Pie & AI Medellín: A Discussion with Andrew Ng and Helmuth Trefftz

We already have a lot of data, but computers, both CPU computing and especially GPU computing, were getting fast enough for us to finely train big enough neural networks to take advantage of the data we already have. So I saw a few different communities flip from using more traditional methods and using deep learning. I think speech recognition was actually one of, the first, one of the first to flip, thanks to work with people like Don B. And Jeff Hinton. Computer vision came a little bit later with imagenet. And then I saw MLP flip after that. And then robotics. You know, I think a lot of healthcare work is not converting as well. Even though the technology has improved dramatically, one of the things that our community still has a gap to fill is to take the great research results and to translate them into more practical and deployed results. The gap between a research paper and a practical deployment is still quite large and our community is still getting better at that in terms of future technologies. It turns out that a lot of the rise of AI was driven by the large consumer Internet companies like Google, Facebook, Microsoft, I do that have a lot of users and therefore have a lot of data. One of the things I'm excited about right now is the rise of small data. You've heard about the rise of big data and if you have a million images with, you know, 100 million user records and three giant networks that runs there. But once you start to look outside consumer Internet, for example, if you have a manufacturing plant that is inspecting smartphones to see if they have scratches on them or not, fortunately no factory has manufactured a million cell phones of scratches on them that need to be thrown away. And some factories just do not have millions of pictures. And if you're able to get new networks of work, even with small data, you know, 10 pictures or 100 pictures or million pictures, it dramatically opens up the set of applications. And so with the advances in things like few shell learning, one shell learning, data generation, meta learning, I'm seeing exciting progress in small data. And actually the work that our team in Medium is doing, a lot of these innovations are allowing us to enter industries that just do not have the job data sets of your consumer Internet. Exactly.

Can we go a little bit in that direction? You know, of course not everything can be disclosed, but could you tell us about some areas that the company Medin is working on? Like broad areas?

Yeah. So let's see. One of the things I'm excited about is a lot of the. I feel like small data is the cutting edge of deep learning. It's not the only one. It's One of the cutting edge areas of deep learning and a lot of the work that our team is doing, a lot of work that navigate AI is doing in small data is done by the team here in Medellin. So I think we're teams doing cutting edge, you know, deep learning work that I don't think is being done in America yet, for example. And I think other than that, the team in Henryji has I think over a dozen projects ranging from have an educational initiative, Chin Justin here that is taking engineers, software engineers, providing machine learning training and then helping them get placements in other companies to do AI work. Doing a range of projects from vision section that you, that you heard mentioned actually this afternoon as you're being with the team, the new networks are building the PCB purchase to a good board manufacturing again, new networks that are different than anything in any published paper. So right backwards cutting edge work. Projects in healthcare, projects trying to optimize fuel savings, some projects in trying to improve job recommendations. So our team here in Medellin includes the launch team from Nadine AI which is helping companies jumpstart AI Adoption AI Fund, which is the 175 million US dollar startup studio that Helman mentioned, which is building many different AI companies including four portfolio companies with a presence in Magellene. And then also Deep Learning AI which is our educational teaching deep learning specialization which as well as AI for everyone. And as part of a commitment to this region, Deep Learning AI has also created and is releasing Spanish Transl so that the Spanish speaking audience can also have easier access.

The Ministry of Technology in Colombia has a plan to form over 4,000 people in artificial intelligence engineers that will be capable of doing machine learning and artificial intelligence. Do you think that plan makes sense? And if so, what do you think will be the impact of such a plan in Colombia if we accomplish to form over 4,000 of machine learning engineers?

I think training 4,000 people is a great start and I would love to see it grow over a few years to 40,000 and then maybe 400,000 is a lot, but I would love to see it eventually end up significantly more than 4,000. You know, today Silicon Valley of Beijing were our tech hubs of AI. They weren't always this way. By the way, in 2014 when I went to China, I joined by you to build AI in China. One of my American friends asked me, they said, hey Andrew, why are you trying to build AI in China? Like how much AI is there in China? But at that time in 2014, I saw all the elements in China And I had conviction. I was convinced that AI could rise quickly in China and maybe hopefully I played a role in helping that out. One of the early encouraging signs I see today is that our office, in addition to attracting, you know, people from Colombia to come join us under Alejandro Betinku's leadership, as you should, hopefully, when you get to wave Alejandro. The team has also attracted people to St. John's office coming from Poland, from Egypt, from Bangladesh and from Chile. So one of the things I'd love to see managing the company is a low, at least local, but maybe even global talent hub where if someone in Poland says I'm gonna work in AI, where do I move? Well, they could move to Silicon Valley, maybe they could move to Beijing, or why not think of moving to Colombia? Some seeing early signs of that now an ecosystem is much bigger than any one company. I feel like for us to be. It's impossible for us to be successful by ourselves, which is why a lot of what we want to do is contribute to the community, want to have multiple successful AI teams and have a lot of people in the community learn about machine learning, deep learning and be able to use these tools effectively. Also been grateful for university partnerships with UTA Unite to explore if they can offer more AI training and I think for AI to be used successfully. I know many of you are engineers, if you learn about machine learning deep learning, the field is slow down enough. This is a great time to break in. But also business leaders, it turns out that for AI machine learning to be successful, company is not sufficient to have a team of 10 brilliant engineers. And then for the rest of the company, they know nothing about AI. And so I think providing education at all levels, which deep learning AI does with the technical courses and the non technical course that AI for everyone, but which I'm very encouraged that the Colombian government is doing as well, will help all of us be more successful. You know, let me tell you a secret. Let me tell you one thing I'm not sharing with other people. One of the things I did at DL AI was create a course called AI for everyone, which was attempted for a non technical audience, right? So you know, for the CEO or vice president or manager, product manager that doesn't know how to code. One of the reasons I did that was because I felt as a CEO of a company or you know, VP or management team takes AI for everyone, they will be in a much better position to set the machine learning team up for success. So part of my secret plan was to get more, you know, company executives to Take AI for everyone. Then this will make them want to hire even more machine learning engineers. And this will maybe even worsen the talent shortage we have with machine learning actually is about life.

Any secrets?

Okay.

Last time you mentioned some of the top hot topics that were there one year ago that were hot topics in machine learning. One of them, for instance, you mentioned that already was being able to train your neural networks with smaller data sets. Colleague, I understand small data. What, what other hot topics in machine learning or AI would you.

Signal as the most significant one?

The ones probably bringing the most important changes.

Yeah. Mention a few things. I'm excited about slightly less technical side. I'm excited about the rise of machine learning as a systematic injury discipline. So to make an analogy, I think thousands of years ago, if you wanted to build a bridge, there used to be these wise old men and wise old women that knew how to build bridges that stay up. If you ask them to take a look at which building, they'll say, hey, I have a piece of wood there. And if you add that piece of wood, for some reason the bridge stays out. But eventually we have to create the engineering discipline of civil engineering. And today people all around the world build reliable bridges. I think machine learning and deep learning is still in that earlier era where, you know, a student may train in your network and for some reason it doesn't work. So they go ask, hey Hellman, what do I do? And Hellman says, oh, change the activation function in the fifth layer from, you know, whatever. And then for some reason, One of the most encouraging things I've been seeing in deep learning is this evolution from a, this form of a tribal knowledge, intuition based discipline to a systematic injury discipline. And our teams, you know, naming AIF and Uganda, I feel like we are one of the teams that are in the process of making up new processes all the time. So for example, one of the process that some of. So for example, in software engineering you may have heard of, you have strong and two weeks strong sprints. One of the very unusual things that some of my teams in machine learning do is run one day sprints, right? Because rather than executing for two weeks and seeing where you are and changing directions, when I work on machine learning projects, sometimes we write code during the day, run an experiment overnight, the next morning, we do error analysis to decide what to do next. Then we based on that, we decide what code to write the second day. And then we run these sprints second day at night. And so this is one example of a very strange process. Sometimes one day sprint that's crazy, right? But in the machine learning world, for some projects, not all projects, but some projects, this allows us to drive much faster process. So this is one. And if you look at software change, it took us a long time to figure out that version control was a great idea. And in fact it took a long time to evolve right from I guess CBSE subversion to git. Maybe there'll be something else. Takes a long time to figure out. That code review is a great idea. And I think in the machine learning where we're still making up those processes, but these processes make machine learning more systematic. Some example of that. Oh, real quick. I was at the ICML International conference machine learning a couple months ago. In terms of techniques and machine learning to how excited about. I'll just mention two. Well, one is never look down into that. But people like Peter P. Chelsea Finn openly are doing great work on that. I think a lot of hyperbole reinforcement learning works well simulated hard to make it work. The real robot. I'm seeing signs of traction meta reinforcement learning. And then I'll mention one other weird thing I'm excited about which is self supervised learning. At icmr I spent a whole day at the self supervised learning workshop. And there's this idea that from unlabeled data you can make up supervised learning tasks that allow you to learn very useful hidden representations you can transfer to new tasks. So for example, one of the tasks you could have is take a picture, cut it into a three by three grid. Here's a picture, cut it so you have nine pieces. I'm going to shackle these nine pieces fantasy and then ask the supervised learning algorithm to tell me what was the original mapping of these nine pieces. So you can make up unlimited amounts of labeled data because you're just taking unlabeled images, cutting it up and then sharpening the nine pieces. You can join infinite amounts of labeled data, but with carefully designed artificial tasks that you create from labeled data seem very encouraging signs. You can train a giant neural network, learn fantastic representation and transfer to Utah. So self supervised learning is an exciting vertical. Sorry, I know that was very technical, but these are maybe in my life I've sometimes been right. Not always. I've been right a couple times about what technologies will be important in the future. So this is one that I'll make sense.

Fascinating.

I hope you wouldn't be right. Again. Again. One more time.

Some of the next questions have to do with the careers of the people who are sitting here. So if you don't Mind, I would like to know a little bit better about the audience. How many here? Please raise your hands if you are students.

Okay.

And all of those who are undergraduate students. Quite a few graduate students.

Okay.

And who are professors, teachers like yourselves.

Okay.

Quite a good number. And the rest, I assume, are practitioners of, you know, professionals, software engineers, or practitioners of any other field.

The rest.

Okay, so here comes the question. If a developer in Colombia wants to pursue a career in AI machine learning, what should this developer do? What should she or he start learning? And what would your recommendation be?

You know, the best time to join a field is often when they're still taking off. AI is still taking off. Huge shortage of people that can work in AI. So this is a great time still to be jumping into it. And one common misconception of building AI projects, this is all machine learning engineers, you know, using Jupiter nodals to train your networks. That's actually not true. A good AI team in order to build practical skills systems to ship products often has a mix of both, let's call machine learning engineers as well as software engineers or software developers, front end, back end, mobile. And it usually takes that type of skill working together in order to build a valuable AI product. In terms of software engineers, it turns out that. Let me, let me tell you what I saw in Silicon Valley, right? Ten years, five, maybe ten years ago, one of the ways to win glory as a backend engineer was just work on bigger data sets and everyone else, right? So, you know, you can imagine in a bar in Silicon Valley, two engineers would get together and would say, I process a petabyte of data. And then you go, Well, I process 10 petabytes of data. I think we're moving into an era where skill as a software engineer is not about who processes the most data, is about other things. And with the rise of the AI era, a lot of the most complicated problems at the intersection of AI plus front end or backend or mobile or other things. And I'll give you one example. Let's say you, we build a lot of AI products, so we see all these rounds. Let's say you work on building a machine learning system that takes us input and X ray image and tries to diagnose if the page has cancer, right? Well, it turns out that the output of the machine learning algorithm is uncertain. Neural learning algorithm may look at chest X ray and say, oh, there's something weird here, there's something weird there. But 20% chance is cancer there, 80% chance I'm weird there. But how do you Design a website. How do you build a website that can display this type of information which previously computers did not have? So I think one of the most exciting work to be done on front end engineering is on thinking through reinventing how front end software engineering should work. In the AI era, where, I don't know, you can sometimes train a, train a neural network right there in your web browser and it uploads these crazy things, they're very hard to display the user on the back end. The way we build machine learning data pipelines, we are building very complicated pipelines that I've not seen anywhere else before. Right. You know, get data from a factory, it comes back latency requirements where you store the data, edge, fog, computing connections. And we're facing very complicated problems driven with the need, driven by the existence of the new AI technology. So I feel like many people will have great careers just being a, you know, software engineer. I think it's a fine thing to do. But the new disruptive technology today is AI and I think a software engineer, front end, back end, mobile, whatever, full stack that figures out how your work needs to change in the AI era will be doing something very unique and very valuable and very intellectually exciting. You don't need to, let's see, machine learning is still immature enough that you take some classes from serum on team AI, you could learn enough to get a job as a, as a entry machine learning engineer. But if you're experienced backend engineer, you don't want to be an entry level machine learning engineer. What's much better for your career is that to learn a little bit about AI and then work on the really hard and really important problems on how back end engineering is changing in the AI era. And that's the cause of problems that very few people are working on. And it gives you more of the opportunity to do very unique and very intellectually exciting and important work.

Very nice.

There's your tip.

Okay, these guys here are already working for, for the local company, Medium, right? Raise your hand. So these people see.

I know, I know.

That's, that's where I went to. Who wants to work for their company, for his company managing.

Okay.

So we hear that the selection process for your company is quite different from the traditional Colombian companies. Because, you know, I was joking with Andrew. I was telling him that in the interviews here in Medellin, when you go to a company, they usually ask you.

What'S your mother's name, what's your father's name, what do you like?

For what kind of movies do you.

Like and things like that.

Whereas in their interview process, for instance, they ask them to go in front of a whiteboard and solve a problem and things like that. So what would be your recommendation? Very practical recommendations for these people who want to apply to your company. How should they prepare?

Yeah, let me. I have two suggestions. You know, if you're interested in working with us. Don't sweat too much. Whether, okay, your friend. Do not do. Don't spend six. Don't feel like you need to spend six months preparing before you get in touch. Get in touch with our team. We'll tell you what our interview process is and we'll just tell you, you know, we think you're great or we might all just tell you, oh, why don't you do this first? And then we apply something. We're actually really friendly and we're bother. So don't wait, don't feel like you need to wait for months before you get in touch with our team. Just get in touch. We'll tell you what our interview process is and then what we do. So just get in touch soon. Our interview process is quite rigorous. I guess we use fairly interview processes in the US and in Colombia. And when people come, we will often they'll always ask technical questions. You know, can you solve these technical problems in machine learning or in software engineering or whatever. We usually ask you to do some kind of things and learn if you could, you know, write code to make things work. And then we'll often also ask you about projects you've done in the past. We'll often say, tell us about the project you've worked in the past and your problems. You can better understand what you've worked on before and how well you know that experience and mindset will translate to the types of doctors we work on here. But my advice is, I don't know, I feel like we do have a rigorous interview process. But don't worry too much about it. Just get in touch if you might be interested.

The question is, as a mother, you.

Can go to hell.

But think, think about the world that you're living for your children, right? And the world that your children will be living in.

Right.

So what kind of world do you envision for Nova with the support of AI?

You know, I think with every technological breakthrough we are given a chance to remake the world. And I think the world is not that bad today, but it also be so much better with AI capabilities. I would love for us to go and transform not just the software industry, but every industry and make human Activity much more efficient and much more valuable. And I think we honor and free up a lot of humanity from repetitive to. And I think one mistake that we, our community made the last time around was with the last, last technological disruption, the rise of the Internet. I think we created tremendous wealth, but we also contributed to wealth inequality. So this time around, with the rise of AI, I would like to put all the pieces in place ranging from better educational systems. And yes, AI will automate jobs or parts of jobs. That is our responsibility to make sure people have training in education and well taken care of and have a chance at a meaningful career. But what I would most like is we will use AI to create training as well. That's already really clear. But what I would really like to do is to make sure this time around that we make sure that wealth is fairly shared and that we take everyone along with us. Beautiful.

Before we go on to the questions, I think it's fair to say that Andrew and his friends at Stanford already.

Changed the world by creating Coursera. Right.

It has really changed the way education is delivered. High quality education is delivered to everybody in the world. Let's hope that you're right again and you will your leadership be creating a network.

So thank you very much. Yeah. You know, the rise of deep learning, something maybe almost 10 years ago, was driven by the rise of scale, right? Scale of data and scale of computation. But if you look at what the subscriptive AI teams do today, it's not that we use deep learning for everything. I think of AI as a portfolio of many different tools. And so, you know, sometimes we actually reduce deep learning a lot, but sometimes. A few months ago I was sitting down with a team debunking implementation of pca, right. Looking at residuals of PC and I think the experience AI teams will have portfolio of tools ranging from machine learning and deep learning to maybe a knowledge draft, practice graph models to planning algorithms. And depending on the problem we're facing, we'll draw from different elements of this portfolio in order to solve the problem at hand. I think so much hype about deep learning because of all of these portfolio of tools. Deep learning is the one whose performance is now so much greater than say five years ago. Whereas in contrast, I don't know that Bayesian networks work that much better. You know, I'm sure because of U or certain AI conference bass work a bit better now compared to five years ago. But it's not that, you know, 100x improvement or something that we saw deeper. You know, I think there is excessive hype about AI, especially about AGI. I think artificial general intelligence hopes or fears or something did cause some amount of irrational exuberance and irrational investments. But if you look across the field of AI, yes, you know the media reported to have estimates of how much money right. Google have spent just on alpha R and D mine. But if you look at the net impact on a company like Google, deep learning has been incredibly profitable for the company as a whole. So I think, let's see. One of the big reasons I don't think that we're in a very bad bubble is because I have seen with my own eyes, you know, the value or the, or the money that some water companies have making with AI and that revenue is significantly greater than his investments, even though the investments are also very significant. So the ROI of these investments is then very positive. And as a community, one thing I, I, one thing I believe in is there are some approaches to AI that I found more promising. I told you some of that very mobile work as small data. He said the substance learning I'm excited about supernaly. There's some things I'm excited about. There's some approaches that I'm personally not as excited about in the short term. So for example, I am making no investments in quantum AI. We're doing almost nothing in deep reinforcement learning. I think, I think it's a great research topic but I think it'd be wonderful and useful. Having said that, I will 100% cheer on someone else doing deep RL or quantum AI or almost anything else for him. Reason I would not cheer on hype but I think we as a community, we are big enough for us to explore many different approaches and a small number of successes has paid out really, really well. I think we as a community we will explore some things that will turn out to be gen ends but our community is now so big that we can afford a 50 person team doing some piece of work that turns out the wrong direction because chances are there'll be a different 50 person team, maybe a different company that does something really amazing. So I think I will tend to cheer on everyone trying to push more at all of the different areas of AI. Even if I think some areas are more promising, I would invest more in those myself. But I'm glad that there are others investing in other areas as well because that's what we mean as a community. I think current government's already made many smart moves and I hope that continues. I think that investment in education is a big deal and I think Government has very strong competing power. So organizations like BTN are very effective. I think I'll actually be back in Colombia next month to speak at Columbia 4.0. But so events like that, the power of the government to compete, events put together technology and businesses is also very powerful at landing AI and AI funds. Our teams are working on products and that serve a global audience. And so, you know, the team is writing software that we use in the United States or using Asia or using other countries or using Latin America as well. But one thing in addition to having ours and serving global market, it turns out that every country of Colombia will have a lot of application scenarios that are very unique to Colombia. So the way that, you know, Colombia does common coffee agriculture, I should care about the Colombian coffee or the way that Colombia thinks about mining or oil and gas. There are actually a lot of problems that Colombia is in a better position to solve than some team sitting off in Silicon Valley. And so I think thoughtful improvements in promoting AI for the industries and the verticals in Colombia is already good at. I don't think Columbia should build another website search engine. You know, that was like the competition from 10 years ago. We already very good research engines, just use all the existing ones. But we don't have very good solutions yet for AI for mining or AI for agriculture, AI for cement manufacturing. But along the Colombian problems will be better small for Colombia than Silicon Valley. And I think those are vertical areas that I hope our government can focus on to rapa who move to the.

Next to the next sections.

Can I just say one more thing?

Please, go ahead.

I have a. I have a request to all of you. Over the last couple days, Alejandra and I travel the book down with my lead speaking of all the people. I think because our team has been operating here for a year and I know how good their work is and because even before we started, you know, we spent six months trying to figure out where to go. I find it maybe a little bit strange that I, as an American executive, I have a lot of faith in energy. And what I want to tell you is I need you to also have faith in medicine. Because if what we're here to do is to not just make, you know, companies neifati ma successful, but if what we want to do is build a community, be a regional or global talent hub, have first 4,000 and then maybe in the future 40,000 or even more people trained with AI and if you want your city to in the future be one of the global AI hubs so that the next time somebody in Poland is looking for a job, it's great that they come and join Landing AI. I think it's also great if they come to Medellin to join a different company. So what I want to do with all of you is work together to build this community, build this ecosystem. And for that to happen, I have f managin and I need you to also have faith in Medellin and to work together with us as a community to tell everyone we are going to be this in. Wonderful. Thank you.