Mads Rydahl, Unsilo

The torch is yours. Okay, so I guess I was actually told that I would have a big introduction of who I am, but since that is not the case, I'll just let you know that I'm a local entrepreneur. We're building a, we're using artificial intelligence or machine learning, or whatever you want to call it, to solve the problem of discovery and text retrieval. Basically very similar to what you might think Google is doing. Today, a lot of the content that is required for innovation and new discoveries in science resides outside of the public space. There are so many things that you cannot find in Google. And, we are working with scientific publishers and with some of the biggest research institutions in the world to make some of this stuff more discoverable. And we're using AI to do it. Let me see if I can, here. So you might say, why don't we just let Google solve that? And no, Google is not the solution. Here's the search I did a couple of days ago for pizza. And as you can see, out of the seven, seven hundred and sixty-three million possible results, it chose to show me pizzahut.com. And as you probably know, it's been, I guess, almost ten years since there was a Pizza Hut restaurant here in Aarhus. So, it doesn't really help me much. These, the results that Google returned are, to the most part, ranked by some algorithms that can be boiled down to just popularity. And it fits a commercial use case that isn't really specific to people who want to do something that's never been done before. We work with professional information providers that have really complex setups today. Here's a query from one of our clients. It's a query to set up a content package, a set of articles that they can redistribute around the topic of inflammation. And it's easy to see how this is really difficult to maintain. You can see all the stuff that you've got. You can see all the stuff that you can find. That's easy, but it's really difficult to see what you're missing. So, what we're trying to do is to use machine intelligence to improve this process. So, we're about 30 people here on the harbor in Aarhus, and we've built really amazing semantic technology, basically computer software that can extract knowledge from text. And we're working with some of the largest publishing houses, Springer Nature of the Nature Journal and several of the other larger companies, to solve this challenge of making it easy to find really relevant information. So, by way of example, here's an abstract from a biomedical research article. And if you are familiar with this topic, subject matter, you can probably tell at a glance what it's about. Most people cannot. And with full text, we can search for, like, some occurrence of words that matches your interest, if we know what to search for. And if we want to find similar content, and if you're looking at this article, what else might interest you? The traditional approach has been to use dictionaries of all the words that we know the meaning of, and then we count the occurrences of these words, and then we use the enema that everyone knows what it means, and that becomes a fingerprint of the article, and then we can use that fingerprint to find other articles. The problem is that many of these words have kind of ambiguous meanings. Serum can be many different things. Water can be part in many different contexts. And so what we do at On Silo is that we look at many, many millions of articles, and then we compare them to figure out which phrases are the common occurrence in this large corpus, and which phrases do we think mean approximately the same thing. That way we can find you articles that are similar to what your research interests are, and we can show them to you, and we can even actually show you exactly what they have in common with the article that you read yesterday, even if the author used different words than the words that you were using. So, that's the problem. And today we're actually moving a little bit further. We're moving into what we call language understanding, which is figuring out how these things are interconnected, what are the causal effects that are being described, and we're working with the foremost researchers at Nature to figure out exactly how to extract the best representation and the key findings in articles. We're actually pretty good at this. A recent study found us to outperform Google, Microsoft, and IBM, so we're very, very proud of what we've done. As Matt said earlier, the potential of AI has been portrayed in popular culture to be very much similar to like this human companion that thinks like us and can perform actions that are very similar to the actions to the actions that we can perform. And as he said, that's probably what our clients expect when we walk into meetings. Miraculous is the new normal. People expect us to be able to deliver on these promises that are completely unrealistic. Some of you may have heard of this moment in time. In 2011, IBM Watson beat the grand champions of Jeopardy. But it was not the end of history, at least not yet. That was 2011 and we still have not solved this open knowledge understanding. We still cannot take any text and tell you exactly what it means. That's a super, super hard problem. I know, of course, our friends at Google are working on that, IBM, Microsoft. Everyone is throwing a ton of resources into this and so are we. We're a small team. We have some really cool people. And we've worked on this for the last six years. The thing is, though, that things haven't changed all that much. We think of machine intelligence as something entirely new, something that's been invented just a couple of years ago. But the basic algorithms that we use haven't changed much in the last 25 years. Matt showed this animation of neural layers. Well, neural networks have been around. They've been around for 25 years. We're stacking them a little bit differently. But we're still using the same type of algorithms to do feedback loops and back propagation and reinforcement. And so the bricks are the same. The tools are the same. But what really has changed significantly in the last 25 years is the data sets that we work with. They're so much larger. They're 10,000 times larger than the stuff I used to work with in the 90s. The computational power that we're working with at Amazon, we can. Actually, throw up 10,000 machines and work on a problem overnight that would have taken months to complete before, even with less data. And we can see many examples of this. Like the most common, I guess, for us is OCR. We used to be able to take typewritten text and import it into the machine. Then we typed numbers in boxes whenever we filled out forms for the government. And now today, computers can easily understand handwriting. There's so many companies working really intently on gathering training data for these algorithms. Tesla cars, I just saw the other week, have now driven 4 billion miles. That's about 6 billion kilometers, which is compared, I guess you can say, to a lifetime of driving for around 10,000 people. A lifetime of driving for 10,000 people. And that data is, of course, training data to algorithms that can drive cars in the future. The same with Google. I guess they made image uploading completely free. No upper limit on the amount of images you can upload to Google Images. And of course, that's to get training data for their algorithms. And all of this, of course, is exponential. But still slightly underwhelming. These are not exactly the things that you see in movies, but these are the real things that we've seen so far as the result of machine learning. Rejections. For your next house loan. Siri, which I was part of making. The Tesla autopilot. And ways for soldiers to kill people without actually being present. But the thing is, we shouldn't be so worried about computers taking over our role in society. I guess these are the classic examples that Jeff's grandmasters, actually, the first grandmaster to be beaten by a computer was Ben Glasson in 98. In Copenhagen. And it wasn't until 97 that Kasparov was then beaten by. I don't know what that says about Danes. But as Matt also said, we don't really, when we enslave or domesticize animals, we don't necessarily compete with them. We don't try to run faster than a horse. We actually put a saddle on it and sit on it. So it's really about extending our understanding. of what it means to be human. And I think that definition is going to slide. And it's going to continue to slide in the future. Whenever chess, in the 70s, chess were on TV. I don't know if you remember. I'm old enough to remember chess was on TV. It was like the battle between Russia and the US. And it was the brightest minds in the world. And they were competing for control of the universe. And then computers got better than us. And that's not the pinnacle of human intelligence. anymore. And so we're constantly redefining this and we actually, we don't know what intelligence is. So calling things machine intelligence or artificial intelligence is kind of ridiculous. And I really think it also kind of ignores the fact that these products that we're building using these tools are designed by people with intent for users with intent. And I think actually one way to think about it is to replace that word intelligent with smart. Smart sort of acknowledges the fact that people thought about this problem and then created a solution. And intent also defines the meaning of all the products that we have around us. Any product design involves the understanding of intent and removal of barriers so that the user can get stuff done with minimum effort. And AI always has intent. There's intent in our biology. It's been billions of years of training to build that intent through evolution. And we can probably also build intent into our AI algorithms, but it's going to require intent on the designers as well. We're a long ways away from actual intelligence in machines. So at UnSilo, we're also user-centered, of course. We're working on discovering what the real problems are that we can help solve. And we're trying to map knowledge and build navigational tools for knowledge that will help researchers of the future. Our responsibility in this case, I think right now we're seeing a lot of bots and conversational UIs, but it's going to change very soon. Right now, you have to sort of know the magic words to make it any fun to talk with a conversational agent, or you have these bots in Slack when you're at work which can respond to sort of somewhat flexible input. But in the near future, we're going to have very robust language understanding. You can say things like you would say them to another person, and the machine will understand it. At least in a way that will allow it to respond sensibly. And that's going to lead to a democratization of access to these digital services that we haven't seen before. And it's going to mean that we can actually ignore this computer literacy problem and get a lot of people on board on these new services. But the scary perspective, of course, is that as we're digitizing all these things... We're also allowing our democratic functions, journalism, news feeds, and critical infrastructure to be open for hackers. So we need to be really careful here. You remember maybe all those tinfoil hats in the 90s that talked about Echelon and how we're being surveilled by this unnamed U.S. agency. Well, it turned out that they were right. They're actually not getting the complete picture. It was much worse than that. It's been like that for a long time now. But the world hasn't ended. I mean, we have to consider this. We have to consider how we build our services in the future to avoid some pitfalls. But we're still here and actually kind of look forward to the time when it's not going to be humans in the top of that pyramid, but maybe some more benevolent machines, if you will. So we need two basic things going forward. We need the right to challenge these algorithms. And I know a lot of people, we're actually working on this right now. This is policy. We need transparency, radical transparency around how we're building these systems. And we need to be able to ask, how did you reach this conclusion? Okay, so my application for a loan was rejected. You need to tell me why. And right now, there is actually no legal right there. And then we need to think about the way our system is set up. We have a lot of concentration right now. With software, you can design a solution that works all the way around the globe. And 11 people at Instagram can out-compete. I wouldn't say they directly out-competed Kodak, but you have the possibility of concentration of capital that's never been seen before. And that's kind of dangerous. So we need to think about how our system responds to these threats and these changes. So stay vigilant and don't take anything for granted.