The truth about AI, automation, and Andrew Yang - Kevin Surace
发布时间 2020-01-10 07:00:00 来源
This isn't your average business podcast and he's not your average host. This is the James Altature Show. Today on the James Altature Show. I was quoted as saying and have been saying since then your privacy is over. From 1995 on just get used to it.
这不是一般的商业播客，他也不是一般的主持人。这是詹姆斯·阿尔图尔秀。今天在詹姆斯·阿尔图尔秀中，我被引用说过，并且一直以来都在说，你的隐私已经结束了。从1995年开始，就要适应它。
Because everyone, every service that you use is highly insented to figure out everything else about your life. And all times. Now yes, the government always knew for the last 50 years when you bought an airplane ticket and when you flew somewhere. If they really wanted to know the FBI could know in 10 minutes, it was not a big deal. So we granted that ability for a government to figure that out anyway. But now we've got private companies that for the last 20 plus years know everything about you.
因为每个人、每个你使用的服务都高度渴望发现关于你生活的一切。不停地。现在，是的，政府在过去的50年里一直知道你什么时候买了飞机票，何时去旅行了。如果他们真想知道，FBI 只需10分钟就能知道，这并不是什么大问题。所以我们授权政府去发现这一切。但现在，我们有了私人公司，过去20年多的时间里他们知道关于你的一切。
I've got Kevin Sirace with me, one of the I will say Kevin, one of the world's experts on AI and we're going to discuss what to expect and what's going to surprise us in artificial intelligence over the next year. What you should be thinking about in terms of artificial intelligence. And Kevin, my first question is, I want to have to find out more about your background and stuff.
我现在和Kevin Sirace在一起，他是人工智能领域的专家之一。我们将讨论未来人工智能发展的预期和惊喜，以及你在人工智能方面应该思考什么。Kevin，我第一个问题是，我想了解更多关于你的背景和经历。
But it's China just going to destroy us because they have no ethical qualms about taking this to the edge. Well, look, China has some amazing programs around AI. They have figured out long before I think our government that they got a lead in this space. And China has led in other things. Of course, solar panels, they just decided they were going to own it. They do LED lighting. They decided they were going to own it. They do. And when the government puts their head to it in China, they just own it. They're going to own it. Lockstock and burn.
但是，中国要摧毁我们吗？因为他们没有对将这一切推向极限抱有道德顾虑。嗯，看，中国在人工智能领域拥有一些惊人的项目。他们在这个领域称霸的时间比我们政府之前很久就已经理清楚了。而中国在其他方面也领先。当然，太阳能板，他们决定要拥有它。他们做LED 灯具。他们决定要拥有它。他们做到了。当中国政府把头放进去时，他们只会拥有它，毫不保留，毫不犹豫。
But look at genomics, for instance, right? So this is an area where we have ethical issues correctly or incorrectly, but there's also science that needs to be innovated. And once you innovate at science, it's going to have enormous ramifications on the entire universe of healthcare. And China just doesn't give a shit. So they'll do whatever they'll clone 5 million babies and then kill 4,900,000 of them and just keep the smartest. You know, I'm just making this up.
看看基因组，比如说，对吧？这是一个我们正确或不正确地有道德问题的领域，但也需要创新科学。一旦你在科学上创新了，它将对整个医疗保健世界产生巨大的影响。而中国完全不在意。所以他们会做任何事情，比如克隆500万婴儿，然后杀死490万，只留下最聪明的。你知道的，我只是在编造。
Yeah, you're right. Look, I have a view here. Once CRISPR was out, right? Once people understood it. And this was, you know, it's widely published how one does this, right? But once gene editing, for example, was out of the bag, it is a given that some countries or some labs somewhere will do things that we are not going to do in the United States, either ethically, morally, legally, whatever the case is, right? They're going to engineer babies. They're going to mix DNA from different species. They're going to do things just to figure out what happens. They are. In China, they're doing it.
是的，你说得对。听着，我有一种看法。一旦CRISPR出来了，对吧？一旦人们了解了它。这也是广为人知的如何进行这项工作，对吧？但是一旦基因编辑，例如，被揭露，有些国家或某些实验室将做我们在美国不会做的事情，无论是在道德上，法律上还是其他方面。他们将工程化婴儿。他们将混合来自不同物种的DNA。他们将做一些事情，以了解会发生什么。他们正在中国这样做。
Well, yes, China seems willing to do it. I am less worried about China. I mean, China's going to do some interesting things and we're going to learn from that as they already have, right? They've engineered babies. They're very clear about that. But what I worry about are actually really bad actors. They're sort of a bad actor, but in the end, you know, they need our economy and they don't want to blow everything up, right? But you've got people like North Korea, right, and Iran, that are truly bad actors. Now they may have and probably do have the capability in some labs to do some things with CRISPR that are really dangerous, like engineering, you know, a bug that you can never kill or reproduces it in immense rate. I mean, they could do some really, really bad things.
嗯，对，中国似乎愿意这样做。对于中国，我不太担心。我的意思是，中国将做一些有趣的事情，我们将从中学习，就像他们已经做的一样，对吧？他们已经工程化了宝宝了。他们非常明确。但我担心的是那些真正的坏人。他们有点像是坏演员，但最终，你知道的，他们需要我们的经济，他们不想把一切都搞砸，对吧？但是，你有像朝鲜和伊朗这样真正的坏演员。现在他们可能已经，在一些实验室里拥有并且可能会使用CRISPR做一些非常危险的事情，例如工程化一种你永远无法杀死或以巨大的速度繁殖的病毒。我是说他们可以做一些非常非常糟糕的事情。
And so let alone just human genomics, right? So I'd be more concerned about the true bad actors, including Al Qaeda, ISIS, blah, blah, blah, as these technologies become more available. Look, when you look at nuclear weapons, we kind of us in just a few countries kind of kept nuclear weapons under wraps as a secret, you know, in terms of how to really make them for more than 50 years. That's a hell of a problem. That's because the resources are rare too. Like uranium, you know, that kind of uranium is rare. It's rare, but still, you know, it turned out it was hard to make one as a third world country until recently, until the last decade or two, and enough of the secrets got out.
所以，不要说人类基因组学了，对吧？所以我更担心真正的坏人，包括基地组织、伊斯兰国等等，随着这些技术的普及变得更加危险。你看，当你看到核武器时，我们几个国家中的大部分都将核武器保密了超过50年，以至于如何制造它们是个大问题。这是因为资源也很稀有，比如铀，你知道，那种铀是罕见的。它很稀少，但是直到最近的十年或两三十年前，足够的秘密被泄露出去之前，第三世界国家很难制造出它们。
In other words, you can't keep anything a secret forever. That's a, right, that's the learning here. And you can't keep CRISPR a secret. It's been around for a decade. People know how to do it. Every lab's experimenting with it. They're doing gene editing. And we're going to, you know, the world is going to make some bad things, including some really gross, crazy, they, and look, and China and others are experimenting with, you know, could we make super humans? You know, what genes, yes, at it to make super humans.
换言之，你不能永远保守任何秘密。这是一个教训。并且你不能保守CRISPR的秘密。它已经存在了十年。人们知道如何使用它。每个实验室都在尝试基因编辑。我们将会看到世界上出现一些坏事情，包括一些真正恶心和疯狂的事情。中国和其他国家正在进行实验，看看我们能否制造超人？你知道，用哪些基因来制造超人。
And, and, and, you know, what's going to happen to the gene pool when you do that? You know, in the past, when we played these games, really bad things happened. We played it with animals, you know, bad things happen. You get a super animal, but it dies in a week, right? Something's wrong. But, but here's, here's where AI comes in.
“而且，你知道吗，当你这样做时基因池会发生什么？过去我们玩这些游戏时，真的发生了很糟糕的事情。我们试着玩动物，你知道的，糟糕的事情发生了。你得到了一个超级动物，但它在一周内就死了，对吧？有问题啊。但是，这就是AI的作用所在。”
So you have, you know, mapping the human genome, and we understand, you know, single gene mutations and how CRISPR can, you know, or enhance CRISPR can, can solve these single gene diseases, like tastes, X disease, and so on.
你们已经开展了人类基因组的绘制，我们也了解单基因突变以及CRISPR如何可以解决这些单基因病症，例如口味、X型疾病等。当然，未来CRISPR可能会被进一步改进，从而能更好地解决这些疾病。
But with deep learning and AI, you're going to be able to map all the permutations, multiple gene, you know, mutations, which ones cause intelligence, which ones cause all these weird diseases, whatever. And that's where AI kind of comes in, and again, China has no qualms about research at all.
通过深度学习和人工智能，你就能够映射出所有排列组合，多基因、多突变，知道哪些导致智力，哪些导致所有这些奇怪疾病，总之，这就是人工智能发挥作用的领域，而中国对研究毫不犹豫。
Look, it's a big data problem, and any big data problem provided it's not noisy. And I think DNA is not, I think the genes are not. I think, I think there's some noise, but I think it's very consistent, at least the human gene pool. And, and the human genome.
看，这是一个大数据问题，只要它不嘈杂。我认为DNA不是，基因也不是。虽然我觉得有一些噪声，但至少人类基因库相当一致。人类基因组也是如此。
And I think that people are already using AI to figure out more and more, you know, what buttons do we push to get a super intelligent, very strong tall human, right? Or, or, you know, the perfect looking human, whatever that means today, right? And people say, I need that one. Yeah, yeah, we all need that. Or human that lives 200 years. Right? Could we do that? Could we turn off the, you know, the telemere thing? Well, maybe. I don't know.
我认为人们已经开始使用人工智能了解更多，知道我们应该按哪些按钮才能得到一个超级智能、非常强壮的高个子人，或者提高外表完美的人，无论现在的标准是什么，对吧？人们会说，我需要那个。是的，我们都需要这样的人。或者能活200年的人，对吧？我们能做到吗？我们能关闭或控制端粒吗？嗯，也许吧。我不知道。
And is that good? Should humans, should humans live 200 years? Well, here's what's going to happen. I'll give you one prediction. Like every other technology, what's going to happen is the really, really good stuff that is useful to you will be available to those with money. And it's going to further separate those with money from those who don't have it. Because, okay, right? They will get access to things that they shouldn't have access to.
这个好吗？人类应该活200年吗？其实，有一件事情可以预测。就像其他的技术一样，好得真的非常非常实用的东西只有那些有钱人才能用到。这将进一步拉开那些有钱人和没有钱人之间的差距。因为，你知道的，他们会获得他们不应该获得的东西的使用权。
Now, I sort of agree. Yvall Horari, who wrote sapiens and 21-polysis of 21st century, he feels the same way. But my feeling is, you know, it's just like computers. When you have an industry that grows exponentially, all right, yes, rich people had a supercomputer for a year or two years. But then they became smartphones three years later. Technology improved so fast. And then everybody gets it.
现在，我有点同意。Yvall Horari写了《人类简史》和《21世纪的21种政治形态》，他也有这种感觉。但是我认为，你知道，就像计算机一样。当一个行业呈指数级增长时，富人的确拥有了超级计算机一两年。但是三年后它们就成为智能手机了。技术进步得太快了。接着每个人都能拥有它。
Yeah. Look, that's certainly been true with the rapid adoption of technology, which doesn't take decades anymore. Or even the old rule was 11 years, right? It takes about 11 minutes now. However, things like CRISPR that have to be done very specific gene editing, they have to be done in a specific kind of lab, while the cost may come down, just like, you know, gene sequencing, the cost has come down, it's heading to zero. You know, there's just certain things that will be available to, you know, to those who have wealth, I suspect, for a period of time.
嗯。看，随着科技的迅速普及，这一点确实是正确的，不再需要数十年。或者说，以前的规则是11年，对吧？现在只需要大约11分钟了。但是，像CRISPR这样必须进行特定基因编辑的事情，必须在特定类型的实验室中进行，虽然成本可能会降低，就像基因测序那样，成本已经下降，即将趋近于零。你知道的，有些东西将只会有那些有财富的人拥有，我猜这段时间会持续一段时间。
Because it involves a lot of steps. And it may involve, you know, labs overseas, can't do it in the US, blah, blah, blah, right? And it's interesting to see how it plays out. We're going to learn a lot from the experiments in China. I have a feeling not all will be kept under wraps, right? And by the way, we can go way back in history, right? You know, I mean, back even to Germany, you know, Nazi Germany that was doing, it wasn't at this level because they didn't have CRISPR. But of course they were trying to make super humans through very, very, very simple means. But nevertheless, they, you know, they had multiple experiments doing it. And this has been going on for, you know, almost a hundred years.
因为这包含了很多步骤，还可能涉及到在美国无法进行的海外实验等等。很有趣的是看它是如何实施的，我们从中国的实验中会学到很多，我有预感其中并不会全部保密，对吧？顺便说一下，我们可以追溯到很久以前的历史，比如德国，就连纳粹德国也试图通过非常非常简单的方法创造出超级人类，当然没有CRISPR那么高级，但是他们进行了多次实验。这种实验已经持续了将近一百年。
So you know, the other, not issue with AI, but again, people, I sometimes people bring up privacy too much, but definitely with advanced facial recognition, the whole privacy thing is just going to be a matter of policy rather than a matter of technology. You know, not only government policy, but Facebook's policy, Google's policy, and so on. And China, having no qualms, they're going to be like the best spies in the world. They're going to know where everybody is at every moment. Absolutely.
你知道吗，除了AI之外，另一个问题就是人，有时人们会过分关注隐私，但随着先进的面部识别技术，整个隐私问题只是一个政策问题，而不是技术问题。不仅包括政府政策，还有Facebook的政策、Google的政策等等。而中国则毫不犹豫地成为全世界最好的间谍。他们会知道每个人在任何时刻的位置。非常确定。
Look, in 1995, and you know, I speak around the world a lot on AI and it's in pick on society. It's up. But from 1995 on, basically when the web browser became popular, I was quoted as saying, and have been saying since then, your privacy is over. From 1995 on, just get used to it because everyone, every service that you use is highly insented to figure out everything else about your life at all times.
看啊，1995年，你知道啊，我在世界各地很多时候都在讲AI和它对社会的影响。它是在增长的。但从1995年开始，基本上浏览器越来越流行后，就引用了我的话，并一直在说，你的隐私已经不复存在了。从1995年开始，要习惯吧，因为你使用的每个服务都高度渴望了解你的生活的方方面面。
Now yes, the government always knew for the last 50 years when you bought an airplane ticket and when you flew somewhere. If they really wanted to know the FBI could know in 10 minutes, it was not a big deal. So we granted that ability for a government to figure that out anyway. But now we've got private companies that for the last 20 plus years know everything about you.
现在，政府在过去50年中一直知道你什么时候购买了飞机票和什么时候飞往某个地方。如果他们真的想知道，FBI在10分钟内就可以知道，这不是什么大问题。所以，我们授权政府具备这种能力。但现在，私营公司在过去20多年中了解你的一切。
And I'll give you an example. And I think you know this. But maybe not all your listeners do. Try this at home sometime if you're married. You got to be married to do it. So the person you're married to, have them start to look at something, you know, trips to South Africa. Just go over there and look at trips to South Africa and look at a bunch of tourist companies and still a bunch of stuff on that. I can guarantee you that within an hour and maybe within 10 minutes on Facebook, you not being that person over there, but related to that person start to see ads for trips to South Africa.
我会举个例子给你听。我觉得你知道这个，但是也许不是所有听众都知道。如果你结婚了，可以试着做一下。那么，你的配偶可以开始看一些东西，比如去南非旅行。可以看看旅游公司，以及其他一些相关信息。我敢保证，在一个小时甚至10分钟内，你就能在Facebook上看到关于去南非旅游的广告，这些广告不是针对你本人，而是针对与你配偶相关的广告。
I can guarantee it because I've seen it over and over and over again. I believe you now. Here's a question. And maybe I'm just, I forgot what it's called. Some kind of bias. I'm just like what's called the Honda effect. Once I buy a Honda, I start seeing Honda's. Sure. I feel like I'm talking about trips to South Africa with my wife and then I start to see the ads. How likely is that already happening?
我能保证这个，因为我已经一遍又一遍地看到了它。我现在相信你了。我有一个问题，也许我只是忘记了它叫什么。一种偏见。就像本田效应一样。一旦我买了一辆本田，我就开始看到本田了。当然。我感觉我在和我妻子谈论去南非旅行，然后我开始看到广告。这种情况发生的可能性有多大？
It's probably not happening. It is because talking on your phone is, you know, it can be monitored, but it's really not being monitored. There's maybe monitored by the government, but it's not being monitored by anyone that is going to do anything to feed you ads.
可能不会发生。这是因为在电话上讲话，你知道的，它是可以被监控的，但它实际上并没有被监控。也许政府有监控，但并没有任何人会通过它来为你提供广告。
Now, yes, Alexa listens, but no, they're not feeding that to some massive supercomputer that's analyzing the words and trying to feed you ads at this time. Could they? Sure. But they're not. That's the thing.
现在，是的，Alexa在听，但是不，他们并没有将这些信息馈送到某个正在分析语言并试图向你推送广告的巨型超级计算机。他们能这样做吗？当然可以。但是他们没有这样做。这就是问题所在。
The technology's there. It would be sort of trivial with the technology we have. Oh, but just turn it on. I mean, you know, so you may not know, but I invented long before there was Siri. I invented the first personal assistant, digital assistant. Her name was Mary. The project was called Portico.
这项技术已经存在了。我们可以轻松地利用我们现在拥有的技术实现它。只需打开它就可以了。我的意思是，你可能不知道，但在Siri诞生之前，我就已经发明了第一个个人助手，数字助手了。她的名字叫玛丽。这个项目叫作门廊。
It was for General Magic. Actually, I'll, yeah, you know, became owned by General Motors for on-stair virtual advisor. But the first virtual assistant, I have all the original pan and all those pans got licensed by Apple and others for Siri and other programs many years later. But we had teams of linguists that had to listen to what you were saying, not to sell you ads, but to make the service better because they would codify what wasn't getting caught by the voice recognition, but the speech recognition, right? That's how you did it.
这是为General Magic设计的。实际上，我会被General Motors收购，用作电梯虚拟助手。但第一个虚拟助手，我拥有所有的原始设计，并且后来很多年被苹果和其他公司使用了，比如Siri等等。但我们有一支由语言学家组成的团队，他们必须倾听你所说的话，不是为了卖广告，而是为了改进服务，因为他们会将没有被语音识别器捕捉到的内容编码，这就是如何做到的。
Now, about a year ago, this all came out about Google, about Apple and about Amazon having people, banks of people in rooms listening to what you said. And I just laughed and said, I invented that method. I hired a linguist. I hired actual linguists to listen to you. And then codify. We gave them a language they could code in. So you would say, we would expect you to say, read my email.
大约一年前，有关谷歌、苹果和亚马逊有一众人员坐在房间里听取您所说的话的消息传出。我当时只是笑了笑，并且表示我发明了这种方法。我曾经请来了一些语言学家来听取您的讲话，然后编码它们。我们为他们提供了一种语言，让他们可以进行编码。因此，您会说，“读我的邮件”，而我们会期望您这样说。
And someone would say, get me my email. Oh, well, we got to code in, get me my email. I don't know what get me my email is, right? We have to tell the system what it is. That's still done today, right? Because you got to have some natural language understanding. And sometimes the system, it got the recognition, right? But it doesn't know what to do with that sentence, right? Because it's kind of slaying. Give me my email. Where's my email? Well, I think all of those mean read my email, right? So we got to codify that.
有人会说，“给我看一下我的电子邮件。”嗯，我们需要编程告诉系统“给我看一下我的电子邮件”是什么意思。直到今天，这仍然是必须要做的，因为我们需要一些自然语言理解。有时候系统可以识别出语句，但是不知道该怎么处理，因为有些用语很口语化，例如“给我看一下我的电子邮件”、“我的邮件在哪？”等，其实全部都是想要读取电子邮件。所以我们需要将这些语句编码化。
So still there's banks of people in rooms listening to what you say. Turns out they don't really care. They're just doing their job. But again, your privacy was gone in 1995. The day the web browser came out and became popular, it was over. It had to be over. Yeah. And I'm okay with that.
所以，仍然有一些人在房间里密切聆听您的话。事实证明，他们并不是真的关心，只是在履行自己的职责。但是，再一次，您的隐私在1995年就已经不复存在。网页浏览器的诞生和普及，就是这个结局。这种情况不得不发生。是的，我可以接受这一点。
So it seems like, you know, it's interesting because you talk about 1995 and you can even go back earlier with, you know, the real beginnings of voice recognition. It feels like until recently, there hasn't really been that many huge innovations in AI. Like I feel, you know, deep mind, you know, Google with their deep mind and AlphaGo program. That was for me, it seemed like the first real innovation in a long time in taking big data and converting it into actionable activity using AI.
看起来你说的很有趣，因为你谈到了1995年，甚至可以回顾语音识别的真正起源。直到最近，似乎还没有出现太多的人工智能创新。我感到像DeepMind和Google的AlphaGo程序那样，将大数据转化为可操作活动的方式，是很长时间以来的第一个真正的创新。
Yeah. Yeah. You're right. I think that I look at AI as augment intelligence. And so if you say AI as artificial intelligence is more or less a marketing term and you say, what we're really talking about is machine learning from large data sets and finding things that would be hard for a human to grok because the data sets so large.
嗯。嗯。你说得对。我认为我把AI看作增强智能。如果你说AI作为人工智能更多地是一种营销术语，你说我们真正谈论的是从大型数据集中进行机器学习，并找到人类难以理解的东西，因为数据集太大。
Well then, and it can keep learning is more data comes in. And you can apply those learnings to the new data. And so an example obviously is facial recognition at Facebook. You know, the first time they turned that on. But then they were the obvious people to do it because they had billions and billions of pictures of faces and names attached to those faces, mostly correct.
那么，如果有更多的数据进来，它可以继续学习。您可以将这些学习应用于新数据。所以一个明显的例子是Facebook的面部识别。第一次他们启用它时，但是他们是明显的人来做它，因为他们有数十亿张脸和那些脸附加的正确大部分的名字。
So you didn't have a very noisy database, but you had a huge database. And over some period of time, they could build a neural net that recognizes your face versus my face. Absolutely pretty much 100% of the time, better than 95% recognition. That's amazing actually when you think about it, right? Huge data problem, such a huge data problem that a human could not look at billions of faces instead of calling out names. It's an impossible problem for our brain.
所以你的数据库不是很吵闹，但很大。而且在某个时间段内，他们可以构建一个神经网络，识别你的脸和我的脸。绝对可以达到几乎100%的识别率，比95%的识别率好多了。想想看，这真是令人惊叹的成就啊！这是一个巨大的数据问题，一个如此巨大的数据问题，即使我们人类看上亿个脸也无法叫出名字。这是我们大脑无法解决的难题。
So we look at that and say, well that seems artificially intelligent. Well no, it's a very, very, very, very good data match. Right? Very, very deep neural net. And it's going to make great decisions based on every shadow in your face, in your classes, in the hair and sort of everything else. But in the end, guys, it's just math. It's just math. All we're doing is trying to find the highest scoring thing that we can match to.
所以我们看到这个，会说，嗯，那看起来很人工智能。但实际上，这只是非常非常非常非常精准的数据匹配。对吧？这是一个非常非常深层的神经网络。它会根据你脸上、眼镜上、头发上的每一个阴影以及其他一切来做出非常好的决策。但最终，各位，这只不过是数学。只不过是数学而已。我们所做的只是尽力找到最高得分的匹配物。
And by the way, if you take that same recognizer that recognizes faces and put a chair in there, it'll say, Jim, it doesn't know a chair, right? Didn't learn a chair. It learned faces. It only knows faces. So this isn't artificially intelligent. We're getting into now where artificial intelligence is going, right? It isn't artificially intelligent like we see in the movies, like X-Mac and I, which is an amazing film or hers, an amazing film. They're so far away. An AI system having a general understanding.
顺便说一句，如果你把识别脸部的那个程序放到一个椅子旁边，它会说：“Jim，它不认识椅子，对吧？没有学过椅子，它只学过脸。”所以这不是人工智能。我们现在正在探讨的是人工智能的未来。它不像我们在电影中看到的那样，比如《X战警》和《她》这些惊人的电影。它们还相差甚远。AI系统需要有一个普遍的理解才能算得上是人工智能。
I'll give you a minute. Yeah, go ahead.
我会给你一分钟的时间。好的，你可以开始了。
Well, no, related to that is I agree with you on the facial recognition. It's similar to voice recognition. But then you have something like a game like, like Go or chess where the computer looked at, you know, a million games and a figured out the rules. And within a few more hours was already the world champion of chess. And then there's kind of a hidden layer. So there's no hidden layer on the facial recognition. We're using a very standard 30 year old statistical technique and, you know, matching faces to what, to a data set we know. But with Go and chess, it's sort of doing several things.
嗯，不，跟那个相关的是，我同意你对于面部识别的观点。它和语音识别很相似。但是你有像围棋或国际象棋这样的游戏，电脑会分析成百上千场比赛，并找出规则。再经过几个小时，就已经成为了国际象棋世界冠军。然后有一种隐藏层。所以在面部识别上没有隐藏层。我们使用的是一个非常标准的30年老技术，将面孔与我们所知道的数据集进行匹配。但是在围棋和国际象棋中，有一些不同的事情正在发生。
It's kind of figuring out the metrics that are important. Like it doesn't know in advance what metrics are important. It sort of figures it out. It's a reward system, right? It's a reward system. But think about it this way. I give some examples in my talks of a reward system that the people at Unity shared with me, some games that they built to show how giving it no rules other than to win, it will figure out how to play the game better.
有点是要找出重要的度量标准。就像它事先不知道哪些指标重要一样，得慢慢摸索。它是一种奖励系统，是个奖励系统。不过，可以这样想，我在演讲中举了些Unity的工作人员分享给我的奖励系统的例子，他们开发了一些游戏，展示了在没有其他规则约束的情况下，这种奖励系统会逐渐发现并完善游戏规则。
And it does this by watching, either watching many games or playing many games, depending on if it's computer game or not or an offline game, right? But either way, think of it this way. If you tomorrow could play, you know, Go or Chester checkers, it doesn't really matter. Let's say it's Chess, but you could play Chess a million times in an hour. And remember, every move that took you further to winning and every move that took you further away from winning and what all the other circumstance is, if your brain could do that within an hour you'd be an amazing player because you could explore the outcome of those million things.
它通过观察比赛或者玩很多游戏来实现，这取决于是电脑游戏还是离线游戏，对吧？但无论如何，就这样想吧。如果明天你能玩围棋或者黑白棋（国际跳棋），都无所谓。假设是象棋，你可以在一个小时内玩上百万次。还要记住，每一步都让你更接近胜利或者更远离胜利，还有其他所有情况。如果你的大脑能在一个小时内做到这一点，你就是一个了不起的玩家，因为你可以探索这些百万种情况的结果。
And if you could play a game on a computer like a Atari game that we used to play or whatever, you know, a simple game that just, you know, you have to, you know, it's Pac-Man or you have to eat the thing. You could very quickly outrun any human in that game simply by playing it about 10,000 times. And a computer could play that in parallel 10,000 times, maybe in 10 minutes or an hour or a few hours. And this is a reward based system.
如果你能在电脑上玩一个像我们过去玩的Atari游戏或其他简单的游戏，你知道，一个简单的游戏，你必须吃东西，就像吃豆人一样。如果你能玩这个游戏大约10,000次，你就会很快赢过任何人。而且电脑可以同时玩10,000次游戏，可能只需要10分钟、一个小时或几个小时。这是一个基于奖励的系统。
And these are very simple because all you're doing is giving the darn thing a reward when it wins and giving it no reward when it loses and it doesn't want to lose. That's all it knows. So it learns how to win. No different, I'll give you an example of the mouse. We all know about the mouse in the lab and you put him in the maze, right? And first you ring the bell and the mouse just sits there going, you're an idiot, right? But over months and months and months of trial and error, it figures out that when you ring the bell there's food in the upper right corner and it runs the maze, it goes and gets the food, it comes back so the bell can ring again. It learns this.
这些都很简单，因为你所做的就是在它赢了时给它一个奖励，而在它输了并且不想输的时候不给它奖励。这是它所知道的全部。所以它学会了如何赢。我来给你举个例子：老鼠。我们都知道在实验室里把老鼠放在迷宫里，对吧？一开始你敲一下铃铛，老鼠只是坐在那里觉得你是个傻瓜。但在数个月的试错中，它逐渐弄明白了：每当你敲响铃铛时，上右角就有食物，于是它跑迷宫，拿到食物，回来让铃铛敲响，再一次拿到食物。它学会了这一点。
Now what's fascinating about that is it tried 10,000 times but after a while it figured out exactly what you're out to take to get the food every time. Now put a second mouse in there.
现在有趣的是，它尝试了10,000次，但在一段时间后它完全知道了每次获取食物的所需动作。现在把第二只老鼠放进去。
The second mouse watches that first mouse every time the bell rings goes and gets the food and thinks the first mouse is artificially intelligent because clearly it's got a much bigger brain than the second mouse that can't, don't even know what the bell means, let alone where to find the food.
第二只老鼠每次听到铃声响起就看着第一只老鼠去拿食物，心里想它太聪明了，因为它的大脑显然比第二只老鼠更发达，连铃声的意义和找到食物的方法都不知道。
We are the second mouse. The first mouse is no smarter than we are, right? In fact, it's totally dumb, it just had thousands of tries at it and it could remember every try. Well, right now though, particularly with these game examples, there's three mice, there's the initial data set, okay, there's humans looking with wonder and there's now the ability to scale the data set because the AI will play itself to create more data.
我们是第二只老鼠。第一只老鼠并不比我们更聪明，对吧？事实上，它非常笨，只是尝试了成千上万次，并且能够记住每一次尝试。然而，现在特别是对于这些游戏示例来说，有三只老鼠，有初始数据集，还有人类惊奇地注视，并且有现在能够扩大数据集的能力，因为AI会自己玩游戏以创建更多数据。
With facial recognition, I can't create more variations. Actually, this is an interesting thing. I can take a picture of your face and now with AI, I don't have to see other pictures of you, I can create other configurations that are probably your face and add to my data set. So, I think AI is also being applied to its own data to generate more data set to learn from. Absolutely.
有了面部识别技术，我不能创建更多的变化。这实际上是很有趣的事情。我可以拍摄你的脸部照片，现在通过人工智能，我不需要看到你的其他照片，我可以创建其他可能是你脸部的配置，并添加到我的数据集中。因此，我认为人工智能也被应用于其自身的数据中，以生成更多的数据集以供学习。完全正确。
In games, that's true. It can be done in facial recognition. Everybody's seen what's happening in deep fakes, I think a great use of this technology, the whole plan it is seeing is the Netflix film, right? Did you see the Irishman? Oh, no, I didn't see it, but they reverse aging, the D aging. The D aging.
在游戏中是这样的，可以在面部识别中实现。每个人都看过深度伪造的情况，我认为这项技术的一个很好的用途就是Netflix电影的整个计划，对吧？你看过《爱尔兰人》吗？噢，我没看过，但他们进行了反衰老，即去老化。去老化。
Now, it's a little freaky because you go, that isn't exactly what they looked like when they were 20 because I knew what they looked like when they were 20 and Scorsese, he wanted them to look the way he wanted them to look, right? The fact of the matter is, aside from the fact they walk like old men, no matter what you do, it is fascinating how good that deep fake technology has gotten. Yeah, particularly they make the clasping back disappear also on the film. Yeah, it's a fact.
现在，这有点可怕，因为你会想，当他们20岁的时候，他们并不是长得那副样子，因为我知道他们20岁的样子，而斯科塞斯想要他们看起来像他想要的那样。实际情况是，除了他们走路像老人之外，无论你做什么，深度伪造技术的进步令人着迷。特别是他们在电影中还使剪切背景消失了。这是事实。
You can take them right out, get them onto the wheelchair. No, it is fascinating what we're doing. And of course, that can be done in a number of AI methods today. Gans, of course, General Iversary on that work, great way to basically now make up people that didn't exist at all simply by comparing the forgery to real things and saying, I still think it's a forgery. And you keep getting better and better. Again, that's a reward system, right?
你可以把他们直接取出来，把他们放到轮椅上。不，我们正在做的事情非常有趣。而当然，今天可以用许多AI方法来做到这一点。如Gans，当然，对那项工作的一般评估，是通过比较伪造品与真实物品来说，我仍然认为它是伪造品，从而基本上制作出了根本不存在的人。并且你一直变得越来越好。再说一遍，那是一个奖励系统，对吧？
The system just wants to get better and get a better reward until they fooled again, right? So it's the inspector sort of. So it's really fascinating what we're able to do. That said, most of what we just talked about is fun and fancy, it's interesting for games, didn't change too many people's lives, didn't change lives.
这个系统只是想变得更好，获得更好的奖励，直到它们再次被欺骗，对吗？所以它就像检查员一样。我们能够做什么真是很有趣。话虽如此，我们刚刚谈的大部分是有趣和花哨的，它们对游戏很有意思，但并没有改变太多人的生活，也没有改变人们的生活。
So what do you think, what's the next thing coming up that either you're worried about or that will change people's lives? So look, a lot is and some people are really worried about AI taking over the world and you know, they're going to be our overlords, etc, etc. I don't see that probably in our lifetime.
那么，你认为接下来会发生什么事情，是你担心的还是会改变人们生活的？看，很多事情都在发生，有些人真的很担心人工智能开始掌控世界，说什么它们会成为我们的主人等等。但我认为这种情况可能在我们的有生之年内并不会发生。
The reason is, as always, with these technologies. Remember, in AI in the 60s and 70s, we made huge leaps and then we made none. And then in the 90s, we made some leaps and then we made none. And then finally around 2012, we got neural nets to work, deeper neural nets, right?
原因就像往常一样，是这些技术的问题。记住，在60年代和70年代的人工智能领域里，我们取得了巨大的进步，然而之后就没有再取得什么进展了。随后在90年代，我们又有了一些进步，但之后又停滞不前了。最终，大约在2012年左右，我们让神经网络开始工作了，深层神经网络，对吧？
And all of a sudden that math was out of the bag and then we made big leaps again and then it flattened out. And what happens is in all these technologies is we far overestimate the short term impact. But we underestimate the long term impact.
突然间，数学问题就解决了，我们又取得了很大的进步。但渐渐地，这种进步趋于平稳。在所有这些技术中，我们总是高估了短期的影响，低估了长期的影响。
So again with AI, we have two, two, three years ago when people were saying, AI is about take over the world, neural nets are the thing, you know, it's over. We overestimated the impact within a three year timeframe because the impact actually wasn't that much. I mean, it's interesting to you and I technically and some of your listeners technically that, yeah, we can blow away Go now. That's amazing. So it's like, it didn't change anyone's life, right? That's your professional Go player. All right.
所以再谈AI，两三年前人们说AI将接管世界，神经网络是未来，你知道，那就完了。我们高估了它在三年时间内的影响，因为实际上它并没有那么大的影响。我是说，从技术角度来看，你和我和你的一些听众可能会觉得很有趣，我们现在已经能够打败围棋了，这很令人惊讶。但是这并没有改变任何人的生活，对吧？那只是你职业的围棋棋手而已。
So, but let's take the one game that they haven't beat. Like how would you go about? Because I get it, like a Go board or a chess board is just essentially a vector of attributes and boom. But how would you take a game like poker that has a lot of hidden information and there's this human component? There's no world, there's no computer that is a world champion poker player.
那么，让我们来看看他们还没有赢过的那一局。你会怎么做呢？因为我明白，围棋或国际象棋棋盘本质上只是属性的向量，轻轻松松就能解决。但是，对于像扑克这样有很多隐藏信息和人的因素的游戏，你会怎样？没有一个世界，也没有一台计算机是世界冠军扑克选手。
Well, I think that's what they grasp. Yeah, that's a complicated issue. So it depends on who you're playing, right? A real poker player. Look at what poker is. Poker is, did you get the right cards?
嗯，我认为这就是他们掌握的。是的，这是一个复杂的问题。所以这取决于你跟谁玩，对吧？一个真正的扑克玩家。看看扑克是什么吧。扑克就是，你是否得到了正确的牌？
First of all, I mean, that's just luck of the draw and there is luck of the draw. And that's just part of the game. That's different than Go and that's different than chess is different than the checkers, right? It's luck of the draw. So that's the first problem.
首先，我的意思是，那只是抽签的运气问题，而且抽签是有运气的。那就是比赛的一部分。这与围棋、象棋、跳棋不同，对吧？这是抽签的运气问题。所以这就是第一个问题。
The second problem is the rest of it is human emotion inference and people learning to read people's faces and people learning how to hide that. Then we teach a computer to read your face and over time guess whether you're bluffing or not.
第二个问题是其中剩余部份是关于人类情感推理以及人们学习看懂他人的面部表情，以及人们学习如何隐藏这些信息。接着，我们教导计算机去读取你的脸庞，并逐渐猜测你是否在虚张声势。
So the answer is absolutely yes, if you kept the same people around the table and they played 500 games. That's enough data to read all the faces to ultimately then figure out who did win, who was bluffing, who wasn't bluffing, right? Totally doable.
所以答案绝对是肯定的，如果您让同样的人坐在桌子旁玩500局游戏。这足够的数据就可以读取每个人的面部表情，并最终弄清谁赢了，谁在虚张声势，谁没有虚张声势，对吧？完全可行。
But if you changed up the people and you got someone who did not give the same facial expressions and maybe didn't give any hint at all, who knows their hints are different, right? Maybe they get fussy, maybe they drink more, maybe whatever.
但是，如果你换了人，遇到一个不给出相同面部表情甚至完全没有任何提示的人，谁知道他们的提示可能是不同的，对吧？也许他们会发脾气，也许他们会喝更多酒，也许什么都有可能。
Now once you change the person, we may not be able to win the game again until we watch that person play hundreds of times, then we can win the game. Unless the AI discovers there's subtle micro expressions that can't be controlled. It could be, we don't know, right?
现在一旦你换了人，我们可能再也不能赢得这场比赛，除非我们观看那个人玩数百次，然后我们才能赢得这场比赛。除非人工智能发现有微妙的微表情是无法控制的。这可能是真的，我们也不知道，对吧？
I mean, I don't know, even professional poker players don't know what those are, but professional, really good poker players still lose. They just win more than they lose and I think they do that partially because they get used to watching the tells on this player, right?
我是说，你知道的嘛，就算是职业扑克玩家也不知道那些是什么，但是职业的、真正优秀的扑克玩家还是会输。他们只是赢得比输得多而已，我觉得部分原因是因为他们习惯了观察对手的表情对策略的影响，对吧？
The eyes get bigger, they get smaller. And they start to memorize a bunch of things that person did, watch them win or lose a few times or watch them bluff a few times. Then finally go, I can read their bluff because that's the trick to the game. The trick to the game is reading one's face or that they drink more. You're bluffing and you always order tea. I don't know, right?
眼睛会变大，也会变小。它们会开始记住那个人做的一堆事情，看他们赢了或输了几次，或者看他们假装几次。最后你就能够读懂他们的假装，因为这就是游戏的诀窍。游戏的诀窍就是读懂一个人的表情或者他们喝酒的多少。你在假装而且总是点茶。我不知道，对吗？
So yes, AI could do that, but no better than a human can do it, provided it's the same set of people. If you never change the people, they will figure out things. A human might go and scratch his leg, doesn't even know he scratches his leg, right? But he does it every time he's bluffing and he's holding the hand that's no good. Yeah, we could apply AI there for sure.
所以，是的，人工智能可以做到这一点，但并没有比人类更好，只要是同一组人。如果你从不改变这些人，他们会自己想出办法。一个人可能会去抓他的腿，甚至不知道他在抓他的腿，对吗？但每次他在虚张声势，手中没有好牌时，他都这么做。是的，我们肯定可以在这里应用人工智能。
So it seems like 90% of the development in AI since the 1980s has just been increased processing speed of computers. So using roughly the same techniques, yes, we'll improve, we'll throw some more layers onto the neural networks, we'll play around with the statistics a little, but it's basically just we can now handle big data. And maybe there might be an innovation if you can use AI to increase your data set in interesting ways like, you know, imagine you have data about self-driving.
看起来自1980年代以来，人工智能领域90%的发展只是计算机处理速度的提升。因此，使用大致相同的技术，我们确实会有所进步，我们会在神经网络上加入更多层，稍微调整一下统计数据，但基本上是我们现在能够处理大数据。也许会有一些创新，如果你能使用人工智能以有趣的方式增加你的数据集，就像你有关自动驾驶的数据一样。
Now you start to imagine scenarios where the same car makes a left turn and you have to kind of simulate that first with AI to create the data set. That seems interesting to me, but what else? What could be, that seems incremental. So what seems like a big change?
现在你开始想象一辆车左转的情景，你需要利用人工智能来模拟这种情况来创建数据集。这对我来说听起来很有趣，但还有什么呢？什么能够是逐步发展的？那么什么似乎是一个巨大的改变？
Well, yeah, so the big change that everybody wants to figure out, yeah, there is no breakthrough in sight because I think we just don't understand it is. If you really want artificial intelligence, it isn't about processing big data. It's about two things.
嗯，是的，所有人都想要弄清楚的重大变化，是的，我们似乎并没有看到突破，因为我认为我们只是不了解它。如果你真正想要人工智能，它并不是关于处理大数据。它关乎两件事情。
General knowledge, which we're not good at. Remember what we're doing today in AI is very vertical. You notice we're going to teach an AI algorithm to play this game. Okay, that AI algorithm can't recognize a dog that's sitting next to him. Doesn't know if the dog pooped on it has no clue, but it does know how to play go brilliantly, but it's all it does, just like facial recognition systems can't recognize a chair. Why? It's not what they do. They do facial recognition. They're very attuned to that.
我们不擅长的是常识。记住，我们今天在AI领域从事的是非常垂直的工作。你会发现我们要教一个AI算法如何玩这个游戏。这个AI算法可能不认识坐在它旁边的一只狗。它不知道这只狗是否拉了屎，也不知道发生了什么，但它确实非常擅长玩围棋，这就是它所能做的。就像面部识别系统无法认出一把椅子一样。为什么呢？因为这不是它们的工作。它们只是做面部识别，对此非常敏感。
So professionals have been working on very, very vertical business-oriented kinds of things. Facial recognition, speech recognition, translation. Driving has turned out, driverless vehicles has turned out to be a huge problem. With every neural net we've got thrown at this with everyone at Google with 15 years experience on the road now, the problem is the following.
专业人士一直在从事非常垂直的面向业务的事情，例如面部识别、语音识别和翻译。无人驾驶车辆已经成为一个巨大的问题。即使我们在谷歌的每个拥有15年驾驶经验的人脑神经网络上投入了所有的资源，问题仍然存在。
There may be an unlimited number of unusual events because if we had to drive on a track and there were no humans allowed on the track, we could have done that 30 years ago. Virtually, you know, with barely a vision recognition system. Vision recognition system can recognize a white line, we could do that. By the way, John Deere has had driverless tractors for the better part of a decade because they're in farm fields. They can mark the boundaries of the farm field and the thing just goes up and down as long as no one gets in front of it. It does its thing, right? Tractor does its thing. Totally doable, by the way.
可能会出现无数不寻常事件，因为如果我们必须在赛道上驾车，并且禁止人类进入赛道，那么30年前我们就已经可以做到了。实际上，你知道的，只需要一个视觉识别系统。视觉识别系统可以识别白线，我们可以做到这一点。顺便说一下，约翰·迪尔（John Deere）在过去十年中一直拥有自动驾驶拖拉机，因为他们在农田中。他们可以标记农田的边界，只要没有人挡路，这个东西就会上下来回地移动。拖拉机会完成它的工作，完全可行，顺便说一下。
When you put cars on the road, you get an unlimited number of unusual events. You see something in front of you through a vision system or a lidar and you don't know the system because it's never seen it before. It doesn't actually know how to respond. Is it a shadow? Is it something that's happening in the rain or the wind? Is it leaves just blowing by and they're gone? Now we as humans have something unique. We have literally, you know, if you're 50 years old, you've got 50 years of taking information into your eyes and you know you can just recognize that thing looks solid. I better stop. Or it is solid, but it's a bag and I can run over it.
当你把汽车放在路上时，你会遇到无数奇怪的事件。你通过视觉系统或激光雷达看到前方的东西，但你不知道它是什么，因为它从来没有见过。它实际上不知道该如何回应。它是阴影吗？是在雨中或风中发生的事情吗？是叶子在飞舞，一下子就消失了吗？现在，我们作为人类拥有独特的东西。我们确实有50年或者更长时间，通过眼睛收集信息，你可以认出事物的外观，知道它们是实体的。我最好停下来，或者它是实体的，但是是一个袋子，我可以碾过它。
Now think about that, right? You see a bag but it's kind of blowing in the wind. You know it's not full of rocks and just run over it. It's fine. You might get caught under the car. It's not going to hurt anyone, right? But an AI system goes, I don't know what that is. In fact, it might be a curled up baby. It's, it's, it's, stop. You remember when we were kids, we'd look up at the clouds and parents would say, oh, what does that look like? Oh, it looks like a dinosaur. It looks like this. It looks like that. The thing looks like a dinosaur to the AI system. It might be a dinosaur. It doesn't know, right? It's, and then you got unknown things.
现在想想看，你看到一个袋子，但它在风中摇晃。你知道它里面不装满了石头，所以你越过去没问题的，你可能会被卡在车底下，但不会伤到任何人，对吧？但是一个AI系统会说，我不知道那是什么。事实上，那可能是一个睡成一团的婴儿。它，它，它，停下来了。你记得小时候，我们经常抬头看云朵，父母会问：“哦，那是什么？是恐龙吗？还是这个？还是那个？”这个东西在AI系统看来可能看起来像恐龙，也可能是恐龙。它并不知道，对吧？它是有未知事物的。
You know, you got, yes, it's granny crossing the road, but, but now you got, you know, granny on a, on a, on a, on a unicycle. And I've never seen a unicycle before. What do I do with that? Do I stop? Do I run it over? So the unlimited events problem is a real problem for AI. It's a, it's a super problem for AI. But, but they haven't driving, they haven't driving on highways, right? And there hasn't been major issues. I think this is fine because very little happens there. It's the straight in the city where you get an unlimited set of crazy events going on, right?
你知道，你碰见了，没错，是老奶奶要过马路，但是，但现在你看，你知道，老奶奶在骑独轮车。我从来没有见过独轮车。我该怎么做？停下来吗？碾过去吗？所以无限事件问题对AI来说是个真正的问题。这是一个超级问题。但是，他们还没有在高速公路上开车，对吧？而且也没有重大问题发生过。我认为这没关系，因为那里很少发生事情。只有在城市中，你会遇到一系列无限的疯狂事件，对吧？
So, so I mean, potentially you could just say, okay, legal on highways. And then last mile, we need bed X or whatever to meet your trucks. But what do you think of Andrew Yang's predictions about, you know, you know, the elimination of millions of truck drivers because highway driving is, is solved. Highway driving is nearly solved.
嗯，我的意思是，潜在地，你可以只是说，在高速公路上合法。然后最后一英里，我们需要协调X或其他东西来满足你的卡车。但你对安德鲁·杨有关数百万卡车司机被淘汰的预测有什么看法呢，因为高速公路驾驶几乎是已经被解决了。
Well look, first of all, there's a shortage of, of, of, of long distance truck drivers in the United States by the tune of 20 to 30%. That is we really need 20 to 30% more than we have. So the first thing that will go is to close the gap on the shortage. That's number one. You can't, not everything will be driverless on day one, but, but I think we will start to close that gap in long distance driving. And then the truck will stop and stop and a human will get in and do the last X miles, right? I think that's going to happen. And then right after that, just to be competitive, you have to do the rest of it because a long haul trucking across the country is, let's call it five or six thousand dollars to take, you know, 40 foot truck across the country.
咳，首先，美国目前缺少长途卡车司机，数量约为20%至30%。我们需要比现有数量再多20-30%。所以首要任务就是弥补这一缺口。这是第一点。一开始并不能把所有事情都交给无人驾驶，但是我认为我们会开始填补长途驾驶的这一缺口。然后卡车会停下来等待，而人类司机将会走最后里程。我想这种情况会发生。接着，为了竞争，你必须开展剩下的技术，因为跨越全国的长途卡车运输费用约为5-6千美元，需要用40英尺的卡车。
It's just what it is today. That's what it costs. Most of that is labor followed by gasoline, followed by right off of the vehicle, right? That those are the three things. And so you've got to get labor out of it to take the $6,000 down to say $4,000 or $3,000. That's how you're going to do it. It's the only way to do it. It's the only way you've got to do it. So that labor is going to come out in the next five or six or seven years and that will displace truck drivers. No question.
这就是今天的事实，这就是它的成本。其中大部分是劳动力、然后是汽油，最后是车辆报废损失，对吧？这就是三个要素。所以你必须从劳动力中省下钱，将6,000美元降到4,000或3,000美元，这就是你要做到的。这是唯一的方法，这是你必须做的唯一的方法。因此，在接下来的五六七年中，劳动力将被替代，这将导致卡车司机失业，毫无疑问。
But let's talk about, I'm going to jump from that for a second because we're talking about jobs. There is going to be job loss in the world, certainly in the United States, from AI in the next decade. But it's not going to hit us as bad as it's going to hit other countries. And let me tell you my theory on why.
让我们谈谈吧。我们正在谈论工作，但我要暂停一下。未来十年，AI将在全球，特别是在美国，导致就业岗位的流失。但它不会像其他国家那样严重打击我们。让我告诉你我的理论。
Because over the last 20 years, the U.S., in much of the West, the U.S., specifically, has shed its lowest end mundane tasks to India, China, Mexico, some other countries. And we did so because we could hire people over there to do the mundane tasks in under $1 an hour. When here they were $10 to $15 an hour. We had 20 years of shedding as many mundane tasks as could practically be shed.
因为在过去的20年中，美国和许多西方国家，特别是美国，已经将最低端的平凡工作交给了印度、中国、墨西哥和其他一些国家。我们这样做是因为在那里雇佣人们做平凡的任务不到1美元一小时。而在这里，这些工作每小时需要支付10到15美元的工资。我们已经经历了20年的摆脱尽可能多的平凡工作的过程。
That includes customer support. Look, you call your bank, I don't care who you call, they answer in India. It's not right or wrong. They do because it's a dollar an hour versus whatever would have been $20 an hour here. That's the fact of the matter. So who's going to get hurt the worst in the first decade? China and India. And maybe Mexico. Why? Because it's factory work and it's customer support. And it's software QA and it's all of these mundane tasks that can be automated with AI at the earliest level.
这包括客户支持。你打电话给银行，我不在意你打给谁，他们接电话的在印度。这并不是对错。他们这样做是因为在印度每小时只需要一美元，而在这里不知道需要20美元。这就是事实。那么在第一个十年里谁会遭受最严重的损失呢？中国和印度，也许墨西哥也是。为什么？因为这是工厂工作，客户支持，软件质量保证，所有这些乏味的任务都可以在最早的阶段通过人工智能自动化。
We're seeing RPA companies like Automation Anywhere automate out customer support now. For 80% of the calls, not 100, but those 80%. It's the 80% that we're sent to India. Microsoft got rid of those 20 years ago. So what happened to job loss then? Did we, the economy suffer? I don't recall it suffering then. That's why I wonder how much of this is fear mongering. It's fear mongering. And for now, it's fear mongering. And the reason is, so for the next decade, most of the jobs that would be lost to AI are offshore already.
我们现在看到像Automation Anywhere这样的RPA公司正在自动化客户支持。对于80%的电话，而不是100%，但是那80%。这是我们发送到印度的80%。微软在20年前就摆脱了这些工作。那么当时失业率发生了什么？我们经济遭受了什么？我不记得当时遭受了什么。这就是为什么我怀疑这有多少是恐吓。这是恐吓。现在，它是恐吓。原因是，在接下来的十年中，大多数将由AI失去的工作已经离岸了。
We sent them offshore. And so they're the easiest to automate. And the first things you automate are the ones that are the easiest. Not because of the most expensive, just the easiest. Like to automate. What about like, what about like middle management or white collar type jobs or you know, radiologists, lawyers. Sure, sure, sure. So there will be some things like radiologists are first going to be augmented by AI.
我们把他们派往离岸地区。所以他们是最容易自动化的。最先自动化的是最容易的事情。不是因为最贵，只是最容易自动化。比如自动化。那像中层管理人员或白领工作，或者放射科医生、律师呢？当然，当然。所以，像放射科医生这样的职业首先会被AI增强。
That's already happening. You can send, you can send a lot of these pictures and these images to the cloud. And the cloud will do a better analysis than the radiologist. But the FDA doesn't allow a system to diagnose today under any circumstances. So the system can only report to the person who will review those, those, that data and diagnose. I'll give you an example. One of the heart monitor companies that sends you a little heart monitor, your home, it replaced the halter monitor, they took all of their data from 53,000 patients and developed a neural net around it to try to identify specific anomalies in the heartbeat, right, in the EKG.
这已经开始了。你可以发送很多这些图片和图像到云中。而云会进行比放射科医师更好的分析。但美国食品药品监督管理局今天不允许系统进行诊断。所以该系统只能向将检查这些数据并诊断的人报告。我给你一个例子。其中一个心脏监测公司会向你家发送一个小型心脏监测仪，代替了传统的心脏监测仪。他们从53，000名患者的数据中提取了所有的数据，并围绕它发展了一个神经网络来尝试识别心脏电图中特定的异常。
And they ended up identifying 12 different ones that ended up being more accurate than 12 the best cardiologists in a room arguing whether that person has this arrhythmia or not, right, the AI is already better than they are. However, the FDA will never allow the AI to diagnose directly to you, not in the next decade. It's going to go to the cardiologist and the cardiologist will look at it and decide if he or she agrees with it and then give you your diagnosis. That's the FDA stand on this right now for a lot of reasons, including the doctors have said, you know, I'm not going to be replaced by some artificially intelligent thing, but it's fine if it augments my work because I don't have any time during the day anyway.
他们最终鉴定出了12个不同的方式，这些方式的准确性比一群心脏病专家一起讨论某人是否有心律不齐时的12种方式要高。是的，AI已经比他们更好了。然而，未来十年内FDA肯定不会允许AI直接对你进行诊断。AI会发送给心脏专家，专家会查看并决定是否同意，然后给出您的诊断。这是目前FDA对此的立场，其中包括医生们表示，我不会被一些人工智能东西所替代，但如果它可以增强我的工作，那也很好，因为我白天已经没有时间了。
And they haven't read a peer review paper and God knows how many years, right? So you want your doctor augmented by AI. You want to take your symptoms, put them in a computer and there's already these systems now. When they put them in a computer, the computer comes back and says, run these five tests, three of them they might not have thought of. And they turn it you and say, let's run these five tests. That's fine. The computer augmented their work. Those doctors are going to go away 10, 20 years, lawyers are going to go away. But lawyers are already doing NDAs with I AI now. Why do I want to review an NDA? It's the most mundane stupid thing. You're paying me $350 to review an NDA. The machine can do it for a dollar. They don't want to do it anyway. So the mundane tests are getting done.
他们已经很多年没有阅读同行评审的论文了，对吧？所以你想让你的医生与人工智能合作。你想把你的症状输入计算机，现在已经有这些系统了。当他们把症状输入计算机后，计算机会回来说，运行这五个测试，其中三个他们可能没有想到。然后他们告诉你，我们来做这五个测试。这没问题，计算机增强了医生的工作。这些医生可能会在 10 年、20 年后消失，律师也一样。但是律师现在已经在和人工智能签署保密协议了。为什么我要审查一份保密协议？这是最乏味、最愚蠢的事情。你花 350 美元让我审查一份保密协议。机器可以用一美元做到。他们也不想做这件事。所以这些乏味的测试正在得到解决。
Lastly, this country right or wrong, I'm not not making a political statement here, is it pretty much full employment and I know there's arguments about people doing two jobs and there are lousy jobs in this and that. But all up, we're 3.5% unemployment, which means there's more demand and in fact there's more job openings out there in a variety of fields than they're ever back. Right.
最后，这个国家对与错，我在这里并不是在发表政治言论。它也几乎达到了全面就业，我知道有人会争论有人需要做两份工作或者这个国家有一些不好的工作。但是总的来看，我们的失业率只有3.5%，这意味着有更多需求，事实上，不同领域有更多的工作机会比以往任何时候都要多。是的。
And the reality is, if you took 100 people, would you say 97.5% of them deserve to be employed? That's part of the problem right now. Right. That's exactly right. But they are employed and that's because we've really got full employment. We've got the best employment picture this country has had. Essentially, it's a record keeping, right. And again, it's not even a political statement, just is.
现实情况是，如果有100个人，你会说有97.5%的人应该被雇佣吗？这就是现在的问题之一。没错，完全正确。但他们已经被雇佣了，这是因为我们现在实现了充分就业。我们拥有了这个国家最好的就业形势。基本上，这是一项记录，不是政治声明，只是事实。
And so that means that as those truck jobs go away, as you lose a million truckers, yes, they may not be driving a truck long distance, but they may be driving more of them locally because more of them are coming in, right? It lowers the cost of transportation across the country. Do you know what that does? Lower consumer prices. Do you know what that does? Spur demand. You know what that does? Improve the economy.
那就意味着随着货车工作数量的减少，失去一百万卡车司机，是的，他们可能不会再开长途卡车，但他们可能会在本地开更多的卡车，因为更多的卡车会进来，对吧？这降低了全国范围内的运输成本。你知道这会做什么吗？降低消费品价格。你知道这会做什么吗？刺激需求。你知道这会做什么吗？改善经济。
Whenever you lower cost, the economy goes up, more money circulates around and there's more jobs. Now, their job might be in a factory, their job might be in local deliveries, their job might be in something else. I don't know, might not be driving a long distance truck. Just as we used to have many more people in agriculture, 90% of the country was in ag and today it's 1%. And yet that 1% feeds the entire country in Nessa. Why? Well, because we've got machines. And employment didn't go down, it's gone up. Why? It brought essentially the net across the food way down. How is it that you can go to the store and buy an e-record for 20 cents? 20 cents.
每当你降低成本，经济就会上涨，更多的钱会流通，就会有更多的工作机会。现在，他们的工作可能是在工厂里，也可能是在本地送货，或者是其他方面的工作。我不知道，可能不会开长途卡车。就像我们以前在农业领域有更多的人，全国90%的人都在农业领域，而今天只有1%。然而，这1%的人可以在内萨带领全国人民吃饭。为什么？因为我们有机器。就业机会没有减少，而是增加了。为什么？因为它实质上将食物价格降低了。你怎么能以20美分的价格去商店买一张电子唱片呢？20美分！
There's got to be 20 cents of water in that corn. How do you do it? Well, you do it because it's a lot of machines and a lot of yield. We've learned how to yield the crops better. Everything has worked better, right? So I think that's going to happen here too. I'm not worried about it. For the next 10 to 20 years.
那些玉米中应该有20美分的水，你是怎么做到的呢？呃，因为有很多机器和很高的收成。我们学会了如何提高作物的收成。所有事情都变得更好了，对吧？所以我认为这里也会发生这样的事情。我不担心这个问题。在未来的10到20年里。
So I feel like there's sort of three conclusions here. First, we're screwed because North Korea is going to make, use AI, big data, and CRISPR to make the worst pandemic in the world. And there's no way to really avoid that. How does North Korea get the technical resources to do that? How do they get educated to do that? You know, the problem...
我感觉有三个结论。首先，我们完蛋了，因为朝鲜将使用人工智能、大数据和CRISPR技术制造世界上最严重的瘟疫，而我们没有真正避免的办法。那么，朝鲜如何获得这些技术资源？他们如何接受相关教育呢？这是一个问题...
I don't want to say the problem is, look, much of that work was not done by governments if it was done by academia. And they published their peer reviewed papers. And they published exactly how to repeat it because they want their experiment repeated. That's part of the whole goal of scientists, right? They please repeat my experiment to validate that it works for you and your lab if you follow all these steps. The steps are out there. And that's the problem. I don't want to say it's a problem. It's one of the wonderful things about academic research is it shared worldwide. One of the negative things about academic research is shared worldwide. So everyone who has a reasonable lab can execute what the academics have done. No question. So that's on the bad side.
我不想说这是问题，你看，很多工作并不是政府做的，如果是学术界做的，他们已经发表了同行评审的论文，也准确地发表了如何重复它，因为他们想让别人重复这个实验，这是科学家的目标之一，对吧？他们想请你重复我的实验，以验证这些步骤是否对你和你的实验室起作用。这些步骤已经公开了。这就是问题所在。我不想说这是问题，学术研究最美好的事情之一就是具有全球共享性，不过它的消极影响也是全球共享，任何一个有良好实验室的人都可以重复学术界所做的事情，毫无疑问，这是不好的。
Second is more, the second clue is more neutral, which is as we were saying and as we've even discussed before, over the past 30, 35 years, maybe there's been incremental improvements in AI, but the big advantage it seems has been computers have gotten a lot faster. So whatever analysis you've been doing on big data before is 20 million times faster than it was 20 years ago.
第二条线索更为中性，正如我们之前所说和讨论过的，过去30到35年中人工智能可能有一些逐步的改善，但似乎最大的优势在于计算机速度大幅提高。因此，你之前对大数据所做的任何分析现在比20年前快了2000万倍。
Yeah, 20 million times faster and better and we're finding more things. We're finding more connections between the data and what's happening in the future. I mean, look, as you know at FFANTS, we're using AI to automatically test software. And it doesn't completely eliminate the humans, but it changes their task from I have to test or write scripts to I'm going to let the machine find the bugs for me. And frankly, the machine is way better at finding bugs than 200 humans are. No question.
对啊，它比人工测试快20倍以上，效果也更出色。我们发现了更多的信息，发现了数据与未来的联系。你知道，我们在FFANTS使用人工智能自动测试软件。虽然它不能完全取代人类，但是它让测试的工作变得容易了，从我要测试或编写脚本，变成我让机器来帮我查找漏洞。说实话，机器在查找漏洞方面比200个人类测试员更出色，没有疑问。
But are you finding now the companies that are your clients, they're not firing QA people, they're just being, they have time to make more applications and make more profits to the company? It turns out that's exactly right. They're not firing the QA people. What they're doing instead is saying, when we get this garage of bugs now, let's prioritize them, let's work in production, let's work on developing more test data to put into the system. Right? And by the way, let's increase the coverage.
你发现你的客户公司现在没有解雇QA人员，他们只是有更多的时间做更多的应用程序，并使公司获得更多的利润吗？实际情况是确实如此。他们并没有解雇QA人员。相反，他们说，当现在有这么多错误时，让我们优先考虑它们，让我们在生产中工作，让我们开发更多的测试数据放入系统中。对吧？顺便说一句，让我们增加覆盖范围。
So I'll give you an example. Very, very typical. We'll have a client of FFANTS that says, I've got 100 people in QA, let's say on this big application. And we've been doing a release every four weeks, we now want to make it once a day, four weeks to once a day. And we want to go from 20% code coverage to 100% code coverage.
所以我会举一个例子，非常非常典型。我们的FFANTS客户会说，“我在这个大型应用中有100个QA人员，我们每四周发布一次，现在我们想每天发布一次。我们还想将代码覆盖率从20％增加到100％。”
Okay. Let's do the multiplication. So four weeks to one day is about, call it 20X improvement in productivity you would need, right? You got to go 22X, right? 22, 22 work days in a month.
好的，我们来做乘法。所以四周相当于一天左右，你需要提高生产力约20倍，对吧？你需要提高22倍，是吗？因为一个月有22个工作日。
So, so let's call 20X improvement in productivity to do that with the same team. If you just left the team. Well, how are you going to make them 20 times more productive? But actually, it's more than that because now I want to go from 20% code coverage to 100 percent.
好的，那么让我们说我们要在同一个团队中实现20倍的生产力提升。如果你刚刚离开了团队，那么你该怎么让他们的生产力提升20倍呢？但实际上，情况更为复杂，因为现在我希望将代码覆盖率从20%提高到100%。
So it's 20 times five of that. I now have to be 100 times more productive as a team to meet management's goal of four weeks down to one day and 20% out, 200% coverage. I've got to be 100 times more productive. So either I'm going from 100 people to 10,000, that's one way to do it. I guess, right? 110,000 people.
那么这就是那个的20倍。现在我们需要比以前更加高效率达到管理层的目标，把四周缩短为一天，减少20%的成本，覆盖率增加200%。我必须变得比以前高效100倍。所以要么我要把团队从100人变成10,000人，这应该是一个方法，对吧？110,000人。
Or I better get AI to augment my 100 person team to make each person worth 100 people. And that's what we're doing. We're using AI to augment what they're doing to make each person 100 times more productive than they were before. And that gets used by shortening the cycle time and increasing the coverage that is finding more bonds. And so you actually need the same size team to meet that. Right.
或者我最好使用AI来增强我的100人团队，让每个人都价值100人。这就是我们正在做的事情。我们正在使用AI来增强他们正在做的事情，使每个人的效率比以前高100倍。这通过缩短周期时间和增加发现更多债券的覆盖范围来实现。因此，您实际上需要相同规模的团队才能满足这一需求。没错。
So this leads to the third conclusion, which is that all the theories that quote unquote, this time things are different because middle class jobs are being outsourced to AI is overblown because of all the historical, we've been through this before historically many times.
因此，这带来了第三个结论，即引用一些理论的那些声称，现在情况不同，因为中产阶级的工作正在被AI所取代，这种说法夸张了，因为从历史上来看，我们经历过很多次这样的情况。
Yeah, look, there is a time in the future. And I'm going to say 100 years, right? It's further out than everybody thinks. Where virtually every job we could possibly imagine probably can be done better by a machine, including jobs that require a high EQ, right? And in AI today, we're not talking about high EQ. We've got essentially high IQ down one little pathway.
嗯，你看，未来有一个时间点，比如说100年后，对吗？它要比大家想象的更遥远。几乎所有我们能够想象到的工作都很可能会被机器更好地完成，包括那些需要高情商的工作。而在当前的人工智能领域，我们并没有谈及高情商。我们基本上只专注于高智商这条路线。
Again, processing big data and making judgment calls as new data comes in great. They're good at that, but that's about it. So we're a very, very long way from having empathy, real empathy, not programmed empathy, but real empathy.
处理大数据，并在新数据出现时做出判断是很厉害的。他们在这方面非常擅长，但仅仅在这方面。因此，我们离真正的同情非常遥远，不是编程的同情，而是真正的同情。
In fact, all the people working in the labs on this and even in academia, when you talk to the scientists about real empathy, they really look at you and go, you're kind of crazy, right? We don't know how humans have empathy. We don't even know why. No, I worked on these types of problems when I was in graduate school for computer science and AI and nothing has changed. Nothing has changed.
实际上，所有在此类实验室和学术机构工作的人，当你与科学家谈论真正的同理心时，他们会看着你，说：“你有点疯狂，对吧？我们不知道人类如何有同情心。我们甚至不知道为什么有同情心。”不，我在攻读计算机科学和人工智能的研究生阶段就一直在研究这些问题，但什么也没有改变。什么也没有改变。
We have a school of 1989. Right. We have no idea how to have empathy other than the program it so that when you ask Alexa to marry you, she says, oh, I'm sorry, I can't do that. I'm a disembodied, you know, whatever. Right? I mean, it's cute, but that's just a programmatic response.
我们有一所1989年的学校。是的。除了程序化响应，我们没有任何想象怎样拥有同情心的办法。当你向Alexa求婚时，她会说“哦，抱歉，我不能这样做。我是一个无形的，你知道的，它。”是吧？我是说，这很可爱，但那只是程序化的回应。
We had that at General Magic with Portico in the 90s, which was the first virtual assistant, with great programmatic responses that made people laugh. But after three or four of them, you'd realize it would recycle, right? They'd cycle back around and it was the same three or four that were programmed in. It's not true empathy.
我们在90年代的General Magic和Portico就拥有了第一个虚拟助手，具有出色的程序响应，让人们开怀大笑。但是，在使用了三或四次后，您会意识到它会重复，对吧？ 它们会重新循环，而这几个程序就是相同的三或四个。这不是真正的同理心。
So how do we get empathy? Well, we don't even understand why humans have empathy. And humans have empathy because at some point in our history, we had to survive and that and it took empathy to survive somehow, right? To keep, probably to keep the group of people around you and having that group gave you a higher chance of survival and we're the offshoot of that, right?
那我们怎么获得同理心呢？其实，我们甚至不理解为什么人类有同理心。人类之所以有同理心，是因为在我们的历史某个时刻，我们必须生存下去，而这需要同理心的支持。可能是因为保持身边的人群，有助于提高生存的可能性，我们就是这一历史事件的后继者。
I wonder if you can pose things like this though as a big data problem. Like, let's say you have a million transcripts between therapists and their patients and you just pattern match now. I go into an AI therapist and I ask a question that's been asked before or similar questions been asked before and here's the therapist response.
我想知道你是否可以将事情作为大数据问题来处理。比如说，你有一百万个治疗师与病人之间的对话记录，现在你只需要模式匹配就行了。我去找了一位AI治疗师，问了一个之前已经问过或类似的问题，然后这位治疗师会给我回答。
And we're here as a couple therapist responses and I'm allowed to respond to any of them. And that could be, again, it's not real empathy. It's again, what we've been talking about was just pattern matching, but that's a fake solution to a real problem.
我们在这里作为情侣治疗师，我可以回应他们的任何问题。然而，这并不是真正的同理心，只是我们之前所谈论的案例匹配，这只是对于真正问题的虚假解决办法。
All you would have done is program a psychologist. Right. That's someone you actually want to live with. But true empathy is when something happens in your life, James, and your partner goes looks at you and starts to tear up and says, I really, really feel for you. What can we do? Can I make you dinner? Can we watch a movie tonight? What can I do to help you feel better? And you go, wow, that's a whole share and empathy thing. That's real. It's real.
你所做的只是编写一个心理学家。没错，那是你真正想要生活的人。但真正的共情是当你生活中发生了什么事情，你的伴侣看着你开始哭泣，说我真的，真的为你感到难过。我们能做什么？我能给你做晚餐吗？今晚我们能看电影吗？我能做些什么来帮助你感觉更好？你感到“哇，这是真正的分享和共情。这是真实的。”
And it turns out that's very important for humans. And it's very important in our work. When we talk about going to work or doing our work or whatever it is, a lot of is the interaction we have with other people. A lot of the reason we go every day is we love that interaction, right? Like the interaction we like, whatever it is, right? If you don't like the people, I don't know why you go. So part of that is that part of it is a sense of purpose. A sense of purpose. A machine has no sense of purpose. It just executes its code, right?
原文：
And it turns out that's very important for humans. And it's very important in our work. When we talk about going to work or doing our work or whatever it is, a lot of is the interaction we have with other people. A lot of the reason we go every day is we love that interaction, right? Like the interaction we like, whatever it is, right? If you don't like the people, I don't know why you go. So part of that is that part of it is a sense of purpose. A sense of purpose. A machine has no sense of purpose. It just executes its code, right?
翻译：
原来这对人类非常重要。对我们的工作也非常重要。当我们谈论去上班，做我们的工作或者做其他事情时，很大程度上是和其他人的互动。我们每天去工作的原因很多是因为我们喜欢和其他人的互动，对吧？不管是什么互动，只要是我们喜欢的就可以了，对吧？如果你不喜欢这些人，那我不知道你为什么去啊。所以，这其中一部分是一种目的感。目的感。机器没有目的感。它只是执行它的代码，对吧？
So there's, this is what I'm trying to say. We're so far away, just like you worked on this in 89 and I worked on it in 98, 99. And people have been working on this since then. And no matter how big and deep and smart we make these systems, we don't understand how to model empathy because we don't even know why we're empathetic. Rather than we did it to survive.
所以我的意思是，我们彼此之间隔得太远了。就像你89年开始做，而我是在98、99年开始的。自那时以来，一些人一直在研究这个。无论我们把这些系统变得多么大、深、聪明，我们都不理解如何模拟同情心，因为我们甚至都不知道为什么会有同情心。和我们生存有关的原因不同。
So really the only thing I get worried about is that, again, processing speeds get faster and faster, maybe some techniques improve. And we find more interesting data sets now because the processing can handle it, more interesting data sets that are dangerous. So for instance, the human genome and looking at permutations of genes instead of single gene mutations, that's potentially helpful but also potentially dangerous. And maybe there's data sets that are bigger, that are more complicated right now that we can't solve that are even more dangerous. I don't know. There are always dangerous data sets, right?
我唯一真正担心的是处理速度越来越快，某些技术也可能得到改进。我们现在发现更有趣的数据集，因为处理可以处理更多，更有趣的数据集也更危险。例如，人类基因组并研究基因变异的排列组合，这有潜力有助于我们，但也可能很危险。也许有更大、更复杂的数据集我们现在无法解决，更危险。我不知道。总是有危险的数据集，对吧？
There are things we're going to learn from large sets of data that I suspect the US government is already doing, right? That you might be able to learn from huge sets of genomic data, how to really create something that will wipe out all life on earth. I mean, it's not an impossibility that you could develop a virus of bacteria or something. It would literally invade everything, every plant, every human, everything, it would wipe out life on earth. That's possible. It's not crazy, right? You could certainly look at data sets on the ocean and say, what could we do to reach a tipping point?
有些事情我们将从大数据集中学习，我怀疑美国政府已经在做，对吧？你可能能够从巨大的基因组数据中学习到如何真正创造一种会灭绝地球上所有生命的东西。我的意思是，你可能会开发出一种病毒或细菌之类的东西。它会彻底侵蚀一切，每一个植物，每一个人，一切，它会抹去地球上的生命。那是可能的。这不是疯狂的，对吧？你可以肯定地看待海洋的数据集，说：我们能做些什么来达到一个临界点？
I mean, maybe we're going to do that with climate change anyway, the way things are going and wipe us all out. But you're right, there's dangerous data sets for sure. There's danger in cybersecurity because people are using AI to hack cybersecurity and then people are using AI on the other side of it to kind of keep the AI out, AI battling AI. We're seeing all those things happen.
我觉得，可能按现在的情况下去，我们最终会因气候变化而被摧毁，这个可能是注定的。但你说得没错，确实存在危险的数据集。网络安全也很危险，因为人们正在使用人工智能来攻破网络安全，然后另一方面的人也在使用人工智能来阻止它的攻击，这就是AI互相对抗。我们正在目睹所有这些事情的发生。
But I think most people today that I talk about with AI say, how's it going to impact my life? What do I need to know? When's it going to impact my life? I've seen facial recognition and I've seen some cute stuff in movies and I hate you know other than that. I don't have a robot in the kitchen, the cook's shut. I don't have a simple robot that only has to do one function, cook me a meal. That's actually a very valuable task. Cooking clean, like cooking clean in a house. Someone's got a real robot that really does that even though it's not empathetic, that's a powerful idea.
我认为今天大多数与我谈论AI的人都会问，它会如何影响我的生活？我需要知道什么？什么时候会影响我的生活？我看过面部识别，也在电影中看过一些有趣的东西，但除此之外，我没有厨房里的机器人，也没有只需要完成一个功能——给我做饭的简单机器人。那是一个非常有价值的任务。像做卫生一样做饭。有人拥有一个真正可以做到这一点的机器人，即使它没有同理心，这是一个强大的想法。
But people have tried. There are lots of cooking robots that people have toyed with but they end up getting back down to basic, basic machinery and basic storage of certain things and a refrigerated section and certain ingredients have to be there and then there's just an oven thing so it has to be cooked in that. I mean, it's a robot that we could have probably built 30 years ago, right? It's just not that smart actually. It's not that interesting and it's very, very expensive. Well, that's not interesting. What I want is a thousand dollar robot that cooks and cleans and everybody wants that. That's a market by the way. We go build that.
但人们已经尝试过了。有很多人玩弄的烹饪机器人，但最终它们都回到了基本机械和某些东西的基本存储、冷藏部分和必备的原料，然后只是一个烤箱，所以必须在其中烹饪。我的意思是，这是一个我们可能30年前就可以建造的机器人，对吧？它实际上并不那么聪明，也不那么有趣，而且非常昂贵。这不是有趣的事情。我想要一个价值一千美元的机器人，它可以烹饪和清洗，每个人都想要它。顺便说一下，这是一个市场。我们去建造它。
It turns out it's really hard. It's really, really hard to replicate all the little things that we do around a house. Think about cleaning in places that are hard to reach or cleaning window cells or cleaning things where you have to lift the blinds and then clean and put it down or taking some books out and clean.
原文：It turns out it's really hard. It's really, really hard to replicate all the little things that we do around a house. Think about cleaning in places that are hard to reach or cleaning window cells or cleaning things where you have to lift the blinds and then clean and put it down or taking some books out and clean.
翻译：原来这真的很难。真的、真的很难模仿我们在房屋中所做的所有小事。想想清洁难以到达的地方、清理窗户边框或者清理需要抬起百叶窗来清理的物品，还有取出一些书来清理。
When you go, people ask me, what's the last job that will ever be replaced by AI? I say a plumber and an HVAC repairman, repair person. That's fine. And Andrea and Osmo agrees with that too. Like this is where we're full circle where here he might be correct. Yeah, yeah, he is correct because every house is different, every plumbing problem is different where the pipes are is different. It would be so expensive to create a robot in a database would be full of so much noise that it's an impossible problem to solve.
当你走的时候，人们会问我，哪项工作将是AI最后能取代的工作？我说是管道工和暖通维修工。这很好。 Andrea 和 Osmo 也同意这个观点。我们又回到了起点，这里可能是正确的。是的，他是正确的，因为每个房子都不同，每个管道问题也不同，管道的位置也不同。创造一个机器人会是一个非常昂贵的问题，而且数据库中会充满很多噪音，解决这个问题几乎是不可能的。
Yet, you can send a human in who is a plumber and HVAC repair person and if they're any good, they will eventually find the problem. They will eventually fix it and they'll charge you a couple hundred dollars to do so.
不过，你可以找一个能修水管和空调系统的人来检修，如果他们很擅长的话，他们最终会找到问题所在，解决问题，并收取你几百美元。
Well for a couple hundred dollars, it's way cheaper to have those people do that work than it will ever be in our lifetime to build a robot that would come to your house and fix your plumbing.
用几百美元雇人来做这项工作要比在我们的一生中建立一个机器人去你家修理你的水管要便宜得多。
Now they're going to happen. It's too complicated. It's too, it's just plumbing and it's too complicated. That should level set everyone listening to this. Plumbing is too complicated for a robot.
现在它们即将发生。它太复杂了。仅仅是水管就太复杂了。这应该让所有听到这个的人都明白，水管对于机器人来说太复杂了。
So what do you think we're going to do?
那么，你觉得我们接下来要做什么呢？
Well, Kevin Sirace, this has been enlightening. Informative particularly on the economic stuff is really fascinating and scary on the pandemic stuff. Although the flip side of that is that AI will get better at drug discovery too for the virus, for any AI developed diseases as well. So it balances out.
哎呀，Kevin Sirace，真是开眼界。经济相关的信息尤其引人入胜，而关于疫情的信息则让人感到恐怖。不过好的一面是，人工智能将会在病毒和任何由人工智能产生的病症的药物研发方面变得更加出色。所以它是有平衡的。
But I really appreciate you coming on the podcast, giving us the state of the world in AI this year. Yeah, great conversation we can talk for hours. I'm sure we'll do it again, but thank you so much for having me.
我非常感謝你來參加播客，並向我們介紹了今年人工智慧領域的最新動態。是的，我們的對話非常精彩，我們可以一直聊下去。我相信我們會再次進行這樣的對話，但非常感謝你邀請我參加。
Yeah, thanks Kevin. I appreciate it.
嗯，谢谢Kevin。我很感激。
Yeah, you back.
是的，你回来了。