2017 could very well be the year of the robot — not the T-1000 or even the real-life anthropomorphic monstrosities that garner viral headlines every few months, but rather stand-alone living-room devices imbued with AI smarts. Google Home and Amazon’s Echo are the most famous, but a whole raft of these gadgets is preparing to flood the market. One of the most advanced will likely come from Baidu, the Chinese tech giant that, like Google, began as a search engine and now has its tendrils in all sorts of digital and physical spaces. Andrew Ng, Baidu’s chief AI scientist, calls these devices “conversational computers,” and he’s a key reason any of them have learned to talk in the first place.
A former AI researcher at Stanford, Ng is best known for spearheading the Google Brain initiative, an ambitious artificial-intelligence project that helped advance Silicon Valley’s understanding of deep-learning techniques. Instead of being programmed to respond to specific actions, a deep learning system is fed massive amounts of data from which it is able to discern patterns over time, loosely mimicking how the human mind absorbs information. Ng’s system at Google famously figured out what a cat looks like after scanning millions of online images.
Deep-learning technology has traveled from the ivory tower of academia to the research labs of Silicon Valley to, now, our living rooms. Baidu is one of the many tech companies unveiling AI-powered robot assistants at this year’s CES in Las Vegas. The company’s new device, a vaguely humanoid-shaped gadget called Little Fish, plays music, dictates the news, and answers trivia, like the other assistants on the market, but it also boasts a small touchscreen. Ng says the screen will help Little Fish deliver certain kinds of information (like a restaurant menu) much faster than screenless competitors. I talked with Ng about how Chinese and American digital assistants differ (Little Fish speaks only Chinese languages for now), why we don’t have to be scared of killer robots, and the possibility of working in a hostile political environment. The conversation has been edited for length and clarity.
A lot of the other home assistants don’t have a screen. Why does this one have a screen?
Ng: The fastest way for a computer to get information to a human is through vision. There are a lot of scenarios where something displayed visually would be more efficient. If I ask you, Victor, what is a hammerhead shark, how are you going to explain that with words? With a picture you get it right away.
If you’re hungry and you want food delivered, you can have a machine read out a list of the 10 nearest restaurants, then say, “restaurant number three,” and then have it read out all 50 items on the menu, and then say, “OK I want item number 47.” You can do that; it’s just relatively inefficient. A screen is a very efficient way for you to quickly browse a lot of restaurants, browse a lot of dishes, see pictures and select what you want.
Do people interact with an assistant differently when it’s on their phone and in their pockets compared to when it’s a home robot?
You can have a lean-back experience when you have these home robots. Picture yourself at home leaning back on your sofa. It’s really nice when you can just call out what you want, such as, “Little Fish, I want a bottle of water,” and just have it get that delivered to you. It’s actually much faster and more natural and more efficient than if you have to go find your cell phone, unlock your cell phone, go find the right app. … I think 2017 will be the year of the conversational computer.
I know Google has kept its AI assistant pretty neutral, whereas Siri has the scripted jokes and is known for at least appearing more human to the average user. What’s Baidu’s approach?
Right now we have both. A lot of our functionality is focused around doing useful things for you, such as “show me a picture” or “tell me about the weather” … or “find a hairdresser to come to my house and cut my hair.” That’s actually a real thing in China. But we also have a service where you just tell it, “I just want to chat.” We’ll have users asking our [digital assistant], “Are you male or female?” “Do you have a girlfriend?” “Would you want to be my girlfriend?” Friendship is one of the use cases we support, and it’s quite popular among the users.
Do you have any sense of whether Chinese and American users have different expectations of a digital assistant?
Because of different market pressures and different opportunities, the Chinese market and the U.S. market have just evolved differently. People in Japan and also China have a much greater willingness to engage with computers as personalities. The culture in Japan is very friendly to robots, and there’s a strong strain of this in China as well. This is why I think personal chatbots have taken off much faster in China than in the U.S. I’m not sure if you’ve heard of the Microsoft Xiaoice chatbot — it’s really taken off much faster in China than the ones named Tay and Zo have taken off in the U.S. By the way, the leader of Baidu’s Duer platform [the name for the company’s Siri-like assistant technology] was actually the creator of Microsoft’s Xiaoice.
I think that on-demand services, because of high population density, have also taken off much faster in China. It’s inexpensive to provide services like — do you want someone to show up and wash your car, or do you want a chef to show up at your house and cook a meal for you, or do you want a hairdresser to come cut your hair? It’s economical to provide a lot of services because of the very high population density in China.
There’s a lot of anxiety right now in the general public about AI-powered automation putting people out of work. Is there a way to avoid that outcome? How can we face that challenge?
I’m so glad you raised jobs as the issue, because I think that’s a real issue, and not the specter of evil AI killer robots or Terminator or whatever. Jobs is a real problem. I don’t think there is a way, nor do we want to stop AI from getting much better. I think over the long term AI will free humanity up from a lot of routine, repetitive work so we can all spend our time doing higher-level things. Just as the Industrial Revolution freed up humanity from a lot of repetitive physical drudgery, I think AI will free up humanity from a lot of mental drudgery. For example, if you’re driving your car in a traffic jam — no one likes that, hopefully AI will do it for you.
Personally, I’m not worried about us not having jobs. Each wave of disruption has created a lot of new jobs. I am worried about the mismatch. I think we need new ways to educate people for these new jobs that are being created. I think MOOCs [massive open online courses] will be part of the solution, but I do think we need more. For example, I support basic income, but even a version of basic income where we don’t pay you to “do nothing.” We pay you with the expectation that you will keep studying to increase the odds that you can get into the workforce and contribute to the tax base that is paying for basic income. I think that there will still be a need for us to figure out a new New Deal, I guess. Just as we build our current educational system to be focused on this economy, we’ll need a new [system] that is better structured to help people engage in lifelong learning … so that people feel refreshed for the constantly changing jobs that the economy creates and that we actually need people to do.
Do you think there could be a point when politicians or the general public will be actively hostile toward AI research and view it as a threat to their livelihood? How do we avoid that kind of outcome?
One thing that concerns me is when AI researchers whitewash the issue. I think there is a temptation to pretend the issue is the specter of these evil AI killer robots. It’s a PR distraction from the real issue, which is job displacement. I think that researchers have a responsibility to talk about the real problem and be honest and transparent about what might need to happen and also to contemplate solutions — which I think is providing [for] the education of communities as well as the support so that every person has an equal path to doing equal work.
I read a Harvard Business Review article you wrote last year talking about how data is a scarce resource in AI development. Does that mean that new startups or even academic institutions won’t be able to compete as well with a Baidu or a Google, companies that have so much data at their fingertips?
Today we train our system on five years of data. That’s a lot of data. There are some verticals — for example, take medical images — where there just aren’t that many images in the entire world, and I think those are verticals where a group with access to even a modest data set could make a lot of progress. But I think for the verticals that are squarely in the bull’s-eye of the large companies, it is challenging for a group without access to significantly large data sets to be as effective.
In Europe right now, there seems to be a lot of hostility toward tech giants by the government, and that might happen in the United States now with the incoming president. How do you think China perceives tech? Is it encouraging innovation or being hostile to the idea that it’s too powerful?
I’m not sure. I will say that a lot of deep-learning research in the U.S. was funded by DARPA. My research at Stanford was funded by DARPA and the National Science Foundation. That allowed us to do a lot of the basic research in deep learning that enabled us to develop the basic technology. I certainly hope that governments worldwide — in the U.S., Europe, China — will continue to support this research.
I’ve been using this analogy that AI is the new electricity. About 100 years ago, electrification transformed agriculture, transportation, manufacturing. Industry after industry was transformed by electricity. I think AI is now in a similar position to transform industry after industry. It’s already transformed the IT industry — web search, advertising, speech translation, speech recognition. Really, Baidu is deeply, deeply powered by AI. The conversational computer, which we’re announcing at CES, will transform a lot of computers in home appliances. Looking forward, there’s a clear path for AI to transform health care, to transform logistics. There will be a lot of industries where the power for AI to transform them is already relatively clear.