by Kai-Fu Lee
FIRE YOUR BANKER
Both China’s corporate data and its corporate culture make applying second-wave AI to its traditional companies a challenge. But in industries where business AI can leapfrog legacy systems, China is making serious strides. In these instances, China’s relative backwardness in areas like financial services turns into a springboard to cutting-edge AI applications. One of the most promising of these is AI-powered micro-finance.
For example, when China leapfrogged credit cards to move right into mobile payments, it forgot one key piece of the consumer puzzle: credit itself. WeChat and Alipay let you draw directly from your bank account, but their core services don’t give you the ability to spend a little bit beyond your means while you’re waiting for the next paycheck.
Into this void stepped Smart Finance, an AI-powered app that relies exclusively on algorithms to make millions of small loans. Instead of asking borrowers to enter how much money they make, it simply requests access to some of the data on a potential borrower’s phone. That data forms a kind of digital fingerprint, one with an astonishing ability to predict whether the borrower will pay back a loan of three hundred dollars.
Smart Finance’s deep-learning algorithms don’t just look to the obvious metrics, like how much money is in your WeChat Wallet. Instead, it derives predictive power from data points that would seem irrelevant to a human loan officer. For instance, it considers the speed at which you typed in your date of birth, how much battery power is left on your phone, and thousands of other parameters.
What does an applicant’s phone battery have to do with creditworthiness? This is the kind of question that can’t be answered in terms of simple cause and effect. But that’s not a sign of the limitations of AI. It’s a sign of the limitations of our own minds at recognizing correlations hidden within massive streams of data. By training its algorithms on millions of loans—many that got paid back and some that didn’t—Smart Finance has discovered thousands of weak features that are correlated to creditworthiness, even if those correlations can’t be explained in a simple way humans can understand. Those offbeat metrics constitute what Smart Finance founder Ke Jiao calls “a new standard of beauty” for lending, one to replace the crude metrics of income, zip code, and even credit score.
Growing mountains of data continue to refine these algorithms, allowing the company to scale up and extend credit to groups routinely ignored by China’s traditional banking sector: young people and migrant workers. In late 2017, the company was making more than 2 million loans per month with default rates in the low single digits, a track record that makes traditional brick-and-mortar banks extremely jealous.
“THE ALGORITHM WILL SEE YOU NOW”
But business AI can be about more than dollars and cents. When applied to other information-driven public goods, it can mean a massive democratization of high-quality services to those who previously couldn’t afford them. One of the most promising of these is medical diagnosis. Top researchers in the United States like Andrew Ng and Sebastian Thrun have demonstrated excellent algorithms that are on par with doctors at diagnosing specific illnesses based on images—pneumonia through chest x-rays and skin cancer through photos. But a broader business AI application for medicine will look to handle the entire diagnosis process for a wide variety of illnesses.
Right now, medical knowledge—and thus the power to deliver accurate diagnoses—is pretty much kept bottled up within a small number of very talented humans, people with imperfect memories and limited time to keep up with new advances in the field. Sure, a vast wealth of medical information is scattered across the internet but not in a way that is navigable by most people. First-rate medical diagnosis is still heavily rationed based on geography and, quite candidly, one’s ability to pay.
This is especially stark in China, where well-trained doctors all cluster in the wealthiest cities. Travel outside of Beijing and Shanghai, and you’re likely to see a dramatic drop in the medical knowledge of doctors treating your illness. The result? Patients from all around the country try to cram into the major hospitals, lining up for days and straining limited resources to the breaking point.
Second-wave AI promises to change all of this. Underneath the many social elements of visiting a doctor, the crux of diagnosis involves collecting data (symptoms, medical history, environmental factors) and predicting the phenomena correlated with them (an illness). This act of seeking out various correlations and making predictions is exactly what deep learning excels at. Given enough training data—in this case, precise medical records—an AI-powered diagnostic tool could turn any medical professional into a super-diagnostician, a doctor with experience in tens of millions of cases, an uncanny ability to spot hidden correlations, and a perfect memory to boot.
This is what RXThinking is attempting to build. Founded by a Chinese AI researcher with deep experience in Silicon Valley and at Baidu, the startup is training medical AI algorithms to become super-diagnosticians that can be dispatched to all corners of China. Instead of replacing doctors with algorithms, RXThinking’s AI diagnosis app empowers them. It acts like a “navigation app” for the diagnosis process, drawing on all available knowledge to recommend the best route but still letting the doctors steer the car.
As the algorithm gains more information on each specific case, it progressively narrows the scope of possible illnesses and requests further clarifying information needed to complete the diagnosis. Once enough information has been entered to give the algorithm a high level of certainty, it makes a prediction for the cause of the symptoms, along with all other possible diagnoses and the percentage chance that they are the real culprit.
The app never overrides a doctor—who can always choose to deviate from the app’s recommendations—but it draws on over 400 million existing medical records and continually scans the latest medical publications to make recommendations. It disseminates world-class medical knowledge equally throughout highly unequal societies, and lets all doctors and nurses focus on the human tasks that no machine can do: making patients feel cared for and consoling them when the diagnosis isn’t bright.
JUDGING THE JUDGES
Similar principles are now being applied to China’s legal system, another sprawling bureaucracy with highly uneven levels of expertise across regions. iFlyTek has taken the lead in applying AI to the courtroom, building tools and executing a Shanghai-based pilot program that uses data from past cases to advise judges on both evidence and sentencing. An evidence cross-reference system uses speech recognition and natural-language processing to compare all evidence presented—testimony, documents, and background material—and seek out contradictory fact patterns. It then alerts the judge to these disputes, allowing for further investigation and clarification by court officers.
Once a ruling is handed down, the judge can turn to yet another AI tool for advice on sentencing. The sentencing assistant starts with the fact pattern—defendant’s criminal record, age, damages incurred, and so on—then its algorithms scan millions of court records for similar cases. It uses that body of knowledge to make recommendations for jail time or fines to be paid. Judges can also view similar cases as data points scattered across an X–Y graph, clicking on each dot for details on the fact pattern that led to the sentence. It’s a process that builds consistency in a system with over 100,000 judges, and it can also rein in outliers whose sentencing patterns put them far outside the mainstream. One Chinese province is even using AI to rate and rank all prosecutors on their performance. Some American courts have implemented similar algorithms to advise on the “risk” level of prisoners up for parole, though the role and lack of transparency of these AI tools have already been challenged in higher courts.
As with RXThinking’s “navigation system” for doctors, all of iFlyTek’s judicial tools are just that: tools that aid a real human in making informed decisions. By empowering judges with data-driven recommendations, they can help balance the scales of justice and correct for the biases present in eve
n well-trained judges. American legal scholars have illustrated vast disparities in U.S. sentencing based on the race of the victim and the defendant. And judicial biases can be far less malicious than racism: a study of Israeli judges found them far more severe in their decisions before lunch and more lenient in granting parole after having a good meal.
WHO LEADS?
So which country will lead in the broader category of business AI? Today, the United States enjoys a commanding lead (90–10) in this wave, but I believe in five years China will close that gap somewhat (70–30), and the Chinese government has a better shot at putting the power of business AI to good use. The United States has a clear advantage in the most immediate and profitable implementations of the technology: optimizations within banking, insurance, or any industry with lots of structured data that can be mined for better decision-making. Its companies have the raw material and corporate willpower to apply business AI to the problem of maximizing their bottom line.
There’s no question that China will lag in the corporate world, but it may lead in public services and industries with the potential to leapfrog outdated systems. The country’s immature financial system and imbalanced healthcare system give it strong incentives to rethink how services like consumer credit and medical care are distributed. Business AI will turn those weaknesses into strengths as it reimagines these industries from the ground up.
These applications of second-wave AI have immediate, real-world impacts, but the algorithms themselves are still trafficking purely in digital information mediated by humans. Third-wave AI changes all of this by giving AI two of humans’ most valuable information-gathering tools: eyes and ears.
THIRD WAVE: PERCEPTION AI
Before AI, all machines were deaf and blind. Sure, you could take digital photos or make audio recordings, but these merely reproduced our audio and visual environments for humans to interpret—the machines themselves couldn’t make sense of these reproductions. To a normal computer, a photograph is just a meaningless splattering of pixels it must store. To an iPhone, a song is just a series of zeros and ones that it must play for a human to enjoy.
This all changed with the advent of perception AI. Algorithms can now group the pixels from a photo or video into meaningful clusters and recognize objects in much the same way our brain does: golden retriever, traffic light, your brother Patrick, and so on. The same goes for audio data. Instead of merely storing audio files as collections of digital bits, algorithms can now both pick out words and often parse the meaning of full sentences.
Third-wave AI is all about extending and expanding this power throughout our lived environment, digitizing the world around us through the proliferation of sensors and smart devices. These devices are turning our physical world into digital data that can then be analyzed and optimized by deep-learning algorithms. Amazon Echo is digitizing the audio environment of people’s homes. Alibaba’s City Brain is digitizing urban traffic flows through cameras and object-recognition AI. Apple’s iPhone X and Face++ cameras perform that same digitization for faces, using the perception data to safeguard your phone or digital wallet.
BLURRED LINES AND OUR “OMO” WORLD
As a result, perception AI is beginning to blur the lines separating the online and offline worlds. It does that by dramatically expanding the nodes through which we interact with the internet. Before perception AI, our interactions with the online world had to squeeze through two very narrow chokepoints: the keyboards on our computers or the screen on our smartphones. Those devices act as portals to the vast knowledge stored on the world wide web, but they are a very clunky way to input or retrieve information, especially when you’re out shopping or driving in the real world.
As perception AI gets better at recognizing our faces, understanding our voices, and seeing the world around us, it will add millions of seamless points of contact between the online and offline worlds. Those nodes will be so pervasive that it no longer makes sense to think of oneself as “going online.” When you order a full meal just by speaking a sentence from your couch, are you online or offline? When your refrigerator at home tells your shopping cart at the store that you’re out of milk, are you moving through a physical world or a digital one?
I call these new blended environments OMO: online-merge-offline. OMO is the next step in an evolution that already took us from pure e-commerce deliveries to O2O (online-to-offline) services. Each of those steps has built new bridges between the online world and our physical one, but OMO constitutes the full integration of the two. It brings the convenience of the online world offline and the rich sensory reality of the offline world online. Over the coming years, perception AI will turn shopping malls, grocery stores, city streets, and our homes into OMO environments. In the process, it will produce some of the first applications of artificial intelligence that will feel truly futuristic to the average user.
Some of these are already here. One KFC restaurant in China recently teamed up with Alipay to pioneer a pay-with-your-face option at some stores. Customers place their own order at a digital terminal, and a quick facial scan connects their order to their Alipay account—no cash, cards, or cell phones required. The AI powering the machines even runs a quick “liveness algorithm” to ensure no one can use a photograph of someone else’s face to pay for a meal.
Pay-with-your-face applications are fun, but they are just the tip of the OMO iceberg. To get a sense of where things are headed, let’s take a quick trip just a few years into the future to see what a supermarket fully outfitted with perception AI devices might look like.
“WHERE EVERY SHOPPING CART KNOWS YOUR NAME”
“Nihao, Kai-Fu! Welcome back to Yonghui Superstore!”
It’s always a nice feeling when your shopping cart greets you like an old friend. As I pull the cart back from the rack, visual sensors embedded in the handlebar have already completed a scan of my face and matched it to a rich, AI-driven profile of my habits, as a foodie, a shopper, and a husband to a fantastic cook of Chinese food. While I’m racking my brain for what groceries we’ll need this week, a screen on the handlebar lights up.
“On the screen is a list of your typical weekly grocery purchase,” the cart announces. And like that, our family’s staple list of groceries appears on the screen: fresh eggplant, Sichuan pepper, Greek yogurt, skim milk, and so on.
My refrigerator and cabinets have already detected what items we’re short on this week, and they automatically ordered the nonperishable staples—rice, soy sauce, cooking oil—for bulk delivery. That means grocery stores like Yonghui can tailor their selection around the items you’d want to pick out for yourself: fresh produce, unique wines, live seafood. It also allows the supermarkets to dramatically shrink their stores’ footprint and place smaller stores within walking distance of most homes.
“Let me know if there’s anything you’d like to add or subtract from the list,” the cart chimes in. “Based on what’s in your cart and your fridge at home, it looks like your diet will be short on fiber this week. Shall I add a bag of almonds or ingredients for a split-pea soup to correct that?”
“No split pea soup but have a large bag of almonds delivered to my house, thanks.” I’m not sure an algorithm requires thanking, but I do it out of habit. Scanning the list, I make a couple of tweaks. My daughters are out of town so I can cut a few items, and I’ve already got some beef in my fridge so I decide to make my mother’s recipe of beef noodles for my wife.
“Subtract the Greek yogurt and switch to whole milk from now on. Also, add the ingredients for beef noodles that I don’t already have at home.”
“No problem,” it replies while adjusting my shopping list. The cart is speaking in Mandarin, but in the synthesized voice of my favorite actress, Jennifer Lawrence. It’s a nice touch, and one of the reasons running errands doesn’t feel like such a chore anymore.
The cart moves autonomously through the store, staying a few steps ahead of me while I pick out the ripest eggplants and the most fra
grant Sichuan peppercorns, key to creating the numbing spice in the beef noodles. The cart then leads me to the back of the store where a precision-guided robot kneads and pulls fresh noodles for me. As I place them in the cart, depth-sensing cameras on the cart’s rim recognize each item, and sensors lining the bottom weigh them as they go in.
The screen crosses things off as I go and displays the total cost. The precise location and presentation of every item has been optimized based on perception and purchase data gathered at the store: What displays do shoppers walk right by? Where do they stop and pick up items to inspect? And which of those do they finally purchase? That matrix of visual and business data gives AI-enabled supermarkets the same kind of rich understanding of consumer behavior that was previously reserved for online retailers.
Rounding the corner toward the wine aisle, a friendly young man in a concierge uniform approaches.
“Hi, Mr. Lee, how’ve you been?” he says. “We’ve just got in a shipment of some fantastic Napa wines. I understand that your wife’s birthday is coming up, and we wanted to offer you a 10 percent discount on your first purchase of the 2014 Opus One. Your wife normally goes for Overture, and this is the premium offering from that same winery. It has some wonderful flavors, hints of coffee and even dark chocolate. Would you like a tasting?”