Kai Fu Lee: A.I. vehicles are 20 years away

So I think increasingly as this technology advances, there will be industrial applications as well as consumer applications that will essentially enable a conversational robot. It still might be text, it might be on your phone, it might be like a chat window, or it might be a speech interface, or it might be a virtual digital human on your screen, or it might actually one day be a full robot. But whatever it is, because it's really gained some mastery of the human language, it's able to converse with us, have a dialogue with us. It feels like the dialogue is reasonable and it helps us get our job done. We can ask questions. So this will really elevate the capabilities of intelligence when AI starts to master language. I should also clarify that the fact that it appears to have a mastery of language doesn't mean that it fully understands. It's merely able to very intelligently map what you said against the gigantic repository of what it knows. It doesn't really understand language the way that we do, the way that we can have emotion, emotive responses, and create connection between people, and draw very vague analogies, and truly understand every type of humor. It's not quite that. But for kind of business interactions or personal agent kinds of things, I imagine there would be very close to having very smart personal assistants that are perhaps like your secretary or your junior entry-level analysts in the work that you do. The people that could find answers for you and solve problems for you without extreme business complexity. I imagine that AI will be able to do that in the 20-year time frame.

The second key point I want to predict is in the transportation area. We all talk about autonomous vehicles. So when will they come about? My prediction is that L5 capability, that is fully autonomous, no steering wheels in the car, can go anywhere to anywhere, that capability of driving is probably about 20 years away. But that doesn't mean we have to wait that long to be sitting in an autonomous vehicle. In fact, we can sit in one right now. The lower left-hand side shows your robo minibus, and that one is in deployment today on special routes, but next year it will be in full deployment in three cities in China. The key thing to recognize is that while to a human driving a bus seems like a challenging task, maybe even more than driving a car because it's bigger, to AI, driving a bus is so much easier because buses drive on fixed routes, typically in large, well-paved roads, and you can also choose cities that have decent weather. It doesn't have to deal with edge cases like a baby crawling across the street. What do you do? Is it the baby or is it a puppy? Can you tell? Because AI learns from data, and no autonomous vehicle has data with baby versus puppy crossing the road, so how is it ever going to learn that? So doing buses, and in fact before that, doing highways, is a very good way to collect data, gain practice, and then the data makes the system better, and so on. Also, the robo bus doesn't have to deal with very tiny roads, windy roads in the mountains. It doesn't have to deal with what happens if there's a natural disaster that causes a small road in the countryside to not be able to see the street signs, or maybe the pavement has been broken due to an earthquake. What about all these cases? If you're in big cities and you're operating on large paved roads, and you're driving fixed routes to maybe just 50 places, 50 fixed stops, then that is much easier. So phase one, maybe the next 10 years, will be increasingly more difficult environments in which autonomous vehicles are applied, all the time gaining data and improving themselves and getting uploads. If you've seen some reports on the Tesla summoning feature, you would know that when Tesla first launched this feature, it's just the capability of having your parked car find you. It's very clever, but when they first launched it, it didn't work out all that well. But as they accumulated more and more data, it began to work extremely well. So that shows how powerful that data collection and then the auto update that Tesla sends out is very critical. So that's kind of phase one, the next 10 years. And then after that 10 years, getting into work on robo taxis, getting people comfortable, passing laws, ensuring the casualties are better than today, can take a long time. That's why I estimated maybe 20 years, maybe even longer. One thing that can make it faster is smart cities and smart highways. Imagine if the road starts communicating with the car that says, 'You're swerving off the road, be careful,' then the car can self-correct. Or maybe the road can tell the car, 'There's an accident up ahead, you can't see it, but slow down, be careful.' So essentially, we're turning the current signs on the street from those telling human drivers what's happening or what to do, to signals that tell the cars what's happening and what to do. When you turn signals native in the car's language, it can of course appreciate it better than reading human signs. But furthermore, cars will be able to talk to each other. One car could say, 'I just blew a tire, so be careful around me.' That could be saving lives just by that broadcast. And one car can tell another, 'I'm in a hurry, the person in my car needs to get where he needs to get to in a hurry, please get out of my way, 10 cents.' And then the other car might say, 'No, I want a dollar.' Then they negotiate, and then they give the right away. So that could happen as well. Also, cars will be able to negotiate their exact path so that they barely miss each other by one centimeter, and humans won't be able to do that. So in 20 plus years, humans will slowly feel that we are unsafe to ourselves. The biggest danger to ourselves is not the autonomous vehicle but our poor ability to drive. And then eventually, people won't be allowed to drive on highways, and perhaps all the roads. That would be beyond 20 years, but it's coming. Once we are able to deliver very capable autonomous vehicles, that can be the next operating system after Windows and Android, because you can use that same capability—the multi-sensor fusion, the AI deep learning capabilities, the real-time processing, integrating knowledge of everything in your environment, recognizing who's moving, what's moving—those very capabilities that make autonomous vehicles work are the same things we need for robotics. So great robots will come after that as well. And we've all heard about using space travel to go up in space and then come down as a faster way to go from, say, London to Tokyo.

The third area is a big one. It is probably the area that will make the biggest difference: applying AI to healthcare. The reason is quite simple. It is the fact that healthcare is in the process of digitizing, and AI feeds on data. Healthcare is now digitizing everything: healthcare records, radiology, pathology, data from our wearables, and data from multi-omics, genetic sequencing, blood tests. All of that are incredibly valuable data. That's basically the same as when I gave you the example of all the data about my income and house and whether I deserve a loan. This is data to determine whether you're ill or how to become healthier. So all this data collected for people longitudinally over years, if not decades, can be tremendous, not only in improving the quality of diagnosis and treatment. Imagine now individualized treatments. We're all used to everyone getting sick and getting the same treatment, and if a drug doesn't work, maybe we try another, maybe there's a second defense. But now that we can collect all this data, we can determine maybe for some illnesses, different people with different history may be more efficaciously treated with different types of treatments. This is called precision medicine, giving different treatment to people, really leveraging the big data and the individuality of treatment in healthcare. And beyond diagnosis and treatment, monitoring, long-term care, prevention are all great ways to use this data. Once this is done, we can see applications everywhere, from imaging, drug discovery, to many, many applications. I'll mention four quickly now. One is, I'll just start from the upper right: in pathology, people really are not the ideal species to do pathology because we can't see all the images very clearly, we can't integrate all the knowledge, we can't combine and compare with all the people. In fact, in certain types of diagnosis, AI is already beating humans, for example in sicknesses related to the eye, and this will just continue and improve. A second area, on the upper left, is AI drug discovery. This is using, for example, DeepMind's protein folding to do the protein folding to begin with, followed by, for example, using AI to help find a target on which to attach a molecule, which is the drug, and then to discover which molecule is most likely to work. One of the companies we invested in, called Silicon Medicine, has actually used this approach with human supervision but AI discovery to discover drugs for two illnesses: one is pulmonary fibrosis and the other is kidney fibrosis, both almost untreatable now becoming treatable. The most exciting thing is that what AI is doing is not smarter than people; it's just helping to reduce the scope and the size and help make discovery faster because it eliminates unlikely candidates and lets the human scientists hone in on the right drug faster. So the real outcome is that the human scientists' productivity increases and is able to perhaps discover a drug three to four times faster and maybe 10 times cheaper. What this means is that currently, pharmaceuticals are not willing to invest in drugs for rare diseases because it could cost two billion dollars for such a drug discovery, and because it's a rare disease, it won't pay for its R&D. But if we are able to reduce the cost by a factor of 10 by using AI tools to help the scientists, we can now go after rare diseases, and that will help us live healthily as well. On the lower left is autonomous robots. We all know about a company called Intuitive Surgical. It is currently still operated by a doctor remotely and does certain types of surgery. In the future, as robotics become finer and finer in motor skills and accuracy and autonomy, it will gradually move from human-operated to human-delegated to parts of it, to let AI operate as an example. You might argue the surgery itself is too important, the surgeon should do that, but what about suturing, sewing the patient up? That is almost like a sewing machine. Probably the robot can do a better job, so why not delegate parts of the surgery to robots? That will reduce the stress and workload on the surgeon, allowing the surgeon to treat more patients. That's a great thing. And then eventually, some surgeries will be potentially performed fully by robots under the surgeon's care and direction, and that probably is around the 20-year time frame, assuming the various laws support that. So these are some of the exciting ideas. One of the more recent ideas I briefly mentioned is: can AI help us live longer and healthier? I'm personally using a technology from a company we invest in called Deep Longevity. It takes all of our data—multi-omics, blood, and radiology data—to identify the attributes of aging and tell us what we should do to slow down the aging, or in some cases even reverse the aging effect. The advice it gives can be related to your diet, it might be related to exercise or sleep, or might be introducing some medicine and nutrients. Of course, this is all supervised by a doctor who will tweak the recommendations. I've been on this regimen for about a year, and I am watching my blood become healthier and healthier. In fact, it's about six years healthier than it was one year ago. So by some metric, I might be five years younger now.

The fourth area is in robotics. I believe the advances in robotics are happening, and China in particular is driving it because China is the world's factory and has a strong incentive to lower the cost of producing goods, because Chinese labor costs are going up compared to countries like India, Vietnam, and Brazil. So in the area of autonomous transport, moving things, we are seeing four forklifts to be the lowest hanging fruit, followed by moving things outside, followed by trucks, followed by buses, followed by cars, all along collecting more data enabling the next more difficult scenario. We've kind of covered that under transportation, but in industrial it's something similar. We want to go after the low-hanging fruits. In a factory, the human capability being used from simplest to hardest are: using our eyes for visual inspection—is the t-shirt proper size or not, proper color or not—then moving things around, kind of like forklifts, but also in this example you see a watering robot in a farm. The third case is a picker, something that can pick merchandise up, imagine using it in Amazon's warehouse. And then the fourth example is a fancy finger that can actually pick up very delicate objects, in this case an egg yolk, and that could be applied to more cases of picking. And then the last case is autonomous flying with a drone. So all of these are enabling cases that will potentially make manufacturing, the production process, much more efficient, replacing parts or all of human labor. This will be further advanced as robots become self-replicating, self-maintaining, as 3D printing starts to take off, and also manufacturing can be extended to homes to be built in modules, extended to energy in terms of making battery and solar panels, extended to materials as new ways of producing goods that are constructed by scientists one molecule at a time.