Dr. Andrew Ng, Founder and CEO, Landing A.I. and DeepLearning.AI
Interviewer: Brian O’Keefe, FORTUNE

Fortune Brainstorm A.I. 2021: Big breakthroughs

Conferences

the development and diffusion of artificial intelligence is transforming businesses and impacting multiple facets of our everyday lives. I mean the conversation so far has been proof of that one of the world's foremost computer scientists and AI pioneers highlights the importance of closing the gap from proof of concept to production in order for a i to to deliver on its full potential outside the consumer internet industry. He may also share the next big breakthrough what the next big breakthrough in AI is likely to be dr Andrew ng is founder and ceo of Landing AI, an adjunct professor of computer science at Stanford University. He joins us remotely in a conversation with fortunes brian O'keefe. Thank you rana And there's Andrew. Hello Andrew. Hi, hi brian. So great that you can join us. This is, we can hear you, great, we can hear you. Great. Can you hear me? Yes and today it's great to be here virtually. And I'm so sorry, I couldn't tell you in person I was excited to come in person to boston but with a big announcement today which is why I had to stay close to home but I'm glad to see you all virtually. Yeah, well you're giving me my segue but we're very happy to have you joining us virtually because it shows that we're very cool. 2021 conference, we can bring you in but let's let's go right to that news that you mentioned today. Um uh big news Landing ai your company announced today, $57 million in venture funding series a funding at an undisclosed evaluation. We assume it's very, very big. Um It was led by MacC Rock Capital intel Capital invested Samsung catalyst fund. So some really big time investors, first of all, congratulations. And I'd like to know what was your pitch to the investors? Um, I couldn't even give my pitch to all of you because it affect many of you watching this as well, which is that whereas we figured out how to get a I to work for consumer software, Internet other industries like manufacturing around focus for other industries as well. Needed different recipes. So then the AI has been building data centers and the attitudes, but to to make it fast and easy for other companies focused on manufacturing to build and deploy successfully. So how did you get this idea? I mean, you're so well known for working in the big consumer tech industry. How did you get the idea to apply AI and computer vision to manufacturing? What what was the spark? Yeah. So my previous engagement is something that the google brain team also ran the idea by do and that whole, respectively, to great internet companies embrace modern AI and deep learning. When I stepped away from my road leading those teams, I decided I've done this to consumers internet companies, someone should keep doing that. But I don't need to do it anymore. And I went to work to take me out to other industries and frankly, it's somewhat naively when I started on the a I I did not understand, I misunderstood that a different recipe than what some friends and I had built up to consumer internet would be needed for these other industries. We re studies like Mackenzie saying $13 trillion of ai value creation in the next year, a handful of years, but casually outside consumer software internet, I'm not seeing this realize much as we realize anywhere yet. So that's what I'm expected to do. So we're gonna just wanna remind people in the room, we're going to have a few minutes for questions in a moment. Andrew. Um but I want to ask you, you know, we've had some some talk earlier today about data. I mean you're looking at a different data pool to solve these specific problems in manufacturing, that's a lot of what you're taking on now. Andrew moore from your old company Google uh google cloud was here earlier today and he was talking about how google has developed a universal defect detector tool, but you know, taking massive amounts of data of every kind of defect, you know, in different types of products and that they could apply that to custom problems. I think you're approaching it a different way. Could you maybe talk about the differences in that approach? Yeah. So, you know, when when I was working in consumer software companies, I once built a computer vision system using 350 million images. And when I gave a talk in the manufacturing conference and asked the audience how much data do you have? The majority of the audience said 50 or fewer images for defense. So 350 million versus 50 or fewer. That's quite a stark contrast. Mhm. And it turns out maybe not surprisingly uh tech bill for 300 million images doesn't work if you have 50 images and that's all the data that exists in the world. You go to smartphone maker, they only made 50 defective phones of that type of defect or go to L. C. D. Panel maker, PCB maker, some content away from maker. You know, automaker, they don't make that many defective parts. So maybe all the data that exists in the universe It's 10 or 20 or 50 or 100 images and you either have to make the tech work for that or not. So I think that maybe I would been pioneering new direction for Ai Data Central Day I we launched the movement but now lots of companies that work on acknowledges that I'm quite proud of the number of countries talking about data century AI which is generating a new version of machine learning to AI to that is increasingly able to work even when the dataset sizes are smaller. And I think that would be critical to break open these applications of Ai into other industries where the data says just on as big as consumer internet industries with a billion, jesus and very large data. So data centric Ai is the key to applying ai to these more specific problems where there's more limited data, it's a new technology for AI and then I can share you one of the barriers I see two ai adoption outside consumer software internet. I need to be useful to many of you watching this as well. Um which is that in consumer software internet, you can build one monolithic web search engine or one Ai to serve ads or one speech recognition system or something of working all of those. But you can build one Ai system to serve, you know tens or hundreds or millions or maybe billions of users and generate massive economic, but in most other industries, the world's much more heather genius. So rather than everyone kind of using one or a small number of West search engines, if you go to manufacturing every manufacturing plant, make something different. And so an Ai system for detecting scratches, the smartphone, well that's not going to find cross lines in the semiconductor wafer and you need something else to find dents in the sheet of steel. So the economics is instead of one Ai model generating a billion dollars worth of value. The manufacturing industry maybe needs just use round numbers. Maybe you need 10,000 AI models, each of which generates 5,000,001 $25 million of value. And the AI world, my community, we figured out how to get dozens or maybe a few 100 engineers to build one monolithic Ai system. And so for many years there were proof of concept that small scale and these other industries, but until recently no one really had a recipe for how to help and industries generate 10,000 AI models issue, which generates a few million dollars worth of value and lets the same mistake. Even though I'm focused on manufacturing, this problem is pervasive. So take health care. Every hospital has kind of their own way of clothing, electronic health records. So I don't think it's possible to build one Ai model to read all healthcare records across our hospitals And the challenges again. Every one of 10,000 hospitals and 10,000 healthcare systems needs their own AI model, how can one company do that without, without someone like me hiring 10,000 machines and engineers. Which is pretty hard even for me. So there there's the pitch, there's the pitch to investors. I think that was a key apartment. So I want to make sure we have time to get a couple questions I think already saw one in the back. If you could just make sure you speak into the mic and identify yourself so Andrew, can hear you and if you can't, I'll make sure you can hear it here. This is the location along from established black ethical. Hello? Uh, so as you have less data at this manufacturing applications. Are you also using this knowledge heavy techniques like causality or other. Knowledge based systems in order to supplement the data. Yeah. So thanks for the questions uh actually uh called out specifically we're not using that but we do a lot of work to make it possible for the domain experts to the subject matter experts to express their knowledge. Um actually one of the standing back executive team members that publicly talked about their work with us. So bring forth of that a partnership. Um but I think that in manufacturing their inspectors with deep domain knowledge about what is the defect? Uh for example what exactly is a defect? And well I guess exactly doesn't love welding or what exactly is a stretch. This scratch so small you can safely ignore it and what is the stretch And this is a stretch of the chip. So there are inspectors and in factoring with domain knowledge and it turns out every factory actually tends to define his own defects difference. So I think the right approach, frankly, I think the only approaches and think of is to build tools and we call them data, centric Ai to that makes it fast and easy. But someone will get deep domain knowledge to express the deep knowledge about what exactly is a defect or well or septic Midway for all the component or whatever in terms of labeling data for Ai so they can show a hand for very clear examples of Ai that can then learn to detect these defects reliably even from just a modest number of example. Great! We have another question in the room from back here again, just speak clearly into the mic. Thank you. Andrew this is real hit. Japan from Mastercard. Next few 100 quick question. So once you want all the work that you guys are doing at landing Ai gets into production, does that give us a glimmer of hope of getting manufacturing back in this country? Great question. Thank you for that. Great question. You know, uh living in the United States, I think a lot about national competitiveness and I think that the best way for America or for any nation to build up in manufacturing and to create jobs, preserve a Torrey jobs is to build for the future rather than only trying to preserve the jobs of 30 years ago. So I think american manufacturing is one of those industries. I would love to see take off and deep learning technologies here. It's going to take off whether any of us like it or not. And I would love for american manufacturing. In fact the all nations around the world to embrace this and I think that to create a lot of jobs because it turns out a I is a very good compliment. So high school manufacturers and America's manufacturing raised relatively high and relatively high skill. I think that's a very strong compliment there for accelerating your manufacturing in this country. Great. We have another question on this side of the room. Sure. Hi Andrew Andy Hawk here from cerebral systems. Maybe speak up a little bit because I want to make sure you can hear Hi Andrew Andy Hawk here from cerebral systems. First of all, congratulations on the round. Uh second many of existing state of the art applications for AI were powered by large public open source datasets. So I'm curious how you think about that in the context of this data centric AI with smaller datasets, Do you plan on making data sets for the kinds of applications you're building available in the open source. Thank you. You know, I think that the types of data potentially open source tends to be um things like images of your text because that's a quite common, right? Like a different countries want to recognize, you know, pets and people and things like that. But I find a lot of in this industrial data is very specific. So for example, I've been surprised working two different manufacturers, the two different students actually work with, they have different ways of defining defects. So it's not even obvious to me that data from the manufacturer. Number one, is that useful for steel manufacturer. Uh Number two. And I think that at least for the foreseeable future, most companies treat their um images of products is highly secretive. You know, you don't want to let pictures frankly defective stuff. Right, go out, that's often a few. And these data sets are often small, proprietary, I would love for more companies to open source their data and I think that will help others. Although I think as a practical matter I think it'll be difficult to get any interaction companies to do that. And that's you know, working with one company at the time with proprietary data says small datasets especially seems like it's going to achieve to move this forward, so Andrew, we're getting close to time, but I just wanted to ask you, you know, if you look back at from the time when you found a google brain and the advancements that we've made from then to now, less than a decade. Um what has surprised you about the ways that ai has developed in that time period and what are you, you know, what would you point to as maybe like what would be very surprising or some areas where we would have these big breakthroughs uh in an unexpected place, you know, going forward if you think about the next 57 years. Mhm Yeah, so maybe in the last decade, I think the biggest movement in a I was a shift Deep learning also called neural networks that created a lot of the upside over the last decade. Um looking forward, I think the next big shift that needs to go through for the next decade is a movement to a data center day. I there's a little bit technical but bear with me a last decades a lot of progress on the software, the code and your natural details to build AI systems. But a lot of teams just took some data and shoved into these new networks. So what's missing, there's a more systematic entering discipline of treating good data defeat of AI. So something technically very deep field. So my team but multiple big companies are now working on this data. Century eye movement and I think this is also key to democratizing access to AI because it goes to manufacturing plants or a hospital or these terms and awesome to invent the next generation of map for a I you know, it's not that easy but lets them express the domain knowledge about what is the defect, what is it good electronic health record by entering the data that seems more feasible. And I think that would be highly recipe. So finally taking out from consumer software to net was worked great all of these other industry. Well, I think that's a perfect vision in with democratizing access to AI Andrew. Thank you so much for making time for us and joining us and we'll talk to you again soon. Thank you. Thank you. All right, join me in. Thank you. Andrea yeah.