?

本文為《經(jīng)濟學人》7月刊文章。

Artificial intelligence

人工智能

The algorithm kingdom

算法的王國

China’s deep pool of data means it has a chance to lead in artificial intelligence

中國數(shù)據(jù)資源豐富，有機會在人工智能領(lǐng)域占據(jù)領(lǐng)先地位

AT THE start of this year, two straws in the wind caught the attention of those who follow the development of artificial intelligence (AI) globally. First, Qi Lu, one of the bosses of Microsoft, said in January that he would not return to the world’s largest software firm after recovering from a cycling accident, but instead would become chief operating officer at Baidu, China’s leading search engine. Later that month, the Association for the Advancement of Artificial Intelligence postponed its annual meeting. The planned date for the event in January conflicted with the Chinese new year.

今年年初，關(guān)注全球人工智能(AI)發(fā)展動向的人注意到兩件小事。先是微軟高管之一陸奇在1月表示，自己從自行車事故中恢復后不會回歸這家世界最大的軟件公司，而是將加入中國領(lǐng)先的搜索引擎百度，擔任該公司的首席運營官。當月晚些時候，國際人工智能協(xié)會推遲了年會舉辦的時間。會議原計劃在1月舉行，但時間與中國新年發(fā)生了沖突。

These were the latest signals that China could be a close second to America—and perhaps even ahead of it—in some areas of AI, widely considered vital to everything from digital assistants to self-driving cars. China is simply the place to be, explains Mr Lu, and Baidu the country’s most important player. “We have an opportunity to lead in the future of AI,” he says.

這兩件事再度顯現(xiàn)了一種可能性：在人工智能的某些領(lǐng)域，中國將緊隨美國而位居第二，甚至可能趕超美國。人們普遍認為，從數(shù)字助手到自動駕駛汽車，人工智能對一切事物都至關(guān)重要。陸奇解釋說，中國無疑是未來的熱點所在，而百度是這個國家最重要的玩家。“我們有機會引領(lǐng)人工智能的未來。”他說。

Other evidence supports the claim. In October 2016 the White House noted in a report that China had overtaken America in the number of published journal articles on deep learning, a branch of AI. PwC, a consultancy, predicts that AI-related growth will boost global GDP by $16trn by 2030; nearly half of that bonanza will accrue to China, it reckons. The number of AI-related patent submissions by Chinese researchers has increased by nearly 200% in recent years, although America is still ahead in absolute numbers (see chart).

其他證據(jù)支持他的這一說法。2016年10月，白宮在一份報告中提到，中國在深度學習(人工智能的一個分支)領(lǐng)域的期刊論文發(fā)表數(shù)量已超過美國。咨詢公司普華永道預測，到2030年，與人工智能有關(guān)的增長將令全球GDP增長16萬億美元。它估計這筆“橫財”有將近一半都將流向中國。近年來，中國研究人員提交的與人工智能相關(guān)的專利申請數(shù)量已增加近200%，盡管美國在絕對數(shù)量上仍然領(lǐng)先(見圖表)。

To understand why China is so well placed, consider the inputs needed for AI. Of the two most basic, computing power and capital, it has an abundance. Chinese firms, from giants such as Alibaba and Tencent to startups such as CIB FinTech and UCloud, are building data centres as fast as they can. The market for cloud computing has been growing by more than 30% in recent years and will continue to do so, according to Gartner, a consultancy. In 2012-16 Chinese AI firms received $2.6bn in funding, according to the Wuzhen Institute, a think-tank. That is less than the $17.9bn that poured into their American peers, but the total is growing quickly.

要了解中國為何占據(jù)優(yōu)勢，不妨試想一下人工智能所需的投入。其中最基本的兩種投入——計算能力和資本，在中國十分充裕。從阿里巴巴和騰訊這樣的巨頭到興業(yè)數(shù)金和UCloud等創(chuàng)業(yè)公司都在盡可能快地打造數(shù)據(jù)中心。咨詢公司高德納(Gartner)稱，近年中國云計算市場保持了超過30%的增速，而且這一勢頭還將繼續(xù)。據(jù)烏鎮(zhèn)智庫的數(shù)字，2012年至2016年，中國的人工智能公司獲得了26億美元的資金投入。這一數(shù)字少于美國同行獲得的179億美元，但總額正在迅速增長。

Yet it is two other resources that truly make China a promised land for AI. One is research talent. As well as strong skills in maths, the country has a tradition in language and translation research, says Harry Shum, who leads Microsoft’s AI efforts. Finding top-notch AI experts is harder in China than in America, says Wanli Min, who oversees 150 data scientists at Alibaba. But this will change over the next couple of years, he predicts, because most big universities have launched AI programmes. According to some estimates, China has more than two-fifths of the world’s trained AI scientists.

然而，真正讓中國成為人工智能福地的其實是另外兩大資源。一是研究人才。微軟人工智能事業(yè)部門的負責人沈向陽說，中國除了數(shù)學很強，還有語言及翻譯研究的傳統(tǒng)。阿里巴巴的閔萬里手下有150名數(shù)據(jù)科學家。他說，在中國，尋找頂尖的人工智能專家要比在美國難。不過他預測這種局面在兩三年里就會發(fā)生改變，因為多數(shù)大型高校都開設了人工智能項目。據(jù)一些估計數(shù)字，全世界培養(yǎng)的人工智能科學家中，超過五分之二在中國。

The second advantage for China is data, AI’s most important ingredient. In the past, software and digital products mostly obeyed rules laid down in code, giving an edge to those countries with the best coders. With the advent of deep-learning algorithms, such rules are increasingly based on patterns extracted from reams of data. The more data are available, the more algorithms can learn and the smarter AI offerings will be.

中國的第二個優(yōu)勢是數(shù)據(jù)，這也是人工智能最重要的“原料”。過去，軟件和數(shù)字產(chǎn)品主要是按代碼編寫的規(guī)則行事，因而擁有最優(yōu)秀的編程人員的國家占據(jù)優(yōu)勢。隨著深度學習算法的出現(xiàn)，這些規(guī)則日益建立于從大量數(shù)據(jù)中提取的各種規(guī)律之上。供試用的數(shù)據(jù)越多，算法能學到的東西就越多，人工智能所提供的服務也就越智能。

China’s sheer size and diversity provide powerful fuel for this cycle. Just by going about their daily lives, the country’s nearly 1.4bn people generate more data than almost all other nations combined. Even in the case of a rare disease, there are enough examples to teach an algorithm how to recognise it. Because typing Chinese characters is more laborious than Western ones, people also tend to use voice-recognition services more often than in the West, so firms have more voice snippets with which to improve speech offerings.

中國的體量和多樣性為這樣的循環(huán)提供了強勁的動力。單靠開展日常生活，該國近14億人口就能產(chǎn)生大量數(shù)據(jù)，幾乎超過其他所有國家的總和。就算是某種罕見病，在中國也能獲得足夠的病例去訓練算法識別該疾病。由于輸入中國的漢字要比西方國家的字母費時費力，人們使用語音識別服務的頻率通常也高于西方國家，企業(yè)也就獲得了更多的語音片段來改善語音服務。

The Saudi Arabia of data

數(shù)據(jù)的沙特阿拉伯

What really sets China apart is that it has more internet users than any other country: about 730m. Almost all go online from smartphones, which generate far more valuable data than desktop computers, chiefly because they contain sensors and are carried around. In the big coastal cities, for instance, cash has all but disappeared for small purchases: people settle with their devices using services such as Alipay and WeChat Pay.

真正令中國與眾不同的一點是，它的互聯(lián)網(wǎng)用戶比任何國家都多：大約有7.3億人。幾乎所有網(wǎng)民都用智能手機上網(wǎng)，產(chǎn)生的數(shù)據(jù)比來自臺式電腦的數(shù)據(jù)有價值得多。這主要是因為智能手機包含傳感器，而且由機主隨身攜帶。以沿海大城市為例，在進行小額交易時幾乎已經(jīng)沒人使用現(xiàn)金了，通過手機上的支付寶和微信支付等服務就可以搞定。

Chinese do not seem to be terribly concerned about privacy, which makes collecting data easier. The country’s bike-sharing services, which have taken big cities by storm, for example, not only provide cheap transport but are what is known as a “data play”. When riders hire a bicycle, some firms keep track of renters’ movements using a GPS device attached to the bike.

中國人似乎不是特別擔心隱私的問題，這就使得收集數(shù)據(jù)變得更容易了。例如，風靡該國各大城市的共享單車服務不僅提供了廉價的交通方式，而且還是個所謂的“數(shù)據(jù)游戲”。一些公司會在用戶租用自行車時利用裝在車上的GPS設備追蹤租車者的活動。

Young Chinese appear particularly keen on AI-powered services and relaxed about use of their data. Xiaoice, an upbeat chatbot operated by Microsoft, now has more than 100m Chinese users. Most talk to it between 11pm and 3am, often about the problems they had during the day. It is learning from interactions and becoming cleverer. Xiaoice no longer just provides encouragement and tells jokes, but has created the first collection of poems written with AI, “Sunshine Lost Its Window”, which caused a heated debate in Chinese literary circles over whether there can be such a thing as artificial poetry.

年輕的中國人似乎尤其熱衷由人工智能驅(qū)動的服務，對于自身數(shù)據(jù)被使用也比較淡然。微軟運營的聊天機器人、陽光正向的小冰如今有超過一億個中國用戶，多數(shù)都是在晚上11點到凌晨3點和它聊天，通常都是談論自己在白天遇到的問題。小冰從互動中學習，過程中又變得更聰明。如今它不再只是給人鼓勵和講笑話，還運用人工智能創(chuàng)作出了首部詩集——《陽光失了玻璃窗》。中國的文學界為此展開了激烈的爭論，探討人工智能生成的詩歌算不算詩歌。

Another important source of support for AI in China is the government. The technology figures prominently in the country’s current five-year plan. Technology firms are working closely with government agencies: Baidu, for example, has been asked to lead a national laboratory for deep learning. It is unlikely that the government will burden AI firms with over-strict regulation. The country has more than 40 laws containing rules about the protection of personal data, but these are rarely enforced.

中國的人工智能還有一個重要的支持力量：政府。在該國當前的五年計劃中，科技是其中的重點。科技公司正與政府部門開展密切的合作，例如，百度已受命領(lǐng)導一個深度學習的國家級實驗室。政府也不大可能用過于嚴格的監(jiān)管增加人工智能公司的負擔。中國有40多部法律包含了個人數(shù)據(jù)保護規(guī)則，但很少執(zhí)行。

Entrepreneurs are taking advantage of China’s talent and data strengths. Many AI firms got going only a year or two ago, but plenty have been progressing more rapidly than their Western counterparts. “Chinese AI startups often iterate and execute more quickly,” explains Kai-Fu Lee, who ran Google’s subsidiary in China in the 2000s and now leads Sinovation Ventures, a venture-capital fund.

創(chuàng)業(yè)者正在利用中國的人才及數(shù)據(jù)優(yōu)勢。許多人工智能公司一兩年前才開始運營，但其中有不少公司發(fā)展速度都快過西方的同類。李開復解釋說，“中國的AI創(chuàng)業(yè)公司通常有更快的迭代及執(zhí)行速度。”他在本世紀初曾負責管理谷歌在中國的子公司，如今領(lǐng)導風險投資基金創(chuàng)新工場。

As a result, China already has a herd of AI unicorns, meaning startups valued at more than $1bn. Toutiao, a news aggregator based in Beijing, employs machine learning to recommend articles using information such as a reader’s interests and location; it also uses AI to filter out fake information (which in China mainly means dubious health-care announcements). Another AI startup, iFlytek, has developed a voice assistant that translates Mandarin into several languages, including English and German, even if the speaker uses slang and talks over background noise. And Megvii Technology’s face-recognition software, Face++, identifies people almost instantaneously.

結(jié)果是，中國已經(jīng)出現(xiàn)了一批人工智能獨角獸公司，即估值超過十億美元的創(chuàng)業(yè)公司。總部位于北京的新聞聚合器“頭條新聞”根據(jù)讀者的興趣及位置，利用機器學習來推薦文章，還運用人工智能過濾掉虛假信息(在中國，這類信息主要是各種可疑的養(yǎng)生資訊)。另一家人工智能創(chuàng)業(yè)公司科大訊飛開發(fā)了一款語音助手，可將普通話翻譯成包括英語和德語在內(nèi)的若干種語言，即使說話者話語間夾雜俚語或背景有雜音也沒影響。曠視科技的人臉識別軟件Face++幾乎可以瞬間將人識別出來。

Skynet lives

“天網(wǎng)”恢恢

At Megvii’s headquarters, visitors are treated to a demonstration. A video camera in the lobby does away with the need for showing ID: employees just walk in without showing their badges. Similar devices are positioned all over the office and their feeds are shown on a video wall. When a face pops up on the wall, it is immediately surrounded by a white rectangle and some text giving information about that person. In the upper right-hand corner of the screen big letters spell “Skynet”, the name of the AI system in the Terminator films that seeks to exterminate the human race. The firm already enables Alipay and Didi, a ride-hailing firm, to check the identity of new customers (their faces are compared with pictures held by the government).

在曠視科技的總部，公司向來訪者做了一番演示。大堂里的一個攝像機省去了展示身份證件的必要：員工不用出示工作牌就可進入公司。辦公樓各處都安置了類似的裝置，其所收集的數(shù)據(jù)顯示在一面視頻墻上。當墻上跳出一張人臉時，周圍立刻就出現(xiàn)一個白色方框，以及關(guān)于這個人的一些文字信息。屏幕的右上角，幾個大大的字母拼寫出“Skynet”(天網(wǎng))，也就是《終結(jié)者》系列電影中那個試圖消滅人類的人工智能系統(tǒng)的名字。曠視科技已向支付寶以及網(wǎng)約車公司滴滴提供支持，幫助它們核查新客戶的身份(將他們的面孔與政府持有的照片做對比)。

Reacting to the success of such startups, China’s tech giants, too, have begun to invest heavily in AI. Baidu, Alibaba and Tencent, collectively called BAT, are working on many of the same services, including speech- and face-recognition. But they are also trying to become dominant in specific areas of AI, based on their existing strengths.

眼見創(chuàng)業(yè)公司取得了成功，中國的科技巨頭也開始大舉投資人工智能。百度、阿里巴巴和騰訊(合稱為BAT)正在開發(fā)許多類似的服務，包括語音及人臉識別。不過它們也在試圖利用現(xiàn)有優(yōu)勢成為特定人工智能領(lǐng)域里的主導力量。

Tencent has so far kept the lowest profile; it established its AI labs only in recent months. But it is bound to develop a big presence in AI: it has more data than the other two. Its WeChat messenger service has nearly 1bn accounts and is also the platform for thousands of services, from payments and news to city guides and legal help. Tencent is also a world-beater in games with blockbusters such as League of Legends and Clash of Clans, which have more than 100m players each globally.

目前來看，騰訊最為低調(diào)，直到最近幾個月才建立起自己的人工智能實驗室。不過它擁有的數(shù)據(jù)比百度和阿里巴巴都要多，勢必會在人工智能領(lǐng)域占據(jù)重要地位。其即時通訊服務微信有近10億用戶，同時還是數(shù)千種服務的平臺，例如支付、新聞、城市指南以及法律咨詢等。騰訊還是全球最大的游戲公司，其大熱游戲《英雄聯(lián)盟》和《部落沖突》在全球的玩家都超過一億人。

Alibaba is already a behemoth in e-commerce and is investing billions to become number one in cloud computing. At a conference in June in Shanghai it showed off an AI service called “ET City Brain” that uses video recognition to optimise traffic in real time. It uses footage from roadside cameras to predict the behaviour of cars and can adjust traffic lights on the spot. In its home town of Hangzhou, Alibaba claims, the system has already increased the average speed of traffic by 11%. Alibaba is also planning to beef up what it calls “ET Medical Brain”, which will offer AI-powered services to discover drugs and diagnose medical images. It has signed up a dozen hospitals to get the data it needs.

已是電商巨頭的阿里巴巴正豪擲數(shù)十億元，力爭在云計算領(lǐng)域取得領(lǐng)先。6月于上海舉行的一次會議上，阿里巴巴展示了一項名為“ET城市大腦”的人工智能服務，可利用視頻識別來實時優(yōu)化交通。該服務運用路邊攝像頭拍攝的視頻片段來預測車輛的行為，即刻調(diào)整交通信號燈。阿里巴巴聲稱，在其總部所在地杭州，該系統(tǒng)使交通的平均速度提高了11%。該公司還計劃加強其“ET醫(yī)療大腦”系統(tǒng)，提供各種人工智能服務來幫助研發(fā)新藥和診斷醫(yī)學影像。為獲取所需數(shù)據(jù)，公司已與十幾家醫(yī)院簽約。

But it is Baidu whose fate is most tied to AI, in part because the technology may be its main chance to catch up with Alibaba and Tencent. It is putting most of its resources into autonomous driving: it wants to get a self-driving car onto the market by 2018 and to provide technology for fully autonomous vehicles by 2020. On July 5th the firm announced a first version of its self-driving-car software, called Apollo, at a developer conference in Beijing.

不過，自身命運與人工智能最休戚相關(guān)的還是百度，這在某種程度上是因為這項技術(shù)也許是它追上阿里巴巴和騰訊的主要機會。如今它將大部分資源都投入到了自動駕駛技術(shù)上，目標是在2018年前將無人駕駛汽車投放市場，到2020年為全自動駕駛汽車提供技術(shù)。7月5日，百度在北京的一次開發(fā)者大會上推出了其首款自動駕駛汽車軟件“阿波羅”。

Getting Apollo right will not only involve cars safely navigating the streets, but managing a project that is open to outsiders. Rivals such as Waymo, Google’s subsidiary, and Tesla, an electric-car firm, jealously guard their software and the data they collect. Baidu is planning not only to publish the recipe for its programs (making them “open-source”, in the jargon), but to share data. The idea is that carmakers that use Baidu’s technology will do the same, creating an open platform for data from self-driving cars—the “Android for autonomous vehicles”, in the words of Mr Lu.

阿波羅系統(tǒng)不僅涉及安全行駛汽車，還要管理一個向外部人士開放的項目。競爭對手如谷歌的子公司W(wǎng)aymo和電動汽車公司特斯拉都對自己的軟件和收集來的數(shù)據(jù)嚴防死守。而百度不僅計劃公開自己項目的“秘方”(用行話說就是將它們“開源”)，還計劃共享數(shù)據(jù)。百度的設想是，采用百度技術(shù)的汽車制造商也會效法，從而打造出一個無人駕駛汽車數(shù)據(jù)的開放平臺——用陸奇的話說，就是“自動駕駛汽車領(lǐng)域的安卓系統(tǒng)”。

Drive like a Beijinger

像北京人那樣開車

It remains to be seen how successful Chinese firms will be in exporting their AI products—for now, only a tiny handful are used abroad. In theory they should travel well: a self-driving car trained on China’s chaotic streets ought to have no problem navigating the more civilised traffic in Europe (in contrast, a vehicle trained in Germany may not get far beyond the first intersection in Beijing). But consumers in the West may hesitate to use self-driving cars that have been trained in a laxer safety environment that is more tolerant of accidents. Chinese municipalities are said to be falling over themselves to be testing grounds for autonomous vehicles.

至于中國公司將自己的人工智能產(chǎn)品推向海外時成績?nèi)绾危杂写^察。就目前而言，只有少數(shù)產(chǎn)品走向了海外。理論上說，這些產(chǎn)品的海外之旅應該很順利，比如在中國混亂的道路上訓練出來的自動駕駛汽車，在歐洲更文明有序的交通環(huán)境中行駛時應該也沒問題(相反，在德國受訓的汽車到了北京，過了一個十字路口可能就寸步難行了)。不過，中國的對道路安全的要求不如西方嚴格，對事故的容忍度更高，西方國家的消費者對在如此環(huán)境中訓練出來的汽車可能有所顧慮。據(jù)說，中國城市都在力爭成為自動駕駛汽車的測試場。

There is another risk. Data are the most valuable input for AI at the moment, but their importance may yet diminish. AI firms have started to use simulated data, including those from video games. New types of algorithms may be capable of getting smart with fewer examples. “The danger is that we stop innovating in algorithms because of our advantage in data,” warns Gansha Wu, chief executive of UISEE, a Beijing startup which is developing self-driving technology. For now, though, China looks anything but complacent. In the race for pre-eminence in AI, it will run America close.

還有一個風險。眼下數(shù)據(jù)還是人工智能最重要的因素，但數(shù)據(jù)的重要性也許會減弱。人工智能公司已開始使用模擬數(shù)據(jù)，包括來自視頻游戲的數(shù)據(jù)。新型的算法也許用較少的數(shù)據(jù)就能變得智能。趨勢科技的CEO吳甘沙警告說：“怕就怕我們會因為自己在數(shù)據(jù)上的優(yōu)勢而停止在算法上的創(chuàng)新。”這是北京的一家創(chuàng)業(yè)公司，開發(fā)自動駕駛技術(shù)。不過，目前來看，中國毫無自滿之意。在爭奪人工智能卓越地位的競賽中，中國將是美國的勁敵。

End.