ANDREW NG: With the rise of technology often comes greater concentration of power in smaller numbers of people’s hands, and I think that this creates greater risk of ever-growing wealth inequality as well. To be really candid, I think that with the rise of the last few ways of technology, we actually did a great job creating wealth in the East and the West Coast, but we actually did leave large parts of the country behind, and I would love for this next one to bring everyone along with us.
KEVIN SCOTT: Hi everyone. Welcome to Behind the Tech. I’m your host, Kevin Scott, Chief Technology Officer for Microsoft. In this podcast, we’re going to get behind the tech. We’ll talk with some of the people who made our modern tech world possible and understand what motivated them to create what they did. So, join me to maybe learn a little bit about the history of computing and get a few behind the scenes insights into what’s happening today. Stick around. Today I’m joined by my colleague Christina Warren. Christina is a Senior Cloud Developer Advocate at Microsoft. Welcome back Christina.
CHRISTINA WARREN: Happy to be here Kevin, and super excited about who you’re going to be talking to today.
KEVIN SCOTT: Yeah. Today’s guest is Andrew Ng.
CHRISTINA WARREN: Andrew is, I don’t think this is too much to say, he’s one of the preeminent minds in artificial intelligence and machine learning. I’ve been following his work since the Google Brain Project, and he co-founded Coursera, and he’s done so many important things and so much important research on AI and that’s a topic that I’m really obsessed with right now. So, I can’t wait to hear what you guys talk about.
KEVIN SCOTT: Yeah. In addition to his track record as an entrepreneur, so Landing.AI, Coursera, being one of the co-leads of the Google Brain Project in its very earliest days, he also has this incredible track record as academic researcher. He has a hundred plus really fantastically good papers on a whole variety of topics in artificial intelligence, which I’m guessing are on the many a PHD student’s reading list for the folks who are trying to get degrees in this area now.
CHRISTINA WARREN: I can’t wait. I’m really looking forward to the conversation.
KEVIN SCOTT: Great. Christina, we’ll check back with you after the interview. Coming up next, Andrew Ng. Andrew is founder and CEO of Landing.AI. Founding lead of the Google Brain Project and co-founder of Coursera. Andrew is one of the most influential leaders in AI and deep learning. He’s also a Stanford University Computer Science adjunct professor. Andrew, thanks for being here.
ANDREW NG: Thanks a lot for having me Kevin.
KEVIN SCOTT: So, let’s go all the way back to the beginning. So, you grew up in Asia? And I’m just sort of curious when was it that you realized you were really interested in math and computer science?
ANDREW NG: I was born in London, but grew up mostly in Hong Kong and Singapore. I think I started coding when I was six-years-old. And my father had a few very old computers. The one I used the most was some old Atari, where I remember there were these books where you would read the code in a book and just type in a computer and then you had these computer games you could play that you just implemented yourself. So, I thought that was wonderful.
KEVIN SCOTT: Yeah, and so that was probably the Atari 400 or 800?
ANDREW NG: Yeah. Atari 800 sounds right. It was definitely some Atari.
KEVIN SCOTT: That’s awesome. And what sorts of games were you most interested in?
ANDREW NG: You know, the one that fascinated me the most was a number guessing game. Where you, the human, would think of a number from 1 to 100, then the computer would basically do binary search but chooses: Is it higher or lower than 50? Is it higher or lower than 75 and so on, until it guesses the right number.
KEVIN SCOTT: Well, in a weird way, that’s like early statistical Machine Learning, right?
ANDREW NG: Yeah, and then, so at six-years-old it was just fascinating that the computer could guess.
KEVIN SCOTT: Yeah. So, from six years- did you go to a science and technology high school? Did you take computer science classes when you were a kid or…?
ANDREW NG: I went to good schools: St. Paul’s in Hong Kong and then ACPS in the Raffles in Singapore. I was lucky to go to good schools. I was fortunate to have grown up in countries with great educational systems. Great teachers, they made us work really hard but also gave us lots of opportunities to explore.
And I feel like, computer science is not magic. You and I do this, we know this. While I’m very excited about the work I get to do in computer science and AI, I actually feel like anyone could do what I’d do if they put in a bit of time to learn to do these things as well. Having good teachers helps a lot.
KEVIN SCOTT: We chatted in our last episode with Alice Steinglass, who’s the president of Code.org, and they are spending the sum total of their energy trying to get K-12 students interested in computer science and pursuing careers in STEM. You’re also an educator. You are a tenured professor at Stanford and spent a good chunk of your life in academia. What things would you encourage students to think about if they are considering a career in computing?
ANDREW NG: I’m a huge admirer of Code.org. I think what they’re doing is great. Once upon a time, society used to wonder if everyone needed to be literate. Maybe all we needed was for a few monks to read the Bible to us and we didn’t need to learn to read and write ourselves because we’d just go and listen to the priest or the monks. But we found that when a lot of us learned to read and write that really improved human-to-human communication.
I think that in the future, every person needs to be computer literate at the level of being able to write these simple programs. Because computers are becoming so important in our world and coding is the deepest way for people and machines to communicate. There’s such a scarcity of computer programmers today that most computer programmers end up writing software for thousands of millions of people.
But in the future if everyone knows how to code, I would love for the proprietors of a small mom and pop store at a corner to go program an LCD display to better advertise their weekly sales. So, I think just as literacy, we found it having everyone being able to read and right, improved human-to-human communication. I actually think everyone in the future should learn to code because that’s how we get people and the computers to communicate at the deepest levels.
KEVIN SCOTT: I think that’s a really great segue into the main topic that I wanted to chat about today, AI, because I think even you have used this anecdote that AI is going to be like electricity.
ANDREW NG: I think I came up with that.
KEVIN SCOTT: Yeah. I know this is your brilliant quote and it’s spot on. The push to literacy in many ways is a byproduct of the second and third industrial revolution. We had this transformed society where you actually had to be literate in order to function in this quickly industrializing world. So, I wonder how many analogues you see between the last industrial revolution and what’s happening with AI right now.
ANDREW NG: Yeah. The last industrial revolution changed so much human labor. I think one of the biggest differences between the last one and this one is that this one will happen faster, because the world is so much more connected today. So, wherever you are in the world, listening to this, there’s a good chance that there’s a AI algorithm that’s not yet even been invented as of today, but that will probably affect your life five years from now.
A research university in Singapore could come up with something next week, and then it will make its way to the United States in a month. And another year after that, it’ll in be in products that affect our lives. So, the world is connected in a way that just wasn’t true at the last industrial revolution. And I think the pace and speed will bring challenges to individuals and companies and corporations. But our ability to drive tremendous value for AI, for the new ideas, the tremendous driver for global GDP growth I think is also maybe even faster and greater than before.
KEVIN SCOTT: Yeah. So, let’s dig in to that a little bit more. So, you’ve been doing AI Machine Learning for a really long time now. When did you decide that that’s the thing you were going to specialize on as a computer scientist?
ANDREW NG: So, when I was in high school in Singapore, my father who is a doctor was trying to implement AI systems. Back then, he was actually using XP systems, which turned out not to be that good a technology. He was implementing AI systems of his day to try to diagnose, I think lymphoma.
KEVIN SCOTT: This is in the late ’80s.
ANDREW NG: I think I was 15 years old at that time. So, yeah, late ’80s. So, I was very fortunate to learn from my father about XP Systems and also about neural networks, because they had day in the sun back then. That later became an internship at the National University of Singapore where I wrote my first research paper actually, and I found a copy of it recently. When I read it back now, I think it was a very embarrassing research paper. But we didn’t know any better back then. And I’ve actually been doing AI, computer science and AI pretty much since then.
KEVIN SCOTT: Well, I look at your CV and the papers that you’ve written over the course of your career. It’s like you really had your hands in a little bit of everything. There was this inverse reinforcement learning work that you did and published the first paper in 2000. Then, you were doing some work on what looks like information retrieval, document representations, and what not. By 2007, you were doing this interesting stuff on self-taught learning. So, transfer learning from unlabeled data.
Then, you wrote the paper in 2009 on this large-scale unsupervised learning using graphical processing. So, just in this 10-year period in your own research, you covered so many things. In 2009, we hadn’t even really hit the curve yet on deep learning, the ImageNet result from Hinton hadn’t happened yet. How do you, as one of the principles, you help create the feel, what does the rate of progress feel like to you? Because I think this is one of the things that people get perhaps a little bit over excited about sometimes.
ANDREW NG: One of the things I’ve learned in my career is that you have to do things before they’re obvious to everyone, if you want to make a difference and get the best results. So, I think I was fortunate back in maybe 2007 or so, to see the early signs that deep learning was going to take off. So, with that conviction, decided to go on and do it, and that turned out to work well.
Even when I went to Google to start the Google Brain project, at that time, neural networks was a bad word to many people and there was a lot of initial skepticism. But, fortunately, Larry Page was supportive and then started Google Brain. And I think when we started Coursera, online education was not an obvious thing to do. There were other previous efforts, massive efforts that failed. But because we saw signs that we could make it work with the conviction to go in.
When I took on the role at Baidu at that time, a lot people in the US were asking me, “Hey, Andrew, why on earth would you want to do AI in China. What AI is there in China?” I think, again, I was fortunate that I was part of something big. Even today, I think landing.ai where I’m spending a lot of my time, people initially ask me, “AI for manufacturing? Or AI for agriculture? Or try to transfer calls using AI? that’s a weird thing to do.” I do find people actually catch on faster. So, I find that as I get older, the speed at which people go from being really skeptical about what I do versus to saying, “Oh, maybe that’s a good idea.” That window is becoming much shorter.
KEVIN SCOTT: Is that because the community is maturing or because you’ve got such an incredible track record that…
ANDREW NG: I don’t know. I think everyone’s getting smarter all around the world. So, yeah.
KEVIN SCOTT: As you look at how machine learning has changed over the past just 20 years, what’s the most remarkable thing from your perspective?
ANDREW NG: I think a lot of recent progress was driven by computational scale, scale of data, and then also by algorithmic innovation. But, I think it’s really interesting when something grows exponentially, people, the insiders, every year you say, “Oh yeah, it works 50 percent better than the year before.” And every year it’s like, “Hey, another 50 percent year-on-year progress.” So, to a lot of machine learning insiders, it doesn’t feel that magical. It’s, “Yeah, you just get up and you work on it, and it works better.”
To people that didn’t grow up in machine learning, exponential growth often feels like it came out of nowhere. So, I’ve seen this in multiple industries with the rise of the movement, with the rise of machine learning and deep learning. I feel like a lot of the insiders feel like, “Yeah, we’re at 50 percent or some percent better than last year,” but it’s really the people that weren’t insiders that feel like, “Wow, this came out of nowhere. Where did this come from?” So, that’s been interesting to observe.
But one thing you and I have chatted about before, there’s a lot of hype about AI. And I think that what happened with the earlier AI winters is that there was a lot of hype about AI that turned out not to be that useful or valuable. But one thing that’s really different today is that large companies like Microsoft, Baidu, Google, Facebook, and so on, are driving tremendous amounts of revenue as well as user value through modern machine learning tools. And that very strong economic support, I think machine learning is making a difference to GDP. That strong economic support means we’re not in for another AI winter.
Having said that, there is a lot of hype about AGI, Artificial General Intelligence. This really over hyped fear of evil killer robots, AI can do everything a human can do. I would actually welcome a reset of expectations around that. Hopefully we can reset expectations around AGI to be more realistic, without throwing out baby with the bath water.
If you look at today’s world, there are a lot more people working on valuable deep learning projects today than six months ago, and six months ago, there were a lot more people doing this than six months before that. So, if you look at it in terms of the number of people, number of projects, amount of value being created, it’s all going up. It’s just that some of the hype and unrealistic expectations about, “Hey, maybe we’ll have evil killer robots in two years or 10 years, and we should defend against it.” I think that expectation should be reset.
KEVIN SCOTT: Yeah. I think you’re spot on about the inside versus outside perspective. The first machine learning stuff that I did was 15 years-ish ago when I was building classifiers for content for Google’s Ad systems. Eventually, my teams worked on some of the CTR predictions stuff for the ads auction. It was always amazing to me how simple an algorithm you could get by with if you had lots of compute and lots of data. You had these trends that were driving things.
So, Moore’s Law and things that we were doing in cloud computing was making exponentially more compute available for solving machine learning problems like the stuff that you did, leveraging the embarrassingly parallelism in some of these problems and solving them on GPUs, which are really great at doing the idiosyncratic type of compute. So, that computer is one exponential trend, and then the amount of available data for training is this other thing, where it’s just coming in at this crushing rate.
You were at the Microsoft CEO Summit this year and you gave this beautiful explanation where you said, “Supervised Machine Learning is basically learning from data, a black box that takes one set of inputs and produces another set of outputs. And the inputs might be an image and the outputs might be text labels for the objects in the image. It might be a waveform coming in that has human speech in it and the output might be the speech.” But really, that’s sort of at the core of this gigantic explosion of work and energy that we’ve got right now, and AGI is a little bit different from that.
ANDREW NG: Yes, in fact to give credit where it’s due. You know actually many years ago, I did an internship at Microsoft Research back when I was still in school. Even back then, I think it was Eric Brill and Michele Vanko at Microsoft way back had already published a paper using simple algorithms, that basically it wasn’t who has the best algorithm that wins, it was who has the most data for the application they were looking at at NLP. And so I think that the continuation of that trend, that people like Eric and Michelle had spotted a long time ago, that’s driving a lot of the progress in modern machine learning still.
KEVIN SCOTT: Yeah. Sometimes, with AI Research you get these really unexpected results. One of those I remember it was the famous Google CAT result from the Google Brain Team.
ANDREW NG: Yes, actually, those are interesting projects, while still a full time at Stanford, my students at the time Adam Coates and others, started to spot trends that, basically the bigger you build in your neural networks, the better they work. So that was a rough conclusion.
So I started to look around Silicon Valley to see where can I get a lot of computers to train really really big neural networks. And I think in hindsight, back then a lot of us leaders of deep learning had a much stronger emphasis on unsupervised learning, so learning without label data, such as getting computers to look a lot of pictures, or watch a lot YouTube videos without telling it what every frame or what every object is.
So I had friends at Google so I wound up pitching to Google to start a project which we later called the Google Brain Project, to really scale up neural networks. We started off using Google’s Cloud, the CPU’s and in hindsight, I wish we had tried to build up GPU capabilities like Google sooner, but for complicated reasons, that took a long time to do which is why I wound up doing that at Stanford rather than at Google first. And I was really fortunate to have recruited a great team to work with me on the Google Brain Project.
I think one of the best things I did was convince Jeff Dean to come and work. And in fact, I remember the early days, we were actually nervous about whether Jeff Dean would remain interested in the project. So a bunch of us actually had conversations to strategize, “Boy, can we make sure to keep Jeff Dean engaged so that he doesn’t lose interest and go do something else?” So thankfully he stayed. The Google CAT thing was led by my, at the time PhD student Quoc Le put together with Jiquan Ngiam, were the first two sort of machine learning interns that I brought into the Google Brain Team.
And I still remember when Quoc had trained us on unsupervised learning algorithms, it was almost a joke, you know I was like, “Hey! there are a lot of cats on YouTube, let’s see this learning cat detector.” And I still remember when Quoc told me to walk over and say, “Hey Andrew, look at this.” And I said, “Oh wow! You had unsupervised learning algorithm watch YouTube videos and learn the concept of ‘cat.’ That’s amazing.” So that winds up being an influential piece of work, because it was unsupervised learning, learning from tons of data for an algorithm to discover concepts by itself.
I think a lot of us actually overestimated the early impact of unsupervised learning. But again, when I was leading Google Brain Team, one of our first partners was the speech team working with Vincent Vanhoucke, a great guy, and I was really working with Vincent and his team, and seeing some of the other things happening at Google and outside that caused a lot of us to realize that there was much greater short term impact to be had with supervised learning.
And then for better or worse, when lot of deep learning communities saw this, so many of us shifted so much of our efforts to supervised learning, that maybe we’re under resourcing the basic research we still need unsupervised learning these days which maybe, you know, I think unsupervised learning is super important that there’s so much value to be made with supervised learning. So much of the attention is there right now. And I think, really what happened with the Google Brain Project was- were the first couple of successes, one being the Speech Project that we worked with the speech team on.
What happened was other teams saw the great results that the speech team was getting with deep learning with our help. And so, more and more of the speech team’s peers ranging from Google Maps to other teams started to become friends and allies of the Google Brain Team. We started doing more and more projects.
And then the other story is after, you know, the team had tons of momentum, thank god, we managed to convince Jeff Dean to stick with the project, because one of the things that gave me a lot of comfort when I wanted to step away from a day-to-day role to spend more time in Coursera was, I was able to hand over leadership of the team to Jeff Dean. And that gave me a lot of comfort that I was leaving the team in great hands.
KEVIN SCOTT: I sort of wonder, if there’s a sort of a message or a takeaway for AI researchers in both academia and industry about the Jeff Dean example. So for those who don’t know, Jeff Dean might be the best engineer in the world.
ANDREW NG: It might be true. Yes.
KEVIN SCOTT: But I’ve certainly never worked with anyone quite as good as him. I mean, I remember there was this-
ANDREW NG: He’s in a league of his own. Jeff Dean is definitely-
KEVIN SCOTT: I remember back in long, long ago at Google. This must have been 2005 or 2005, right after we’d gone public, Alan Eustace who was running all of the engineering team at the time would, once a year, send a note out to everyone in engineering at performance review time to get your Google resume polished up so that you could nominate yourself for a promotion.
First thing that you were supposed to do was get your Google resume, which is sort of this internal version of a resume that showed all of your Google specific work. And the example resume that he would send out was Jeff’s, and even in 2004, like he’d been there long enough where he’d just done everything.
And, you know I was an engineer at the time. I would look at this and I’m like, “Oh my god, my resume looks nothing like this.” And so I remember sending a note Alan Eustace saying, “You have got to find someone else’s resume. You’re depressing a thousand engineers every time you send this out.” Because Jeff is so great.
ANDREW NG: We’re just huge fans really of Jeff. So me, you know, fans of Jeff among them and just, not just a great scientist but also just an incredibly nice guy.
KEVIN SCOTT: Yeah. But this whole notion of coupling world-class engineering and world class-systems engineering with AI problem solving, I think that is something that we don’t really fully understand enough.
You can be the smartest AI guy in the world and you know just have this sort of incredible theoretical breakthrough, but if you can’t get that idea implemented, not that it has no impact it just sort of diminishes the potential impact that the idea can have. That partnership I think you have with Jeff is something really special.
ANDREW NG: I think I was really fortunate that even when I started the Google Brain Team I feel I brought a lot of machine learning expertise and Jeff, and other Google engineers early team members like Rajat Monga, Greg Corrado, just thought a percent project for him. But there are other Google engineers– really first and foremost Jeff–they brought a lot of systems abilities to the team.
And the other convenient thing was that, we were able to get a thousand computers to run this. And having Larry Page’s backing and Jeff’s ability to marshal those types of computational resources turns out to be really helpful.
KEVIN SCOTT: Well, let’s switch gears just a little bit. I think it was really apt that you pointed out that AI and machine learning in particular are starting to have GDP scale impact on the world. Certainly, if you look at the products that we’re all using everyday, there’s many levels of machine learning involved in everything from search to social networks to-
I mean, basically everything you use has got just a little kiss of machine learning in it. So, with that impact and given how pervasive these technologies are, there’s a huge amount of responsibility that comes along with it. I know that you’ve been thinking a lot about ethical development of AI and what our responsibilities are as scientists and engineers as we build these technologies.
ANDREW NG: I’d love to chat about that for a few minutes. Yeah. There’s potential to promulgate things like discrimination and bias. I think that with the rise of technology often comes greater concentration of power in smaller numbers of people’s hands. And I think that this creates greater risk of ever-growing wealth inequality as well.
So, we’re recording this here in California, and to be really candid, I think that with the rise of the last few waves to technology, we actually did a great job creating wealth in the East and the West Coast, but we actually did leave large parts of the country behind, and I would love for this next one to bring everyone along with us.
KEVIN SCOTT: Yeah. One of the things that I’ve spent a bunch of time thinking about is, from Microsoft’s perspective, when we think about how we build our AI technology, we’re thinking about platforms that we can put in the hands of developers. It’s just sort of our wiring as a company.
So, the example you gave earlier and the talk where you want someone in a mom and pop shop to be able to program their own LCD sign to do whatever and everybody becomes a programmer, we actually think that AI can play a big role in delivering this future. And we want everybody to be an AI developer. I’ve been spending much of my time lately talking with folks in agriculture and in healthcare, which again you’re thinking about the problems that society has to solve.
In the United States. the cost of healthcare is growing faster than GDP which is not sustainable over long periods of time. Basically, the only way that I see that you break that curve is with technology. Now, it might not be AI. I think it is. But something is going to have to sort of intercede that pulls cost out of the system while still giving people very high-quality healthcare outcomes.
And I just see a lot of companies almost every week, there’s some new result where AI can read and EKG chart with cardiologists’ level of accuracy, which isn’t about taking all of the cardiology jobs away. It’s about making this diagnostic capability available to everyone because the cost is free and then letting the cardiologist do what’s difficult and unique that humans should be doing. I don’t know if you see that pattern in other domains as well.
ANDREW NG: I think there’ll be a lot of partnerships with the AI teams and doctors that will be very valuable. You know, one thing that excites me these days with the theme of things like healthcare, agriculture, and manufacturing is helping great companies become great AI companies.
I was fortunate really, to have led the Google Brain team which became I would say probably the leading force in turning Google from what was already a great company into today great AI company. Then, at Baidu, I was responsible for the company’s AI technology and strategy and team, and I think that helped transform Baidu from what was already a great company into a great AI company.
I think it really Satya did a great job also transforming Microsoft from a great company to a great AI company. But for AI to reach its full potential, we can’t just transform tech companies, we need to pull other industries along for it to create this GDP growth, for it to help people in healthcare deliver a safer and more accessible food to people.
So, one thing I’m excited about, building on my experience, helping with really Google and Baidu’s transformation is to look at other industries as well to see if either by providing AI solutions or by engaging deeply in AI transmission programs, whether my team at Landing.AI, whether Landing.AI can help other industries also become great at AI.
KEVIN SCOTT: Well talk a little bit more about what Landing.AI’s mission is.
ANDREW NG: We want to empower businesses with AI. There is so much need for AI to enter other industries than technology, everything ranging from manufacturing to agriculture to healthcare, and so many more. For example, in manufacturing, there are today in factories sometimes hundreds of thousands of people using their eyes to inspect parts as they come off as the assembly line to check for scratches and things and so on.
We find that we can, for the most part, automate that with deep learning and often do it at a level of reliability and consistency that’s greater than the people are. People squinting at something centimeters away your whole day, that’s actually not great for your eyesight it turns out, and I would love for computers rather than often these young employees to do it.
So, Landing.AI is working with a few different industries to provide solutions like that. We also engage companies with broader transformation programs. So, for both Google and Baidu, it was not one thing, it’s not that implement neural networks for ads and so it’s a great AI company.
For a company become a great AI company is much more than that. And then having sort of helped two great companies do that, we are trying to help other companies as well, especially ones outside tech become leading AI entities in their industry vertical. So, I find that work very meaningful and very exciting. Several days ago, I tweeted out that on Monday, I literally wake up At 5:00 AM so excited about one of the Landing.AI projects, I couldn’t go back to sleep. I started getting and scribbling on my notebook. So, I find these are really, really meaningful.
KEVIN SCOTT: That’s awesome. One thing I want to sort of press on a little bit is this manufacturing quality control example that you just gave. I think the thing that a lot of folks don’t understand is it’s not necessarily about the jobs going away, it’s about these companies being able to do more.
So, I worked in a small manufacturing company while I was in college and we had exactly the same thing. So, we operated a infrared reflow soldering machine there which sort of melts, surface mount components onto circuit boards. So, you have to visually inspect the board before it goes on to make sure the components are seated and the solder has been screened and all the right parts. When it comes out, you have to visually inspect it to make sure that none of the parts have tombstoned. There are a variety of like little things that can happen in the process. So, we have people doing that.
If there was some way for them not to do it, they would go do something else that was more valuable or we could run more boards so actually, in a way, you could create more jobs because the more work that this company could do economically, the more jobs in general that it can create.
And I’m sort of seeing AI in several different places like in manufacturing automation as helping to bring back jobs from overseas that were lost because it was just sort of cheaper to do them with low cost labor in some other part of the world. They’re coming back now because like automation has gotten so good that you can start doing them with fewer more expert people but here, in the United States, locally in these communities where whatever it is that they’re manufacturing is needed. It’s like these really interesting phenomena.
ANDREW NG: There was one part of your career I did not know about it. I followed a lot of your work at Google and Microsoft, and even today, people still speak glowingly of their privacy practices you put in place when you’re at Google. I did not know you were into this soldering business way back.
KEVIN SCOTT/ANDREW NG: Yeah, I had put myself through college some way or another. It was interesting though. I remember one of my first jobs, I had to put brass rivets into 5,000 circuit boards. Circuit boards were controllers for commercial washing machines and there were six little brass tabs that you would put electrical connectors onto and each one of them had to be riveted.
So, it was 30,000 rivets that had to be done and we had a manual rivet press and my job at this company in its first three months of existence right after I graduated high school was to press, rivet press 30,000 times, and that’s awful. Automation is not a bad thing.
ANDREW NG: In a lot countries we work with we’re seeing, for example Japan, the country is actually very different than the United States, because it has an aging population.
KEVIN SCOTT: Yeah.
ANDREW NG: And there just aren’t enough people to do the work.
KEVIN SCOTT: Correct.
ANDREW NG: So, they welcome automation because the options are either automate or well, just shut down the whole plant because it is impossible to hire with the aging population.
KEVIN SCOTT: Yeah. In Japan, it actually is going to become a crucial social issue sometime in the next 100 years or so because their fertility rates are such that they’re in major population decline. So, you should hope for really good AI there, because we’re going to need incredibly sophisticated things to take care of the aging population there, especially in healthcare and elder care and whatnot.
ANDREW NG: You know, I think when we automated elevators. Right? Once elevators had to have a person operating them, a lot of elevator operators did lose their jobs because we switched to automatic elevators. I think one challenge that AI offers is that there will be as connected as it is today, I think this change will happen very quickly, or the potential for jobs to disappear is faster this time around.
So, I think when we work with customers, we actually have a stance on wanting to make sure that everyone is treated well, and to the extent, we’re able to step in and try to encourage or even assist directly with retraining to help them find better options, we’re truly going to do that. That actually hasn’t been needed so far for us because we’re actually not displacing any jobs. But if it ever happens, that is our stance. But I think this actually speaks to the important role of government with the rise of AI.
So, I think the world is not about to run out of jobs anytime soon, but as LinkedIn has said through the LinkedIn data and many organizations, and Coursera has seen and Coursera’s data as well, our population in the United States and globally is not well-matched to the jobs that are being created. And we can’t find enough people for- we can’t find enough nurses, we can’t find enough wind turbine technicians, a lot of cities, the highest paid person might be the auto mechanic and we can’t find enough of those.
So, I think a lot of the challenge and also the responsibility for nations or for governments of a society is to provide a safety net so that everyone has a shot at learning new skills they need in order to enter these other trades that we just can’t find enough people to work in right now.
KEVIN SCOTT: I could not agree more. I think this is one of the most important balances that we’re going to have to strike as a society, and it’s not just the United States, it’s a worldwide thing. We don’t want to under invest in AI in this technology because we’re frightened about the negative consequences it’s going to have on jobs that might be disrupted. On the other hand, we don’t want to be inhumane, incompassionate, unethical about how we provide support for folks who are going to be disrupted potentially.
ANDREW NG: Yeah.
KEVIN SCOTT: I think Coursera plays an incredibly important role in managing this sea change in that we have to make reskilling and education much cheaper and much more accessible to folks. Because one of the things that we’re doing is, we’re entering this new world where the work of the mind is going to be far, far, far more valuable even than it already is than the work of the body. So, that’s the muscle that has to get worked out and we’ve just got to get people into that habit and make it cheap and accessible.
ANDREW NG: Yeah. It is actually really interesting. When you look at the careers of athletes, you can’t just train them in great shape at age and then stop working out. The human body doesn’t work like that. Human mind is the same. You can’t just train, work on your brain until you’re and then stop working out your brain. Your brain you go flabby if you do that.
KEVIN SCOTT: Yes.
ANDREW NG: So, I think one of the ways I want the world to be different is I want us to build a lifelong learning society. We need this because the pace of change is faster. There’s going to be technology invented next year and that will affect your job five years after that. So, all of us had better keep on learning new things.
I think this is a cultural sea change that needs to happen across society, because for us to all contribute meaningfully to the world and make other people’s lives better, the skills you need five years from now may be very different than the skills you have today. If you are no longer in college, well, we still need you to go and acquire those skills. So, I think we just need to acknowledge also that learning and studying is hard work. I want people if they have the capacity.
Sometimes your life circumstances prevent you from working in certain ways, and everyone deserves a lot of support throughout all phases of life. But if someone has the capacity to spend more time studying rather than spend that equal amount of time watching TV, I would rather they spend that time studying so that they can better contribute to their own lives and to the broader society.
KEVIN SCOTT: Yeah, and speaking again about the role of government, one of the things that I think the government could do to help with this transition is AI has this enormous potential to lower the costs of subsistence. So, through precision agriculture and artificial intelligence and healthcare, there are probably things that we can do to affect housing costs with AI and automation.
So, looking at Maslow’s Hierarchy of Needs, the bottom two levels where you’ve got food, clothing, shelter, and your personal safety and security, I think the more that we can be investing in those sorts of things, like technologies that address those needs and address them across the board for everyone, it does nothing but lift all boats basically.
I wish I had a magic wand that I could wave over more young entrepreneurs and encourage them to create startups that are taking this really interesting, increasingly valuable AI toolbox that they have and apply it to these problems. They really could change the world in this incredible way.
ANDREW NG: You make such a good point.
KEVIN SCOTT: So, the last tech thing that I wanted to ask you is, there is sort of just an incredible rate of innovation right now on AI in general, and some of the stuff is what I call “stunt AI” not in the sense that it’s not valuable but it’s-
ANDREW NG: No go ahead. Name some names. I want to hear.
KEVIN SCOTT: No, so I’ll name our own name. So, we, at Microsoft did this really interesting AI stunt where we had this hierarchical reinforcement learning system that beat Ms. Pac-Man. So, that’s the flavor of what I would call “stunt AI.” I think they’re useful in a way because a lot of what we do is very difficult for layfolks to understand.
So, the value of the stunt is holy crap, you can actually have a piece of AI do this? I’m a big classical piano fan and one of the things I’ve always lamented about being a computer scientist is, there’s no performance of computer science in general, where a normal person can listen to it or if you’re talking about an athlete like Steph Curry, who has done an incredible amount of technical preparation and becoming as good as he is at basketball, there’s a performance at the end where you can appreciate his skill and ability.
And these “stunt AI” things in a way are a way for folks to appreciate what’s happening. Those are the exciting AI things for the layfolks. What are the exciting things as a specialists that you see on the horizon? Like new things and reinforcement learning, coming, people are doing some interesting stuff with transfer learning now where I’m starting to see some promise that not every machine learning problem is something where you’re solving it in isolation. What’s interesting to you?
ANDREW NG: So, in the short term, one thing I’m excited about is turning machine learning from a bit of a black art into more of a systematic engineering discipline. I think, today, too much of machine learning among a few wise people who happen to say, “Oh, change the activation function in layer five.” And if for some reason it works, then that can turn into a systematic engineering process that would demystify a lot of it and help a lot more people access these tools.
KEVIN SCOTT: Do you think that that’s going to come from there becoming a real engineering practice of deep neural network architect or is that going to get solved with this learning to learn stuff or auto ML stuff that folks are working on, or maybe both?
ANDREW NG: I think auto ML is a very nice piece of work, and is a small piece of the puzzle, maybe surrounding, optimizing [inaudible] preferences, things like that. But I think there are even bigger questions like, when should you collect more data, or is this data set good enough, or should you synthesize more data, or should you switch algorithms from this type of algorithm to that type of algorithm, and do you have two neural networks or one neural network offering a pipeline?
I think those bigger architectural questions go beyond what the current automatic algorithm is able to do. I’ve been working on this book, “Machine Learning Yearning” mlyearning.org, that I’ve been emailing out to people on the mailing list for free that’s trying to conceptualize my own ideas, I guess, to turn machine learning into more of the engineering discipline to make it more systematic.
But I think there’s a lot more that the community needs to do beyond what I, as one individual, could do as well. But that will be really exciting when we can take the powerful tools of supervised learning and help a lot more people are able to use them systematically.
With the rise of software engineering came the rise of ideas like, “Oh, maybe we should have a PM.” I think those are Microsoft invention, right? The PM, product manager, and then program manager, project manager types of roles way back. Then eventually came ideas like the waterfall planning models or the scrum agile models. I think we need new software engineering practices. How do you get people to work together in a machine learning world? So all sorting it out to Landing.AI ask our product managers do things differently, then I think I see any other company tell their product managers to do. So we’re still figure out these workflows and practices.
Beyond that, I think on a more pure technology side [inaudible] again as I do transform entertainment and art. It’ll be interesting to see how it goes beyond that. I think the value of reinforcement learning in games is very overhyped, but I’m seeing some real attraction in using reinforced learning to control robots. So early signs from my friends working on projects that are not yet public for the most part, but there are signs of meaningful progress in the reinforced learning applied to robotics. Then, I think transfer learning is vastly underrated.
The ability to learn from- so there was a paper out of Facebook where they trained on an unprecedented 3.5 billion images which is very, very big 3.5 images is very large, even by today’s standards, and found that it turns out training from 3.5 billion, in their case, Instagram images, is actually better than training on only one billion images. So this is a good sign for the microprocessor companies, I think, because it means that, “Hey, keep building these faster processes. We’ll find a way to suck up their processing power.”
But with the ability to train on really, really massive data sets to do transfer learning or pre-training or some set of ideas around there, I think that is very underrated today still. And then super long term- We used the term unsupervised learning to describe a really, really complicated set of ideas that we don’t even fully understand. But I think that also will be very important in the longer term.
KEVIN SCOTT: So tell us something that people wouldn’t know about you.
ANDREW NG: Sometimes, I just look at those bookstore and deliberately buy a magazine in some totally strange area that I would otherwise never have bought a magazine in. So whatever, five dollars, you end up with a magazine in some area that you just previously knew absolutely nothing about.
KEVIN SCOTT: I think that’s awesome.
ANDREW NG: One thing that not many people know about me, is I actually really love stationery. So my wife knows, when we travel to foreign countries, sometimes I’ll spend way too long looking at pens and pencils and paper. I think part of me feels like, “Boy, if only I had the perfect pen and the perfect paper, I could come up with better ideas.” It has not worked out so far, but that dream lives on and on.
KEVIN SCOTT: That’s awesome. All right. Well, thank you so much, Andrew, for coming in today.
ANDREW NG: Thanks a lot for having me here, Kevin.
CHRISTINA WARREN: That was a really terrific conversation.
KEVIN SCOTT: Yes, it was a ton of fun. It was like all of my best conversations, I felt like it wasn’t long at all and was glancing now at my phone and I’m like, “Oh, my god. We’ve just spent 48 minutes.”
CHRISTINA WARREN: One of the questions that you asked Andrew was, what technology is he most impressed by and excited by this coming down the pike with AI? I wanted to turn that back on you because you’ve been working with AI for a really long time at Google, and at LinkedIn, and now at Microsoft. So what have you seen that really excites you?
KEVIN SCOTT: Several things. I’m excited that this trend that started a whole bunch of years ago, more data plus more compute equals more practical AI and machine learning solutions. It’s been surprising to me that that trend continues to have legs. So, when I look forward into the future and I see more data coming online, particularly with IoT and the intelligent edge as we get more things connected to the Cloud that are sensing either through cameras or far field microphone arrays or temperature sensors or whatever it is that they are, we will increasingly be digitizing the world.
Honestly, my prediction is that the volumes of data that we’re gathering now will seem trivial by comparison to the volumes that will be produced sometime in the next - years. I think you take that with all of the super exciting stuff that’s happening with AI silicon right now and just the number of startups that are working on brand new architectures for a training machine learning models, it really is an exciting time, and I think that combo of more compute, more data is going to continue to surprise and delight us with interesting new results and also deliver this real world GDP impacting value that folks are seeing. So that’s super cool.
But I tell you, the things that really move me, that I have been seeing lately are the applications into which people are putting this technology in precision agriculture and healthcare. Just recently, we went out to one of our farm partners. The Microsoft Research has been working with the things that they’re doing with AI machine learning and edge computing in this small organic farm in rural Washington state is absolutely incredible.
They’re doing all of this stuff with a mind towards “How do you take a small independent farmer and help them optimize yields, reduce the amount of chemicals that they have to use on their crop, how much water they have to use so you’re minimizing environmental impacts and raising more food and doing it in this local way?” In the developing world, that means that more people are going to get fed. In the developed world, it means that we all get to be a little more healthy because the quality of the food that we’re eating is going to increase.
There’s just this trend, I think, right now where people are just starting to apply this technology to these things that are parts of human subsistence. Here’s the food, clothing, shelter, the things that all of us need in order to live a good quality life. I think as I see these things and I see the potential that AI has to help everyone have access to a high quality of life, the more excited I get. I think in some cases, it may be the only way that you’re able to deliver these things at scale to all of society because some of them are just really expensive right now. No matter how you redistribute the world’s wealth, you’re not going to be able to tend to the needs of a growing population without some sort of technological intervention.
CHRISTINA WARREN: See, I thought you were going to say something like, “Oh, we’re going to be able to live in the world of Tron Legacy or the Matrix or whatever.” Instead, you get all serious on me and talk about all the great things that in the world changing awesome things that are going to happen. I’m going to live in my fantasy but I like that there are very cool things happening.
KEVIN SCOTT: I did over my vacation read “Ready Player One” and despite its mild dystopian overtones.
CHRISTINA WARREN: It’s a great book. I like the book.
KEVIN SCOTT: That’s a damn good book. I was like, “I want some of this.”
CHRISTINA WARREN: I’m with you. I’m with you. I was a little disappointed in the movie but I loved the book. Yeah. We can talk about this offline but we’ll end this now.
KEVIN SCOTT: Yeah. Well, awesome Christina. I look forward to chatting with you again on the next episode.
CHRISTINA WARREN: Me too. I can’t wait.
KEVIN SCOTT: Next time on Behind the Tech, we’re going to talk with Judy Estrin who is a former CTO Cisco, serial entrepreneur, and as a Ph.D. student, a member of the lab that created the Internet protocols. Hope you will join us. Be sure to tell your friends about our new podcast, Behind the Tech, and to subscribe. See you next time.