HEADLINER - AI for Now, Dr Werner Vogels, Amazon

Good morning, Amsterdam! Oh, come on, come on, hey, light them up, get some energy, some oxygen in you. Yeah, thank you, Alexander, for sticking that IX knife in my back. So, as the CTO of Amazon for the past 20 years already now, you would think I would have a really good crystal ball to predict the future. And actually, I'm pretty good at it, not necessarily because I think about the future—I do—but because we literally have millions and millions of businesses running on our cloud platform. And by seeing what their challenges are, the kind of things that they're interested in, and the things that they are building, I have a really good idea where often technology is going.

Now, as you saw maybe on the intro movie, like you know, Blade Runner and stuff like that, if that's our future, if we think that the future of AI is what we saw in science fiction movies, yeah, we're probably going to be wrong. But I'm not going to talk about all sort of the new AI, because everybody suddenly starts to think that AI started about six months ago. That is absolutely not the case. And actually, I think maybe quite a few of you, maybe Amazon retail customers, you've been using AI for the past 25 years already. You just don't know it. If you used recommendations or similarities or fraud prevention or things like that, now that was all AI already.

And so it's not only that. You know, if you go back well over two and a half thousand years, if you look at Plato and Aristotle, both being students of Socrates, they were already reasoning about machines that would have human capabilities in their brain to help us out. They had differences, you know, Aristotle clearly thought that everything was sitting in the brain, where Plato had a more complete vision of all of that. Actually, if you read the Republic by Plato, he will describe a city-state where there were machines going around, autonomous machines helping you cleaning, helping you prepare your food, things like that. So the vision was there two and a half thousand years ago. It's nothing new.

However, I think in the 1940s and 50s, when we started building computers, suddenly that whole idea became realism. Plato and others were actually thinking that the brain was something symbolic, and as such, it was all about if we can sort of emulate the reasoning in the brain, we may be able to do all these other things. That turns out that was not necessarily the right approach. Definitely, robots and embodied AI sort of went out of the window at that moment. Things started to change in the mid-50s. There was a workshop at Dartmouth where a bunch of researchers came together and actually for the first time coined the term artificial intelligence. And it was really done because these researchers still thought that it was all about mathematical reasoning in the brain. After all, computers were built to do mathematics, and as such, was the ideal tool to think about how AI would evolve in terms of reasoning.

And again, you know, they still thought that the brain was something symbolic, that it would have mathematical structures in it, and as such, that was how you would actually approach the problem. And it was sort of called symbolic AI, still a top-down approach: how do we reason, the rest will follow. One part of that actually that became relatively successful in those days already were expert systems. And maybe this is before your time, but you know, I programmed some stuff in Prolog to actually build sort of a large rule-based system which incorporated sort of the knowledge of, let's say, doctors in it, and then follow a tree to actually find answers. Now, sort of one of the first real applications of artificial intelligence. Now, it didn't really go anywhere because it turned out that top-down approach didn't deliver anything. Although by now, automated reasoning and all the other tools that came out of that strand actually are really, really important, for example, in proving to yourself that your security features that you've built actually work.

Then the shift starts, and robotics start to arrive, and people start to think that artificial intelligence, just like Plato was thinking, actually should have a body. And we maybe, if the top-down approach doesn't work, maybe we start doing bottom-up. So what do robots need? Robots need vision, they need speech if they want to be able to understand and communicate with humans, they need to be tactile, all sorts of other things. And so the bottom-up approach came: let's see how we can make these robots work by developing these technologies like, you know, image recognition, text-to-speech, speech-to-text, things like that. And that's a field in AI that's been existent for the past 30, 40 years and been working on that. Actually, if you look at the Amazon fulfillment centers, we literally in each fulfillment center have running tens of thousands of robots around. They're not running into each other. And you're able to deliver packages and be organizing and things like that for the past 20 years already. This is nothing new.

And so all this idea about sort of now that we've had technologies, and if you look at the past technologies, probably I think the rise of hardware, both in terms of GPUs and other training mechanisms, combined with, let's say, deep learning, reinforcement learning, all of that has given us a world of AI that works right now. However, you know, foundational models, transformers, factors, things like that, have also given us a new strand which allows us to train massive data sets, ends up with these LLMs and things like that. And I'm pretty sure that all the other 50 speakers at this conference will talk about that. I won't, because it's equally important to think about what is actually in AI right now that absolutely works and that we know how it works and that we can build great applications with. I could easily do a presentation of an hour or two about all the different amazing things that my customers have been building in the past years.

And AI for now, we really have to think about that. It's not necessarily all the LLMs and things like that that you need to solve the problems that you would like to do with your company right now. And as I said earlier, you know, these things, we've been using it for a long time now. If you think about a tool like Alexa or any of the other voice assistants, or, you know, your hands-free stuff in your phone, which has been around also for 20 years, all of that would classify as AI. Why? Because it's natural language processing, text-to-speech, speech-to-text. You can put some translation in the middle if that's what you want. But also things like forecasting, personalization, fraud detection. Now, Amazon sits literally on billions and billions of orders from the past. We know exactly which ones are fraudulent. Now you can build a model out of that. If a new order comes in, you push it through there, it will give you a score what the likelihood is that this is also a fraudulent order. You don't decline it, you give it to a human to investigate. Remember, all of these systems, current or future AI systems, are not to make decisions for you. They make predictions for you. You still make the decisions. It doesn't evolve, it doesn't mean that you no longer have to use your brain when these things arrive. You still have to investigate.

Yeah, I'm just going to pick a few of my customers that have built amazing stuff with AI that already really, really works. And I've done this TV series called Now Go Build, where I basically go visit startups around the world that actually solve hard human problems in healthcare, in food, in all sorts of areas. And so I'll give you a few of these examples. Now, AI for food is an important problem. If the world is going to, say, grow with another two billion people in the coming, until what is it, 2050, I believe, you know, we need to really think about sort of what are we going to do about our food. You know, what are going to be economic situations, what are we going to do about health, all these things are suddenly being pushed to scale. So there's a number of organizations out there working really hard on that.

And if you look at rice, for example, there, 50% of the world depends on rice as a major food source. So there's an institute just outside Manila called the International Rice Research Institute, and their task is to eradicate hunger and poverty in the world by actually promoting the good use of rice and good farming around it. So they have 200,000 strands of DNA of rice in the freezer, and there's a backup in Norway. However, the problem is that when these seeds come in, humans have to look at them, have to see which of the seeds are good, which are future things like that. It creates a massive backlog. That means many of these seeds get destroyed before they can actually be put in the freezer. So this is a human process. Uses computer vision. Why don't we? We can build, we just put a camera on top of that and make sure that AI starts sort of signaling which of these seeds are not good. Easy system to build, 30% to 40% productivity increase. Not only the productivity increase, there is no backlog anymore, which means I can keep all the seeds. This is a very simple application of AI technology, but it works really, really well already.

They have a massive data set, since the 3,000 Rice Genome Data Set, it's for free available on AWS if you want to do research with that. And it's really important because many researchers out of the world are using this data set to create new insights into rice farming. And also using drones. Now, think about drones. Do you think they fly by themselves? Yes, they do. They're not piloted. If you think about sort of one of these organizations that actually using drones to deliver medicine in remote areas in Africa, not only medicine but, for example, vaccinations, and then bring back tests that need to be tested. These drones, you know, you tell them where to go and then they need to figure it out by themselves. There's about 50 different sensors on that thing, but it has to think for itself. It doesn't think for itself, it goes through the different rules and models that it has to get to the direction. And so in this particular case, it's really important that sort of these massive rice farms that are out there, how is actually fertilizer applied. These farmers have never been educated on it, and they just use way too much fertilizer, get runoff, and they get much of algae sort of growth in rivers and things like that, massive impact. So most of these farmers, they actually can't read and write, so an app on your phone doesn't work. They can call into a system, explain the patch of land where it is, how big it is, and the system will actually tell you what fertilizer to buy and when to apply it. Simple systems that can use human technology like voice to actually get access to great digital systems.

Precision agriculture is a really big thing. AI plays a massive role in that. But let's look at aquaculture. There's another problem around food is that we depend on protein, but the way that we grow protein at the moment is extremely inefficient. Fish farming, however, is, I think the number is that you need 10 kilos of food to create 1 kilo of protein in regular farming. Fish farming, however, 1 kilo of food results in 1 kilo of protein. Extremely efficient. But how do we grow them? I mean, in the wild, it's really hard to do. So if you put many of them together, you know, in a farm, you get all these other diseases. And so what this company, which is called Akva Group, does, have massive pens in the fjords of Norway, 200,000 salmon in them, and actually use computer vision to track the health of each of those salmon. How are they growing? More importantly, how is the lice on their bodies? Because if one of them gets massive lice, all 200,000 need to be destroyed. So they make use of these cameras and a whole bunch of other sensors in and around their system to make sure that these can grow. And they scored well over, I think next to the IoT sensors, they scored well over 1 billion fish by now, have a data set that you can use to think about other ways of growing fish and protein. So in that, computer vision plays an enormous important role. That's sort of the thing that sort of binds all of this together, which are really sort of the examples of what we saw earlier of embodied AI. Basically, our sensors, the ways that we operate, we mimicked these in technology, which we now call AI.

This is a maybe more sobering example. Thorn is an organization that has as a goal to eradicate sexual exploitation of kids. And so as you know, many of these imagery ends up on social networks. Social networks have people looking at these imagery and then sort of deleting them if they're not appropriate. What Thorn has done is build a massive data set of these images and such that gives tools to social networks to sort of automatically process whether images fall in that particular category. Again, there's still humans that have to look at it, but what Thorn does is actually sort of blurring these images so there still a decision can be made, but you know that not everyone looking at these images need to have a whole evening or therapy after that. And so now they have this massive database on that. They also have a text database: what are common patterns in grooming, and be able to alarm on that.

Already, all of this modern AI. Now, important in all of this, of course, is that, you know, we can think about sort of how sort of all these new technologies, AI technology, machine learning, different techniques, things like that, and then the insights that you can get out of that or help you with efficiency and all that kind of things. None of that works if you don't have good data. Yeah, all of this starts with data. You know, and then already the tools help you there, eliminating bias or eliminating hate speech or things like that, all of those which if you train on massive data sets that you actually don't know what's in there will actually arrive. And so AI needs really good data. And to be honest, these massive data sets that we have by now, they need AI to get insights into it.

Now, in the past, and if you think about the '90s, relational databases were capable of storing your data. And that meant, why is that? Because on forehand, you really thought about what you wanted to know. That meant you knew what kind of data you needed to collect, and as such, you know, you already knew what kind of queries, and that was only the data you collected. With cloud storage became so cheap that you basically keep all your data around, whether it's structured or not. Now, if you have a lot of unstructured data, which are compared to sort of a haystack, now if you want to look for a needle in a haystack, how do you do that? You use a magnet. The magnet is machine learning. The tools underneath AI allow you to look for these pots of gold at the end of this rainbow in this haystack. So it's all about data, because without data, we don't have anything like this in artificial intelligence. It is really the final frontier. Can you make sure that your company has data where it needs to be? And most of our companies have gone through mergers and acquisitions, you have multiple silos, all this data sits everywhere. Can you bring the data together in one place? Can you sort of fine-tune the models that you have with your company data such that you get unique answers for your organization, not for someone else's? Can you maintain privacy? Can you maintain safety? One of the reasons why we've invested in Anthropic is because our major concern with newer technologies is can we keep our customers safe, can we make sure they make the decisions on data that is relevant for them. So it's not only sort of these big models that we have, but be able to fine-tune the models with your unique application is something that we can do already really well.

Yeah, so I believe that, you know, as much as you can get excited in these days about sort of the future that generative AI brings, and it will be very exciting, it will have major impact on many different business organizations. But there's a lot of technology that is called AI, that is AI for now, it works really, really well. There is a in Sweden a research study that, you know, they have a public health service that allows women to be scanned for breast cancer every two years. Yeah, so to get a mammogram, radiologist look at this mammogram, raise a flag if they see something suspicious. Now these are literally tens of thousands of mammograms that they have to look at. Your eyes aren't that great anymore at the end of the day. It turns out in this study that AI scores 30% more cancers than one radiologist looking at the images, and actually it's just as good as two independent radiologists looking at an image. That already shows where this is going. It will be your assistant, it will be assisting you in finding things and won't replace the radiologist. Yeah, that's AI that works right now. So don't only stare yourself blind at sort of gen and things like that. Look at what is really right now. This is not a field that's going to go away. Image recognition, natural language processing, all the things that you want to build in your applications work right now. But as soon as it works, you don't call it AI anymore.

Yeah, now with all of this, if you're interested in the TV series, follow the QR code there. And with that, thank you, and go build your AI systems.