Six Months After Taking Over Baidu, Lu Qi Lays Out His Plan to Remake It

Voice assistants moving into home appliances, and an "Android for the auto industry" — this is Lu Qi's plan to remake Baidu, six months into the job

It has been six months since the biggest personnel shake-up in Baidu's history.

In January this year, Baidu announced that Lu Qi, formerly an executive vice president at Microsoft, would become the company's senior president and chief operating officer, with every one of Baidu's business heads reporting to him.

Lu Qi replaced Robin Li as the company's de facto top executive. Since then, several of Baidu's businesses have been restructured, and a number of executives — Andrew Ng and Wang Jin among them — have left.

Robin Li brought Lu Qi in to put out fires. Baidu spent 2016 mired in scandal, with revenue stalling, profit falling, and its core advertising business shrinking for the first time since the company went public.

Last month, JD.com's market value came within a hair of overtaking Baidu's. Baidu's fall out of "BAT," the trio that has ruled the Chinese internet, looked like a formality away.

Lu Qi is the most experienced manager Baidu has ever recruited, having run core divisions at both Yahoo and Microsoft.

Six months into the job, he announced Baidu's brand-new plan to remake the company.

This past Wednesday, Baidu rented nearly all of floors one through three of the National Convention Center for an AI developer conference. More than 200 media outlets and over 4,000 guests and developers filled the venue.

"AI is Baidu's opportunity," Lu Qi said onstage. Baidu, he said, is an artificial intelligence company, and it will build an "AI ecosystem" — that is his plan for remaking Baidu.

The conference looked a lot like a product launch from Microsoft, Google or Apple. Lu Qi carried the whole event the way Satya Nadella, Sundar Pichai or Tim Cook would, rattling off announcements and introducing partners and colleagues.

Some notes were different, though: "China will surely have the chance to lead the world!" and "In three to five years, make China a powerhouse in the auto industry." Lu Qi did not talk this way in his earlier public appearances. It is unclear whether this is simply who he is, or whether Baidu has changed him.

The plan has two parts: intelligent driving, which Baidu has now opened up, and DuerOS, which grew out of the voice assistant Duer. Baidu also brought in a long list of "partners" — Nvidia, Intel, Ford, Daimler, Lenovo and others — to go "all in on AI" together.

The new plan says nothing about Baidu's most basic business, and its only profitable one: search advertising.

Voice Assistants Move Into All Kinds of Home Appliances

Baidu's first idea is to let everyone — "whether a large company, a small company, or even an ordinary developer — build their own smart voice-interaction device with a low barrier to entry." As Robin Li put it, Baidu's new mission is to go "from connecting information to awakening everything."

He means DuerOS, Baidu's new voice-technology platform. Jing Kun, general manager of Baidu's Duer Business Unit, called it "the Android of the AI era."

Duer began as a voice-assistant product under Baidu's search business, embedded in Baidu Search's mobile app to handle voice queries much the way Siri does. In January this year, the team built a lower-level platform, DuerOS, and worked with a startup to put out hardware. In February, Lu Qi, newly in office, promoted Duer to a full business unit.

Now DuerOS is meant to move into more hardware. Baidu is open-sourcing its development kit, so any developer can use Baidu's algorithms and technical resources free.

The chipmakers Nvidia and Intel are supplying hardware too — that is, chip "reference designs" — which means that, in theory, any hardware developer can use these tools to build a device carrying Baidu's voice assistant.

The pitch to developers is that DuerOS understands human speech better and is faster to deploy. Jing Kun demonstrated onstage that a developer could change just 17 lines of code and convert a Raspberry Pi running Amazon's Alexa voice assistant (English only) into a device running DuerOS's Chinese voice interface with a single click — "one minute to get your device into the China market."

The point was how easily DuerOS slots into hardware. Another example: an engineer modified a Lego robot, turning the toy into a "smart robot" that could hold a conversation.

Baidu also stressed that its speech recognition is 97 percent accurate, picking up a voice at a distance or in a noisy room, and answering questions that depend on what came before. You can ask, in sequence, "Find me a popular home-style restaurant in Beijing," then "Is there one near the Lama Temple?" then "Does the second one have Wi-Fi?"...

To show how much it has put into semantic training, Baidu announced it had acquired an American startup called KitAi. The company's CTO came onstage to demonstrate its technology, a tool for quickly training "wake words." Developers can customize the phrase that starts DuerOS — "Hey Siri" is a wake word, the phrase a device listens for before switching on — and then load it onto their hardware.

Robin Li said: "Our competitors never acquire technology companies. Only Baidu is willing to pay for technology."

Lenovo, HTC, Xiaomi and others are already working with Baidu on hardware. We saw some of the results in the exhibition area at the AI developer conference.

Lenovo used DuerOS to build a television with a microphone in the remote, so voice replaces most of the typing: "search for popular recent shows," "I want to watch Nirvana in Fire," "fast-forward 10 minutes." Another feature runs on image recognition — when a celebrity appears onscreen, you can ask "who is this?" and a short bio pops up in the bottom-right corner.

Haier has a smart refrigerator that uses Baidu's voice assistant, though its features look duller. The fridge has a large screen, and you can speak into the microphone to ask for recipes, order fresh groceries, adjust the temperature inside. A Haier employee at the event told us the company will eventually put a camera in the cooling compartment as well, giving the fridge image recognition so it can tell on its own how much food is inside and what it is.

The startup Armoto made an educational toy for children that uses DuerOS's speech recognition to let the machine talk with people — you can ask it to play a song, put a question to it, have it raise the cannon on its back or close the visor on its helmet.

Baidu showed more than a dozen such products, but none of them talk to each other. That is not quite what Amazon Echo, Google Assistant and Apple's AI plans are after; all three want a hub — a speaker or some other control device — connected to the appliances and fixtures around your home.

How to make devices work together, how to build situations that would make a smart home genuinely useful: Baidu said nothing about either. What Baidu can do right now amounts to changing the way you already operate a piece of hardware into voice.

Baidu's Driverless Platform, Apollo, Also Wants to Be the "Android of the Auto Industry"

Self-driving cars are not new territory for Baidu. A year and a half ago the company put out a press release saying a self-driving car it had developed with BMW had begun road testing. Baidu's test car then was a red BMW 3 Series GT, with a lidar unit sitting conspicuously on the roof.

At this conference, boss Robin Li took the stage himself with what looked like the most important reveal: a self-driving car being tested on public roads.

On the big screen, President Lu Qi held a video call with Robin Li, who was sitting in a Jeep Cherokee. Li rode in the front passenger seat of the red Cherokee test car; the engineer in the driver's seat was not holding the wheel. Li said the car had just gotten onto the Fifth Ring Road and was heading to the venue.

Lu Qi said the car was developed jointly by Baidu and Bosch. Li said it drove smoothly even on the heavily trafficked Fifth Ring.

But the way the car actually behaved on the road was worrying. In under a minute of footage, it changed lanes on the Fifth Ring with a following car right behind it, then changed lanes twice in a row — one of those merges crossed a solid white line. It never signaled once.

Had a traffic officer been watching, Baidu's driverless car would have collected three penalty points and a 200-yuan fine inside that minute. Changing lanes twice in a row is not illegal, but it would certainly have drawn curses from other drivers.

Traffic rules in mainland China do not currently allow self-driving cars on public roads, and it is unclear whether Baidu filed for approval before this test. Staging a demonstration like this for marketing is not good news for a young industry already facing regulators.

Lu Qi announced that in September the Apollo platform will open up simulation data for fixed-lane autonomous driving, delivering partial self-driving capability, with fully driverless capability around 2020. Baidu said it would work with roughly 50 companies, more than a dozen of them automakers.

Baidu has changed the direction of its autonomous-driving research. The car that appeared this time had no lidar on the roof; instead, a camera-like component was mounted behind the windshield.

During a later demonstration, a Baidu staff member told QDaily that the car drove itself using visible-light radar alone.

It looks like Tesla's approach.

The hardware costs less than the lidar approach. Lidar units run to tens of thousands of dollars apiece; judging by Tesla's prices, a full set of visible-light-camera self-driving hardware costs a few thousand.

Last year, Baidu's self-driving cars did carry lidar on the roof. In August last year, Baidu and Ford together invested $150 million in Velodyne, the best-known lidar company.

Judging by yesterday's demonstration, which Robin Li endorsed in person, the lidar route has been dropped.

Baidu also offered rides in a "trajectory-following self-driving car" outside the venue. This sort of self-driving car needs a human to drive the route once first, while the system uses GPS and 4G to record the vehicle's path and speed; after that a computer can steer the car through exactly the same drive.

With this demonstration, Baidu wanted to show how easy the Apollo platform makes it to develop self-driving technology.

But trajectory-following driving proves at most that a vehicle's automatic controls work. For real autonomous driving it is not much use.

Baidu Is More Practical Than Before, But What It Is Doing Isn't Any Simpler

Baidu had never held an artificial intelligence conference before; the annual event where it describes the future is "Baidu World," each September.

Lu Qi was not yet at Baidu for last year's Baidu World, but the company had already made artificial intelligence a core strategy. Robin Li stood onstage and walked through eight specific applications in detail: a sales system that understands what people say and offers advice; an algorithm that mimics Leslie Cheung's voice; an augmented-reality marketing plan; voice translation to help people hail cabs; user analytics to help clients sell more movie tickets...

Every demo looked like it could make Baidu money. Not one of them went anywhere.

Drawing a pie that never comes out of the oven has been the house style of Baidu's conferences these past few years — there were also chopsticks that detect gutter oil, a driverless bicycle, AR glasses, all of which came to nothing.

This year Lu Qi did not bring up any of last year's items. In fact he said nothing at all about how to make money from artificial intelligence. That is probably because Baidu's AI technology is still early, and the more urgent question is how to get people using it in the first place.

Before a smart home can have a platform, it needs people using it. That is why the smart-home platforms of Google, Apple, Samsung and the rest have gone nowhere, and why they are all now following Amazon into smart speakers — a smart speaker can at least do one thing well, which is play music.

This time Baidu has stopped talking about a platform. It is selling appliance makers a voice assistant that is easy to integrate, and hoping those partners can find a smart home people will want to use.

What Lu Qi has brought is a lower-level, more elementary development plan. The open-source tools and documentation are already on GitHub, and the driverless cars and smart hardware at the conference could be tried and ridden in rather than standing there as empty display shells. That is a positive change.

As Lu Qi said onstage, what Baidu wants to build is an artificial intelligence like Android, and hand it to its partners.

The doubt is just as obvious: very few things appear on their own once a platform exists. Hardware least of all.

Looking at Android's success today, it is easy to forget how it began. Google launched Android in 2007 and pulled phone makers into the Open Handset Alliance (OHA). No Android phone appeared that year.

The first real Android phone, the HTC Dream, was not the product of HTC taking Google's system and building something on its own. It came out of rounds of negotiation between the two companies, more than a year of overtime by both engineering teams, and one complete teardown and restart.

Platforms are all built this way. Earlier still, Microsoft built Windows on top of an existing PC industry, not the other way around.

Before Baidu's voice assistant can be everywhere, whether it copies Amazon Echo or does something else, Baidu has to build a piece of hardware consumers want to use.

Cars are the same, and perhaps worse. Unlike the smart home, where nobody has found a clear direction, self-driving technology is already on the road.

Tesla has sold tens of thousands of cars that can largely drive, avoid obstacles and steer themselves on public roads, and next year a software update will give them fuller autonomous-driving capability.

Tesla not only has more cars on the road and years of collected data, it has a clear business model. Self-driving is a major selling point, and an owner pays $5,000 to switch on a car's autonomous-driving feature (rising to $6,000 if activated afterward; in China, 48,700 yuan and 58,000 yuan) — as much as Baidu earns selling several hundred thousand ads.

Fortunately for Baidu, Tesla does not plan to let other carmakers use its self-driving technology. Carmakers need a self-driving solution, and that is the opening Baidu is counting on.

The first large company to work on this was Google. It started earliest and has the most advanced results so far: a self-driving fleet now up to 300 vehicles, and rides opened to ordinary members of the public.

Google's internet services cannot be used in China. Its car technology is another matter. Self-driving technology has to work offline; what matters most is reading the road and controlling the car, not internet services.

Autonomous driving is not a "feature." It is a subsystem of the whole car. Beyond computation, it has to turn its results into actual operations — accelerating, braking, turning, switching on the indicators. In other words, it is hard to make an ordinary car into a dependable self-driving car by bolting features onto the outside; autonomous driving has to be designed in from the start.

It is a core component as important as the engine, not a third-party service. Lives depend on it, so carmakers need extensive testing — which means supporting several self-driving systems would cost enormously.

Most cars sold in China are built by domestic joint ventures but designed abroad. Google's self-driving system is unlikely to be kept out at the border, just as every phone maker uses Android. To push its self-driving system, Baidu will have to compete with companies like Google. Beating domestic rivals, as it did in search, is no longer enough.

Test cars, algorithms and data are all indispensable to autonomous driving. Test cars can be bought — Baidu has deals with Bosch and Nvidia, so buying the compute modules and other parts is no problem. Data takes years to accumulate. Google has been driving in Mountain View for at least six years; Baidu is registered with California's DMV to test self-driving cars too, but its test cars have not been seen on the road.

Judging by QDaily's earlier visit to CES Asia and this test ride, the demo cars Baidu is willing to show in public run only in closed environments, on roads with freshly repainted, clear markings, free of pedestrians and other vehicles, at what looked like under 30 kilometers an hour. The cars on display still drive themselves less capably than the demonstration we saw at Ford's lab.

Is There Still Time for Baidu?

After the open AI strategy was announced, Baidu's Nasdaq-listed shares rose as much as 3 percent — a positive reception for the new story at a moment when tech stocks were falling overall. Deep partnerships with big companies like Nvidia and Intel clearly made that story more appealing.

Baidu has reached the point where it has no choice but to change.

After the medical-advertising problems broke out in 2016, followed by regulatory scrutiny and personnel turmoil, Baidu's revenue stalled. Its advertising business began shrinking in the third quarter of 2016, the first decline since Baidu went public in 2005. More than 90 percent of Baidu's revenue comes from advertising.

The problem has not stopped. Advertising revenue in the first quarter of this year was $2.14 billion, 1.3 percent lower than the same period in 2016, while Baidu's overall operating profit fell 9.3 percent. It had 23.2 percent fewer advertising clients than a year earlier.

As the traffic gateway of the PC era, Baidu never carried its successful advertising model over to the phone. People now open apps directly, WeChat above all. The national apps with huge user bases have become the new advertising giants — Toutiao, for one.

Google faces slowing ad growth and falling prices per ad as well. Unlike Baidu, though, Google has several mobile services used every day by more than a billion people: Gmail, Android, YouTube. Google also does not lean on medical advertising, and its ad revenue is still growing.

That is exactly why Google can keep the blood flowing to its artificial intelligence story.

Lu Qi is rebuilding his team around artificial intelligence in a direction no different from Google's, and it may also be the thing he is best suited to do. Early in his career he did engineering at Yahoo, managing teams of thousands and presiding over the birth of Hadoop, the foundational technology of the big-data industry, at Yahoo. He then went to Microsoft to run Bing, managing an even larger engineering team.

The difference is that Yahoo then, Microsoft later, and Google — the company Baidu now measures itself against — all had plenty of profit to fund technology research.

None of them had to hurry, to rush their future technology into money. That was especially true at Microsoft, where the Office division alone made over $10 billion in net profit a year, more than five times the profit of all of Baidu.

Baidu cannot expect that kind of luxury.

And whether it is Apollo in cars or DuerOS in the smart home, both are still laying the first infrastructure and using open source to attract developers. Nobody knows what year they start making money.

Baidu is betting its future on artificial intelligence. Is there still time?

Business Adjustments and Executive Departures Since Lu Qi Joined Baidu

1.19

Lu Qi becomes president and chief operating officer of Baidu Group.

2.08

Baidu shuts down its medical division.

2.16

The Duer Business Unit is created, reporting directly to Lu Qi.

2.16

Baidu fully acquires the startup Raven Tech; founder Lyu Cheng becomes general manager of Baidu's smart-home hardware unit.

3.01

The Intelligent Driving Group (IDG) is formed, merging the former Autonomous Driving Division, Intelligent Vehicle Division and connected-car business.

3.16

Zeng Liang, general manager of Baidu Nuomi, is fired for disciplinary violations.

3.22

Andrew Ng, Baidu's chief scientist and head of Baidu Brain, resigns.

3.27

Wang Jin, senior vice president of Baidu, announces his resignation.

4.28

Li Xinzhe, Baidu's chief financial officer, moves to become CEO of Baidu Capital.

5.26

Lu Fubin, vice president in charge of Baidu's content ecosystem, announces his resignation.

Cover image from Baidu