Published November 11, 2016
Kristinn R. Thórisson is a professor of computer science at the Reykjavík University. He is also the founding Director of the Icelandic Institute for Intelligent Machines (IIIM), and has been working on the bleeding edge of this industry for three decades. We sat with Kristinn amid a sea of computers to talk about what artificial intelligence is, where it’s heading, and the implications it holds for the rest of us humans.
What were the beginnings of the IIIM?
In 2004 and 2005, we laid the groundwork for the artificial intelligence lab at Reykjavík University. This brought forth the Centre for Analysis and Design of Intelligent Agents (CADIA), which is now in its eleventh year. It’s a university research lab, so it works mostly on basic research questions. With education being the primary objective of an academic research lab, education always takes precedence, which limits severely their ability to commit to dates and deliverables. IIIM is a deliberate attempt at making up for these limitations by building a bridge between academic research—and researchers—and industry. We do a lot of prototyping and feasibility analysis, and our people are very good at delivering products on time.
Shortly after the financial crash, and before tourism took off, everyone was talking about Iceland’s tech industry being the Next Big Thing. Do you feel as though the spotlight has moved away from tech?
Yeah, it seems that somehow our thoughts tend continuously return to the older industries, like farming, fishing, and energy production. We still haven’t adopted the kind of thinking you need to have in order to make the startup mentality a part of your everyday life. So this idea that a startup that fails is a failure has to go away. People forget that IBM used to be a startup. But how many startups were tried and failed for every IBM? With the recent events in startup culture in Iceland, people write negative things about the startups that fail to sustain themselves, e.g. past a second infusion of financing. This is a serious mistake, because by their very nature, startups are an attempt to make something new. And since no one knows perfectly what the future holds, you can’t be right every time. That doesn’t mean you shouldn’t try.
And it doesn’t mean it’s a total failure when it folds. This is a basic conception of the startup process which is missing, I think, in the way people think about the workforce in Iceland. Just because a startup folds doesn’t mean startup culture is a bad idea, or somehow irrelevant to Iceland. Startups are a primary way for how we sustain innovation and our competitiveness internationally.
The people in the companies that fail are still around, and will be so much more experienced the next time they try. We need to take the next step and recognise this fact if we want a vibrant environment where Icelanders have a chance at being at the cutting edge of numerous fields, rather than just a few. The people in those companies that fail will be so much more experienced the next time around.
One of the things that rekindled our interest in this subject was the recent news of the creation of an artificial intelligence transcriber for Parliament. What inspired the need for such a programme?
In a general sense, people’s awareness has been raised significantly over the past five years about the possibilities that various AI technologies harbor. I think people’s intrigued is piqued. Artificial intelligence has stopped being a Hollywood sci-fi concept and has moved closer to reality, and with that, all sorts of ideas spring forth that people would have dismissed or not understood otherwise. And the possibilities have indeed become quite enormous for AI, and it’s foolish not to at least consider them. But it’s also foolish to jump too quickly. I would, for example, not trust my life to an AI right now.
In an Icelandic context, what sectors are people most interested in when it comes to the application of artificial intelligence?
There’s quite a range, considering our manpower. For example, at IIIM we are working with a number of startups and more established companies, including Össur, Mint Solutions, Suitme, Rögg, Svarmi, Costner, and many more are in the pipelines targeting a variety of technologies and business solutions. At CADIA we have Jón Guðnason, who has been working very closely with Google to create speech recognition for Icelandic. We have Hannes Högni Vilhjálmsson, who’s been working with [game company] CCP, and so have I in the past. Kamilla Jónsdóttir has been working on a project in aviation. I have been directing a long-running project on artificial general intelligence (AGI), over the past seven years. This has to do with moving away from the current “black box” design of AI, towards more capable, continuously learning agents.
Can you elaborate on what AGI is?
The mechanisms that we use in AI now, typically referred to as “machine learning,” has very little in common with human learning. The learning is prepared in the lab, and then it’s turned off when the product ships. So what you get is a machine that did learn at some point, but now it’s out of the lab and can’t learn anymore. The reason why you can’t “leave the learning on” is that there are no ways of ensuring that what the AI might learn in the future is going to be useful or sensible. What we’ve been doing in CADIA over the past five years is to come up with machine designs where we have a better understanding of the direction that the learning will go in, so you could leave the learning on. Such a machine will become safer and safer over time.
How does an AI learn “wrong”?
Say you’re in a self-driving car that has learned, for example, what a stop sign looks like. You’re riding along, and encounter a stop sign with bullet holes in it. You don’t know if the machine is going to understand it’s a stop sign, or confuse it with some other sign. What happens then? You can’t predict it. Even if the machine has been taught very thoroughly, you still can’t foresee all the variations that it might encounter, and therefore your certainty is only as good as the researcher’s ability to think up scenarios beforehand where things could go wrong. How good are we at thinking of things that could go wrong? Not very.
So how would AGI approach a situation where the stop sign is rusted, upside-down and has bullet holes in it?
The way that our machine operates, and the ways in which we think it will be much safer, is that this machine has the ability to assess the reliability of its own knowledge. If it encounters such a stop sign, it could say, “What the hell is that? I better do what I consider to be the safest maneuver in this circumstance.” These machines would still go through the training that regular AI goes through, but their predicted behaviour to new things will be much more sensible, predictable and reliable. At present the only place you can find prototypes of such a machine on the planet is at RU. We’ve done some interesting things with this in the lab, but we still need more time, more testing, and more funding.
On that subject, how solid are we when it comes to having people in this country who can work in this field? Are we losing our best and brightest?
Yeah, I think there’s been “brain drain” in this country over the past five or six years. And in the years leading up to the financial crash, we also had another kind of brain drain in that students in computer science were being gobbled up by the banks. That’s not exactly the place you want them to be if you want them to innovate—if you want a vibrant startup community. Closely related to that is that we’ve not seen the necessary increases in funding for the universities.
Looking forward, what are the projects in AI that you are most excited about?
Well, my own, of course! We now have this project on “machines that understand”. I think this is where we have to go to be able to trust machines with more sophisticated tasks, and with our lives. Understanding is the way to go if you want intelligent machines. People might say, “So you want them to take over the world?” I would say: not in any way that software isn’t already taking over the world. If you look at the current application of software, it’s used to reduce cost, to improve efficiency and so on.
In which aspects of our daily lives over the next five years do you think AI is going to come up most?
It’s not like everyone’s going to have a robot in their homes in the next five years. I don’t think it’s going to be that obvious. It’s going to be mostly invisible or partly invisible. A lot of it will be online, and used for things like finding documents, blocking ads, and so forth. These will be increasingly driven by AI. But there’s a lot of untapped opportunities for applying artificial intelligence and we have only begun to scratch the surface.
Why do you think we’re afraid of the idea of intelligent machines that are capable of learning? Where does that come from?
I think it’s built into the human psyche to have this continuous evaluation of “Us and Them”—a self-protection mechanism. If something surprises you, you want to classify it as Friend or Foe as quickly as possible. AI is surprising us now—to have to think of machines in a way that makes it easy to anthropomorphise them. We’re comparing their behaviour, that we’re very familiar with, to something that we’ve never had to compare anything non-human or even non-animal to. And that I think raises this red flag. The self-analysis and introspection that we’re now building into our machines will by all measures so far make them safer and more predictable, and that is an obvious benefit over the present state of the art. Hopefully it becomes sufficiently obvious to most people over time to drive this fear away.
Gallery of Kristinn at his lab, by Art Bicnick: