Building An AI-Powered Society with AI Fund's Andrew Ng | Greymatter

Thanks, Sarah. Great to be here.

So AI today is creating tremendous economic value, but it's actually a large set of tools that we apply in different circumstances to different problems. So I think the vast majority of the economic value created by the modern rise of AI today is through one technique called supervised learning, which just means learning input to output mappings. Such as input an email and output is this spam or not, or input an ad and output do you think the user will click on this or not, or input a picture and tell me if your face is in this picture, or input a picture of what's in front of your car and tell me where are the positions of the other cars for a self-driving car, and so on and so on. And it turns out that when you can find the right business context to find these input to output mappings, this is called supervised learning, this turns out to be incredibly valuable. The other big categories of techniques that I think a lot of companies find very useful are transfer learning, unsupervised learning, and reinforcement learning. So what is transfer learning? Unless you want to read x-ray images for a radiology problem but you don't have that many x-ray images, then it might be difficult to get a supervised learning algorithm to do well on it just because maybe you don't have enough data. Transfer learning is a technique that lets you learn from a huge data set of say pictures of cats and people and pedestrians and cars, normal terrestrial images, because you can get a lot of images of those, and take the knowledge you learn from those millions of images and transfer it so that when finally your AI system looks at a few x-ray images, it can learn to diagnose say from x-ray images even from a much smaller data set. So this transfer learning where you learn, transfer from one task like recognizing cats and people and cars and dogs to a different task like recognizing or diagnosing from x-ray images. Unsupervised learning is another technique that people are excited about. So that's having a computer look at a lot of data and just tell us, you know, hey computer, look at this data and figure some stuff out. This is useful for a few applications. For example, today a lot of web search algorithms or really algorithms that manipulate text such as web search or chat bots use one type of unsupervised learning algorithm that has read a ton of text on the internet. And through just reading lots of text on the internet without anyone kind of really telling it anything, has figured out, you know, a lot of things about the meaning of words. And through the understanding of meaning of words that has learned off the internet, it turns out that these AI algorithms then become much better at your web search or chat bot and so on. And the final category, reinforcement learning, is a type of technique where the press coverage I think is vastly greater than the actual instances or production products deployed using reinforcement learning. But reinforcement learning I liken to how you would train a cat or how you train a dog really, where you let the dog do whatever it does and whenever it does something good you go 'oh good dog' and whenever it does something bad you go 'bad dog' and then over time your dog figures out how to do more of the good things. So it turns out reinforcement learning is great for learning to play games like chess and go and checkers and so on because whenever it wins a game you go 'oh good computer' and when it loses you go 'bad computer' and then over time it learns to play these games really well. So in terms of economic value created today, I would rank supervised learning, input output mappings, as the first by far, followed by transfer learning, transferring from one task to another, followed by unsupervised learning, and then with reinforcement learning being really the distant last in terms of economic value created today. Having said that, the AI world is advancing rapidly and so who knows, there might be breakthroughs and this rank ordering might totally change in the next two years.

I think there are a ton of exciting areas in machine learning. So just within deep learning, I think that recent advances in data synthesis through GANs, generative adversarial networks, it's been amazing how accurately we can synthesize pictures of people in novel situations or synthesize audio clips of people, you know, saying stuff that they never actually said. And I think that will certainly affect the entertainment industry. It might affect other industries as well, although I think that maybe remains to be seen. For many years there's been a lot of work on meta-learning. There's actually very good work out of Harvard University several years back and then my old team, the Google Brain team, is also doing a lot of work on this, but really using machine learning to speed up other machine learning algorithms is a cool idea. It's a very nice tactic. I don't know if it will fundamentally change everything but it is a nice project to be working on. The AI community is so innovative and so creative. Many of these things you kind of see the seeds of it a few years earlier. The tech community is so global and we all talk to each other all around the world that the right ideas have the opportunity to catch on very quickly and really go from zero to 100 in like a year. So this is an exciting time both to be doing basic research in AI and to be looking at all these great techniques and trying to figure out how we could use them to help people.

Yeah, AI is the new electricity and like electricity 100 years ago, I think it'll transform every industry. At Baidu I was fortunate to work with a fantastic team including Robin Li, the CEO, who's just a phenomenal executive, very smart about AI, but also many, the whole company up and down was just phenomenal in terms of developing technology and also figuring out new products. So the AI group that I led at Baidu worked both on supporting existing businesses, so we had many, many dozens of projects supporting existing businesses, everything from web search to advertising to maps to on and on and on, to use AI to improve these different products. And then one thing that Baidu as a company I think was also very good at was to use AI to create brand new categories of products. For example, when I was still there, Baidu launched the DuerOS team which provides a voice control platform for a lot of consumer electronic devices, for smart speakers and for other things. And it's actually in a very strong position of leadership in China right now. So I guess the American version of that might be Amazon Alexa, although one piece of history that's not widely known is that Baidu announced a smart speaker in public actually before Amazon announced the existence of Amazon Alexa. So it's very interesting how a lot of these ideas contemporaneously sometimes appear a bit earlier in the US, sometimes appear a bit earlier in China. But I think at Baidu the AI group really, two of the main activities were to support existing businesses and I think it helped many businesses, as well as use AI capabilities to start up new lines of businesses such as DuerOS, the smart speaker voice control platform, or the autonomous driving team from Baidu which also came out of my team. When you look more broadly across society, across industries, I think that the opportunities for AI actually mirrors these two categories I just named at Baidu, which is that existing incumbents will have fantastic opportunities to transform using AI, that's actually what Landing.ai is focused on, but then there'll also be many new opportunities for brand new startups or maybe brand new industries to be created with this wave of new technology we now have access to.

I think the Google Brain team really helped Google play a non-trivial role in the transformation of Google into the great AI company a lot of people perceive it to be today. And I think the AI group at Baidu were responsible for the AI strategy and technology platform for Baidu and that helped transform Baidu. But I think that for AI to reach its full potential we need to transform not just great internet companies but really many companies in many different categories across many sectors of industry. I've been speaking with CEOs from multiple industries about recommendations I make to help really large incumbent companies transform for the age of AI. A couple recommendations I tend to make: one is form an AI team that matrices AI capabilities into multiple business units. You know, if you're a CEO of a large enterprise and you have say five or 10 business units reporting to you, it'll be very difficult to get all five or all 10 business unit leaders to build their own strong AI team. So I think from a perspective of CO, the odds of success will be higher if you can build an AI team that leverages both internal capabilities, build the things you need to build, but more importantly leverages external capabilities as well. Don't buy the things that will become industry standard anyway or that an external party can develop much more efficiently, and then matrices those capabilities into all of the different business units. So I think this is a playbook that will help many companies transform with AI. The tech world is evolving so fast, it's important that companies help your employees to keep learning because the tech will keep on shifting. Now one other change over the last decade is that it's now possible to educate people much more conveniently and at lower cost than ever before. So 10 years ago if you want to train people on AI, you know, well what could you do, right? You could try to get me to show up and give a talk or get you Sarah to show up and give a talk, but that isn't the most scalable way to help companies transform. But with the rise of online digital content, companies like Coursera, Coursera is seeing a ton of companies using this type of online digital content to very inexpensively train up huge numbers of employees on modern technologies.

I think there is a huge difference in culture both at the executive level as well as at all levels of the companies I think between the US and China. A lot of companies in China are still younger, often founder-led, and so on average companies in China tend to make decisions very quickly whereas some of the more mature, you know, board-controlled publicly traded companies, it's harder for say the CEO to very quickly change the organization to embrace a brand new direction. So that's one difference at the executive level. And really I think in China so many people, you know, have seen fortunes made and seen fortunes lost all in the last 10 years that a lot of CEOs and chairpersons, they know that when the next technology wave comes, if you don't jump fast enough it might be your turn to see your company go down in this way. So I find willingness to adopt new technology very high in Asia. But having said that, I think the US has a stronger basic research and technology capability than China. So the US has at least today a significant advantage there. But I think continued investments seem like good investments really for many countries.

With the rise of AI technologies we can now make computers do things that, you know, no one could two or three years ago. In fact, I'm actually very inspired that today there are high school students that can download open source code and probably do stuff that the best research group in the world could not have done three years ago. So what an exciting time to be alive. Now in terms of the opportunities this creates to help people, there are opportunities for incumbent companies to transform as well as to create brand new startups. So in fact if you look at the last, maybe one of the last major waves of technological transformation, the rise of the internet, the rise of the internet allowed the creation of few great startups like Google, Facebook, Baidu. Some of the incumbents also did a great job transforming. I think Apple and Microsoft did a great job transforming from a non-internet company to an internet company. And so some of what I do actually tries to mirror that, where Landing.ai is focused on helping incumbent companies transform for the age of AI, but also for AI to reach its full potential there will be many startups that I think need to be created to use these new abilities of, you know, machine learning and AI to do great new things. So the AI Fund is focused on that. The AI Fund's primary business activity is a bit different to that of most venture capitalists and our primary business activity is to initiate and to really build startups, AI-powered startups from scratch. A lot of people have asked if we're a startup incubator and we actually say no, we don't think of ourselves as an incubator, we think of ourselves as a startup initiator. But most commonly we engage with entrepreneurs, AI engineers, and various engineers and business people before their startup has been incorporated, before the team has joined, and sometimes even before they have a specific idea. And then we work with them to explore different ideas and then go from there to build the startups.

So Landing.ai is helping companies transform to adopt AI and we decided to start in manufacturing. Before focusing on manufacturing we did a systematic scan of multiple different industries, you know, including healthcare and education and a few others, and we thought that manufacturing is an industry where we can make inroads quite rapidly and help maybe as the first industry that we'll focus on transforming. Silicon Valley has done a great job transforming the digital universe, but really to your listeners, if your listeners look around, I think most of the pixels that hit their eye were actually created by the manufacturing industry. AI going to manufacturing, I'm excited about implementing a digital transformation of the physical universe around you. And I think maybe in some parts of the country people have enough stuff, but if we could lower the cost of physical goods, what better way to give everyone in the world a 10% raise other than if we can make a lot of things 10% cheaper. So I'm excited about working with the manufacturing industry to improve access to everything from bicycles to drugs to washing machines. I spend a lot of time visiting factories and I just find that exciting. But I think also we can help both the industry as well as the people that the manufacturing industry serves.

You know, so answering even more broadly than manufacturing, I think that I've actually yet to run across one industry vertical that is fully satisfied with its own data. So even look at the great AI companies, they do have better data infrastructure than many other companies or maybe than all other companies, but I don't think even any of them are fully satisfied with the quality and the maturity of their data. One pattern I'm seeing across a lot of industries is that often first comes the IT revolution or the digitization revolution and that creates data, and then a little bit behind the IT revolution comes the AI revolution because AI can come and eat that data to create value. And over the last 20 years because so much of society has been going through the digitization revolution, for example if you take an x-ray, that x-ray is now much more likely to be a digital image like a computer picture rather than physical x-ray film. So that's data that AI can come in to eat. Or take the supply chain, everything from x-ray images to what you just ordered online, this is creating data. In industry after industry is enabling AI to enter all of these industries.

Yes, I've spent a lot of time in manufacturing plants and no one's satisfied with the quality of the data, right? Some of these plants, Wi-Fi isn't up and running, some of these machines aren't even hooked onto the internet and we're plugging USB drives into computers to carry data around. So it's just, so there's a huge variety. And there are also highly automated factories where you walk to the plant and it's a giant plant and there are like 30 people in this huge building because it's all robots. So there's a huge range in terms of the degree to which different plants have been digitized. But one of the exciting things is also working with all this messy, dirty, incomplete data and sometimes even partnering with them to figure out how to make the data better. You know, some CEOs have come to me and said 'hey Andrew, this is what I'm going to do, give me three years to build the perfect IT system then I'll come back to you and I'll do AI then.' And this is a terrible strategy, so just don't do that. Usually advice, if you collected some data, it's probably worth getting an AI team to take a look at it because often the AI team can give feedback to your digitization team or your IT team on where to collect more data and maybe where to invest more and less, that makes sense to figure out where you can get the value and then take a targeted approach to then collecting the data you do need.