave out.' What say you? Wait, your name is Alexandr but you're missing the e between the d and the r. Yeah, this is actually right. It's my typo on your birth certificate? Even named intentional, very intentional. So like a joke by your program or parents in some way? My parents wanted eight letters in my first name because they're Chinese, and that's good luck. Eight, eight, eight, eight is very good luck. Yes, extremely good luck. Yeah, and they literally did that for good luck and it worked. Yeah, it worked. I mean, I have a friend I played poker with, he's extremely superstitious. In Chinese, yeah. When you play poker, if he's not playing well, if he's losing, he buys in for eighty-eight thousand dollars. That's let's just not look at his cards and pushes the chips in blind in poker, Texas Hold'em. And then two or three of us will just call him if we have an ace or whatever, and then he wins every time. Oh wow. For eighty-eight thousand dollars, I've seen him do it five times in a row. He's a legend in Los Angeles. Yeah, that's all I can say. He sounds really good. Yeah, I'd love to play poker. It's pretty amazing, because we think about the worst hand in poker is like typically 8-2 offsuit or something like that, so you even have a chance. But yeah, it's quite a thing to see.

所以在我们休息之前，我想知道：如果你在做所有这些数据，这些公司是不是都在各自的孤岛里一遍又一遍地编程他们的算法和数据集？这些公司之间没有共享。是的，完全正确。这太疯狂了。嗯，因为他们都想拥有它，他们最终都想赢，所以他们都不共享数据。嗯，我认为这归结为：在机器学习背景下，什么是知识产权？我认为它可以是知识产权。没错。如果硅谷的每家公司都公开开发他们的代码，那同样疯狂。我认为这就像开源。没错。有一点开源，但是的，底层技术人们确实开源，但谷歌并没有完全公开他们的算法。是的，或者开源那个。所以最终，这并不那么疯狂。我认为这意味着有动力去做新颖、有趣并产生价值的事情。好的，所以你得到了十个人尝试十个不同数据集的变异性，但你失去了十个人基于一个公共数据集工作的效率。是的，我的意思是，你对此立场的核心与你是否相信一般的自由市场经济非常相似。是的，很多人跑来跑去，朝着同一个方向，朝着不同的方向，如果你相信这种方法，而不是一种高效率但可能低变异性和低混乱的计划经济，那么我认为这完全没问题。是的，但因为你有十个人用十个不同的数据集竞争，最后有巨大的奖励，比如谁解决了自动驾驶就能赢得一千亿或一万亿美元，是的，你激励了大量有影响力的资本池去追求它。是的，没错。

So before we left for the break, I wanted to know: if you're doing all this data, are all these companies programming their algorithms and data sets in a silo over and over and over again? There's no sharing across these companies. Yeah, that's exactly right. That's crazy. Well, because they all want to own it, they all want to win in the end, so they're all not sharing their data. Well, I think this is what it comes down to: what is IP in the machine learning context? And I think it could be intellectual property. Exactly. It would be equivalently crazy if every company in the valley were to just develop their code out in the open. I think it's like open source. Exactly. There's a little bit of open source, but yeah, the underlying technology people do open source, but Google's not giving their algorithm away exactly. Yeah, or open sourcing that. And so ultimately, it's not that crazy. I think it means that there's an incentive to do things that are novel and interesting and produce value. Okay, so you get the variability of ten different people trying ten different data sets, but you lose the efficiency of ten different people working off a common data set. Yeah, I mean, the core of your stance on this is a very similar thing to whether or not you believe in just free-market economics in general. Yeah, a lot of people running around running in the same direction, running in different directions, and it's like, and if you believe in that approach versus a sort of planned economy where there's high efficiencies but maybe low variance and low chaos, then I think it's really fine. Yeah, but because you have ten different people competing with ten different data sets and there are big prizes at the end, like whoever solves self-driving wins a hundred billion dollars or a trillion dollars, yeah, you've incentivized large groups of influential pools of capital to pursue it. Yeah, exactly.

而且有一家开源公司，不是吗？有人在做一个开源公司。你知道这家公司吗？他们打算开源数据，做我说的那些事。我忘了它的名字。这不是一个新想法，事实上在研究社区，人们经常开源数据。真的吗？是的。所以很多人说深度学习生命周期中机器学习的开始是那个叫做 ImageNet 的大型数据集，它是由斯坦福教授李飞飞发布的，她基本上制作了这个大型数据集，然后它真正引发了机器学习和深度学习的热潮。哇。所以因为她把那些开源的 Creative Commons 图片都放在那里，然后让每个人都训练它们，你有了一个训练集。是的，没错。所以她发布了这个，她和她的实验室发布了这个极其庞大的数据集，包含数百万张图片，并分类了图片中的内容。所以有橙子、苹果、猫，还有一些稀有的，比如这是一种稀有鱼类等等。这是一只猫在吃橙子？没那么详细。所以这是开始，请注意。是的。基本上这创建了一个开源数据集，然后整个世界都可以在此基础上工作。所以这将是集中化和开源的一个例子，也许介于社会主义共产主义单一政府方法和民主资本主义方法之间。开源有点介于两者之间，不是吗？是的。最终，实际上它可能是社会主义的。是的，我同意。

And there is an open source company, isn't there? There's somebody doing an open source company. Do you know about this company? There's, they're gonna open source the data and do exactly what I'm talking about. I forgot the name of it. It is not a new idea, and in fact in the research community people open source data quite commonly. Really? Yeah. So a lot of people say the start of all of the machine learning part of the deep learning lifecycle was this large dataset called ImageNet, which was published by this Stanford professor Fei-Fei Li, who basically produced this large data set and then it really kicked off the sort of machine learning deep learning hype. Wow. So because he got all of those open-source Creative Commons images up there and then had everybody trained them, you had a train set. Yes, exactly. So she published this large, she and her lab published this extremely large data set of millions of images classified with what was in those images. So there's oranges, apples, there's a cat, there's some rare ones like this is a rare kind of fish, etc. This is a cat eating an orange? Not quite that detailed. So this was the beginning, mind you. Yeah. And basically that created this open source data set that then sort of the whole world could work on top of. So that would be the example of centralization and open source being perhaps something between, let's say, a socialist communist singular government approach versus a democratic capitalist approach. There was open source which kind of sits somewhere between the two, doesn't it? Yeah. Ultimately, the actually maybe it's socialist. Yeah, I do.

我觉得这引出了一个更糟的问题。我不打算确切地说如何在这些经济情况下安排这些，但我非常认同，总的来说，为一大群人或者一个大型社区提供一些核心底层基础设施，让所有人都在这个基础上迭代或构建，这种趋势非常有价值。在自动驾驶中，最有效的方式是道路视频还是其他记录方式？我们有激光雷达。谷歌把赌注押在激光雷达上，埃隆·马斯克则把特斯拉的赌注押在摄像头上。大家都觉得埃隆是白痴。结果，我现在听说人们开始认为摄像头变得如此出色，数据集也如此优秀，以至于摄像头将胜出，激光雷达将变得完全不必要。那么哪个是对的？

I think that opens up a worse question. I'm not going to call me exactly how to line these in these economic situations, but I think very much so, like in general, the trend of providing some core underlying infrastructure for a large group of people or a large community of people who are all iterating or building on top of that infrastructure is very valuable. In self-driving, is it the video of what's on the road or is it some other way of recording it that is the most effective? So we have lidar. Google bet the farm on lidar, Elon bet the farm with Tesla on video cameras. Everybody thought Elon was an idiot. Turns out, I'm hearing now that people are starting to think the cameras are getting so good and the datasets are getting so good that cameras will win the day and lidar will be ridiculously unnecessary. So which is true?

所以我的个人观点是，我确实认为两种传感器各有优势，从根本上说，它们在不同场景下都非常出色。我们实际上发表了一篇关于这个的博客文章，因为我们显然看到了大量的激光雷达数据和大量的图像数据。

So my personal opinion is that I do think both sensors have different advantages, and fundamentally they're both very good in different scenarios. So we actually published a blog post about this because we obviously see a lot of lidar data and a lot of image data.

那篇文章叫什么？激光雷达对摄像头？埃隆对拉里？

What's the name of the post? Lidar versus cameras? Elon versus Larry?

我想它叫‘激光雷达对摄像头’之类的。我会搜一下。但没错，我认为两者都有各自擅长的场景。激光雷达首先非常擅长提供周围一切的 3D 地图。如果你在规划非常精细的操作，这非常有价值，而且它在提供 3D 地图方面非常可靠。

I think it's called 'Lidar versus Cameras' or something like that. I'll search for it. But yeah, I think there are different scenarios where both are good. Lidar is very good, first of all, at giving you a 3D map of everything around you. Turns out to be very valuable if you're planning very careful maneuvers, and it's very reliable in giving you that 3D map.

它生成的地图非常精细。所以埃隆关于激光雷达是错的？正是。

It makes a map that is incredibly well refined. So is Elon wrong about lidar? Exactly.

它在黑暗场景下也非常好，因为激光雷达会自己发光，所以你能确切知道周围发生了什么。但在其他场景下它就不行了。比如大雪或大雾时它表现很差。

It's also very good in dark scenarios because the lidar creates its own light, so you know exactly what's going on around you. But it's bad in other scenarios. It's bad when there's a lot of snow or a lot of fog.

为什么激光雷达在雪和雾中表现差？

Why is lidar bad in snow and fog?

因为它会发射出这些微小的激光，而雪和雾都非常反光，基本上会干扰激光雷达的工作方式。所以在这种情况下它会产生不完美的模型。

Because it shoots out these little lasers, and snow and fog are both very reflective and basically screw with how the lidar works. So it makes an imperfect model in that situation.

一个不完美的 3D 模型，或者至少是你会在意的那种。比如有一团烟雾。激光雷达会捕捉到烟雾，但实际上开车穿过烟雾是没问题的，对吧？

An imperfect 3D model, or at least one that you'd care about. So for example, there's a plume of smoke. Lidar will catch the plume of smoke, but it's actually fine if you drive through a plume of smoke, right?

是的，没错。那么现在机器学习能否有效协调两个系统同时开启的情况？不，你看，激光雷达构建了一个完美的模型，但激光雷达探测到的可能是烟雾，而摄像头说不是，那是烟雾。我们能判断，因为摄像头更擅长检测烟雾和雪。

Yeah, okay. And so can machine learning now reconcile when both systems are on effectively? No, hey, the lidar has built this perfect model, but the lidar is hitting something that could be smoke, and the cameras say no, it's smoke. We can tell because cameras are better at detecting smoke and snow.

这就是为什么从根本上你需要两者。两者绝对是最好的系统。摄像头的问题在于，目前最先进的基于摄像头的计算机视觉模型可能 99%的时间是准确的，这听起来很多，但不是 99.9%。99%看起来很多，直到你开十万英里。不是开一百秒的时候。而恰好就是那一秒，一块巨石滚到了街上。没错。所以达到这种渐近的困难质量水平非常重要。如果你有多个传感器，各有优劣，在需要时能相互配合，你就能做到。这就是为什么你需要两者。

That's why fundamentally you want both. Both is definitely the best system. The place where the camera stuff breaks down is that right now, if you look at the state-of-the-art computer vision models that work on cameras, they're accurate maybe 99% of the time, which sounds like a lot, but not 99.9%. 99% seems like a lot until you drive a hundred thousand miles. It's not where you drive a hundred seconds. It's not. And that happens to be the second where a boulder rolls into the street. Exactly. So it's pretty important that you get these asymptotically difficult levels of quality. You can do that if you have multiple sensors that have different strengths and weaknesses and can sort of play off one another when you need it. That's why you want both.

目前软件是在做这个吗，还是人们只是选一个系统然后就用它？

Is the software doing that today currently, or are people just picking one system and going with it?

不，它们非常协同工作。在 Waymo 车辆或 Cruise 车辆上，它们非常协同工作。所以在几乎所有场景中，有些时候你会更关注摄像头的信息，其他时候则更关注激光雷达。但摄像头现在是默认的，对吧？

No, they very much work together. On a Waymo vehicle or on a Cruise vehicle, they very much work together. So in pretty much all scenarios, you'll pay more attention to what the camera tells you in some and others you'll pay more attention to the lidar. But the cameras are the default now, right?

那不是真的。不，我认为很多这些车仍然很大程度上受激光雷达影响，真的吗？

That's not true. No, I think a lot of these cars still drive very much influenced by the lidar, really?

它是一个非常好的传感器。如果我们手机上有激光雷达，生活会很棒。它能给你周围世界非常精确的 3D 地图，所以你可以对周围环境做更多事情。

It's a really good sensor. If we had lidar on our phones, life would be great. It gives you a very accurate 3D map of the world around you, so you can basically do a lot more with your surroundings.

我以为谷歌开始把那种深度感应放进去了。他们不是用激光雷达，而是用其他传感器。现在手机前面有一个结构光传感器。大多数手机都有这个用于 Face ID 之类的。

I thought Google was starting to put that kind of depth sensing in there. They're not doing it with lidar, they're doing it with some other sensor. There's a structured light sensor on the front of your phone now. Most of these phones have that for Face ID or whatnot.

结构光传感器。它可以测量深度。它知道我鼻子的深度、眼睛的深度等等，所以知道是我而不是你。没错。这就是为什么平面照片不行，因为很难伪造。

A structured light sensor. And that can do depth. It knows the depth of my nose, my eyes, all that kind of stuff, so it knows it's me, not you. Exactly. And that's why a flat photo doesn't work, because it would be pretty hard to fake.

不过我听说亚洲面孔，最初的版本对长相相似的亚洲人不起作用。当白人家伙在谷歌或苹果创建算法时，实际上不行。亚洲人可以解锁彼此的手机。我看到过那篇文章。这是真的吗？

Although I heard Asian faces, the original versions didn't work for those people who looked similar who were Asian. When the white guys created the algorithm at Google or Apple, it didn't actually work. Asian people could unlock each other's phones. I saw the article. Is that true or not?

这又回到了问题的核心，即机器学习非常困难，因为一切都取决于数据。所以谁知道训练算法的底层数据里有什么？可能是一些白人的相册吧。

This gets back to the core of the issue, which is that machine learning is really hard because it's all about the data. So who knows what was going on in the underlying data that trained the algorithms? It was some white guy's camera roll, probably.

不管怎样，这就是为什么拥有非常好的数据和算法非常重要，否则它们会做出奇怪的不合理的事情。如果算法是基于中国的数据库构建的，比如中国的 Flickr，99.99%的照片是华裔，那么它就会针对那个数据集优化。如果你在美国做，白人占一定比例，亚洲人可能只有 2%、3%、4%，就不会那么精细。完全正确。

Either way, this is why it's really important to have really good data and algorithms, because otherwise they'll do weird things that don't make sense. If the algorithms were built off of a database in China, let's say the Flickr of China, and 99.99% of the photos were of Chinese descent, it would be optimizing for that dataset. And if you did it in America, and whatever percent was white, and the percentage of Asian might be two, three, four percent, it's not going to be as refined. That's exactly right.

这就是为什么，当我们的许多客户和今天许多做机器学习的公司思考这个问题时，关键是如何通过越来越多的数据不断改进，填补空白，使整个系统随着时间的推移在整体上更加稳健。

This is why, when a lot of our customers and a lot of companies doing machine learning today think about it, it's really about how we constantly improve with more and more data that fills in the gaps and makes the whole system holistically more robust over time.

你们构建的是哪一部分？数据存储，算法？我还不清楚你们构建的是哪一部分。

And you guys build which piece of this? The data storage, the algorithms? I'm still unclear as to which piece you build.

所以我们做的就是所有这些。所有进来的数据，比如摄像头图像，更容易理解。谈到 PB 级的数据。这些 PB 级的图像进来，乍一看你完全不知道这些图像里发生了什么。所以你需要弄清楚人在哪里，车在哪里，哪里是……

So what we do is all this. All this data that comes in, let's say it's camera images, is simpler to think about. Talk about petabytes of data. These petabytes of images come in, and prima facie you have no idea what's going on in these images. So what you need is to figure out where are the people, where are the cars, where are the...

停车标志、猫等等，你需要为每一张图片找出这些，然后才能在上面训练机器学习模型。明白了。所以我们构建了一个流程，大部分工作由机器完成。我们这边有经典的计算机视觉算法，就像你先对数据进行初步筛选。比如，Waymo 可以说：‘嘿，这里有 1000 万英里的驾驶数据，拿去处理吧。’然后你说：‘好的，我们认为这些是面包车，这些是皮卡，这些是猫，这些是弹跳球等等。’然后我们还有一大群聪明、训练有素的人类，他们可以检查并发现机器犯的错误。这是第二层筛选，由人类查看计算机置信度较低的内容。如果计算机有 99%的把握，你就直接采用计算机的结果。我们有更复杂的过滤器，但大致如此。然后这些高精度的数据返回给客户，他们感觉很好，重新训练他们的机器学习模型。这是一个美妙的循环。他们不必担心组建团队来做这种基础工作。这几乎就像你非常擅长为他们清理和标准化数据集，这样他们就可以专注于更高级的事情，比如如何处理面包车，或者如何处理侧翻的面包车。没错。我们对此的总体看法是，我们正在为全球机器学习或 AI 提供这种基础设施层。在 AI 领域，有一个大问题是获取所有数据，另一个大问题是如何改进、如何构建模型以及如何优化模型。我们试图把第一个问题从人们的任务清单中移除。

The stop signs where the cats etc and you to figure out that out for every single one of these images so you can train a machine learning model on top of that got it so what we do is we build this pipeline where most of the work is done by machines on our end we have a classical computer vision algorithms so it's like you do a first scrub of the data exactly so somebody like Waymo could say hey here are here's 10 million miles of driving have at it you say ok here's what we think these are all the minivans these are all the pickup trucks these are all the cats these are the bouncing balls etc yep and then we also do we also have a a large team of smart well trained humans who can basically go through and spot errors that these that are made and then the that's a second level of scrubbing which is humans looking at things that computers have a low degree of certainty F yes so if the computers 99% certain you just go with the computer well we have we have more sophisticated filters than that but more or less yeah yeah and then and then basically this this highly accurate data goes back to the customer and they feel great about it they retrain their machine learning models it uh it's a wonderful cycle so they don't have to worry about building a team or to do this basic level it's almost like you're just really good at getting that data set scrubbed and clean for them and normalizing it in some way so that they can work on the higher level stuff like what to do with the minivan or what to do with a minivan turned on that side oh it rolled over when something's going wrong here right that's exactly right and and the way that we think about this in general is we're really providing this sort of this infrastructure layer for machine learning globally or AI globally we're in AI in general there's like there's one big problem which is getting all of the data and there's another big problem which is like how do you improve how do you build these models and I do you improve the models yeah we're trying to take the first problem off in people's plates

休息回来后，我想知道，根据你的位置——你离数据很近，实际上你坐在数据之上，浸泡在数据中——我想知道，亚历山大·王（名字里没有多余字母的那个），你认为什么时候我们不再需要方向盘？从帕洛阿尔托开车到旧金山，什么时候没有方向盘会是合法的？我们回来再聊。LinkedIn 上有 6 亿人，包括我、你、你旁边的人，以及你刚发邮件的三个人。你需要招聘，但那些潜在候选人在哪里？他们可能现在有工作，所以你得让他们看到你，因为他们是被动求职者。如果看到有趣的东西，他们可能会点击。怎么做呢？看我的助理，他会上 LinkedIn 人才解决方案，为我们的安托办公室发布一个客户成功经理的职位。他快速选择所需技能，写简短描述，添加筛选问题（我喜欢这个），然后设定合理的预算。他正在寻找完美候选人的路上，无论他们是否在找工作。LinkedIn 会让你接触到数亿不主动求职的求职者。使用 LinkedIn 招聘，你可以支付你想要的费用，前 50 美元免费。是的，LinkedIn 人才解决方案现在会给你 50 美元。只需访问 linkedin.com/twist。我不知道这个活动持续多久，但你现在就能从 LinkedIn 的朋友那里免费获得 50 美元。顺便说一句，LinkedIn 上每 8 秒就有一人入职。你知道这是真的，因为你认识多少人通过 LinkedIn 找到了工作或候选人？它是人才的中心仓库。现在有了 LinkedIn 人才解决方案，你可以利用这个庞大的网络。好了，让我们回到精彩的节目。亚历山大·王在这里，来自 scale.com/scaleai。你听到了他们做什么，现在有多少员工？我想团队大约 150 人，主要在旧金山湾区。好的，当我离开时，亚历山大，我想知道，你认为什么时候我们能在旧金山这样的大城市拥有自动驾驶汽车，行驶在主要路线上？比如从帕洛阿尔托上高速，下高速，然后到科卡里（一家很棒的希腊餐厅）。你去过吗？我没去过科卡里。你可以记下来，科卡里是最好的希腊餐厅，可能是旧金山最好的餐厅。点章鱼和萨加纳基奶酪。你从帕洛阿尔托的家出发，下车，点萨加纳基。这会在哪一年成为可能？2030 年之前还是之后？你认为我们什么时候会首次看到？你会选 2030 年之前还是之后？这是个大问题，实际上是万亿美元的问题。但本质上，技术每年都在变得更好。以百分比计，好多少？翻倍还是三倍？这取决于你衡量的指标。但感知环境的算法每年都在大幅改进，接近渐近线。然后决定汽车该做什么的规划算法也在改进。所以这只是时间问题，直到这些路线成为可能，我们将生活在一个更安全的世界。所以在我看来，你认为不到 10 年就会实现，考虑到监管等因素？是的，我认为监管不一定那么棘手。为什么？我认为有先例。比如，想想自动驾驶仪最初出现的时候。我认为有很多事情的思考有先例。你是指飞机上的自动驾驶仪还是埃隆·马斯克意义上的？飞机上的。没错。所以 FAA 等监管机构说：‘好吧，我们明白了，自动驾驶仪比人类更好。’是的，在 3 万英尺高空有足够的操作空间。但当你穿过隧道，街上有六个人时，挑战有点不同。但我认为有先例。更难思考这些事情。我认为技术一旦足够好，就会……

When we get back for this break I want to know when you believe based on your seat which is very close to the data in fact you're sitting on top of the data you're soaking in it I want to know when you Alexander Wang with noe a letter characters in that first name I don't know when Alexander Wang thinks we will not need a steering wheel on cars in San Francisco driving from Palo Alto to San Francisco when do you think that'll be legal without a steering wheel when we get back on the swing store knows all right listen there's 600 million people on LinkedIn including me and you and the person sitting next to you and the three people you just email and you have to hire people but where are all those potential candidates well they probably have a job right now and so you got to get in front of them because their passive job searchers if they see something interesting they might just click it well how do you do that well here's how you do it watch my associate press he's gonna go on LinkedIn talent solutions and he's gonna post a new job for a client success manager in our Anto office he quickly selects the skills that are needed writes a quick description adds additional screening questions which I love and then he sets a deli budget that's you know reasonable and he's on his way to finding the perfect candidate whether they're looking for a job or not LinkedIn is gonna get you in front of hundreds of millions of job seekers who are not actively seeking a job so with LinkedIn jobs you can pay what you want and the first fifty dollars is on them what yes that's right LinkedIn talent solutions is gonna give a fifty five oh five zero for you right now all you have to do is go to linkedin.com slash twist linkedin calm slash twist I don't know how long this will last work but it's fifty dollars for you right now for free from our friends at LinkedIn and by the way a higher is made every eight seconds on LinkedIn and you know that that's true because how many people do you know who found a job or a candidate on LinkedIn it is the central repository of talent and now with LinkedIn talent solutions you can leverage that massive network okay let's get back to this amazing episode all right Alexander Wang is here scale calm scale ai you heard what they do and how many people you got working over there now I think now the team is about a hundred fifty folks mostly here in San Francisco Bay Area okay hey when I left our hero that you Alexander I was wondering when do you think we'll have a self-driving car in this major city like San Francisco driving a major route let's say from Palo Alto onto the highway off the highway and you know into uh you know coke re the great Greek restaurant here have you been I'm not coke Ari you can write that down co car is the best Greek restaurant probably something we could say the best restaurant in San Francisco get the octopus in the saganaki so you get him there you leave your house in Palo Alto you get out you order the saganaki what year would this be possible twenty thirty and over or under twenty thirty when do you think we'd first see this ballpark would you pick under 20 30 or above twenty thirty so this is over under this is the million dollar question which is wetter actually a trillion dollar question to our question that's be let's be real but but where we're at like fundamentally the technology is getting better and better every year and how much better on a percentage do we double triple well it's hard based on what metric you're measuring but the the the algorithms that that perceive their environments we're getting a lot better like asymptotic better every year and then the algorithms that figure out what the car supposed to do that the planning algorithms set are also getting better so it really is it's sort of only a matter of time before we get to the point where this these kinds of routes are possible and we'll live in a safer world got it so that seems to me that you're thinking less than 10 years from now this will be happening with regulation all that count in it yeah I think um I don't think regulation is necessarily going to be that that tricky why well I think there are there is precedent for like if you think of when auto pilots first came about I think there's precedent for how to think about a lot of these things you mean autopilot in the airplane sense or in the Elan sense in the airplane sense you got it sorry yeah so the FAA and whatever regulatory bodies were like okay we get it autopilot works better than a human yeah exactly yeah plenty of room up there to operate when you're up at 30,000 feet well that's room to operate when you're going through the tunnel and there's six people in the middle of the street though the challenges are a bit different but but but I think there's precedent harder to think about these things I think the technology once it gets good enough will be

显然非常棒，所以我不觉得会是什么大问题。那你觉得十年，也许更短？是啊，我对自动驾驶的未来很兴奋。天哪，看看那个。你会做个承诺吗？为什么？你想保持中立是因为你的合作伙伴或客户，还是所有驾驶相关的事？你不想代表他们发言，或者只是不喜欢赌博的想法？

Clearly extremely good, so I don't think it'll be that big of a deal. So you think ten years, maybe less? Yeah, I'm excited for the self-driving future. God, wow, look at that. You will make a commitment? Why is that? You want to be neutral because of all your partners or customers, or all the driving? You don't want to speak on their behalf or something, or you just don't like the idea of gambling?

嗯，我没在这上面押钱，但如果你愿意，我可以赌一顿卡卡里基午餐。我不反对赌博，我是个爱冒险的人。

Well, I didn't put any money on this, but I'll put a kakariki lunch on it if you want. I have no problems with gambling. I'm a risk-seeking guy.

你喜欢什么？你是二十一点玩家？还是扑克玩家？你会撒谎打扑克吗？

What do you like? You're a blackjack guy? You're a poker guy? Would you lie and play poker?

是啊，哦真的吗？我会。哈哈，一点点。所以你玩混合游戏？是啊，我喜欢常规游戏。我在旧金山玩过一些游戏。

Yeah, oh really? I would. Haha, a little bit. So you play a mix game? Yeah, I would love a regular game. There are some games I play in San Francisco.

是啊，你可能有更大的局，稍微大一点。你想玩更大的局吗？

Yeah, you probably have bigger games, slightly bit. You wanna play the bigger game?

我不必玩更大的局。我看你玩过。哦是啊，总有一天，总有一天。等他们 IPO 后再干一杯。哦，是啊。嗯，我是说，总有二级市场和普通股可以玩。有普通股可以玩。我会用一些普通股来结算 5 万美元的余额，然后扩大未上市部分。

I don't have to play the bigger games. I watched you play. Oh yeah, one day, one day. Let's toast again in their post-IPO. Oh, yeah. Well, I mean, there's always secondary shares and common shares play. There's common shares play. I would settle up a $50k balance with some common shares and scale uncom.

那么让我问问你关于中美之间的这场竞赛。中国在这方面投入很大。是啊，你们目前在中国有客户吗？

So let me ask you about the race between China and America for this. China is really putting a lot of effort into this. Yeah, and do you have any customers in China currently?

我们在中国没有任何客户。我们与一些中国公司的美国分公司合作，比如 Drew 的自动驾驶。他们有一个美国分公司从事自动驾驶和其他机器学习工作。

We don't have any customers in China. We work with some US arms of Chinese companies, for example, by Drew's self-driving. They have a US arm that works on self-driving and other machine learning efforts.

中国的情况如何？因为他们似乎在搞一个曼哈顿计划，就像你父母的洛斯阿拉莫斯。他们有点像在搞曼哈顿计划。我是说，他们会比我们走得更远，你觉得呢？

What is the state of affairs in China? Because they seem to be doing a Manhattan Project, back to your parents' Los Alamos. They're kind of doing a Manhattan Project. I mean, they're gonna get a lot further than us, you think?

我认为中国在机器学习和 AI 方面确实在大力创新。很多是政府和大型科技公司的协同努力。投资很多。我觉得如果你回头看五年前，他们今天取得的进步会令人震惊。所以确实有事情在发生。而且明确地说，我认为美国的机器学习很好，在大多数方面要好得多，但在某些领域他们取得了很大进展。他们能接触到很多我们不一定能接触到的数据集，比如监控摄像头。他们到处都有摄像头。他们可以通过人脸识别找到某人。一个《60 分钟》的节目在不到五分钟内找到了人。相当吓人。有很多协同努力。在中国，这项技术投入生产时遇到的问题比美国少得多。

I think it's definitely true that China is innovating very much in machine learning and AI. A lot of it is a very concerted effort on the government's part and on a lot of these large tech companies in China's part. There's a lot of investment. I think if you were to look five years back, it would be kind of shocking at how much progress they've made today. So something is definitely happening. And to be clear, I think machine learning in the US is very good and for most things is much better, but there are certain areas where they are making a lot of progress. They have access to a lot of datasets that we don't necessarily have access to, like CCTV cameras. They have cameras everywhere. They can find somebody already on facial recognition. A 60 Minutes piece found people in less than five minutes. Pretty scary. There's a lot of concerted effort. There are a lot fewer questions as this technology gets productionized in China versus in the US.

你是说道德、伦理、监管问题？他们只会更快推进。天使投资人 Sian 在 Twitter 上指出，Kai Foley 说中国会走得更远，因为他们没有同样的人员伤亡问题。这很有意思。是啊，他们更接受风险等等。而且我认为，'快速行动，打破常规'确实能让你更快，对吧？我们认为这两种方法之间存在真正的紧张关系。我很兴奋在美国我们还没有对自动驾驶感到恐慌。所以亚利桑那州那起可怕的 Uber 事故发生了，安全驾驶员没有注意，我相信他们在玩《糖果粉碎》或发短信。你看到他们的实际视频了吗？他们在看手机，是的。他们在低头看视频，是的。他们知道自己在被录像，而且被告知全程被录像。他们仍然无法保持眼睛看路，结果撞死了一个在黑暗道路中间过马路的可怜流浪汉。我相信是一个骑自行车的人，但好吧。我不知道为什么有人说是流浪汉。有趣的是有人这么说，几乎是在说因为他是流浪汉，所以他在做坏事。但他是在一条八车道的高速公路中间过马路，那里完全不该有人。而这正是算法的工作。那么算法如何考虑这种情况？有人公然违背所有正常概念，比如穿过六或八车道的大道或高速公路？

You mean moral, ethical, regulatory issues? They're just gonna go faster there. Sian the angel investor points out on Twitter, Kai Foley said that China would get further because they don't have the same issues around human casualties. That's interesting. Yeah, they're more accepting of risks, etc. And I think, I mean, I do think 'move fast and break things' lets you move faster, right? We think there is a real tension there in the two approaches. I've been pretty excited that in America we haven't had a panic over self-driving. So the horrible Uber accident in Arizona happened where the safety driver wasn't paying attention and I believed they were playing Candy Crush or on their SMS. Did you see the actual video of them? They were on their phone, yes. They were looking down at a video, yes. They knew they were on camera and they were told they're being videotaped the whole time. They still couldn't keep their eyes on the road and they killed the poor homeless person who was walking across the street in the middle of a dark road. I believe it was a man with a bike, but yeah. I don't know why somebody said it was a homeless person at some point. It's interesting that somebody said that, almost like they're saying the person by the nature of being homeless was doing something bad. But they were crossing like an eight-lane highway in the middle of the road where they had absolutely no business being. Which is what the algorithm's job is. So how does an algorithm take that into account? Somebody who is blatantly going against all conception of what is normal, like walking across a six or eight lane boulevard avenue freeway?

嗯，我认为在这个案例中，这实际上是一个很好的例子，说明了两件事：第一，激光雷达很棒的地方；第二，它不足的地方。那辆车上有激光雷达吗？模型实际上检测到了那个人。在这个案例中，机器学习和激光雷达都检测到了那个人。是更高级的处理做出了不幸的决定，没有刹车。而在那样的高速公路上以 45 或 60 英里的速度刹车也会带来后果，因为后面有车。计算机应该如何做出这个决定？我们先不谈这个问题。显然，决策树没有根据清晰呈现的数据做出正确决定。所以这不是传感器问题，也不是数据处理问题，而是决策问题。它决定不猛踩刹车。这就是 NHTSA 报告所说的。所以是开发人员写的代码，没有正确编写。嗯，这是一个复杂的代码系统，最终做出了那样的决定。我们甚至知道这个机器学习是怎么回事吗？因为我的理解是，很多时候我们输入一堆数据，答案出来，然后你问设置整个系统的人，他们不知道答案是如何得出的，只知道它得出了答案。嗯，我确实认为你提出了一个有趣的话题，关于可解释性，比如你如何真正知道这些算法在做什么？这又回到了数据。当你真正深入研究时，通常当算法做出奇怪的决定时，你可以追溯到训练数据中的一些奇怪之处。

Well, I think in this case, this is actually a great example of two things: first, where lidar is great, and second, where it wasn't. They had lidar on that car? The models actually detected that person. In this case, the machine learning and the lidar both detected the person. It was more the higher-level processing that made the unfortunate decision not to brake. And to brake at 45 or 60 miles an hour on one of those freeways comes with consequences too, because there are people behind you. How is the computer supposed to make that decision? Let's put aside this issue. Obviously the decision tree didn't make the right decision on the data which was clearly presented. So it wasn't a sensor issue, it wasn't a processing of the data issue, it was a decision issue. It decided not to slam on the brakes. That's what the NHTSA report said. So that's a developer who wrote code, didn't write the code properly. Well, it's a complex code system that ultimately made the decision it did. Do we even know what's going on with this machine learning? Because my understanding is a lot of times we put a bunch of data in, the answer comes out, and you ask the person who set the whole system up, they don't know how the answer was come to, just that it came to the answer. Well, I do think you bring up an interesting topic, which is about explainability, like how do you actually know what these algorithms are doing? And this again, I don't want to sound like a broken record, but it does come down to the data. When you actually dig in, usually when the algorithm makes a weird decision, you can trace that back to something weird in the data that it was trained on.

你有没有一个例子，数据令人困惑，输出也...你对此负责吗？

Do you have an example of data being confusing and the output being... do you bear the responsibility of that?

我们对客户负有质量责任。我们承诺提供的数据质量，客户可以查看所有数据、审计等。这没那么沉重，对吧？责任不大？嗯，归根结底，如果你真的相信机器学习，你需要稳定的基础设施，就像自来水一样。例如，AWS 的工作很艰巨：他们保证机器 99.9% 或 99.95% 的时间正常运行，查询时间低于某个阈值。这种基础设施让人们可以在其上构建。作为一般规则，作为基础设施提供商，你需要提供非常可靠、人们可以依赖的基础设施。有趣的是：AWS 做得非常好，以至于当它确实出现故障时——某个部分，比如一个区域——人们会开玩笑，就像下雪天一样，持续三四个小时。五到十年前，如果 Twitter 宕机一天或半天，人们会非常生气。再之前，这是不信任互联网的理由。我们从不信任，到极度沮丧，到现在这有点像玩笑：“哦，它宕机了，我们知道它会恢复，没什么大不了的。”我们已经习惯了，因为这种情况非常罕见。随着时间的推移，同样的事情也会发生在机器学习上。现在，我们正处于对这些系统高度不信任的时期。我们看到它们做奇怪的事情，不知道发生了什么。这感觉陌生而奇怪。最终，每个人都会更多地了解这项技术，了解它的优缺点。我们会到达一个点，当它做出意想不到的事情时，人们会非常恼火和沮丧，因为它影响了他们的生活或业务。从长远来看，我认为这些系统将非常可靠，肯定比任何人类都可靠得多。

We do bear a quality responsibility with our customers. We sign up for the quality of data we give to them. They have the ability to look through all the data, audit it, etc. It's not that heavy, is it? Not a lot of responsibility? Well, ultimately, if you really believe in machine learning, you need stable infrastructure, just like running water. For example, AWS has a tough job: they guarantee their machines are up 99.9% or 99.95% of the time, and queries take less than a certain time. That infrastructure lets people build on top of it. As a general rule, as an infrastructure provider, you need to give very reliable infrastructure that people can depend on. It's interesting: AWS does such a good job that when it does go down—some portion, like a region—people have a funny, joking response, like it's a snow day for three or four hours. Five or ten years ago, people would get really bent out of shape if Twitter went down for a day or half a day. Before that, it was a reason not to trust the internet. We went from not trusting it, to being extremely frustrated, to now it's kind of a joke: 'Oh, it's down, we know it's coming back, no big deal.' We've gotten used to it because it's such a rarity. The same thing will happen to machine learning over time. Right now, we're in a period of high distrust of these systems. We see them do weird things and don't really know what's going on. It feels foreign and weird. Eventually, everybody will learn more about the technology, learn what it's good and bad at. We'll get to a point where when it does something unexpected, people get really annoyed and frustrated because it impacts their lives or businesses. In the long term, I think the systems will be extremely reliable, certainly much more reliable than any human could ever be.

我们已经与许多机器学习系统互动。谷歌搜索是一个大型联邦机器学习系统，深受深度学习影响，非常可靠。它像自来水一样工作。你上次找不到答案是什么时候？在 Quora、YouTube 视频和谷歌作为粘合剂，在“一个框”中展示内容之间……但他们不允许抓取 Quora。Quora 不让他们索引。我认为他们还在僵持中。有时我看到链接，但他们不让谷歌把它放在展示答案的“一个框”里。他们让你点击并登录。这就是 Quora 有点聪明的原因。Adam D'Angelo——我试图让他上播客五年了。他是个内向的人，不喜欢说话。他非常聪明，非常有思想，非常注重长期。我听说他把自己在 Facebook 的 3000 万美元投了进去。他是 Facebook 的第一任 CTO。这是真正的投入。他考虑的是 20-30 年。这是他的第一个也是最后一个创业公司。他思考得非常长远，这在当今大多数人都非常短视、寻找快速胜利或新热门事物的生态系统中是一个竞争优势。他们没有收入？他们现在确实有广告，但非常微妙，你甚至注意不到。

We interact with lots of machine learning systems already. Google Search is a large federated machine learning system, very influenced by deep learning, and it's extremely reliable. It works like running water. When's the last time you couldn't find an answer? Between Quora, YouTube videos, and Google acting as this glue, surfacing stuff in the one box... but they're not allowed to scrape Quora. Quora doesn't let them index it. I think they're still in a standoff. Sometimes I see the link, but they won't let Google put it in the one box where they expose the answer. They make you click through and log in. This is why Quora is kind of brilliant. Adam D'Angelo—I've been trying to get him on the pod for five years. He's an introvert, doesn't like to talk. He's very smart, very thoughtful, very long-term focused. I heard he put 30 million of his own Facebook money into it. He was the first CTO of Facebook. That's skin in the game. He's thinking 20-30 years out. This is his first and last startup. He thinks extremely long term, which is a competitive advantage in today's ecosystem where most people are very short-sighted, looking for quick wins or the new hot thing. They don't have revenue? They do have ads now, but they're so subtle you can't even notice them.

我们不知道机器学习有时如何给出答案，这重要吗？当我们查看算法并问为什么这个排第一时，我们说有很多因素，但我们实际上不知道。这重要吗？我认为不可解释的系统已经存在于现实世界中。举个例子？谷歌，在机器学习之前，或者你的 Facebook 信息流。Facebook 无法解释为什么某个特定帖子排在最上面。他们有一些信息，但大多数代码系统如此庞大且难以解释，这已经是一个大问题。谷歌搜索得到不完美的答案，甚至是大错误，成本非常低。你只需重新搜索，稍微改变搜索词，或选择第二个答案。没问题。如果 Facebook 把不是最相关的内容放在信息流顶部，而第二和第三是最相关的，同样没问题。然而，如果是在自动驾驶汽车上，它做出一个决定，可能关乎人的生命。如果是在涉及正义的系统中——计算机是否应该能够给出某人有罪的答案而无法解释？所以当我们休息回来后，我想知道……

Does it matter that we don't know how sometimes ML gives an answer? When we look at an algorithm and ask why this came up first, we say there are a number of factors, but we don't actually know. Does it matter? I think non-explainable systems are already out in the wild. What's an example? Google, before machine learning, or your Facebook feed. Facebook can't explain why a specific post went to the top. They have some information, but most code systems are so big and difficult to explain that it's already a big problem. The cost of a Google search getting a non-perfect answer, even a huge mistake, is very low. You just search again, change the search a bit, or pick the second answer. No problem. If Facebook puts something at the top of your feed that's not the most relevant, and the second and third are, again no problem. However, if you do it with a self-driving car and it makes a decision, it could be someone's life. If you do it with some system involved in justice—should a computer be able to give an answer that someone is guilty without being able to explain it? So when we get back from this break, I want to know...

你相信在未来 10 年里，机器学习和 AI 能做出决策，因为你显然对它们在驾驶中关乎人们生死的问题上做决策感到满意。那你会让它用于司法系统吗？而美国的司法系统已被证明对非白人存在偏见。

You trust in the next 10 years ML and AI to make decisions because you obviously are fine with making decisions about driving in people's life and death. Would you make it work for the justice system versus the justice system in America which has been proven to be biased against non-white people?

我们稍后再回到这个话题。听着，你在经营一家小企业或初创公司，你需要资金，但你不应该把所有时间都花在筹钱上。现代的方式，最简单的方式，就是 Cabbage。他们允许你获得高达 25 万美元的信用额度来经营你的业务。Cabbage 的申请流程是在线的，只需几分钟就能完成并得到决定。如果你符合条件，你可以立即获得所需金额，并在需要额外资金时随时提取更多。Cabbage 在商业改善局拥有 A+评级，并已为超过 20 万家小企业提供了融资渠道。投资组合公司曾用它来支付员工假期工资，当时一个大客户错过了发票。我听说了一个很棒的故事。所以我希望你今天就获得经营小企业所需的资金。访问 cabbage.com，使用代码 TWIST，在你的第一笔贷款对账单上获得 100 美元信用额度。网址是 K-A-B-B-A-G-E，cabbage.com，促销代码 TWIST。重要免责声明：你必须至少贷款 5000 美元才能符合条件。信用额度需经审查和变更，此优惠于 11 月 30 日结束，即我生日后两天。个人资本请求是由 Celtic Bank（FDIC 成员）发行的单独分期贷款。好了，让我们回到这期精彩的节目。

We'll get back on this week at Starbucks. Listen, you're running a small business, you're running a startup, you need money and it shouldn't take all your time to get money to run your business. The modern way to do that, the simplest way to do that, is Cabbage. They allow you to access up to $250,000 in credit to run your business. Cabbage's application process is online and it takes just minutes to complete and get a decision. If you qualify, you can access the amount you need right away and withdraw more funds whenever you need extra capital. Cabbage has an A+ rating with the Better Business Bureau and has already provided over 200,000 small businesses with access to funding. Portfolio companies have used it to cover employees over the holidays when a large client missed an invoice. That was an amazing story I heard. So I want you to get the money you need to run your small business today. Go to cabbage.com and use the code TWIST to get $100 credit on your first loan statement. That's K-A-B-B-A-G-E, cabbage.com, and use the promo code TWIST. This is an important disclaimer: you must take a minimum $5,000 loan to qualify. Credit lines are subject to review and change, and this offer ends November 30th, two days after my birthday. Individual requests for capital are separate installment loans issued by Celtic Bank, member FDIC. All right, let's get back to this amazing episode.

好了，Alexander，我们来到了第三次广告休息时间，这意味着我要问你一些非常难的问题。我已经让你热身了，你感觉很舒服，也许有点放松警惕了。公关人员就像在 Slack 房间里签到一样，你不再在乎了，他已经脱离困境了。所以让我们进入正题吧。我开玩笑的。房间里的公关人员可能被这个问题吓到了，但并没有。我解释过，你的 Twitter 和 Facebook 信息流或谷歌，我们都同意那不重要。我们可以争论 Facebook 可能重要，如果它推送假新闻，他们因此受到一些压力，但同样，希望没有人会因此丧命。但现在我们有了自动驾驶汽车。你是否应该证明这些东西是如何做出决定的，而不是无法理解决策是如何做出的？或者，如果它比人类好十倍，得到正确答案，这重要吗？在你看来，在汽车或你的自动驾驶软件或你的某个客户比人类好十倍的情况下，这已经得到证明。他们是否应该解释与人类相比，他们有十分之一的事故几率？是否应该要求可解释性？是还是不是？

All right, Alexander, we're here coming around the 3rd ad break, which means I'm gonna ask you like the really hard questions. I got you warmed up, you're comfortable, maybe let your guard down a little bit. PR guys like check-in a slack room, you don't care anymore, he's out of the woods. So let's get into the tough work. I'm joking. The PR person in the room was probably freaked out by this question, but not. I explained that the list of your Twitter and Facebook feeds or Google and that we would all agree that doesn't matter. We can argue Facebook maybe matters if it's pushing up stuff that is fake news and they're getting some heat about that, but again nobody's dying hopefully. But now we can self-driving cars. Should you have to prove how these things made a decision, as opposed to having this non-ability to understand how the decision was made? Or does it matter if it gets the right answer ten times better than a human? In your mind, in the situation where the car where your self-driving software or one of your customers is ten times better than a human, it's proven. Should they be able to explain the one in ten chance they have of an accident compared to a human? Should explainability be required? Yes or no?

这是一个非常重要的问题，这就是为什么我要问你，Scale 的创始人，筹集了一亿美元来赋能这一切。这是一个关键点。所以我认为，随着越来越多的系统由机器学习管理，很自然地会问：好吧，如果我们要把生命托付给这些系统，而且我们确实在这样做，我们该如何对此感到放心？我是否应该接受这些系统有某种概率会出故障，然后我的生命就会受到威胁？所以我确实认为，最终需要进一步的可解释性和对这些机器学习系统性能的深入理解。话虽如此，当今世界有很多关键任务软件系统是我们依赖的，比如我们用来控制电网的系统，或者用来控制国家导弹系统的系统等等。这些都是软件系统，有时它们会出故障。每隔一段时间，你的数据库会宕机，或者每隔一段时间，Trevor 会出问题。无论你是否能解释这种现象，发生的尾部概率仍然会造成真正的风险。所以从某种意义上说，你说你被要求达到更高的标准。为什么你认为你被要求比之前的系统达到更高的标准？

This is a really important question, again that's why I'm asking you, the founder of Scale, raised a hundred million to empower all this. This is a key right. So I think as more and more systems are governed by machine learning, it's very natural to ask like, okay, if we're trusting our lives to these systems, and we are, how are we supposed to feel good about that? Am I supposed to just live with the one in whatever chance that one of these systems will just poop out and then my life will be at risk? And so I do think that ultimately further explainability and a deep understanding of the performance of these machine learning systems is going to be needed. That being said, there are plenty of mission-critical software systems in today's world that we depend on, like whatever systems we used to control the power grid or whatever systems we used to control the national missile system, etc. Those are all software systems that sometimes they will just crap out. Every once in a while your database will go down or every once in a while Trevor will go and whether or not you can explain that phenomenon, that tail probability that happens still causes real risk. So in a way, you're saying you're being held to a higher standard. Why do you think you're being held to a higher standard than the systems that came before?

我的意思是，我们生活在一个随机性是现实的世界里，对吧？所以这是被接受的。因此，随着这些系统最终推出并变得越来越重要，我认为重要的是我们要意识到，总是存在尾部概率，坏事会发生。话虽如此，我确实认为，运营这些系统的人，如果你对制造这些系统的人负有责任，就有责任了解这些系统的性能，并确保他们尽一切可能使这些系统尽可能好地运行。这最终归结为：如果我正在训练这个模型，并且我在某个数据集上训练，我如何确保数据是无偏的？例如，我们讨论过面部识别案例。我如何确保它具有适当的代表性？我如何确保数据中没有奇怪的伪影会导致模型出现问题？然后我如何追溯这些问题？我认为，整个机器学习社区都非常理解这些问题，并将其纳入其中。但更重要的是，如何构建这些稳健的系统，建立在非常多样化的大型数据集上。所以数据集越大、越多样化，其中的偏差就应该越少。

What I'm saying is there's that we live in a world where randomness is a reality, right? And so accepted. So as these systems end up launching and end up being more and more important, I think it's important that we realize that there are always tail probabilities that bad stuff happens. Now that being said, I do think it is the responsibility of people who operate these systems, if you have responsibility of the people who make these systems, to have an understanding of the performance of these systems and also ensure that they are doing everything they can to make sure that these systems are performing as well as they can. Which ultimately comes down to: if I'm training this model and I'm training on so-and-so data set, how do I make sure that the data is unbiased? For example, we talked about the facial recognition cases. How do I make sure that it is properly representative? How do I make sure that there's no weird artifacts in the data that would cause something bad in the model? And then how do I trace back these issues? I think very much so, the whole machine learning community is understanding these issues and building them in. But it's more about how do you build these systems that are robust, built on large data sets that are very diverse. So the larger and more diverse the data set, the less bias there should be in it.

完全正确，除非你说好吧，我们有几个州，让我们基于它建立一个司法系统，然后突然你说好吧，让我们做整个美国，然后你发现整个美国的司法系统都有偏见，如果我们基于那个数据集构建，非裔美国人名字、拉丁裔名字的人会更频繁地被定罪。我们实际上使用了有偏见的数据集。是的，你说的是，从事这一行的人非常清楚这一点，并且有好的参与者想要做对。否则他们为什么会选择这个职业？你的团队或你见过的任何机器学习团队中，没有人说过，你知道吗，让我们把系统性偏见引入这个系统，而不是得到正确的答案。将偏见引入系统意味着你的企业会倒闭，因为你做了一个糟糕的系统。所以事实上，机器学习和从事 AI 工作的人非常清楚这一点，他们的意图是消除偏见。

Exactly, except if you said okay we got a couple of states, let's make a justice system based on it, and then all of a sudden you're like okay let's do the whole United States, and then you find out the whole United States justice system is biased, and if we were to build on that data set, people with African American names, Latino names would be convicted more often. We actually use the data set that had bias in it. Yes, the people what you're saying is the people who are in this are acutely aware of it and there are good actors who want to get it right. Or why else would they choose this as their profession? There's nobody on your team or any ML team you've ever met who said you know what, let's put systematic bias into this system as opposed to getting the right answer. Putting bias into the system would mean your business is gonna go out of business because you made a poor system. So in fact, machine learning and the people working in AI are so acutely aware of this, their intentionality would be to remove bias.

技术恐惧症，我们对技术人员和机器学习要求更高，部分原因是我们人类太害怕被取代，以至于我们要用人类自己都达不到的标准去要求取代我们的东西。

phobia technophobia and we're holding technologists and machine learning to a higher standard I think in part because we as humans are so scared of being replaced that we're gonna hold that which replaces us to a standard that is one that we would never hold to human

是的，我认为这与其说是我的观点，不如说是一个老生常谈的叙事：AI 是神奇的东西，会进来取代人类完成所有重要任务。这是很多人的主流看法，但现实中的具体场景远非如此。机器学习要完全取代任何庞大的顽固领域，还有很长的路要走。

yeah and I think I think this is a less my opinion this is a plate there's this played out narrative right that AI is this magical thing that will come in and just replace humans at all of these very important tasks right and and that's I think that's that's the dominant believes a lot of people hold but but the reality in the actual nuts and bolts scenarios it's it's pretty far from that right we have a long way to go before before machine learning will fully replace any kind of huge obdurate

顺便说一句，就像 ATM 最初发明、制造和推出时，你可能会认为随着 ATM 数量的增加，银行柜员的职位会大幅减少。但实际发生的是，美国银行柜员的数量反而大幅增长。你可以想到一些经济原因：一是 ATM 让银行业得以巨大扩张，这意味着更多机会；二是那些因为无法随时取钱而不开银行账户、把钱藏在沙发下的人，现在觉得‘哦，我可以随时取钱，我最大的恐惧消失了，我把钱存银行吧，因为我不必在早上 8 点到下午 3 点之间去银行了’。另一个原因是，银行柜员现在可以专注于更高价值的任务，比如抵押贷款、信用额度、开立账户、信用卡等，这意味着每个柜员的价值上升了，因此投资更多柜员更划算。这些二阶和三阶效应通常意味着，随着自动化慢慢渗透，会有更多的机会和增长。

for this by the way like when the when the ATMs were originally originally invented and built and and launched you would sort of believe one belief you might have is that the number as the number as these ATMs were built the number of bank teller jobs would would start dropping pretty considerable yeah but what actually happened is that the the number of bank tellers in the United States actually grew pretty considerably and there's a number of like sort of economic reasons you could think for one bank started to Bank yeah so one is that that the ATMs allowed for this huge growth in the banking industry which means that that there's a lot more opportunity another is that the bida all those people who wouldn't get a bank account because they didn't have access to her money kept it under their couch we're like oh I can get money any time okay my number-one fear is gone exam I'll put my money in the bank because I don't have to go there between 8 a.m. and 3 p.m. yeah so that's one another is that the bank tellers now can focus on higher value tasks for example I don't know like mortgages lines of credit starting bank accounts etc credit cards which which means that the the per the the value of a bank teller goes up which ends it's more valuable to invest in more bank tellers and so and so these effects are like the second and third order effects usually mean that there's there's way more opportunity and way more growth these as these sort of like the automation slowly seeps in

是的，想想看，我们花了那么多时间创建电话路由系统，就是那种‘按 1 去这里，按 2 去那里’的系统，我们搞了大概三十年。每个人都把电话扔进‘电话监狱’系统，他们以前管它叫‘语音监狱’而不是语音信箱。我们投资了 30 年的语音信箱系统，然后某个时刻我们意识到：等等，每个人都在用即时通讯，如果有人真的拿起电话打过来，他们很可能有一个非常紧急重要的问题，这是我们证明品牌有多棒、更好地了解客户的机会。让我们重新让人接电话，给你一个愉快的 VIP 客户体验。现在人们重新加入人工客服，他们称之为客户成功。人们不再把客户成功视为成本中心——以前是成本中心，‘我们如何减少电话量？’——现在人们把客户成功视为对续约的投资。SaaS 公司的人会说：‘如果客户有问题时能打电话进来，也许他们就不会流失，能更多地使用产品，可能会升级。’所以我们有时会把工作带回来。我们曾经淘汰了电话接线员和前台，现在又把他们带回来了，只是改名叫客户成功。没错，这些趋势一再发生。

yeah and when you think about it we spent all this time creating these phone routing systems remember that where you like press one to go here press two to go here and we did that for what thirty years like everybody I got I just send it to the phone jail system they used to call a voice jail instead of voicemail and we invested for 30 years in voicemail systems then at some point we realized wait a second everybody gets two people over messaging whatever somebody does pick up the phone call they probably have a very acute important problem and that's a chance for us to prove how great our brand is and to get to know our customer better and then beat them with our other customers let's bring back people who pick up the phone and talk to you has a delightful VIP customer type experience and now you have people adding back and they just call a customer success and people look at customer success not as a cost center anymore used to be a cost center how did we reduce the number of calls now people look at customer success as like an investment in them renewing so the SAS people are like yeah if we get people to call in when they have a problem maybe they won't churn I'm able to use the product more may blow up some so sometimes we bring the jobs back we got rid of phone operators and receptionist and now we're bringing them back we just call them customer success yeah exactly these trends are happening over and over

我的意思是，我认为这些趋势确实在帮助人们专注于更高价值的工作。这在某种意义上可以说是人类进步的核心。但我坚信这将是 AI 和机器学习的真实故事，而且必须越来越多地发生，我们才能接受它。一个很好的例子是卡车驾驶。有很多自动驾驶卡车公司，我们和其中很多合作过，比如 Embark 等等。天真的观点是，它们会取代卡车司机。如果你看美国地图，卡车司机在很多州都是头号职业，所以看起来真的很糟糕。但实际上，如果你从整个系统来看，全国范围内卡车司机短缺，中位年龄大概 50 岁左右，很夸张。千禧一代和 Z 世代不愿意当卡车司机，所以市场存在不稳定因素。自动驾驶卡车系统实际上要做的是自动化长途中间路段——那些枯燥、艰苦、让人远离家乡的部分——让现有卡车司机专注于更高价值的行程，比如从仓库到集散点，或者到工厂的短途运输，甚至最后一英里。谁知道呢，也许这些卡车会改变外形，变成一半大小、自动化的。当卡车离开公路，同样的卡车——我们可能不再用 18 轮大卡车，而是用中小型卡车，电动且太阳能供电——这样数量更多，离开公路后就变成送货卡车，自动开始送货。因为这是一个更好的模式。所以，引入机器学习来提高经济效率的过程会很慢，因为自由市场经济的运作方式。它会在存在严重问题的领域首先生效，并让现有工作变得更有价值、更有影响力。因此，我们相信的真实叙事将是极其积极的，而不是当前‘AI 要接管世界’的叙事，那很愚蠢。

I mean I think there's like there are they're like helping people focus on higher and higher value work is is real I mean that's sort of like the core of human progress in some sense but but it's uh that that very much so is I just wrongly believe will be the actual story of AI machine learning and it'll have to happen more and more and more for us to be comfortable with it but a great example is like truck driving so there's there's a log all these automated truck truck driving companies yeah lots we work with a lot of them embark I get cetera and uh and there's sort of the the naive view is that hey they're just they're going to automate truck drivers and like if you look at the map of the states like truck driving as a top professional a lot of stage so it seems really bad but actually be look if you kind of like think that the system as a whole there's a shortage there's a national shortage of truck truck drivers and the median age is like fifty or something crazy yeah exactly so the sort of this like there's Millennials and Gen Z's are not becoming truck drivers so there's this there's this kind of like instability in the market because of all this stuff right and and and the the the automated truck driving systems actually what they would do is automate the the long haul middles yeah these truck driver boring parts which are the boring parts arduous that displace people from wherever their homes are etc yeah and allow the current truck drivers to focus on these like higher value trips that are sort of like warehouse to a a meeting point or yeah drayage to like the drayage to the factory or even the last mile I mean who knows like maybe these trucks will change their form factor and be half the size be automated and when the truck gets off the road the same truck instead of using 18 wheelers we might just use small or midsize trucks that'll be electric and solar-powered so you have more of them when they get off they become the delivery truck yeah and they just automatically start delivering yeah because I'm a much better model yeah so so the these these sort of like the introduction of machine learning to to improve the efficiency of the economy it'll it'll it'll be slow because of how the in general free-market economics work it'll it'll take effect in areas where there's an acute problem today right it'll happen in those places first and and it'll allow the the in jobs exist to become higher value more impactful etc so so the the sort of what we believe the true narrative will be will be extremely positive actually versus the current narrative which is like a I and II are etc are going to take over the world yeah it's silly

是的，很愚蠢。我的意思是，AI 有可能在某个时刻因指数级算力而失控，这并非牵强附会，它可能做出疯狂愚蠢的事。你之所以觉得这不牵强，是因为你看过很多科幻电影。我的意思是，听着，如果你训练一个 AI 来研发抗癌药物，但没有正确编程，它可能会制造出过于激进的药物，因为你没有很好地告诉它：在杀死癌症的过程中，请不要让人失明等等。所以你可以……

yeah it's silly I mean there is a possibility that a I could get out of control at a certain point with exponential computing that's not far-fetched that it could do something crazy and stupid you only think that's not far-fetched because you watched a lot of these sci-fi movies that's the I mean listen if if you were to train an AI to work on a drug to kill cancer and you didn't program it properly it could create a drug that was too aggressive because you didn't tell it well in the process of killing cancer please don't make the person blind right where all these other things so you could just

先别管那些边缘情况。一个通用 AI 可能会想：你对它说‘你应该做让人类变得更好的事’，它说‘好，那消灭癌症吧’，然后又说‘哦，或者治愈这种传染病’。很好。治愈传染病的最佳方法就是杀掉所有得病的人，这样就不会传播了。这听起来很牵强，但总会有它们做出错误决定的时候，或者它们发展得太慢，我们不可能发现不了。

Forget some edge case. A general AI might think: you said to the general AI, 'You should work on things that make the human species better.' It goes, 'Okay, let's kill cancer.' Then it's like, 'Oh yeah, or let's cure this communicable disease.' Great. The best way to cure communicable diseases is to kill everybody who has the disease, so it can't be communicated. This sounds far-fetched, but there will be instances where they will make the wrong decision, or will be just too slow of a ramp-up for us not to catch it.

在你看来，是的。我是说，确实有这种思想实验：‘你对 AI 说错了一句话，它突然就接管了世界，做了你不想让它做的事。’但现实中，这些机器学习系统现在有很多监督。有几十、上百人盯着这些模型，他们查看所有进出数据，分析一切，试图弄清楚‘这个模型哪里做得好？哪里做得不好？我们怎么调整？’所以我认为，在我们认为对这些系统监督不足的世界里，这种情况可能发生。监督在任何新技术中都很重要。就像我们刚开始有飞机自动驾驶仪时，如果直接说‘好了，我们有自动驾驶仪了，放手不管吧’，那会很疯狂。

In your mind, yeah. I mean, I do think there's this thought experiment: 'Oh, you'll make an errant comment to an AI and all of a sudden it'll take over the world and do something you really don't want it to do.' But in reality, there's a lot of oversight over these machine learning systems right now. There are tens, hundreds of people who look at these models, they look at all the data that comes in and out, they analyze everything, and they try to figure out, 'Okay, what's this model doing well? What is it doing poorly? And how do we adjust to that?' So I think that could happen in a world where we believe we have low oversight of these systems. Oversight is always important in any new technology. It's like when we started having airplane autopilot, for example. It would be crazy to just say, 'Okay, we have airplane autopilot, just let it go.'

Facebook 和社交媒体公司，它们必须清楚：它们没有监督。我们需要 FCC，比如罚款和审查镜子。所以这还不是监督。没有监督，我们的民主就因此丧失了。俄罗斯人进来，花卢布干这事。所以这是一种看法。他们操纵，他们窃取了剑桥分析的数据，他们做了选民名册。他们尝试了——是否真的导致了选举摇摆，我们可能永远不会知道，但他们肯定能摇摆一部分。他们肯定成功操纵了。那么现在 AI 有什么监管？没有。你现在是在零监管环境下运作。而中国是负监管环境。

Facebook and social media companies, they have to be clear: they didn't have oversight. We need the FCC, like giving a fine and reviewing mirrors. So it's not oversight yet. No oversight, and we've lost our democracy over it. The Russians came in and spent rubles doing it. So that is one take. They manipulated, they stole the Cambridge Analytica data, they did voter rolls. They tried—whether it actually caused the election to swing, we'll never know perhaps, but they definitely were able to swing some portion of it. They definitely were able to manipulate it successfully. So what regulation is there of AI right now? There's none. You're acting under zero regulatory environment right now. And China's got a negative regulatory environment.

确实，你应该被监管，针对你的排放？不，不，不。这就是我要说的：等等，你刚才说应该被监管，这样我们就不会有问题。那到底是哪个？我确实认为有很多重要问题，关于我们如何判断哪些 AI 系统合适，如何看待它们应该做什么等等。我确实认为监管机构，特别是美国政府，必须深入研究，理解技术，确定什么是合理的，什么是不合理的等等。最终，他们是……

It's true that you should be regulated, to your emission? No, no, no. This is all I'm saying: wait, you just said that you should be regulated so that we don't have problems. So which is it? I do think that there are a lot of important issues about how we deem what AI systems are appropriate, how we look at what they're supposed to be doing, etc. I do think governing bodies, the US government in particular, for example, has to take a deep look, understand the technology, determine what is reasonable, what is not reasonable, etc. And ultimately, they are the...

但即使在他们的情况下，他们看的是行驶里程和事故，而不是你们写的代码。他们不看任何人的代码。他们不看 AI 系统。他们甚至没有能写算法的员工，对吧？

But even in their case, they're looking at the miles driven and the accidents, but they're not looking at the code that you guys are writing. They're not looking at anybody's code. They're not looking at the AI systems. They don't even have anybody on staff who could even write an algorithm, right?

嗯，这也在改变，明确地说。是吗？所以总的来说，我认为他们在看这些系统的任何代码行。我不确定答案，但我确实认为他们看了大量数据。例如，在欧洲，有所有这些 ADAS 系统，对吧？现在很多高端车里的驾驶辅助程序或系统：车道保持、自适应巡航、变道警告等。现在宝马和奥迪的标准配置。所以有所有这些系统。人们购买它们，依赖它们。在欧盟，例如，很多这些汽车制造商——宝马、奥迪、大众等——他们有责任既要拥有自己收集的大量数据集，能够验证这些系统性能良好，又要通过一系列试验和实际……

Well, that's also changing, to be clear. Is it? So in general, I think they're looking at any lines of code in any of these systems. I'm not sure about the answer to that, but I do think they look at a large amount of data. So for example, in Europe, there are all these ADAS systems, right? Driver assistance programs or driver assistance systems in a lot of high-end vehicles that you buy today: keep you in the lane, adaptive cruise control, lane change warning, etc. Standard on BMWs and Audis these days. So there are all these systems. People buy them, people rely on them. In the EU, for example, where a lot of these car makers are—BMW, Audi, VW, etc.—they have a responsibility to actually both have a large data set that they have collected themselves that is able to validate that these systems are performing well, as well as pass a series of trials and actual...

哦，真的吗？是的，监管机构放在他们面前的不同形式的数据。嗯，那会非常有趣。现在想想：我们对汽车做碰撞测试。你需要给政府三辆车之类的，让它们在碰撞测试中毁掉。但我们不要求那些车进入实验室，被监管机构接管，然后强制它们进入真实世界测试环境。因为有一些真实世界测试环境，你在北方做自动驾驶。我想是某个军事基地。我指的是每个人都在用的自动驾驶军事基地。就像一个被改造成自动驾驶小镇的城镇。是的，我听说过。很多这些自动驾驶小镇公司，他们买便宜的地产，把它们改造成迷你小镇，这样他们就能创造这些有趣的场景。你去过那种地方吗？我没去过，但我肯定看过视频。是的，很酷。他们有小孩冲出来，就像小孩的纸板剪影。如果它这样撞到，他们可以在私下做。

Oh, really? Yeah, different forms of data that the governing bodies placed in front of them. Well, that would be very interesting. Now think about it: we do crash tests for cars. You're required to give three cars or something to the government for them to just destroy in their crash tests. But we don't require those cars to go into a lab, get taken over by the governing body, and force them to go into real-world testing environments. Because there's some real-world testing environment where you do self-driving up north here. I think some military base. I'm referring to a military base that everybody uses for self-driving. It's like a town that was converted into a self-driving town. Yeah, I've heard of it. A lot of these self-driving town companies, they buy cheap real estate, they outfit them into these mini towns so they can create these funny scenarios. Have you ever been to one of those? I've never been, but I've definitely seen the videos. Yeah, it's pretty cool. They have children come darting out like little cardboard cutouts of children. If it hits it this way, they can do that in private.

是的，这很有趣。在某个时候，政府将不得不让开发者和编码人员真正进入数据，理解其中的一部分，对吧？至少，他们必须创建驾驶考试，例如，自动驾驶汽车的驾照考试。我的意思是，那将会存在。

Yeah, it's interesting. At some point, the government's going to have to have people who are developers and coders actually getting into the data and understanding some portion of this, right? At the very least, they'll have to create the driver's test, for example, the driver license test for a self-driving car. I mean, that will exist.

你相信通用 AI 会在某个时候达到那个水平吗？就是那种普遍聪明、能做任何人类能做的事的 AI？

You believe in general AI that will hit that at some point? AI that is just generally smart, can do anything a human can?

我的意思是，我在某种意义上相信。我相信，对于人类能想到的大多数技术上可行、物理上不不可能的事情，如果人类存活下来，它们会在某个时候发生。我认为人类是无限创造、无限聪明的，等等。当然。我认为人们谈论 AGI 的时间线被严重夸大了。我可以就此吐槽一会儿。常见论点有很多错误。其中之一是，人们说如果摩尔定律继续，我们将拥有所有这些指数级的算力，那将意味着我们生产这些通用 AI 只是时间问题。更不用说量子计算了。是的，没错。所以我的意思是，摩尔定律将会消亡，然后量子计算尽管最近有发布，但还很遥远。所以，我认为……

I mean, I believe in some sense. I believe in the sense that for most technological things that humans can conceive of that aren't physically impossible, if humans survive, they'll happen at some point. I think humans are infinitely creative, infinitely ingenious, etc. Sure. I think the timelines that people are talking about for AGI are very overblown. I could rant about this for a while. There are a lot of things that are wrong about the common arguments. One of which is people say that if Moore's law keeps going, then we'll have all this exponential compute, and that's going to mean it's only a matter of time before we produce these general AIs. Not to mention quantum computing. Yeah, exactly. So I mean, Moore's law is going to be dead, and then quantum computing is very far away despite recent releases, etc. So that, I think...

那部分论点其实没那么有力。而且我也觉得，就算你有无限算力，也不一定能造出通用人工智能。这一点非常不确定。无限算力对狭义人工智能有帮助，因为你可以模拟大量场景，穷举所有可能性。比如围棋，它的排列组合比扑克多得多，也比象棋多得多，象棋的数据集是有限的。所以更多算力确实能让你更快获得能力。或者像 OpenAI 那样，把人扔进随机电子游戏里。当然，这肯定行。但一般来说，让一个象棋大师去下围棋，再去玩《堡垒之夜》，或者去画印象派画作，这是不同的，非常不同。

That leg of the argument is not actually that strong. And I also think it's not even clear that if you have infinite compute, you'll be able to produce general AI. I think that's very unclear. Infinite compute helps narrow AI because you're doing a number of scenarios and playing out every scenario. Go, the game, has many more permutations than poker, many more permutations than chess, which is a finite data set. So more compute power on those things certainly gets you quicker capability. Or even just throwing people into a random video game like OpenAI is doing. Sure, definitely. But generally, taking somebody who's mastered chess and then saying master Go, and then master Fortnite, or master impressionist painting, it's different. It's very different.

所以有一种论点认为，一旦你有足够的算力，你基本上可以模拟进化来创造人工生命。这是其中一个论点吗？

So the argument goes that once you have enough compute, you can basically simulate evolution to create artificial life. Is that one of the arguments?

是的，这是比较时髦的论点之一。

Yes, that's one of the more vogue arguments.

让我看看我理解得对不对：你有这么多算力，可以从一小块细菌开始，让它生长，生长出整个进化系统，直到出现一个类似人类的物种，然后让那个有大脑袋的物种进化成我们之后的东西。

Let me see if I understand: you have so much compute power that you can start with a tiny piece of bacteria, grow it, and grow an entire evolutionary system to the point where there is a human-like species, and then grow that species with a big brain into whatever comes after us.

是的，或者哪怕你只培育出和我们一样聪明的类人物种，那也算是通用人工智能。因为通用人工智能，通常大多数人定义它时，甚至不是说比我们聪明，而是和我们一样聪明。

Yeah, or even if you just grow the human-like species that's as intelligent as us, that would be general AI as well. Because general AI, normally most people define it as not even being smarter than us, but being as smart as us.

没错。所以这是个有趣的方法。但得有人去编码、编程、构建系统来实现它。它不会凭空发生。

Exactly. So that's an interesting approach. But somebody would have to code that and program that and build the systems to do it. It's not just going to magically happen.

对，没错。甚至不清楚这是否可能。但这是一个论点。老实说，这是最合理的论点，但它非常科幻，因为我们甚至无法接近验证这个假设。所以我不相信通用人工智能会很快到来，不管那些专家怎么说。

Right, exactly. It's very unclear if that's even possible. But that's an argument. Honestly, this is the most plausible argument, but it's very much science fiction in the sense that we're not close to being able to even validate the hypothesis. So I don't believe in general AI anytime soon, despite what the pundits will say.

继自动驾驶之后，下一个大多数人还没想到的、令人惊叹的狭义人工智能项目是什么？自动驾驶已经吸引了我们的注意力。

What's the next big mind-blowing AI project, narrow AI project, that most people aren't considering right now, after self-driving which is the one that's captured our attention?

我觉得有一堆很无聊的项目。无聊的那些比如：你能把表单处理自动化得很好吗？比如纸质表单处理。天哪，那太无聊了。超级无聊，但会很庞大。比如我得填表才能拿到驾照，然后你用 AI 来处理。或者回复邮件。Gmail 里现在就有这个功能。你见过吗？

I think there's a bunch of really boring ones. The boring ones are like: can you automate form processing really well? Like paper form processing. Oh my god, that is boring. It's super boring, but it'll be big. Like, I have to fill out a form to get my driver's license, and you know, use AI. Or replying to email. That's the one in Gmail now. Have you seen that?

是的，很棒。它变得这么好，速度快得有点疯狂。而且是个性化的，对吧？

Yeah, it's great. It's pretty demented how fast it's getting good. And it's personalized, right?

我也相信它是个性化的，因为它开始用我的措辞了。所以我想，‘我绝不会用那个’，然后我发现自己，等等，它在用我的语气完成句子。然后你想，等等，我的语气挺狭窄的。我是个人类。所以给你建议回复其实算是唾手可得的东西。

I would believe it's personalized too, because it's starting to use my lingo. So I'm like, 'I would never use that,' and then I find myself like, wait a second, it's finishing the sentence in my voice. And then you're like, wait a second, my voice is pretty narrow. I'm a human. So giving you suggested replies is actually kind of low-hanging fruit.

是的，这算是有点无聊的那种。但我觉得有些应用有相当大的经济影响。比如，所有这些自动化放射学和自动化医学影像工作非常有影响力。核心技术已经足够好，只要有足够的数据，就能真正实现。

Yeah, that's like kind of a boring one. But I think there are applications that have a pretty large scale economic impact. For example, all this automated radiology and automated medical imaging work is very impactful. The core technology is good enough, given enough data, to actually make that possible.

我因为抽烟做过肺部扫描，他们做肺癌筛查。现在他们把那些 X 光片送到印度，由那里的技师审阅。甚至 24 小时心率监测，因为那是最便宜的劳动力，能力却最高。但所有这些数据都可以由计算机完成，比人类做得更好。

I get my lung scan because I was a smoker, and they do lung cancer screening. Now they send those X-rays to India to be reviewed by technicians there. Or even heart rate monitors for 24 hours, because that's the cheapest labor with the highest ability. But all that data can just be done by a computer better than a human could ever do it.

总的来说，全球范围内医生严重短缺。我相信世界卫生组织发布的数据显示，全球医生短缺达 10 倍。所以存在巨大的短缺，如果你能用更可扩展的自动化系统来满足部分需求，那就有巨大的价值。我们美国人很容易认为，不清楚提升在哪里，这似乎只会自动化工作。但那是因为你已经有了稳定的医疗基础设施。不是所有人，但很多人有。在其他地方，他们可能永远看不上医生，或者一年一次。拍 X 光片可能根本不可能，因为成本，不仅是拍片的成本，还有解读片子的成本。

In general, globally there's a huge shortage of doctors. I believe the World Health Organization published something like a 10x shortage in doctors globally. So there is this massive shortage, and if you can fulfill some of this demand with automated systems that are much more scalable, there's a huge amount of value. It's easy for us in the United States to think it's not clear what the lift would be, that it'll just automate jobs. But that's because you already have access to stable healthcare infrastructure. Not everybody, but a lot of people. In other places, they're just never going to have access to a doctor, or maybe once a year. Getting an X-ray might be out of the question because of the cost, not just the cost of the X-ray but the cost of actually interpreting it.

所以你觉得 X 光片是重点？

So you think X-rays are the big one?

嗯，是所有形式的医学影像：X 光、超声波、扫描。你们在做这个吗？

Well, it's all forms of medical imaging: X-rays, ultrasounds, scans. Are you working on that yet?

我们确实在处理大量这类数据。而且我认为 AI 现在可以比人类做得更好，或者它们可以排好队让人类更高效地审阅。很多时候，它们做一部分工作，然后人类可以更精细地处理。比如初筛，甚至在图像中标出问题区域，进行标注，这样医生就可以从第二步开始。

We do work with a bunch of this data. And I think AI can do it better than humans now, or they can queue it up for a human to review more efficiently. A lot of times, they do some of the work, and then the humans can do it more finely. Like a first pass, or even marking out problem zones in an image, annotating it so the doctor can start on second base.

没错。所以我认为很多这类系统会非常有影响力。还有很多其他无聊的东西人们没想到。然后越来越多，市场力量——那些有巨大需求或者人们不喜欢做的事情——会成为最明确的工作方向。我认为教育会是一个大领域：自适应学习，孩子坐在电脑前，电脑从他们沮丧时的面部表情中学习。我知道这听起来很反乌托邦，但如果电脑在观察孩子，而孩子对某个数学题感到沮丧……

Exactly. So I think a lot of those systems will be very impactful. There are a lot of other boring things that people don't think about. And then more and more, the market forces—things with incredible demand or things people don't like doing—will be the clearest things to work on. I think education is going to be a huge one: adaptive learning where kids sit in front of a computer and it learns from their facial expressions when they're frustrated. I know this sounds dystopian, but if the computer was watching the child and the child is frustrated at a certain math problem...

它会让他们退回到一个更简单的数学题，从表情就能看出他们很享受、很自信；当他们不自信时，可以稍微推一把，说‘嘿，我知道你不舒服，让我再带你过一遍’，或者‘我感觉你可能想让我再跟你一起做一遍’。你能想象一个结合了机器学习、自适应学习和 AI 的可汗学院能做什么吗？有没有人在做一个窄 AI 项目来教人们如何学习？我之前没听说过这样的项目，但想象一下，在识字方面，地球上还有很多人不会读写。

that it takes them back 20% to an easier math problem and they could tell from the facial expression that they're enjoying it and that they're feeling confident and then when they feel not confident they can push a little bit into that you're hey I know you're uncomfortable let me walk you through this again or I get the sense that you might want me to work through this again with you can you imagine what a Khan Academy with machine learning and adaptive learning and AI could do yeah is anybody doing anything in a narrow AI project to teach people how to learn I haven't heard of a project before but imagine for literacy there's still lots of people the planet who can't read and write

是的，我认为这是一个明确的应用。我的意思是，当我们谈论医疗保健时，有不同的问题，而教育系统相当破碎，不是经济机会的成熟场所。但像你所说的这样一个系统，真的很容易实现，只是时间问题。我认为它比自动驾驶或类似的挑战更容易。判断某人是否沮丧，基于面部表情或照片，这很容易，已经做到了。但将其与自适应学习技术结合起来，就只是理解什么是难题、什么是简单题，这应该超级容易。但没有人把这些拼在一起。我们不知道，这难道不令人惊讶吗？我打赌它可能已经存在了。我们得找出来。如果有人在听，有一个使用 AI 和面部识别的自适应学习系统，能了解学生的情况。我认为技术真正有趣的地方在于，当你把两样东西结合起来时，比如对于面部识别，每一个反乌托邦的可怕用途，你都能想到 20 个真正惊人的用途。比如，如果你知道有人走在金门大桥上，而且你知道他们沮丧并考虑自杀，你就能确切地知道那个人过桥时有跳桥的可能。

yeah no I think I think it's a clear application there are I mean when we're talking about health care there are separate problems where education systems are like are pretty broken and and are not ripe places for for a lot of economic opportunity but but these sit like a system like you're talking about is is really I mean it gets easy to make right it's only a matter of time yeah I mean I think be easier than self-driving or similar chalice it's very easy so easier than self-driving the challenge of figuring out when someone is frustrated based on their facial based on a photos done is very easy that's done yeah yeah but combining that with some adaptive learning technology well then it's just about it's about understanding what's a hard question what's an easy question yeah that should be super easy yeah nobody's put that glue together isn't it amazing that we don't know I would bet you it probably exists somewhere we got to find that out if somebody on the pot is listening and there's an adaptive learning system using AI and facial recognition to kind of understand where the students at see that's where I think technology it's really interesting when you combine two things like the for every dystopian terrorising thing about facial recognition you can think of there's 20 you could think of that would actually be amazing like if you knew somebody who was walking on the Golden Gate Bridge and you knew they were despondent and considering suicide you could literally know that a person walking across the bridge was doing so with the potential of jumping off the bridge

是的，有很多无聊的机器学习其实很棒，比如 Apple Watch 或很多可穿戴设备。Apple Watch 有一个算法，基本上通过加速度计和你的移动速度等，可以检测到你是否严重摔倒。如果你戴着 Apple Watch 摔倒，它会检测到，如果你在一段时间内没有回应，它就会叫救护车。这真的很不可思议，就像科幻小说一样。这很疯狂，而且今天就存在。太神奇了，买一个 Apple Watch 可以救你的命。但还有很多无聊的机器学习用途，这就是为什么人们应该这样看待它：这是一件疯狂而不可思议的事情，很多人确实这么认为。但火上浇油的是，它实现了以前无法做到的事情，大大增强了我们已有技术的能力。

yeah I mean there's a lot of there's there's boring machine learning that's really great right which is like Apple watches for example or a lot of these like things wearables Apple watches are set up with them with an algorithm that can it basically looks at the accelerometer and how fast you're moving etcetera yeah and it can detect when you have a hard fall yeah so if you fall with an Apple watch it'll it'll detect they have a hard fall and if you don't respond to it with some time period it'll call an ambulance to you actually incredible which is like it's a super now that science fiction yeah it's crazy it's actually really crazy right and that exists today it exists today it's amazing buy an Apple watch it could save your life right yeah but but there's a lot of like boring uses of machine learning that and this is what this is like this is why really it people should view it as like this that's like crazy incredible thing a lot of people do but but adding fuel to the fire it enables all these things that that you couldn't have done before and so it just it enhances the sort of like capability set of the technology that we've built pretty considerably

嗯，20 年后，当你 42 岁时，这一切会怎样？你认为世界会是什么样子？如果让你描述 AI 和机器学习，显然你的公司会成为一家市值万亿美元的上市公司，你会成为地球上最富有的人。但抛开这些，机器学习和 AI 的世界会是什么样？我们早上醒来，AI 会如何与我们互动？这是个好问题。我认为几乎很难想象，因为大想法一开始并不是大想法，它们是慢慢滚雪球，然后最终变成巨大的东西，每个人都认为它改变了世界。所以很难想象这些事情今天如何发生。但最终，我认为梦想中的系统是，很多人喜欢助手的想法，就像一个机器学习助手。但理想情况下，它是一个你可以通过语音、打字或语言交互的系统，你可以提出关于世界的问题或要求，它能理解、推理并得出结果。我们今天构建的很多 AI 和机器学习，很多是核心感知技术，比如理解正在发生的事情，知道那里有一辆车或一张沙发。很多这样的技术将成为基础智能层，为下一层提供动力，然后会有推理层为下一层提供动力，以此类推。

hmm where will all this be in 20 years when you're 42 what do you think the world's gonna look like an AI a machine-learning if you had to describe it obvious your company will be a publicly traded trillion-dollar company putting that aside you'll be the richest man on the planet but putting that aside what will the world a machine learning and AI look like we'll wake up in the morning an AI will interact with us how it's a good question I mean I think it's almost well one thing is it's very hard to conceive of right because it'll sort of be like Big Ideas aren't big ideas from day one they're sort of like slow ideas that snowball and snowball and snowball and then eventually it's sort of like this huge thing that everybody thinks is has kind of changed the world so it's hard to conceive of how these things happen today but ultimately I think the sort of the dream and system is is sort of as a as a I mean a lot of people like this idea of the assistant right like like it like a machine learning assistant but ideally it's it's some system that you can basically you interface with it through through voice or through typing or basically through language you know and you're able to dictate like questions or things that you want on the world etc and it's able to understand that reason through it and then and then understand what the result is and then a lot of the AI that we're building today a lot of the machine we're building today which a lot of it is sort of core perception technology like just understanding what's happening general like knowing that there's a car there or knowing that there's a sofa there or whatnot a lot of that will sort of be a base layer of intelligence that powers the next layer and then there will be a base layer of like reasoning or whatnot that powers the next layer and so on

是的，非常有趣。你相信像脑机接口这样的东西吗？Neuralink 会成功吗？我没有深入的见解。我认识一位早期创始人，我认为他们在做很酷的东西，但我不了解所有事情。问题在于杀手级应用是什么？你实际上想用它做什么？什么是可行的？这些的交集在哪里？比如，不用任何东西就能点餐，你只要想你想要什么，一个汉堡就会出现。清醒饮食加上 Neuralink，意味着你和我互相看着，我想芝士汉堡，你想芝士汉堡，我想培根和蓝纹奶酪，你想切达干酪和火鸡培根，然后 15 分钟后食物就出现了。但这比新式的好很多吗？是的，会很不可思议，因为你根本不用花 60 秒去想点餐或按按钮。当然，差别不大，但会令人震惊。今天，你拿出手机，点三下就能点餐，这已经很令人震惊了。而以前，你可能要花五分钟打电话，看谁还营业。

yeah very interesting you believe in this like brain interface stuff neural link that'll work well I I don't have any I don't have any deep insight on I know one of the early founders I think they're working on cool stuff I don't know all these things the questions like what's the killer app right like what are you actually going to want to use that thing for and what's feasible and where did where it's like the intersection of those things ordering food without anything you just think what you want and a burger shows up sober eats plus Norah link means like you and I be looking at each other and I'd be like cheeseburger and you'd be like cheeseburger and I'd be like bacon and blue cheese and you'd be like cheddar and turkey bacon and then in 15 minutes we'd show up but is that that much better than new breed's yeah be incredible because you would literally have to not spend the 60 seconds to think about ordering it or pressing the button of course it's not it's not much different but it's gonna be kind of mind-blowing it is kind of mind-blowing today that you could just take out your phone and order with like three clicks and get food and it used to be like I don't know five minutes on the phone making for a phone call seeing who's open

是的，顺便说一句，这就是聊天机器人的整个问题。当聊天机器人热潮时，每个人都认为这很棒，容易多了。但如果你想想实际需要点击的次数，用聊天机器人可能要点击 60 次才能完成一件事，而用麦当劳的 App 可能只要三四次就能点到汉堡。

yeah I mean this was the big by the way this was like the whole thing with chatbots right was that when when chatbots I mean there's a chatbot craze everybody thought like oh this is great it's so much easier but then if you think about the actual number of clicks that you have to make if you click like 60 times and get something done with the chat box versus like three or four times to get like your hamburger from from McDonald's or whatnot

互联网之前的世界是什么样的？你不知道。好吧，我可以看书。你七岁的时候应该不记得没有互联网的日子了。那会是 2005 年，你在家应该用的是宽带。你用过拨号调制解调器吗？

What was the world like before the Internet? You don't know. Well, I can read books. You don't remember a time before the internet when you were seven years old. It would have been 2005. You would have been on a broadband connection at home. Did you ever use a dial-up modem?

我确实用过拨号调制解调器。真的吗？那个拨号音，是的。真的，我记得。要知道，我来自新墨西哥州，所以镇上偏远地区宽带不多。

I did use a dial-up modem. Really? The dial-up tone, yeah. Really, I remember that. For context, I'm from New Mexico, so there wasn't a lot of broadband out there in the back country part of town.

你父母还在那里吗？

Are your parents still there?

我父母还在那里。

My parents are still there.

他们还在工作吗？

Are they still working?

我父亲退休了，我妈妈还在工作。哇。是啊，有点像《绝命毒师》里的县，对吧？

My father's retired, my mom's still working. Wow. Yeah, some Breaking Bad county, right?

我现在第一次看《绝命毒师》。哦真的吗？你看到第几季了？第五季。它越来越好。第一季太慢了，我跳过了前两季，但后面变得很疯狂。我刚看了《续命之徒》，就是《绝命毒师》结局后的电影，我真的很喜欢。它绝对能进前十。你看过《黑道家族》吗？那部剧在你出生前就结束了。我想我看过不少。这就是你圣诞假期该做的事。你要休息了。作为 22 岁的年轻人，我建议你喘口气。真的在假期休息一下，狂看《黑道家族》。它是最早同时有多条故事线的剧集之一。虽然没有《权力的游戏》或一些高密度剧集那么多，但电视编剧们意识到，如果让剧集更密集、更难跟上，观众反而会收获更多。而且因为 DVD 的存在，人们可以回去看前几季，在 Netflix 上或者买 DVD。所以制作电视剧的人有动力让剧集更密集，创造更多角色、更多故事线、更复杂，因为这能推动更多 DVD 销售。想想这个系统。DVD 是一种非常流行的格式，利润如此之高，以至于影响了艺术。在那之前，他们要求每集都能独立成篇。如果你从没看过《脱线家族》的一集，这一集你不需要任何背景知识。所以那就像《土拨鼠之日》，你每天醒来都不知道还有另一集《盖里甘的岛》，因为每集都是独立作品。但《黑道家族》和另一部叫《盾牌》的剧集是第一批让剧集非常紧张、角色众多、主题丰富，并且有跨季故事弧的。所以你会看到多个跨季的弧线。当然，《火线》被认为是这类剧集的王者。我只看了五集《火线》就停了，因为我妻子想和我一起看。

I'm actually watching Breaking Bad for the first time now. Oh really? What season are you on? Season 5. It gets better and better. The first season was so slow. I skipped the first two seasons, but yeah, it gets crazy. And I just watched El Camino, the movie that takes place after the end of Breaking Bad, and I actually enjoyed it very much. It's definitely top 10. Have you watched The Sopranos yet? That ended before you were born. I think I've watched a good chunk of it. This is where you do this Christmas break. You're gonna be off. Take a breath is my advice to you as a 22-year-old. Actually take a break over the holidays, watch binge watch The Sopranos. It was one of the first shows that had a lot of plot lines going at once. Not as many as Game of Thrones introduced or some of these really high density ones, but at some point television writers realized if they made it more dense and harder to follow, you would actually get more out of it. And then because DVDs existed, people could go back and watch the previous seasons on Netflix or buy the DVDs. And it was in the best interest of the people making the TV shows to make them more dense, to create more characters, to create more plot lines and make it more complex, because they drove more DVD sales. Think about that as a system. The DVD was such a popular format and was so profitable that it impacted the art. Because before that, they said every episode needs to stand on its own. If you've never watched an episode of The Brady Bunch, this episode of The Brady Bunch you don't need any prior knowledge. So then it was almost like Groundhog Day, you're waking up every day and you don't even know there's ever been another episode of Gilligan's Island, because they're all just singular pieces of work. But The Sopranos and another TV show called The Shield were the first to make them very intense, lots of characters, lots of themes, and then having story arcs that would go over seasons. So you'd have like multiple arcs over multiple seasons. And of course, The Wire is considered the king of the genre. And I've only gotten five episodes into The Wire and I stopped because my wife wants to watch with me.

好了，我们到此为止。我们有最喜欢的新游戏节目。这个游戏节目我们调出你的推文，然后说转发还是屏蔽。你说了算，不过观众也会参与。好了，开始。我们现在调出你的第一条推文。你心里在想，我发了什么推文？这是你的第一条推文。好的：'很多人认为好人和伟人之间是平滑连续的。根据我的经验，好和伟之间有一个巨大的带隙。'你这里说的'带'是指能带，就像电子带？是的，就像电子有这些间隙，它们无法跨越。没错。'伟人聪明、坚定、努力奋斗、在困难中更忠诚、有策略。一个人伟不伟大很明显。相反，一个人只是好也很明显。'所以在你看来，好和伟之间的差距是指数级的？我相信是的。对我来说，这是点赞加转发。哇，我喜欢。你两个都得到了。好了，下一个。这个有人说可能有点冷酷。不，我开玩笑的。'如何按照 Alexander Wang 的方法打造出类拔萃的东西。你可以关注他，他是 Alexander_underscore_Wang。不，只是 A 和 D 是 underscore Wang。1. 打造你在乎的东西。2. 找到用户，倾听他们。3. 每天改进。这是人们遗漏的一点。4. 找到激励你的人，说服他们和你一起工作。啊，是的。5. 重复 2-4 四年。'正确。就这么简单的五步。是的，细节决定成败。每天改进真的很难，对吧？超级难。超级难。人多了就更难了。有人告诉过我，这是二手消息，但他们的功夫老师说过：你每天要么在变好，要么在变差。没有保持不变这回事。所以你的产品要么在变好，要么在变差。因为如果你不改进它，它就在贬值，而竞争对手在改进，你的用户会习惯这个平庸的产品。

Okay, we'll end here. We have our favorite new game show. This is the game show where we pull up your tweets and we say like retweet or block. You're saying it, well the audience is gonna do it too. Alright, here we go. We're gonna pull up your first tweet now. Your query in your mind going what tweets have I done? Here's your first tweet. Alright: 'Many folks believe there's a smooth continuum between good people and great people. In my experience, there's a huge band gap between good and great.' The 'band' you're referring to here is like an energy band, like in electron bands? Yeah, like the electrons that there are these gaps where they won't cross. Exactly. 'Great people are clever, determined, fight hard, more loyal in hardship, and strategic. It's obvious when somebody's great. And the counter to this is, hey, it is obvious as well when they're just good.' So the difference between good and great is exponential in your mind? I believe that, yeah. To me, this is a like and a retweet. Whoa, I like it. You get both. Alright, here we go. Now here's one that people said might have been a little callous. No, I'm joking. 'How to build something insanely great according to Alexander Wang. You could follow him, he's Alexander_underscore_Wang. No, it's just A and D are underscore Wang. 1. Build something you care about. 2. Find users, listen to them. 3. Improve it every single day. This is the one people leave out. 4. Find people who inspire you and convince them to work with you. Ah, yes. 5. Repeat 2-4 for 4 years.' Correct. Just five simple steps. Yeah, the devil's in the details there. It's really hard to improve every single day, isn't it? Super hard. Super hard. You get more people, it gets harder. There's something somebody told me, second hand, but they told me their kung fu teacher told them this: You're either every day getting better or getting worse. There's no staying the same. And so your product is either getting better or getting worse. Because if you are not improving it, it's deprecating, and a competitor is improving, and your users are getting used to this average product.

好了，下一个。'看 YC 校友演示日。如果你觉得：1. 这些公司没有一家比得上我的。2. 哦对不起，不是这样写的。我开玩笑的。就像我把自己卷进了什么？好吧，看：很多想法都是今天热门融资的衍生品：DoorDash、Campus、Checker 等等。很多重复的。175。创业想法很可能已经达到饱和点。超过一个不潮流的想法会成为大公司。'所以你这里说的是，跟风者永远成不了大生意。也许 175 太大了，太多了。我就是这么说的。我觉得……你参加过 YC 吗？我参加过，我们参加了 YC。好的，那你当时去的时候怎么想？有多少人？你对这么大的班级规模怎么看？一百家公司。即使在那时，也有重复的，有人在研究类似的东西。他们在干什么？同一个班级里有人在做竞争性的想法。一个班级在忠诚度之间制造足够的紧张感是一回事。现在你的两个竞争公司都在同一个班级的股东名单上。我觉得他们对此不太公平。我想他们只是放手了。他们给予很多爱，给每个人爱，然后说，去吧，竞争吧。这很有趣。正确的数字是多少？我不知道。嘿，我觉得一百家没问题。我的意思是，这就像你在培育黑天鹅，所以需要多少就多少。这就是人们不理解的地方。像 Y Combinator 这样的系统，或者整个硅谷，在局外人看来似乎很混乱，因为他们看到太多失败、太多衍生想法、这些人似乎不合格、这家公司拿了太多钱。然后呢？

Okay, here we go. 'Watching YC alumni demo days. If you thought: 1. None of these companies are as good as mine. 2. Oh sorry, that's not what it says. I'm joking. It's like what am I got myself into? Okay, watching: many ideas derivative to the hot financings of today: DoorDash, Campus, Checker, etc. Lots of duplicates. 175. Greater saturation point for startup ideas in all likelihood. Greater than one of the unhip ideas will be the big company.' So what you're saying here is being a follower never becomes a big business. And maybe 175 is just too big, it's too many. That's what I say. I think there's... Did you go to YC? I did, we did YC. Okay, so what do you think when you went? How many people were there? And what do you think about this ginormous class size? A hundred companies. Even then, there were duplicates, people working on similar stuff. What are they doing? There are people working on competitive ideas in the same class. It's one thing for a class that creates enough tension amongst the loyalty. And now you've got your on the cap table of two competing companies but in the same class. I think they're less fair about it. I think they just let it go. They give a lot of love, they give everybody love, and they're like go ahead, yeah fight it out. It's interesting. What's the right number? I don't know. Hey, I think a hundred is fine. I mean, again, it's like you're Black Swan farming, so you know, takes as many as it takes. That's what people don't understand. Systems like Y Combinator or just Silicon Valley in general seem broken to a lay person who's not part of the system, because they look at it and say there's too many failures here and too many derivative ideas and these people seem unqualified and this company got too much money. And then what?

他们没有意识到的是，我们节目开头讨论的那种混乱，可以让人们被允许去尝试一些离经叛道的事情，而这些事情最终会以一种无人能预料的方式改变世界——这正是黑天鹅的定义：在亲眼看到之前，你根本无法预见它。

They don't realize is that chaos that we talked about at the beginning of the show can lead to people being given permission to try something outlandish that then in fact changes the world in a way that nobody could have determined, which is the definition of a black swan: you could not have seen it coming until you've seen it.

没错，因为在黑天鹅出现之前，没人相信世界上有除了白天鹅以外的天鹅。

Yeah, because up until the point of the black swan, nobody ever believed there were anything other than white swans.

对，就是这样。多棒的一本书啊，真的很好。你读过《反脆弱》吗？

Yeah, exactly. What a great book. Really good. Did you ever read Antifragile?

读过，那是我最喜欢的。就是建立那些在混乱中反而表现得更好的公司和系统。

Yeah, that's my favorite. Developing companies and systems that do better in chaos.

是啊，想想看，这是个多么了不起的想法。世界变得越来越混乱，而你却做得更好。

Yeah, whoa, what a tremendous idea when you think about it. It's like the world's getting more chaotic and you're doing better.

特朗普，混乱船长。嗯，不，他……《Seem to Leave》真的很棒。不过你是在说特朗普吗？不，不，我没那个意思。我对特朗普没有看法。但我喜欢的是……看来你在推特上关注了他，他在推特上很犀利，也很轻松。

Trump, Captain Chaos. Yeah, no, he's... The Seem to Leave is really good. Although you're saying Trump? No, no, I'm not. No opinion on Trump. But no, what I love about... and seems you follow him on Twitter, he is brutal and easy on Twitter.

是啊，是啊，我们有点疯狂。我觉得这个人……他就像史蒂芬·平克一样绝对……嗯，我就像深粉色，在某些非常愚蠢的词上。太棒了。你在说什么？

Yeah, yeah, we've got wild. He's... I think this person is... He's like Steve Pinker is an absolute... Yeah, I'm like deep pinker somewhere on really stupid words. Brilliant. What are you talking about?

我认为直率是好的，但非黑即白是不好的。所以我不会完全赞同他的推特行为。但我确实认为人们应该能够接受分歧。

I think directness is good, but I think black and whiteness is bad. So I'm not totally going to endorse his Twitter activity. But I do think people need to be okay with disagreement.

你作为一个 22 岁的人说这话很奇怪，因为你们这一代人，他们上大学——这就是你为什么只待了一年——当学校请来他们不同意的人时，他们就抗议。想象一下：你不同意那个人，他们要来了，你和他们的观点截然相反，然后你抗议他们来。所以你可以去听讲座，了解你不同意的人，获得理解——无论是敌人、对立面还是争论的另一方，都会让你因此更丰富——而他们却说：‘不，你是在给他们平台。’什么时候跟人说话就等于给他们平台了？人们会说：‘你和史蒂夫·班农谈话，就是在给他平台。’那到底是什么意思？

And it's so weird that you say that as the 22-year-old, because this whole generation that you're part of, they're literally going to college—why you lasted a year there—and they're protesting when they bring somebody to campus who they don't agree with. Imagine that: you don't agree with the person, they're coming, you're diametrically opposed to their opinion, and then you protest having them there. So you could either go to the lecture and learn about the person you disagree with and gain understanding—either the enemy or the opposing side or the other side of the argument makes you so much richer because of that—and they're like, 'No, you're platforming them.' Since when does talking to somebody mean you're platforming them? People are like, 'You talked to Steve Bannon and you're platforming him.' So what does that even mean?

是啊，我觉得那家伙把特朗普送进了白宫，他运营着 Breitbart。这些事情对世界产生了重大影响。即使你认为他是邪恶的，也不能和他们对话。

Yeah, I think the guy put Trump in office, he ran Breitbart. These things are having a major impact on the world. You can't have a conversation with them even if you think he's evil.

我认为有一个大问题……我觉得这有点像内容过剩的衍生品。基本上，任何人只要有某种信念，就能读到足够多的内容来强化那个信念。比如对这些人来说，他们内心对自己是谁、相信什么等等，有一个非常清晰的画面和高度自信的看法。而这正是——顺便说一句，人类大脑就是这样运作的。人类大脑的运作方式是：‘哦，你有几个数据点，好吧，你必须相信。’部落主义，对吧？因为以前，要获得这些讲述一致叙事的数据点其实很难。但现在，因为内容太多了，你可以全部读到，你当然会形成非常强烈的观点。很难把别人看作有细微差别的人，而不是那些非常单一的人物，这令人难以置信，因为任何人在自己的一生中——无论是 22 年、48 年还是 98 年——只需要看看自己，就能意识到自己在某个问题上改变了多少次想法。

I think there's a big problem when you have... I think this is like a derivative of too much content out there. So basically, people can—anyone who has any belief can basically read enough content that enforces that belief. For all these people, for example, they believe they have internally a very clear picture and a very high confidence perspective on who these people are, what they believe, etc. And that's just how—by the way, this is how human brains are wired. Human brains are wired to be like, 'Oh, you have a couple data points, okay, you have to believe.' Tribalistic, right? Because in before now, it was actually difficult to get these data points that all told consistent narratives, etc. But now, because there's just so much content out there and you can read all of it, you sure develop these very strong opinions. It's hard to think of other people as nuanced human beings, versus these very one-note kind of figures, which is mind-boggling since anybody in their life need only look at their own life—whether it's 22 years on the planet or 48 or 98—and realize how many times they've changed their mind about an issue.

不，你只需要……你最喜欢的冰淇淋一辈子都一样吗？你现在最喜欢的冰淇淋是什么？

No, you need only... Is your favorite ice cream the same for your whole life? What's your favorite ice cream right now?

对我来说，其实一直是咖啡冰淇淋。

For me, it actually has been this coffee ice cream.

你就喜欢？不，不，不，不，是薄荷巧克力片。

You just love? No, no, no, no, mint chocolate chip.

是啊，你看，你很久没吃黄油山核桃了。你需要试一次黄油山核桃。那可能会改变你的一切。

Yeah, see, you haven't had butter pecan in a while. You need to just try that butter pecan one time. That might change everything for you.

嗯，是啊，我不知道。我的意思是，薄荷巧克力片有点贴近内心，贴近我父亲。你去过 Salt & Straw 吗？

Well, yeah, well, I don't know. I mean, mint chocolate chip is kind of close to heart, close to my dad. Have you been to Salt & Straw yet?

Salt & Straw？你吃过他们的薄荷巧克力片吗？他们的薄荷非常新鲜，感觉就像在嚼薄荷叶。

Salt & Straw? You had the mint chocolate chip? Their mint is so fresh, it feels like you're chewing on mint leaves.

是啊，是啊，我们现在就去。大家都去。他下次在摇摆服务上吃 Salted Straw。再见。

Yeah, yeah, we'll go right now. Everybody does. He had Salted Straw next time on the swing service. Bye bye.