导语：Eye on A.I.是由纽约时报资深记者 Craig S. Smith 主持的一档双周博客节目。每一期节目，Craig 都将与这一领域有影响力的人物进行交流，推进广义环境中的机器智能新发展，思考技术发展新蕴意。
机器之心为此系列对话的中文合作方。以下为此系列内容的第四篇，Craig Smith 与李开复先生展开的对话。
Hi, this is Craig Smith with a new podcast about artificial intelligence. I’m a former New York Times correspondent now focused on AI. I have been talking to the people who are making a difference in the space and am bringing the most interesting of those conversations to you. This week, I talk to Kaifu Lee, a thought leader on AI in China who just published the book, AI Superpowers: China, Silicon Valley, and the New World Order just came out. Kaifu, who did his PhD in speech recognition AI at Carnegie Mellon founded both Microsoft Research and Google in China. He now runs Sinovation Ventures, a venture capital fund in Beijing. We talked about the premise of the book, which is that China has already caught up with the US in the field of AI and is poised to surpass it. I hope you find Kaifu as interesting as I did.
大家好，我是 Craig Smith，这是一个有关人工智能的新播客。我之前是《纽约时报》的记者，现在专注于 AI，我将与致力于该领域的人对话并呈现最有趣的内容。这周与我对话的是李开复先生，他是中国 AI领域的一位思想领袖，最近刚发表了新书《AI Superpowers: China, Silicon Valley, and the New World Order（人工智能超级力量：中国、硅谷和新世界秩序）》。李开复先生，博士毕业于卡内基梅隆大学，研究方向为语音识别 AI。他曾创立了微软研究院和谷歌在中国的分支机构。他现在北京经营一家风险投资公司“创新工场”。我们谈到了这本书的前提，即中国已经在 AI领域赶上了美国，并终将超过美国。
CRAIG: Going through the book it looks like you're talking less about national competition than about national competency. Which are very different things because competition implies that one country will dominate another at the end of the day. But what you're talking about really is China developing an AI competency that's equal to or ahead of that of the United States. Can you talk a little bit about that?
KAIFU: I think the two are very correlated. If you have extremely strong competency you will become the leader in the world. So, China has a number of unique advantages, the greatest of which is the huge amount of data. And then the great engineers, companies, entrepreneurs, who are using it to find holes in its sometimes backward traditional economy. And when the economy's backwards you can have a late mover advantage and reinvent retail schools communications, health care, and so on.
Also, I think the government has taken a very techno utilitarian approach which is really to let technologies to be tried before going to regulation. And not working out the kinks before getting the technology to launch. And these factors will propel China forward. While the U.S. does have much deeper research bench, China is rapidly catching up and developing a young cadre of very smart AI engineers who arguably matter more than having a small number of AI superstars because we're now in the age of implementation.
另外，我认为中国政府已经采纳了一条非常技术实用主义的道路，也就是让技术在得到法规管理之前先进行尝试。而不是在技术发布之前就先设定好条条框框。这些因素会推动中国向前发展。尽管美国确实有更深度的研究环境，但中国正在快速追赶并正在培养一个由非常聪明的年轻 AI工程师构成的核心群体，很多人认为这样一个群体比拥有少量 AI超级明星更重要，因为我们现在正处于“实现”时代。
In the age of discovery the US brilliant researchers clearly have an edge. Geoff Hinton, Andrew Ng and so on. The research ruled. But now it's really more about landing the technology in a real business and making money with it. And for that the Chinese make up in AI companies is much more formidable, much more scalable, combining fearless hardworking tenacious entrepreneurs with hard working AI engineers - not writing so many papers but just getting things done, getting problems solved, improving efficiency, profits for banking, reducing costs for factories. It's happening everywhere around us. So, these are the conditions that make up an AI superpower in a very different way than, you know, when you think about national competition. It's more who can build more nuclear warheads, but here we have a very different set of competencies that makes China way ahead in some areas and still way behind in others. But the ones that really matter for the foreseeable future are the elements in which China is strong.
在“发现”时代，美国的杰出研究者显然掌控了前沿——Geoff Hinton、吴恩达等等。他们的研究优势是压倒性的。但现在实际上更多的是将这项技术落地成真正的业务，并用它赚钱。而在这方面，由中国人组成的 AI公司要强大得多且更具可扩展性，更何况还有无所畏惧的勤奋顽强的企业家和尽职尽责的 AI工程师——他们不会写那么多的论文，但会把事情做成，把问题解决；他们能帮银行提升效率和利润，还能帮工厂降低成本。这样的事情发生在我们周遭各个地方。所以，这就是构成一个 AI超级力量的条件，非常不同于我们对国家竞争的看法。国家竞争更多的是看谁能制造更多核弹头，但这里我们看到的是一组非常不同的竞争条件——这些条件会让中国在某些领域领先，而在其它方面仍然可能更落后。但在可预见的未来里，那些真正重要的条件正是中国的强项。
CRAIG: To date the developments in AI are largely confined to their respective markets. The Chinese are developing technology for implementation in China. The U.S. is developing new technology for implementation primarily in the English-speaking world. Or at least the American and European worlds. Can you say that China is ahead if its implementations are restricted to its own market. And one of the things that I've been waiting to see: When are these technological super companies in China going to succeed outside of China. And that's something we really haven't seen.
KAIFU: I have a two-part answer to that but I really think it's not the right way to look at the problem. People think of China as one of the hundreds of markets so you only have one market. What good is that. But my answer is that China is larger and more important and more valuable than the rest of the world combined. And so yes, it's not the whole world. But it may be half the world. So why is it half the world? Well if we look at all the mobile payments that are gathered, which form the strongest basis for AI learning, China has more than the rest of the world combined. If you look at computer vision, gathering of images, face recognition and so on, China has more than the rest of the world combined.
I think the evidence is that the Chinese computer vision companies are worth ten billion dollars. The American ones are worth hardly one billion. So I think to position it as China has only one of the markets and what about the rest of the world, is missing the very facts in front of us, which says China has so much more data and so many more ways of gathering data - and that includes a larger market and more advanced digital collection system, putting sensors and inputs from everything from retail to airports, gathering information from payment activity, this amount of data is phenomenal. And I would argue, if China companies were restricted to stay in China for the next 10 years the total capitalization of Chinese AI companies will still be larger than the rest of the world combined. So that's the first half of my answer.
The second half is that the Chinese companies are looking to go abroad, but they're going abroad in a very different way and arguably a smarter way than the Facebooks in Googles. History of tech colonialism is such that America dominated the world. So, Windows and Intel took over the world and had a 100 percent and demanded adoption whether their products were well localized or not. And China was one of the technically colonized countries in the sense of using PC hardware and software. However, in the age of mobile Internet that's not the case anymore. The American technologies are reaching less and less of the world. Chinese technology is actually reaching more and more of the world but not in the same way.
下半部分答案是中国公司正在积极出海，但它们出海的方式非常不同，可能也比 Facebook和谷歌等所采用的方式更聪明。在技术殖民主义的历史中，一直是美国主宰世界。所以，Windows和英特尔几乎掌控了全世界的计算机，它们的产品会得到 100%的采用，不管它们的产品本地化做得好不好。而且在 PC的硬件和软件方面，中国也是一个被技术殖民的国家。但是，在移动互联网时代，情况已然改变。美国技术触及世界的范围越来越小。中国技术实际上已经越来越多地走向世界，但方式却不一样。
You don't see anyone using DiDi in Indonesia. However, DiDi is very cleverly partnering with all the locals so as to form an alliance of the insurgents against the American hegemony. Uber is trying to dominate the world using one brand, one platform, one world. That's the typical American way. Windows, Microsoft, Intel and then going on to attempts by Yahoo, Google, Amazon, with less success but still decent success. But that method is not going to be good enough anymore because technologies now touch physical aspects of our world. Putting Uber in Brazil is not a trivial matter. There's government relations, there's usage patterns, there's taxi coalitions. So, it requires a local to be successful. So, the Chinese AI company approach is 'Let's partner with the locals.' So, Didi has partnered with locals in Southeast Asia, South America and is greatly expanding its footprint against Uber. And it doesn't own the local partners. It owns maybe 20 percent, 30 percent, maybe with Softbank, maybe by themselves. And is forming a very powerful alliance where the local companies now feel they have a chance at building products for their own country.
在印度尼西亚，你不会看到任何人使用滴滴。但是，滴滴非常聪明地与所有当地公司建立了合作关系，形成了一个针对美国霸权的反抗联盟。Uber试图使用一个品牌、一个平台、一个世界来主导这个世界。这是典型的美国做法。Windows、微软和英特尔都成功了，雅虎、谷歌和亚马逊也一直在这样做，虽然没有那么成功，但成绩也很不错。但这一方法再也不会那样有效了，因为现在的技术已经触及到了我们的物理世界。把 Uber投放到巴西不再简单轻松。会有政府监管，会有使用模式，还有出租车联盟的问题。所以，这需要一家本地公司才能取得成功。所以，这家中国 AI公司的方法是“让我们与当地公司合作吧”。滴滴已经在东南亚和南美与当地公司建立了合作关系，针对 Uber极大地扩展着自己的足迹。而且滴滴并不拥有当地合作伙伴。它可能拥有 20%或 30%，也许和软银一起，也许只靠自己。它们正在构建一个非常强大的联盟，当地公司现在觉得他们有机会为他们自己的公司开发产品了。
So, China has been through the technical colonialism. So, it's empathetic to other countries and develops ways to work with them to give them money, business, knowhow, experience and perhaps even AI technology and maybe sharing of data parameters at the end of the day. Tencent and Alibaba are among the largest investors in tech in the world. So, it's just the method is different. See the American companies really want one brand, one technology, one platform, own it all. But I think the days of that may be over.
CRAIG: We're talking about corporate AI not national or military. And increasingly these companies do not have a particular national identity. I mean, they may be based in one country or another. Google's a good example. It owns Deep Brain and Geoffrey Hinton works for Google in Canada. Is that happening with China as well? Is that what you're suggesting. That these Chinese giants are first of all private companies or at least quasi private companies and their influence or their profile is becoming increasingly multinational?
Craig：我们谈的是企业 AI，而不是国家或军用的 AI。而且这些公司渐渐地不再具有单一的国家身份，虽然它们的总部可能在某个特定国家。谷歌就是个很好的例子。该公司拥有 DeepBrain，Geoffrey Hinton 在加拿大为谷歌工作。中国也有这样的情况吗？你说的就是这个吗？那些中国巨头是私营公司或者准私营公司吗，它们是否正在具备越来越大的国际影响力？
KAIFU: Clearly, Baidu, Tencent and Alibaba all have Silicon Valley offices in which they each employee hundreds if PhDs. So, there is certainly a desire to take talent from all over the world. They do have plans to go into other markets. So, yes, they want to become global companies in both senses of that.
CRAIG: Again, in terms of the framework of your book, AI Superpowers, are the superpowers the companies or the superpowers the nation states.
KAIFU: The collection of companies in each country. Yeah, I was not making a case of a nationalism or military. I was not going into that. That's not my expertise. I have no visibility into either U.S. or China, NSA or you know People's Liberation Army efforts. So, I don't cover that.
CRAIG: No certainly. But what I'm saying is that as long as it's not a government-owned effort, as long as it's within the sphere of private enterprise, increasingly they are not national efforts. They're multinational efforts. Geoff Hinton is a good example, I mean he's a Briton, I think he may now have Canadian citizenship, but he was educated in the United States and worked for a long time in the United States and is now living in Canada. You know that's not American AI, the things that he's developing. It's very international and even with the Chinese PhDs, they study in the U.S., they do research in the U.S., they go back to China. They work in China or they come back to the U.S. and work. It just seems like it's becoming very fluid. And it is going to be increasingly difficult to talk about national strategies other than at the level of education.
Craig：当然。但我的意思是，只要这不是政府在做，只要这还在私营企业的范围内，它们渐渐地都会模糊国家的界线。这是跨国性的工作。Geoff Hinton就是个很好的例子。他是个英国人，我想他现在可能有加拿大国籍，但他是在美国接受教育的，并且曾为美国工作过很长时间，现在住在加拿大。你知道他开发的东西并不是美国 AI。这是非常国际化的，甚至有些中国人博士也是在美国学习的，然后在美国做研究，或回到了中国。他们也可能在中国工作之后又回到美国工作。看起来有很强的流动性。而且除了教育水平方面，谈论国家战略将越来越困难。
KAIFU: Yes I agree with all the points you made and I think that all these things you mention will expand. But at the same time, you know China as well as Canada have had fairly effective national policies that build up national infrastructure for investment, advancement, education, training, improved roads for autonomous testing, help with the VC funding of AI companies. So, each country is rightfully trying to create more tax paying AI companies that will bolster the country's competitiveness. So, I think that's not directly in conflict with the globalization effort you talk about.
李开复：是的，我认同你所有这些观点，我认为你提到的这些还会继续延展。但与此同时，你知道中国和加拿大都有相当高效的国家政策，能为投资、发展、教育、培训等构建国家基础设施，能为自动驾驶汽车测试改善道路，帮助 AI公司获得风险投资。因此，每个国家都在努力创造更多会交税的 AI公司，这能提升其所在国家的竞争力。所以，我认为这与你谈到的全球化方面的事情并没有直接冲突。
CRAIG: To me that is where the idea of national AI strategy has its greatest impact, is in the economy. So, it's not that China is competing with the U.S. to become some sort of a master of AI or that the U.S. is competing with China to dominate in AI. It's that the implementations and the basic research that's coming out of each country will have an impact on each country's economy and in the Internet age and with repositories like arxiv, the developments in research on one side of the world are disseminated at lightning speed all over the world. So, it's very difficult to have an edge in one country or another. So, what you're talking about really is the entrepreneurial ecosystem that drives unique implementations, not so much about developing algorithms or systems that are unique to one country or the other.
Craig：在我看来，国家 AI战略思想影响最大的地方是在经济方面。所以，中国与美国竞争并不是为了变成 AI的主人或美国与中国竞争是为了主宰 AI。而是说各个国家的 AI实现和基础研究将会影响每个国家的经济状况。在互联网时代，有了 arXiv这样的存储库，在世界一端的研究进展能以闪电般的速度传播到整个世界。所以一个或另一个国家很难占据前沿。你说的实际上是推动特有实现的企业生态系统，而不是开发对某个国家特定的算法或系统。
KAIFU: Yes, I think the academic parts of the AI community is very naturally transparent, helpful, honest, use the common data set with the experiments being replicable. So, it's quite different from other sciences where it's not always easy to replicate say a clinical trial. Because of AI's digital nature allows it to be validated, tested and therefore people are basically standing on the shoulders of giants who are eagerly publishing in order to get academic credit but not extending any sort of, you know, national edge, which may or may not be wanted by the government. But if it did want it, it's very hard to actually consummate that.
李开复：是的，我认为 AI社区的学术界部分自然是非常透明的、互助的、诚实的，他们会使用常用的数据集来进行实验，使得实验可复现。这一点不同于其它某些科学，比如临床试验的成果并不总是容易复现。因为 AI的本质是数字的，使得它可以轻松得到验证、测试，因此人们基本上都是站在巨人的肩膀上。他们都急于发表自己的研究成果，以提升自己的学术影响力，而不会延展成任何形式的国家优势。有的政府希望有这样的优势，有的则不然。但如果政府确实想要这个优势，实际上却会变得很难实现它。
Having said that, there are some possibilities where the U.S. or any other country could take significant leadership. For example, you know, a lot of the world's best AI people are in Google. And if they make a breakthrough and choose not to publish, well they as a company will have a leadership position to build products others may not be able to replicate. And that would be indirectly a national advantage, should that happen.
CRAIG: Again, I'm not sure whether it's a national advantage or a corporate advantage. In your book you talk about the risk of hermetically sealed corporate environments. Already just going to conferences I can see there is some grumbling about papers being presented, being awarded prizes, that do not give enough of the code to be reproducible.
KAIFU: So even when the original publisher doesn't want to give away source code other people can build it. Because machine learning code is not very large or complex, the replication once the algorithm is known is not an extremely time intensive kind of thing.
CRAIG: The three elements of successful AI are the code, the algorithms, the data, but also the computing power. The computing power in the United States or in the West is really controlled by large corporate interests. Is that the same in China or is there a government aspect to that that helps individual companies.
KAIFU: There is no government subsidy for computing per se, but obviously if you receive some subsidy you can use it on computing if you want. But for a lot of the common big data types of AI, you really don't need that much computing. What we generally talk about are the most complex forms of computing.
If you have a powerful single server with a couple of GPUs that will take care of most computer loads for anything up to video computer vision types of applications.
CRAIG: You talk about data, certainly that's one place China has a clear advantage. Partly, as you note because of the different privacy environments, but also because Chinese society is so interconnected that there's a lot of data being collected all the time. Is that something, again, that you think is giving China an advantage globally? Or is that only in its own market? Or is that allowing them to develop implementations that they can take outside.
KAIFU: China's approach going globally is largely through partnerships. So, for example DiDi's partner in Indonesia, Singapore or India will apply local privacy data restrictions and policies.
So it's not up to Chinese companies to decide. Just like when U.S. companies go to Europe they have to follow GDPR now. So is it very much controlled at each country and given Chinese AI companies approach to going abroad, it is not doing so by itself, partnering locally, it will get taken care of by the local partners.
所以这不是由中国的公司决定的。就像现在美国公司进入欧洲都必须遵守 GDPR一样。因此这方面基本上是由各个国家控制的，鉴于中国 AI公司出海的方法，它们不会自己去做这些事，而是会与当地公司合作，这些事情也会由当地合作伙伴处理。
CRAIG: I meant more that in developing implementations, there's tremendous data available and it's much easier to get your hands on in China than it is in the U.S. For example in medical implementations, it's a problem because medical data is so tied up in privacy laws. But in China I've been told that there's a lot more opportunity because a lot of the medical data is more readily available to developers.
KAIFU: That is a possibility. The issue is also it has to be quality data. So, because the quality of healthcare is significantly lower in China compared to the U.S., China's data quality is not at the U.S. level. So, we'll have to see how this plays out with potentially much larger group of lower quality data. Whether that's good enough to build systems or not, I think that remains to be seen. But the theoretical relative openness to data sharing is certainly an advantage for Chinese AI companies.
We should also clarify that it's not like Chinese users don't care about privacy. It's just that there is a greater degree of openness to using some data if there is a clear benefit to the user, such as better treatment, safety, or convenience or monetary savings. There is a greater willingness to do that. Companies generally do have to disclose to users that they're collecting data. More users may say Okay. And also, the Chinese laws prohibiting sale or transfer of private data to other companies. So, taking to Facebook-Cambridge Analytica example, people would actually potentially be put in jail for doing what they did. So, it's not like the laws are loose and people are copying data everywhere. I think people get that impression.
我们应该澄清，并不是说中国用户并不在乎隐私。只是说如果使用某些数据对用户而言有明显的好处，比如更好的治疗、安全、便利或省钱，他们对此的态度就会更加开放。中国用户分享数据的意愿更高。一般来说，公司都必须向用户说明它们正在收集数据。大部分用户可能都会“同意”。另外，中国法律禁止将私人数据销售或转移给其它公司。所以，如果中国出现类似 Facebook和 Cambridge Analytica 那种事，相关责任人实际上是有可能进监狱的。因此，并不是说法律宽松，人们到处复制数据。我认为人们有这样的看法。
CRAIG: When you talk about superpowers, you're talking about China as a nation or Chinese companies as a superpower. But there is an economic impact to the development and implementation of AI. Do you think that that impact or that effect on the Chinese economy is going to help China close the economic gap with the United States.
Craig：当你谈到超级力量时，你谈到了中国或中国的公司是一大超级力量。但经济也会影响 AI的开发和实现。你认为 AI对中国经济的影响和效果能帮助中国拉近与美国的经济差距吗？
KAIFU: Well given both countries are superpowers, it's hard to predict how the numbers will go. I think it's probably safer to say that U.S. and China will increase their gap with the rest of the world.
CRAIG: The other advantage that the U.S. has is in being an anglophone nation and so much of education in science and technology is in English. But in terms of being a power globally, is that a restriction do you think?
KAIFU: I do think there is a big advantage for the U.S. What you describe in the language is a part of that. But I think, really, it's the U.S. research and university system that draws the world's smartest people to study in the U.S., many of whom stay in the U.S. afterwards. And that ability to really bring in the world's smartest people helps the US, despite its much smaller population to China, to actually end up with a larger technology elite class than China. And that has been the U.S. advantage going back decades, if not centuries. And that advantage will continue for the US.
CRAIG: Is there a way for China to balance that, because you're talking about, you know, this age of implementation. But that's very different than, as you called it, the deep bench research that is happening in the United States. Will that eventually happen in China? Is there a move toward building that sort of capability?
KAIFU: The Chinese central government would love to dramatically improve universities and research and in fact they have improved a lot over the last 20 to 30 years.
李开复：中国中央政府希望能极大提升大学水平和科研能力，事实上过去 20到 30年里已经提升了很多。
But it takes maybe a century for any country to elevate its universities to be best in the world. It took America that long. So, this isn't something that can be quick-fixed. We see small efforts, such as allowing universities to pay more to bring in really smart international talent in AI and setting up research institutes and things like that. But these are all just Band-Aids not the ultimate solution. The ultimate solution is, you have to make teaching and research a very respected and well-paid job. And also, you have to divert people from going to Alibaba, Tencent and startups and stay at universities and you have to make that job career interesting and pay competitive. And then there's also, on top of that, the problem of attracting global students to study, whether that would happen some day or not. So, I just think it's 50 years of slow progress kind of thing because you just don't see any country elevate its education system and research capacity that fast.
但对任何国家来说，要让自己的大学成为世界顶级大学都可能要花费一个世纪的时间。美国就用了那么久。所以，这并不是能一蹴而就的事情。我们看到了一些较小的努力，比如允许大学支付更多钱来引进 AI领域的真正聪明的国际人才，以及设立研究院等等。但这都只是“创口贴”，而不是最终解决方案。最终解决方案是必须让教育和研究成为一个受人尊敬和高薪的工作。另外，也必须让人们不被阿里巴巴、腾讯和创业公司挖走，而是让他们呆在大学里；你必须让他们对这些工作有兴趣，并且支付有竞争力的薪资。然后，在此基础上，还要吸引全球的学生前来学习。这未来某天可能会发生，也可能不会。我只是认为这是一个需要 50年缓慢发展的事情，因为我们看到没有任何国家能快速提升其教育系统和科研能力。
CRAIG: When you talk about China catching up or surpassing the U.S., you're talking about implementation. One of the reasons China has been able to build up corporate entities that now are strong on their own is because they were operating in a closed market. I mean that may be less important now than it was but certainly Baidu - you know this better than anybody - would probably not have been able to compete with Google had Google been able to operate in an unrestricted way in China. Tencent may not have developed WeChat had Facebook or Twitter been able to operate in an unrestricted way in China. Do you think that sort of national protection is still important? And do you see that changing at all?
Craig：当你谈到中国赶超美国时，你说的是实现方面。中国能够自己建立当前的强大企业实体的一个原因是它们在一个封闭的市场中运作。也许这个问题现在已经没有那么重要 了，但过去确实很重要，比如很显然，你实际上知道得比任何人都清楚，如果谷歌能够不受限制地在中国运营，百度很可能无法与谷歌竞争。如果 Facebook或 Twitter 能在中国不受限制的运营，腾讯也可能无法开发出微信。你认为那种形式的国家保护仍然很重要吗？你认为这会改变吗？
KAIFU: I think at this point and probably for the last five or 10 years it's really a non-issue. That is, China has developed into a parallel universe as I described in my book.
So take a U.S. company and say, let's have you do a China version completely unencumbered. They will almost certainly fail. Because the entire building blocks are different. The users are different. Their habits are different. And they're working against incumbents that have massive brand, technology, user loyalty as well as local knowledge and the huge amount of data that it takes to build AI and then the large data gives the local companies better AI and better products. And all of which, really, I think make it very difficult for a U.S. company to come in at this point. I would also add the reverse is also true because in a parallel universe it's going to be just as hard for a Chinese company to do something in the US as well.
所以假设有一家美国公司，我们让它完全不受阻碍地做一个中国版本。他们几乎肯定会失败。因为整个建构模块都不一样。用户不一样。他们的习惯不一样。而且要与那些有大品牌、技术、用户忠诚度以及本地知识和巨量数据的本地巨头竞争，这些巨头可以使用它们的数据来构建 AI，而大数据又会给这些本土公司提供更好的 AI和更好的产品。综合这些因素，我认为美国公司现在已经很难进入中国了。我觉得反过来也是一样，因为在平行宇宙里，中国公司进入美国也会一样困难.