Skip to Content

A Chinese Internet Giant Starts to Dream

Baidu is a fixture of online life in China, but it wants to become a global power. Can one of the world’s leading artificial intelligence researchers help it challenge Silicon Valley’s biggest companies?
August 14, 2014

Punk bands from Blondie to the Ramones once played in Broadway Studios, an age-worn 95-year-old neoclassical building surrounded by strip clubs in San Francisco’s North Beach. But early on this bright June morning, a different sort of rock star arrives. A small crowd attending a tech startup conference swarms around a tall, soft-spoken man in a blue dress shirt and navy suit who politely poses for photos. Andrew Ng, newly appointed chief scientist at Baidu, China’s dominant search company, is here to talk about his plans to advance deep learning, a powerful new approach to artificial intelligence loosely modeled on the way the brain works. It has already made computers vastly better at recognizing speech, translating languages, and identifying images—and Ng’s work at Google and Stanford University, where he was a professor of computer science, is behind some of the biggest breakthroughs. After his talk, the audience of about 200 entrepreneurs, venture capitalists, and tech workers sends him off with two rounds of applause.

The avid reception helps explain why Baidu has made Ng, 38, the linchpin of an effort to transform itself into a global force. The company hired him in May to head its research organization, which includes a new artificial-intelligence lab in Silicon Valley and two labs in Beijing, one focused on deep learning and the other on large-scale data analysis. Often called China’s Google, the company plans to invest $300 million in the new lab and a development office on the same floor over the next five years. Ng (it is pronounced “Eng”) aims to hire 70 artificial-intelligence researchers and computer systems engineers to work in the new lab by the end of 2015. “It will really target fundamental technology,” says Kai Yu, the director of Baidu’s Beijing deep-learning lab, a friend of Ng’s who urged him to join the company.

Baidu, which hopes to get half its revenue from outside China by 2020, is just one of several large Chinese Internet companies now looking abroad for talent and customers, seeking to make the most populous nation on earth more than just the world’s factory. With 632 million citizens online, China claims four of the planet’s 10 most-visited Internet properties, up from just one a year ago. The top 20 Chinese Internet companies listed on public exchanges outside mainland China have a combined market value of about $340 billion. The social-networking giant Tencent, whose WeChat mobile messaging service has 100 million registered users from outside China, accounts for almost half of that. And in September, the e-commerce group Ali­baba was expected to complete what could be the world’s largest initial stock offering ever. Its debut on the New York Stock Exchange could value it at $150 billion.

Andrew Ng hopes to lure AI talent to Baidu’s new Silicon Valley research lab.

As they look beyond China, Baidu and other Chinese companies find themselves on a collision course with the established U.S. Internet leaders. It’s unlikely that companies such as Google, Facebook, and Amazon will be in danger in Western markets anytime soon. But the field is wide open in much of the rest of the world, where billions of people aren’t yet online. Here, companies like Baidu believe they have an advantage because of their experience with customers who are relatively new to the Internet, says Jixun Foo, a managing partner at the venture capital firm GGV Capital and an early investor in Baidu at a previous firm. “Chinese companies are starting to dream,” he says.

Cool Things

The first thing you notice about Andrew Ng is his voice. Extraordinarily gentle, it is almost a whisper, and his speech carries traces of his birth in London and childhood in Hong Kong and Singapore. As he patiently explains the nuances of deep learning, he sounds as if he’s reading a bedtime story to a child. At times, he’s scarcely audible above the clack of billiard balls as engineers on a break play pool in Baidu’s still largely empty Silicon Valley lab, a 15,000-square-foot office space in Sunnyvale, a few minutes southeast of Google’s headquarters. But when Ng turns to his mission at Baidu, his voice rises above the background noise.

Maybe that’s because the mission is a grand one: to change the world with artificial intelligence. Ng says he will focus on projects that could “significantly influence” the lives of at least 100 million people. That, he adds pointedly, means more than creating “shiny” apps that rise and fall on the whims of teenage fashion. “Who knows who’s going to be the next—boy, I’m even losing track—Snapchat?” he says in a rare flash of snark. “When you build some of the hard technologies that companies like Baidu try to, it gives you a more lasting base to build on.”

Ng’s work on artificial intelligence has shaken up a major search company before. He is best known for a project referred to as the Google Brain, which he helped set up inside the secretive Google X research lab in 2011. The project was designed to test the potential of deep learning, which involves feeding data through networks of simulated brain cells to mimic the electrical activity of real neurons in the neocortex, the seat of thought and perception. Such software can learn to identify patterns in images, sounds, and other sensory data. In one now-famous experiment, the researchers built a “brain” with one billion connections among its virtual neurons; it ran on 1,000 computers with 16 processors apiece. By processing 10 million images taken from YouTube videos, it learned to recognize cats, human faces, and other objects without any human help. The result validated deep learning as a practical way to make software that was smarter than anything possible with established approaches to machine learning. It led Google to invest heavily in the technology—quickly moving the Google Brain software into some of its products, hiring experts in the technique, and acquiring startups (see “: Deep Learning,” May/June 2013).

Baidu is one of many ­Chinese Web ­companies on a ­collision course with Internet leaders such as Google, ­Facebook, and Amazon as they look abroad for new customers.

Ng, who calls deep learning a “superpower,” will build a new generation of such systems at Baidu. Services that may result remain in the brainstorming stage, but he will hint at what they may be. He dreams of a truly intelligent personal digital assistant that puts Apple’s Siri to shame, for example. Looking further ahead, the technology could transform robotics, a pet subject for Ng—his engagement photos were taken in a robotics lab—and make autonomous cars and unmanned aerial vehicles much more capable. “We’re going to do some cool things here,” he says with a grin.

They’ll have to if they are to compete: Google, Facebook, Microsoft, and others have been hiring lots of deep-­learning experts for their labs, sometimes even from each other. And Baidu still has a lot to prove. Fairly or not, it has the reputation many Chinese companies do for copying the products and business models of U.S. Internet leaders. It’s a process cynics dub C2C—“copy to China.” Baidu has seemingly tried to emulate Google in countless ways over the years, from its spare search homepage to a head-mounted computer, Baidu Eye, that looks a lot like Google Glass. Baidu has even begun working on self-driving cars. With its new star hire, it appears to be following Google’s lead once again.

Ng insists that the C2C stereotype is no longer accurate, particularly for his new employer. “I used to work for the USA’s Baidu,” he jokes. Then he picks up his phone and says in English, “Please call a taxi for me.” A moment later, Baidu’s translation app utters the same phrase in Mandarin Chinese and shows the equivalent ideograms on the screen. It’s slick—but is it better than Google’s translation app, which appears to do the same thing? That’s not clear. It’s Ng’s job to develop cutting-edge technologies that will leave no doubt who is ahead.

Out into the World

Baidu’s Silicon Valley lab is led by Adam Coates, a 32-year-old who stumbled into artificial intelligence quite by accident. As a Stanford computer science student in 2002, he got talking with Ng, who mentioned that he was working on a project involving remote-controlled helicopters. Coates had built and flown them while at high school in California’s Napa Valley resort town of Calistoga. Ever since, the two have done research together, writing papers on using machine learning for unmanned helicopters, household robots, and image recognition. When Ng left Stanford for Baidu, Coates, then a postdoctoral researcher in Ng’s lab, followed. By then, he had begun to see that machine learning would be crucial to just about everything. “It doesn’t matter whether you’re really excited about language or helicopters,” he says. “You can use it to solve any problem.”

Ng and Coates have one key quest for their new lab: creating software that can, in a real sense, learn on its own. Until recently, most improvements in areas like speech and image recognition came by training software with data that had been laboriously labeled. For example, teaching software to spot cats would require a database of thousands of images, with any cats identified by humans. You don’t have to be an artificial-intelligence expert to see the main drawback of that approach, known as supervised learning. No human child needs to see 50,000 labeled images to recognize a kitty. “We wander around the world and see how things work,” Coates says. “The hope is that we can find algorithms that learn the same way.” Deep-learning systems might still need to see a lot of cats to spot one on their own, but they can be much more useful because they need minimal human help.

Software intelligent enough to understand the images, text, and sound in our lives could make decisions for us and take on jobs such as answering simple e-mails.

Software smart enough to understand the images, text, and sound in our lives could use that information to make decisions on our behalf—and transform our relationship with technology, says Coates. For instance, it might analyze your vacation photos and recognize the people shown in each one, identifying what they’re doing and recognizing landmarks. Then you could find an old shot later by asking for, say, “photos of Mom on the beach.” Or you could snap a photo of a shirt with your phone and ask it to find others like that, trusting that instead of just seeing an arrangement of colored pixels, it would apply an understanding of clothing styles, fabric, and your personal taste. Ng envisions our cell phones being able to recognize speech as well as humans can, so you could at last reliably dictate text messages even in a noisy car. He hopes to see e-mail apps that can learn from your interactions with friends and colleagues and then start answering some simple messages on your behalf. Looking further ahead, Ng and Coates may also get a chance to continue their research on robotics, says Yu. “We’re not only interested in cyberspace but physical space,” he says.

First, however, the Baidu lab in Silicon Valley will try to make it easier to test out deep-learning software, which requires enormous computing power. Training a new speech recognition model can take a week or more, a period Ng would like to cut in half. Last year Coates led a Stanford team to a breakthrough that makes that goal realistic. They built a neural network that roughly matched the Google Brain system for a 50th of the cost—only $20,000—using off-the-shelf graphics chips from Nvidia. That approach could help Baidu get powerful deep-learning infrastructure running at relatively low cost. And it fits well with the company’s existing work in Beijing, where simpler clusters of graphics chips have already been used to train deep-learning systems for image and speech recognition.

Air of Mystery

Walking around Baidu’s headquarters along the technology corridor in Beijing’s Haidian district, you might be excused for thinking you had somehow teleported to the fabled Googleplex in Mountain View, California. Free cafeteria? Check. On-site gym? Check. Sleeping pods? Check. Jeans and shorts, T-shirts, flip-flops? Check, check, check. About the only thing breaking the illusion is a giant Baidu bear-paw logo sculpted into the lobby ceiling. It all seems to reinforce the C2C stereotype Ng and others try so hard to quash. And Kai Yu happily boasts that the similarities to U.S. Internet companies are more than skin deep. Like them, Baidu favors flat management, small teams, fast product cycles—and, he adds, his whole face brightening, cool technologies. “Baidu is not so different from a Silicon Valley Internet company,” says Yu, who ought to know: he spent six years working at NEC Labs America in Cupertino, two miles from Apple headquarters.

Dig into the history of Baidu, however, and you’ll find it has Valley roots of its own. CEO Robin Li cofounded the company in 2000 with biotech salesman Eric Xu, after a stint as an engineer at the Sunnyvale-­based search engine Infoseek. Li was armed with a patent for a way to rank sites in search listings by the number of incoming links—filed in 1997, a year before Google cofounders Sergey Brin and Larry Page patented their similar PageRank algorithm. As China’s Internet population grew, so did Baidu, enough to attract a $5 million investment in 2004 from Google itself—which later tried to buy Baidu for $1.6 billion in an attempt to head off the Chinese company’s IPO, according to Bloomberg Businessweek. Instead, Baidu went public in August 2005, and shares rocketed 354 percent the first day. Much as Google had done in the United States, Baidu quickly solidified its hold on China’s search market and used the profits to expand into a range of other online services.

Baidu has even beaten back Google, albeit with what some observers believe is an assist from China’s government, which blocks access to many Google services inside its borders. And the Chinese company has continued to invest in new ideas, according to early investor Jixun Foo. “Baidu has put a lot of emphasis on the underlying technology, compared with Tencent and Alibaba,” he says. That doesn’t mean its products are all unique: it offers many Google analogues, including maps, a browser, and cloud storage. Hiring Ng might seem to be another “me too” move. But the company had already invested heavily in deep-learning research and achieved results that rival—perhaps even exceed—Google’s.

For instance, the Baidu Translate app has a feature that can, in seconds, identify an object in a photo and name it in written and spoken English. The company’s mobile search app can understand what’s depicted in a photo snapped on your phone and then find images that are similar. Rather than simply matching colors and patterns, the app knows, for example, whether a photo shows a church or a soccer team. At conferences, Yu likes to demonstrate how that feature beats a comparable one from Google. One slide shows that Baidu found photos similar to one of a dog with a bow on its head. Google returned mostly photos of scantily clad women.

Cherry-picked comparisons aside, the technology has paid rapid dividends for Baidu. In November 2012, only four months after Yu opened his lab in Beijing, the company began using deep-learning technology for voice search. Speech recognition errors fell by a quarter. A similar change helped reduce errors in optical character recognition by almost a third. That made its translation app much better at decoding things like restaurant menus, says Haifeng Wang, Baidu’s VP in charge of machine translation.

Yu’s neural networks have even boosted Baidu’s bottom line. One system learns which qualities of an ad make people click on it more often, selects ads to meet those criteria, and runs them at the most opportune moments. That lets Baidu charge higher prices. Li told investors in April that the technology had helped lift first-quarter profits and revenues.

Still, like Google, even a growing, profitable Baidu faces constant challenges from smaller upstarts and established rivals. Most concerning for the company, its comfortable lead in search has declined in the past year. Baidu’s share of searches made in China on desktop computers fell from 80 percent to 75 percent, according to Bloomberg Intelligence. New challenger, launched in 2012 by the mobile software firm Qihoo 360, now has 16 percent of desktop searches, up from 10 percent a year ago.

The rapid shift of Internet usage to mobile devices, a change that’s bedeviled many established U.S. Internet companies, has been particularly dramatic in China, where many people now get their first taste of online life on a smartphone, not a PC. Some 83 percent of people in China now use a mobile device of some kind to get online, and Baidu was caught flat-footed. In the past year, it has moved quickly to reverse the misstep by paying carriers to distribute its mobile apps, spending $1.9 billion to buy the Chinese app distributor 91 Wireless, and redesigning its services and ad formats to work better on phones. All that helped boost the average number of daily users of Baidu’s mobile search app to 160 million in the first quarter, up from 130 million six months before. But Baidu must constantly battle native mobile companies and apps to stay relevant.

Breakthroughs from Ng and his researchers might help. The sweeping transition from traditional computers to smartphones and other mobile devices has produced an explosion of sensory data such as images, video, and sound—the kind of data that stumps conventional software but that Ng has shown deep learning can comprehend. His new employer sees an opportunity to leap ahead of its mobile competitors with services that can understand the world.

“Just as the Industrial Revolution freed a lot of humanity from ­physical drudgery, I think ­AI has the ­potential to free humanity from a lot of the mental drudgery.”

The same technology might also help Baidu win over many of the planet’s five billion people who are not already online and are unaccustomed to the computer technology the developed world has had for 20 years. They will use mobile devices before—likely to the exclusion of—anything else, and deep learning could provide intuitive interfaces that will be attractive to computing beginners. Those newcomers to the Internet—like all of us, really—will not want to learn new modes of interaction, says Ng. They will prefer to speak naturally to their devices to get the information or translation they want. This type of technology might also help Baidu tailor its search results and apps to different languages and locations. That’s something the company has struggled to do, limiting previous efforts to expand outside China. A foray into Japan in 2008 went nowhere because Baidu’s search engine failed to cater to local needs. For now, the company has picked a few less-develop­ed regions to focus on: Southeast Asia, the Middle East and North Africa, and Latin America. It launched a search engine for Brazil in mid-July.

For Baidu, becoming more global is also crucial to its ambition to be a leader in technology. Many people outside China, especially in the West, know little or nothing about the company. That Baidu retains an air of mystery among foreigners was apparent at a cocktail party it hosted for the International Conference on Machine Learning, a prestigious annual gathering of artificial-intelligence experts that was held in Beijing for the first time this past June. Jet-lagged researchers from Google, Microsoft, Facebook, and leading universities mingled in the top-floor Happiness Lounge of the 21-story Pangu 7 Star Hotel, taking in dramatic views of the Bird’s Nest stadium and Olympic Park. Some said they hadn’t heard of Baidu until a couple of years ago and started paying close attention only when Ng joined.

A Baidu employee walks past the “space capsules” where workers can rest at the company’s headquarters in Beijing.

That lack of attention from foreigners is part of a larger problem for Baidu. Its inwardlooking culture and the Chinese technology industry’s reputation for unoriginality limit the company’s ability to compete with Google and other U.S. tech leaders, whose workforces are drawn from around the world. Earlier attempts to change that culture have stumbled. Yong Liu, who left Baidu in January after a short stint as its Silicon Valley–based director of open innovation and partnerships, says he was surprised to learn how Chinese-­centric the company was. He joined a small Silicon Valley lab that Baidu opened in 2013 and found that all of the 30 or so senior engineers and research scientists there were Chinese. “The purpose of a Silicon Valley R&D lab is to attract the best talent, not just the best talent from a minority ethnic group,” he says. Baidu’s leaders concede the point. “We’re making efforts to become a more cosmopolitan company,” says Kaiser Kuo, director of international communications. By rebooting the lab Liu joined, hiring Ng and Coates, and expanding its size and scope, they hope to make Baidu’s research groups—and eventually the rest of the company—more diverse. For all that executives bristle at comparisons to Google, they are actively trying to act a little more like the icon of Silicon Valley on the global stage.

Culture Shift

Back at the Silicon Valley lab, Andrew Ng is trying hard to embrace his dual role as cultural catalyst and technical visionary. He used to have no patience with people who talked about what he regarded as the “fluffy stuff” of organizational culture. Now he can’t get enough of it. His favorite book lately, to his mild embarrassment, is Eric Ries’s The Lean Startup, a management handbook for entrepreneurs. He has also turned to serial entrepreneur and startup guru Jerry Kaplan, who says Ng grilled him for advice on hiring engineers and rallying them behind a mission, and held staff meetings to discuss hiring and lab culture. “Now that I’m older, I really appreciate culture and the importance of being thoughtful about it,” says Ng.

With his global upbringing, Ng makes a good nucleus for a more diverse research group, says Sebastian Thrun, a Stanford research professor and Google Fellow who started the company’s driverless-car project. And Ng is open about the fact that he is a crucial talent magnet for Baidu. He already seems to be attracting a very different kind of person. Among the new hires is Bryan Catanzaro, a graphics chip architect and former Nvidia research scientist—the kind of Berkeley-educated Silicon Valley technologist who otherwise might have joined Google, Facebook, or a hot startup. Ng says he also aims for Baidu Research to be “a little bit porous,” sharing ideas with other researchers and the software developer community and becoming as embedded in the Silicon Valley community as its American rivals are. “There’s an opportunity to create a culture that’s great for research and great for changing the world,” he says.

If Ng’s plans work out, the world will indeed change in some ways. Baidu will have proved that China’s Internet companies can do more than just follow those from the U.S. And perceptive computers will have taken over many tasks we humans must do for ourselves today, perhaps freeing our minds for more creative activities. “Just as the Industrial Revolution freed a lot of humanity from physical drudgery, I think AI has the potential to free humanity from a lot of the mental drudgery,” Ng says. It’s a Google-worthy goal. But to pull it off, Baidu must chart a path indisputably its own.

Contributing editor Robert D. Hof wrote about neuromorphic computing in May/June. Christina Larson contributed reporting from Beijing.

Keep Reading

Most Popular

How to opt out of Meta’s AI training

Your posts are a gold mine, especially as companies start to run out of AI training data.

The return of pneumatic tubes

Pneumatic tubes were supposed to revolutionize the world but have fallen by the wayside. Except in hospitals.

Why does AI hallucinate?

The tendency to make things up is holding chatbots back. But that’s just what they do.

How a simple circuit could offer an alternative to energy-intensive GPUs

The creative new approach could lead to more energy-efficient machine-learning hardware.

Stay connected

Illustration by Rose Wong

Get the latest updates from
MIT Technology Review

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

Explore more newsletters

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at with a list of newsletters you’d like to receive.