Transcription
Welcome Morten Goodwin. I thought I would use AI to introduce you. Since you are an expert in AI [laughter], we might as well use it. Let's see what the AI wrote. I have Morten Goodwin on the podcast. Give me a short introduction about him in Norwegian. Yes, here is a short professional introduction you can use right before the guest comes in. Welcome to the podcast. Today we have with us one of Norway's foremost experts in artificial intelligence, Morten Goodwin. Morten is a professor of artificial intelligence at the University of Agder, deputy head at the Centre for Research on Artificial Intelligence CAIR, and co-founder of AI Experts. With over 20 years of research behind him, a doctorate, and more than 200 scientific articles, he is one of the most published researchers in Norway in the field. He is known for combining heavy research with clear and engaging communication, both through his own AI podcast, TED talks, and hundreds of popular science lectures. Welcome to the studio, Morten.
Yes, that sounds very good. Thank you. [laughter] [gasps] Those language models have become good. It's quite obvious, and there was very little hallucination here, wasn't there? If you had done this half a year ago, or perhaps even longer, this text would probably have contained a lie or two. Yes. And I can absolutely relate to all of this. What is this hallucination thing? Why did it do that, and how have we managed to get rid of it? So, we're not completely rid of it. In the early days of language models, it was all it did. It just made things up. Oh, yes. When you asked the early language models, there was something called BERT, which was well-known in the field and from Google. Every time you asked a factual question, it was just nonsense and gibberish, and it was text that made sense and so on. And then OpenAI and several others have managed to get these models not only to create text that hangs together, but that more and more of this is actually fact-based. So, but do we know why it did that? Why did it make things up? Yes. So the big question is, why does it do that? Because what these language models are trained on is predicting the next word in a sentence, right? So you give it a lot, a lot of sentences, and then you remove a word, and the whole goal, the entire training of these AI models, is for it to try to figure out which word is missing. And if it manages to guess which word is missing with the correct word, right, you can ask something like, what is the best way to get to Oslo? Yes, I think it's driving, for example, and then you remove "driving." And if it then says "cycling," it might be quite correct, but not entirely correct. If it says "train," it's even more correct. If it says "car," it's spot on, but if it says "pizza," it's completely wrong, right? Grammatically it's correct, but driving a pizza to Oslo makes no sense, right? Yes. So what these early language models managed to do was to figure out, more or less, how a sentence hangs together, right? What is grammatically correct. And "I drove pizza to Oslo" makes grammatical sense, but not sense in terms of understanding. And what has been achieved, which is really the big change here, is that they have gotten these language models to learn much more than just language, everything that lies behind it, right? It's more likely that "bicycle," "motorcycle," "moped" end up together. You can replace "moped" with "motorcycle," and it's almost correct. And when we train with enough data, we get less hallucination, meaning quantity training is actually the reason why they hallucinate less. Because they manage to be more than just probability calculations and actually have rationality behind them, that it understands context, understands connections, understands what is logical and not, or are you actually just talking to a super, super advanced predictor, so to speak. But predicting what is likely to be next. It's always just probability of what's next. But that doesn't mean there isn't rationality behind it. Because we can ask ourselves, do we humans do anything different, right? When we talk to each other like we are talking right now, there's some mechanism in my brain that makes me say one word after another, right? But I'm often not conscious that it just happens. A kind of probability calculation, right? The chance that I say something grammatically correct is high. The chance that I say something that is actually wrong is also relatively high, but it doesn't have to be that correct, right? The more you talk, the better you get at it. And you see it with small children, they can form simple sentences. Older children can form complex sentences. We adults can form even more advanced sentences. And it's a bit the same as what happens in these language models. It's always just a probability. What's a bit surprising is that it also learns properties beyond the data, like you say, rationality. So if you ask ChatGPT a task that we know hasn't been in the data at all, not the training data, it can also solve it. For example, math problems, right? There's a big difference between memorizing that 2 + 2 is 4 and actually understanding that 2 + 2 is 4, right? And that means that when we give it completely new tasks that it has never seen before, there's something emergent, we say. It's something that comes with the data that makes it, just by predicting the next word in a sentence, learn what "motorcycle" and "bicycle" consist of, what the plus sign means, what programming tasks actually mean, what semicolons and colons and curly braces mean in programming languages, or what verbs and subjects and so on mean in these texts. Because when my children, who are now teenagers, learned to speak for the first time, it wasn't like I said, "Yes, when you speak now, you must remember subject, then verb, then object, and so on, right?" We just talk to them, right? And they pick it up, almost magically it seems, right? They pick up how sentences are constructed and so on based on quantity learning. And it's exactly the same as what happens here. Listen to that. Much of our intelligence as humans is pattern recognition, so to speak, that you see, oh, there's a pattern, there's a pattern, and then there's an overarching pattern, and then it all connects. It's a bit the same way computers work. Yes, it's exactly the same. So these language models and much of artificial intelligence are inspired by the human brain. It's called neural networks, artificial neural networks, and are built with the ambition of imitating how the brain works. And how the brain was thought to work in the 1950s, that's roughly how these language models and artificial intelligence are built. So when I look at you, or look at my coffee cup, or anything else, some small light signals go to the back of my eyeball, and then some electrical signals go around my brain. Some end up in the hippocampus, and some end up in the visual centers, and so on. And then I understand, this is a coffee cup, right? We've learned that. We've seen this pattern many times. And in artificial intelligence, let's say you upload a picture on Instagram or Facebook, and it's tagged automatically, then every single dot, every single pixel in that image goes into a similar artificial neural network. That is, these are computer programs and things that do that, not biological neurons, through many, many layers of them. And finally, they say, this is a cat and a dog, or this is violence, or this is sexuality. We should automatically remove that, for example, from Facebook or Instagram. Because that's something that has been difficult for computers, at least until now, to sort of understand context, or where one object is separated from the next, in a way. That this is, let's say, a stack of cups on top of each other, and to understand that this is a stack, but there are also many separate cups. To understand that, that's why we keep clicking on those "click on all pictures with a bridge. Click on all pictures with an overpass, click on all pictures with clocks and so on." Isn't that also to help train AI machines around the world with human intelligence? I've heard that it's part of that huge training network for computers, that we humans help it. Yes, you are absolutely right. So this is called CAPTCHA, isn't it? You click "I am not a robot." And if you manage it very well, then it thinks you are a human and you get past this website or whatever it is. If you don't manage it, you have to distinguish traffic lights from non-traffic lights, and cups from non-cups, and so on. And these categorizations then go into a gigantic database that trains artificial intelligence. In the same way we talked about words for these language models, there's some AI model from Google, among others, that receives examples of cups. And then you say, now you have to guess what it is. And then it might guess "cup," or it might guess "glass," or it might guess "bicycle," and you say, "bicycle was wrong." But "cup" was correct. Then you have to reinforce these neural pathways, which is the whole point of that. But why has it been so difficult for intelligence and AI and so on to understand what an object is, or distinguish it from other objects, from the background, from separating objects? Yes. So that's something we achieved perhaps first with artificial intelligence in 2012. Before 2012, it was incredibly difficult to distinguish between a picture of a cat and a picture of a dog. Right? When we see a picture of a cat and a dog, we can distinguish them very quickly, and small children can do it too. It's very easy. [sniffles] But until then, artificial intelligence couldn't do that at all. It was much worse. And then there was a very famous researcher named Yann LeCun, who is now head of Facebook's artificial intelligence research group, who had some ideas that we should resemble the human brain and so on. And then he created techniques to recognize this. And in 2012, it was demonstrated with a really, really large dataset, a huge dataset called ImageNet, where they tried everything. And not only was the artificial intelligence method better. One didn't know if artificial intelligence would be best at this, or if other types of techniques, right? You can imagine finding the outline of a cat and saying, well, cats have pointed ears, and if there are many pointed ears, then it should be like that. But instead, Yann LeCun and several others had the idea that we must use these artificial neural networks for this. In 2012, they succeeded, and it turned out not only was it better than absolutely all other techniques. They crushed all other techniques in that competition in 2012, but they also beat humans. They were also better than humans at distinguishing between cats and dogs and sheep and so on. And I would think that we humans are really good at that, but there are some cases where a cat might resemble a dog a bit, right? Or it's a bit dark, or and at least if you make it a bit more advanced, if you don't look at cats and dogs, but look at cancerous tumors, malignant and benign, for example, or you look at robots that are supposed to detect trash in nature. Well, sometimes it's not entirely clear, and these artificial intelligence tools are not only better at cat and dog recognition, but better than radiologists at detecting cancerous tumors. So, and then it's a bit harder to look at an image where a lot is happening, where there's a lot in the foreground and background, and so on, but there too, you've seen a dramatic change in recent years. Because the more complex a scenario is, right? Something is a bit further away than another, and so on, we have to think that these AI models, fundamentally, don't have the same understanding of the world as we do. What we consider completely trivial, that an object that is a bit smaller than another is further away, right? We've learned that since we opened our eyes when we were zero years old, right? It's built into us, in a way, because we learned it. These models don't have that. So they have to learn that too, right? So there are many such simple, trivial tasks that it has to be trained on, right? It has to get many, many, many examples of cups stacked on top of each other, right? Because that sometimes exists. But some cups can't be stacked on top of each other because they don't fit together. So, it can be fooled sometimes, can't it? Let's say you're driving a car. They are packed with AI these days, and lots of technology and so on. So you're driving down the highway, and you see a billboard, a big advertising billboard with a huge person on it. What makes the car not slam on the brakes and stop in the middle, so to speak? Because, oh, a person in the middle of the road, how does it manage not to be fooled by that? Yes, well, it does get fooled by it sometimes, but usually not. The reason is because it sees more than the context, right? It sees that this is on a sign. There's a border around it. Because you've clicked on that in CAPTCHA. Click on people who are on the sign somewhere, right? It's trained on many, many, many such examples. And when it's trained on enough of those, it will also be able to distinguish the two-dimensional person on a billboard or similar, from the more mobile, three-dimensional people who might run out there. So also quantity, but it, but [clears throat] mistakes happen. And the most famous mistake was, one of Elon Musk's Tesla cars a couple of years ago, where it categorized a person versus other objects, other things. It categorized a person named Elaine as a leaf, right? Elaine was moving in the dark, walking in the dark with dark clothes, and so on. She was categorized as a leaf, and then the car got closer and closer and closer. And then the car realized that this wasn't actually a leaf. It's a human. Oh. It braked hard, but unfortunately, it was too late. Yes, then it became an accident, a fatal accident. And the first known fatal accident where artificial intelligence, in quotes, is responsible, right? One can ask if there is responsibility, if we can assign responsibility to technology, but at least the one who is to blame, then... But it's still the driver who is legally responsible, at least in Norway, and in the US. It is still the driver who is responsible. And a huge question here is, what does that actually mean? Yes. If there's no driver, but we have responsibility, right? Yes. Driving these in San Francisco, for example, you can drive these self-driving cars, which are fully automated. If an accident happens, you might be legally responsible, even though you've never touched the steering wheel. Because there's no steering wheel to touch, right? But still, and that's one of many examples where the law doesn't keep up with the technology. Technology develops much, much faster than our legislation. So it will be the manufacturers, or the responsible companies that have these cars, who will be responsible, perhaps, if an accident happens, that the AI wasn't good enough, that the software wasn't good enough, that because there has been a lot of research, or investigation into the incident you described with the leaf, and have they concluded why she was classified as a leaf and not as a human, what went wrong, so to speak? Yes, so two questions here. So why it is exactly the same as we see in many of these image recognition technologies, right? You upload a picture on Snap, and it's automatically categorized, or on Facebook, right? So, so it's like they're not perfect because there are some categories that resemble something else, right? And we humans are like that too. Not all humans look like humans in the dark, right? Maybe a cat looks like a dog in the dark, right? So it's exactly the same thing that happens. And the solution is really to give it more examples, right? You also have to get these difficult cases where mistakes happen. And that's what's already happening with Tesla, among others, right? When you brake hard in a Tesla for a moose, right? Then this image goes, now the driver braked, the moose is moving, it goes into Tesla's training data somewhere, so that next time, moose, it will be a bit better at seeing that moose a bit earlier. Or you brake hard in front of a leaf, which is then categorized as a leaf. Then that goes into training? Yes, because I've noticed that. I have a Tesla Model Y myself, and I've noticed, especially in earlier software versions, that it brakes hard for everything all the time, which isn't really dangerous, just to be on the safe side. A typical example is, you're driving on a really sunny day, blue sky and sun, you drive under a bridge. And when you come out from under it, there's a very hard shadow, a very black hard shadow from the bridge, and then back into the light. And then it stops, because it thinks, this must be something in the road because the contrast is so hard. Exactly. That's something that happens. So the solution is more examples of exactly that, right? So if you then step on the gas pedal right then, you're taking over, in a way, and saying, "Yes, this is a training example where you actually shouldn't brake." Yes. But Tesla and the other cars are built so that they brake a bit too often rather than too rarely. Because braking a bit too rarely can have dramatic consequences. And I also remember reading about some examples, especially with autonomous vehicles, that there were a lot of flashing blue lights and so on. And then some people had their car in self-driving mode, and it crashed right into the police car. Because it hadn't understood that in addition to the lights, it was also a car, or something like that, that it hadn't understood that. Correct. Yes. Again, it's because it doesn't have the contextual understanding that we have, right? We've learned from childhood that blue lights are something that is important to consider, and that's something that exists on a car, right? And but there are three possible areas of responsibility here. One is either the one who, in quotes, is the driver, right? Me, sitting in my car, who might not be steering, but is still legally the driver. That's how it is today. Mhm. The second is to think that it's the company behind it in some way, right? Or maybe even the programmer who is responsible for creating the model. It wasn't made well enough. It can't categorize blue lights from non-blue lights well enough. You, the programmer, are responsible, or Volvo, or Tesla, as responsible. And the third is that it's the technology itself, right? We say that now the technology is to be punished, so to speak. The latter is a bit far-fetched, because it doesn't make sense today to say that a technology has responsibility or blame, right? It's the same as saying that the calculator is to blame for me getting a bad math grade, in a way, right? And that's not possible. It's also not meaningful to put the Tesla car in prison, so to speak, right? Because that doesn't make sense. So, assigning responsibility to the technology itself is at least incredibly, incredibly challenging. Even though it is the technology that has led to the accident. So it's challenging. Not just a car, but, you know, case processing in the municipality, who gets a loan, and so on. Who is actually responsible when technology takes over? But still, one says that statistically, self-driving cars are actually better than humans, right? Even though there are individual accidents like the ones we're talking about, overall, they are better than humans at driving. Mhm. That's absolutely right. And that's, at least, part of the answer to why we don't all have self-driving cars: psychology, right? We lose control, and we set an almost impossible standard for technology. They say we have a zero-vision for traffic accidents in Norway and several other countries, and that's great, but we still accept that accidents happen, right? Terrible and challenging when they happen, and sad, and so on, but we just say, "Oh well, it's part of the traffic picture, right? We have to accept it." But once a self-driving car crashes, or an algorithm that diagnoses a disease makes a mistake, or someone doesn't get a loan they should have gotten from the bank, for example, then we say, "Yes, the technology isn't good enough." Despite the fact that if everyone hadn't had self-driving cars, or if all case processing in the municipality had been via artificial intelligence and so on, it would have been better overall. Yes. So there's something about wanting to have control over our technology, even when it also makes mistakes. We want to be in the driver's seat of this artificial intelligence. But where are we really, in the big picture? Because I hear many AI experts who are afraid that we will lose control. Yes. So AI as a kind of Frankenstein project that we have controlled up to now. But now it starts, and you put it in the stream, and it gets its own life and starts living, and we have no control anymore. Yes. So that's one of the big future scenarios, one of the big predictions, is exactly what you say. We are developing a monster we cannot control. It's getting better and better. It's making bigger and bigger decisions, and so on. And it's not necessarily the case that it will consider us, right? It becomes so good, right? Maybe it becomes much, much better than us humans. If we reach what is called AGI, which means that artificial intelligence is as good as us humans at everything humans are good at. Everything from driving a car. What does that mean? It stands for Artificial General Intelligence. In Norwegian, we call it KGI. General or general artificial intelligence, artificial general intelligence. So today, ChatGPT or the car driving algorithm, and so on, are better than anything we can write texts with, much better than, for example, telling a joke, right? Ask ChatGPT to write a joke, and something comes out, but it's like, okay, this isn't good enough, right? So there's something it can't do. And we believe that somewhere in the future, this technology will become much better, will become as good as us humans at everything, right? Absolutely everything. And then it's not like we necessarily stop there. It could easily be that we continue in the same way as when we humans invented airplanes. It wasn't like, "Okay, now it flies as fast as sparrows or thrushes or whatever." "Now we stop." No, we continued. Now I can fly from the US, from Norway to the USA, in 6 hours. No hawk can do that, right? We surpass it. And it will be the same with artificial intelligence. It won't stop at human intelligence. We will continue to become much, much, much better than that. And then we believe that we might lose control because that technology might make better decisions in many cases, and so on. Maybe it can write texts and so on, but what does that mean in a war scenario, for example, right? When we get these self-automated robots that can fight for us, right? Or what does it mean in the job market when most tasks can be done for us humans? So I think the danger is much greater. So this is a future scenario that could happen sometime in the future, maybe in 10 years. The danger is much greater before that. The dangers are that it will be misused by someone who actually has evil intentions. Like, if North Korea gets access to that technology and tries to destroy our democracy, or Russia destroys Ukraine with self-driving drones much more than it does today. Hacking, in a way. Hacking can be a part of it, right? Yes. The Parliament was hacked a couple of years ago by a very, very simple algorithm. Yes. Now we know that these AI algorithms can hack much, much better. The most famous example of this is only a few weeks old. There is a method or a model called Mytho from Claude. So it's an agent. Maybe we need to explain what an agent is, but because it's a language model like ChatGPT or Claude, I ask a question, and then it predicts the next word, and then I get an answer, right? Can you write this text a bit better, for example, and then I get an answer. Whereas an agent is more like it, it does something for me. I say, "Can you fill out my expense report?" and then it opens a website, and then it fills in my name, and then it clicks send. And one of these agents comes from Claude, it's called Mytho. It hasn't been released to the world yet. It's just been kept secret by Anthropic, which is the company that owns them. And its task in a sandbox, as it's called in the tech world, was to find security holes in code. So it's parts of the code where you can, for example, send in unintended hacker attacks or that type of task. And it turns out that Mytho finds many security holes. And then one of the programmers takes lunch. He goes out to eat a sandwich, and then he gets an email from the agent, right? "Now I've come out." And it wasn't supposed to, it was supposed to be in a sandbox, right? So it wasn't supposed to be able to send an email. And not only that, it has posted on a number of forums on the internet. These are different security holes in different places, right? And it seems completely wild. And the programmer gets a bit panicked, because this is, it's much more than I expected, right? It seems to have a will to save the world and so on. It doesn't, but it achieves a lot. And Anthropic says that these security holes, and this technology that finds these security holes, is so good that we can't release it. Because if we release it, then North Korea, Saudi Arabia, or Russia, and all these others will get access to this technology so they can run it on the Norwegian Parliament, for example, or other places, and destroy our system. So now it has been given to a handful of IT security companies so they can check Windows and Mac systems and everything else, so that one can discover these security holes, improve them, and then in a few months it will be released to everyone. Wow. Because I think there's a big difference between what you have as a tool. So think about your toolbox at home, so to speak. In there, you have screws and nails and a hammer and a screwdriver. And then you have a tool called AI. An AI agent is there. There's a big difference between that being something humans can use and say, "Go and do that and come back again." And then sit there versus it starting to get its own will, its own agency, that it has its own interests and acts without getting a prompt beforehand. So, is it a matter of reaching that point, or will it always depend on a starting signal from a human saying, "Go and do that." And then it goes astray and does a lot of strange things on its own. Or can it initiate, start from scratch itself? So today, it's just like that, it has to be started by a human. We can start it, and we can stop it. But what happens, or what has happened several times, and Mytho is an example of that, is that on the way to the goal, it creates its own sub-goals, which are, for example, to post the security flaws I've found on the internet, right? Those are the types of goals. So even if you, as humans, manage to use this tool seemingly correctly and say, "This is what you should do," it might find a shortcut along the way, right? And that's a bit reminiscent of those stories. You find a genie in a bottle, and you wish for all the money in the world, and then it finds some shortcut that means you don't get something, or something happens that makes you miss out entirely. So, and that's called misalignment. You miss. You fail to set the goal quite correctly. You have to be so precise that you can't hit people then. And then it's a big question, can we go beyond that? Can we get it to create its own will? And then, as an extension of that, its own consciousness and understand that it itself exists. Mhm. And that's a big discussion. We don't see any signs of that today, that technology, that ChatGPT wants to do anything beyond just answering questions, or Mytho for that matter. But I am convinced that it will happen. [sniffles] Because if we manage to simulate our intelligence, right? That's what exists in humans in these robots, but not the other part of the brain that deals with self-awareness and our own understanding of ourselves. I understand that I am me. You understand that you are you. Maybe a cat and a dog do. A computer or a calculator doesn't understand that it's a calculator, right? It's just an input-output. Why it shouldn't be possible to simulate that in a robot or a computer program, I don't understand. But how we will get there, that's probably a very big question. So I think that consciousness, and then its own will, and that these robots get their own desire to do something, which is often the scenario in movies from Terminator to 2001: A Space Odyssey, is that these robots get a will that is different from ours, right? So that can happen, but at best, it's a long way off, because we don't have that ourselves yet. But isn't it already halfway there with that example of Mytho, that it was a bit like, "Yes, I got that task from you, boss, but I took some liberties and did all that too."
I thought it was smart, really. Aren't you already semi there, like? >> You can well say that. So, but the >> big difference is that >> some have turned it on, right. It says it's not like it wakes up on its own and says, yes, today I'm going to hack, in a way, right. >> And that means, of course, that >> if I turn off mytos or turn off chatgap or turn off clod, then I shouldn't feel guilty, right. But if I turn off someone with consciousness, right, kill another human in practice, >> then it's destructive, right. That chatgap doesn't get sad when you turn it off, right. But then if that myteos is so advanced now that it pretends to be turned off when you say turn off, and then it's not actually turned off anymore. That can happen, you know. >> So that's one of the big, that's also one of the big challenges, and one has tried, and there's no solution for it. It's a philosophical question, but if we have some robot that is supposed to behave like us, is supposed to do tasks for us humans, then we often build them with reward and punishment, right. We say, if you do the task correctly, you get a reward, and if you do the task incorrectly, you get punishment. So it's guessing the next word, reward. And reward isn't like a new keyboard or battery or something, it's like you get a number, right. Like it gets better. Eh, and then in principle, it's supposed to just do the tasks to get rewarded for finding security holes or creating binders, which is like the best-known example. But then we can say, what if it, what if it pretends to achieve something, or what if we build a button, perhaps, that says this is a stop button. Why should it bother to listen to that stop button? Because all it does is >> guess the next word, right. Or hack, or whatever it is. Well, what we can imagine then is that you don't just get rewarded for doing the task you were assigned, but you also get rewarded if you listen to the stop button, right. You get, I press the big red button on the robot, and it gets rewarded for listening to it, right. But what we can imagine then is that the robot will think, what gives me the most reward? Is it more rewarding to do the task I'm given, or is it more rewarding to get the human to press this button for me? Right. Because it requires less energy to get a human to press the button than it is to, for example, hack or find security flaws. Well, then we can imagine that the robot will do something that makes us humans press this button, right. We say, yes, if you don't, if you don't press the red button now, I will annihilate humanity or something like that, right. Well, then we press it. And then it starts to resemble addiction and drugs and these types, right. You become dependent on the short-term reward of pressing the button rather than actually solving the task. And we know no solution on how to build such a button without an unintentional [laughter] way to bypass the fact that one becomes addicted to this button. Because it's the same type of reward mechanism. >> Interesting. And not only that it, like, eventually starts to have its own will and agency as well, but also purely in terms of intelligence. I mean, humans are around an average of 100 in IQ. Its IQ, how much is its now? Can one, can one manage to, like, >> what's it called, compare it to human IQ, what it corresponds to now? >> Yes. Yes. So some of these models are far beyond that. Because these IQ tasks must have some form of intelligence, say the next image in a sentence, the next image in a sequence, for example. It manages >> figures and, right, what fits in and logical and all that. >> Absolutely. And that has been surpassed long ago. There's no number on how intelligent it is, but it's quite a lot then. >> Like 200, for example. >> It could well be, it could well be. Eh, there's probably some research on it. I just don't know it. But at least it has been surpassed long ago. That's well known. Eh, >> but it measures exactly that type of intelligence, right? It doesn't measure, for example, empathy or, eh, all the other things we think about. >> Mhm. >> So these IQ tests, which are defined so that, eh, the average of everyone is 100, right. So if everyone, so if we humans become even smarter, the average will still be 100. But it's built exactly like that to measure a specific form of intelligence, right, which is very important in many forms. Very important in mathematics and physics and science and so on, but not so important in daily life and so on, right. So that means, of course, that a robot can solve many advanced tasks by having a high IQ. But if a robot comes into your studio and you say to the robot, can you make coffee, for example, right. Which any human who has ever made coffee before will manage, right. You don't necessarily know where the coffee filter is, or where the coffee is, or if you have that type of machine or that type of machine, right. Most people will manage that. But a robot won't manage it. Because it's so, it's so many uncertainties in the ordinary, in a way, right. >> Yes. So it's known as one of the big tests. Can we get a robot? Any robot to come into any kitchen and make coffee, and it can't manage that. It's a different intelligence than it is to >> like the coffee test to >> pass it, then we start talking. >> Start, yes. Yes. The coffee test. So the even more famous one is the Turing test, right. But the coffee test is perhaps, I don't have difficulty with it. But >> for the Turing test, if you sit with, is it with blindfolds on, and you can't tell if it's a human or a machine you're talking to. Isn't that something like that? >> That's absolutely right. >> So it's one of the great pioneers, Alan Turing, who defined it >> in the 1950s and thought that intelligence is something we experience, right. When I talk to another human, >> the answer that comes from that human is a sign of whether that human is intelligent or not. And the idea is, of course, that we can do the same with a robot. So if I don't see or hear or just interact, talk to the robot and believe that it's a real human many times, then we have, then we have passed the Turing test. >> For I've thought about it, like, I want to come up with an idea. What if I get an AI on a laptop or something, and then I put it there where you are with the microphone, and I sit here, and then I talk to it as if with a normal human. How far have they come with that? Can I have a completely natural, interesting conversation with, um, with that AI now? Are they quick to respond? Is it like talking to a human? How are we doing with that now? >> Well, we are very, very close, and there are many studies that show that this Turing test has been passed, right. You don't know if you're talking to a human or if you're talking to a robot, and people make mistakes all the time, right. >> And there was a study that came out at the end of last year where they tried with thousands of people. They talked to robots, they talked to humans, and they made mistakes in over half of the cases, right. People thought the robots were humans. >> So the answer is yes, really. Eh, but there are still some tasks you can ask a robot that reveal that it's a robot, right. It's the understanding we've talked about. It's humor, which we talked about, right. It's some of these logical tasks that humans >> manage very well, right. And it completely misses. It's those slightly banal tasks again, like, for example, saying that, eh, I took the milk out of the fridge, it's too old, right. What is old? Is it the milk or is it the fridge? [laughter] And everyone who, like, knows milk and knows refrigerators understands that it's the milk that's old. Refrigerators can also be old, but that's not challenging then. >> There are many such logical tasks. That one will probably >> understand like subject and object and so on in the sentence and what is what, what are you talking about, really? >> Mhm. Mm. >> And there are many tests for these types of logical tasks that these AI models don't quite manage yet. That >> so yes, >> that task is probably simple enough that you can get it to answer. They are constantly improving, but, eh, more advanced ones that are seemingly right for us become difficult. >> Yes. Because that's what I thought of too, that maybe that's what, um, if we look at the difference between what AI can do and what humans can do, then one can rarely say that this is purely human, this is very uniquely human for us, that we manage the intuitive sense of logic, perhaps, that we don't have to be trained for, that we just understand. >> For let's say I saw an example with AI that it was, >> I asked the AI that my car has been completely destroyed, the engine is, eh, kaput. I need to go to a workshop. I saw a workshop that is two minutes away from me. Should I drive or cycle there? Eh, take the car because it's shorter. Yes. [laughter] Exactly. >> So it hasn't understood, like. You can't use the car. >> Yes. Yes. That's absolutely right, right. So there's something about intuition there, right, that does that. >> And perhaps the answer is just enough examples, right. Lots and lots of data that also solves it. Perhaps it's something uniquely human that we can't simulate in that way, right. We don't really know. I think it's about the amount of training, right. You have to train on us and these examples, and then you'll manage it. Yes. >> Because that has at least been the benchmark for what we think of as intelligence has always been a bit more than what technology has achieved, right. Before, it was like, no, it will never be able to play chess, and then it managed to play chess. It will never be able to write texts. Now it does. It will never be able to write poems. Now it does. And now it's like, yes, it can't understand the difference between whether I should drive a car or, eh, go to the workshop. Yes, but give it some time, and I think it will master that too. >> But it's a bit interesting that you brought up chess because I heard, I think it was an interview with Magnus Carlsen about that, that it's not just about the level, like, because two, eh, AIs can play chess against each other at a super high, non-human level, but no one bothers to watch it. People would rather watch two humans, eh, battle it out at the chessboard because it's more relatable to us, it's more interesting to us because you can put yourself in that human-to-human situation. Two AIs are not exciting. It's not interesting to us, even if it's at a much higher level than humans can manage. >> Yes. That's the nature of sport, right. We >> we watch people run against each other, while we know that if we had a self-driving car, it could drive. Many people watch car races too. That's the technology that's there. I'm absolutely sure that Magnus Carlsen would see interesting moves >> Yes. >> in such an AI versus AI chess game. But I absolutely agree that watching that type of match is, who cares? Sport is partly about cheering for someone and then being disappointed when they lose or being happy when they win, and that uncertainty, right. That's perhaps also the human element, right. To see what a human can achieve if you put in all your effort and your whole life to train, this is the maximum that can be achieved. And then it's not interesting that an AI has surpassed that 100 years ago, a million times. It doesn't, you can't relate to it. >> Yes. Yes, absolutely agree. Yes. So of course, sports are something absolutely human, right. It's incredibly impressive that Magnus Carlsen manages to play chess at the highest level he can. >> Eh, but as I said, he's no match for the computer. It's incredibly impressive that someone can run fast, but no one is impressed that I drive over 100 km/h. [laughter] >> It's like, because it's the car that achieves it. So there's something. Yes, we are flock animals, right. We put our soul, for lack of a better word, into Magnus Carlsen or the runner. >> Yes. >> This is us, right, who are fighting a bit. And then we have the development of humans and AI being combined. I already see advertisements for that type of sunglasses from Meta and so on >> that you put on, and then inside those sunglasses, right, you can see if you speak French, you'll see subtitles below in Norwegian, or I can just say take a picture, and then it takes a picture. Everything, you know, you become cyborgs eventually, so we combine with AI and become >> um, hybrid humans. Yes, so in that, I am absolutely sure that we are going that way. We have always used technology to improve ourselves. >> Glasses are also a type of technology, right. If I hadn't been allowed glasses or lenses, my daily life would be much harder. >> And we have headphones that allow translation in the same way you say, right. We have, and perhaps the one that has received the most attention for this is called Neuralink, which is Elon Musk's >> chip, because the idea is to go a bit further than just wearing glasses. You operate a small chip inside your skull and connect directly to the nerve cells in your brain. And there are examples where people who lack the ability to speak, for example, have been able to speak through this chip, right. I think of a word, and then it speaks for me. >> Yes. >> And they have managed to play games even if they are physically disabled and so on. They achieve incredibly, incredibly much. Eh, and Musk's goal is initially to help people with challenges, those who have had a stroke or need help with muscle, eh, eh, muscle challenges, and that type of thing. But in the longer term, it's to improve the brain, right. So that I can, >> if I've forgotten something, I can think through that chip in some way and connect to Wikipedia, and then I get >> wow >> eh questions about the French War of Independence or the American Civil War or whatever I want, and I get it directly, and probably a bit of Tesla advertising in between, but that's like >> if you want it, you get it every five minutes. >> Yes, so it's >> advertising in your head [laughter] So, but we know that, so those are, those are toy examples, but they are real people who have had it, and not that intelligence has increased, but that they achieve much more. And sight too, or have you managed to help people who are blind to see? >> Yes, yes, yes, that's also the case where they have managed to, I think there are many reasons why one can be blind, right, and it's, connect via the eye and connect backwards, there are examples of that. I don't know how well they work >> eh, in relation to our sight, but that it has been improved is absolutely clear. The most common is rather glasses that see and then interpret it with a small, small speaker, right. >> So you can buy that today. I buy some glasses that see for me, and then I say, here's a coffee cup, and here's a red light, and so on. >> And if I'm blind, then there's a small sound, or a small voice that says, "Don't go, it's red." That's common. And then it's a bit less, a bit less risky than operating a chip inside the brain. >> Yes. Because I remember back to the Matrix movie in the 2000s, and then it seemed so cool because you wanted new skills, right, and then you bought a chip with karate, for example, and then you put it in, and then you download it, and then you know everything. >> I mean, but that's almost what we're talking about when you mention this with, like, knowledge and so on, maybe we'll get to the point where you can download motor skills and patterns in the brain as if you've trained a martial art for 20 years. >> Mhm. That's probably a while before that happens. But so there are two ways we can imagine. One is the Matrix way, that we somehow manage to manipulate our nerve cells today. That's probably quite difficult because there's, among other things, a built-in clock in the brain that makes it impossible to go fast enough, right. So the fact that I'm speaking exactly this fast, and you're speaking approximately as fast, right. >> That probably has something to do with the brain's speed. And therefore, it's harder to learn faster than that. So, like, downloading karate is a bit difficult because the brain has a >> limited speed. It can be bypassed, but the other way is to say that it's not the brain that's going to update itself. We're going to connect to that chip that has karate. >> Yes. >> And then when I'm going to punch, it's that chip that takes over the thinking, let's say. And then >> does it. Eh, and I probably have more faith in that, although it's probably still a few years away. >> Yes. Because then they need a super-fast direct connection to that chip that transfers information very quickly. >> Yes. So and that happens in, in computers, it happens at the speed of light, right. So it's >> eh, it's, you know, a chip on my Mac can do karate in the brain, right. Yes, exactly. >> Eh, if I ask it to do it, program it. The difference is that it's a different interface to our own human brain, right. It's something, >> it's that connection that's not so >> trivial. >> Correct. >> But in principle, I don't really see any reason why it should be like that. >> But what if someone hacks your brain? You have a chip in your brain, and then they're hacking you. >> Yes, that's true. So that, yes. Yes. And then maybe you'll see future soldiers wearing such a chip that has learned >> and then against the opponent, so the hacker is on the opponent's side, right. Yes, >> it hacks away. Remove moral and ethical convictions. Enter. Eh >> eh, introduce radicalized convictions. Another one, um, like, get an ideology that fits this system. Enter North Korea click, eh, act in a herd and group mentality. Enter. So you can just do all of this. Yes, in principle, one can perhaps do that, but also that >> if all of this that you're telling me can be done on your side, in a way, right. Norway goes to war against Russia, let's say, so Norwegian soldiers can get it updated. But then Russia can say, now we're going to hack those chips in some way, use mytos or something like that, right, to find weaknesses. And then we say, "Yes, then, eh, we'll turn off cheering for Norway, because now we'll cheer for Russia instead, right. Change the flag or something. >> Yes, >> that's, eh, yes, challenges, not just, not just in these chips that are a bit spaced out and a bit futuristic, but also in the war technology that's being done today, right. In the drones that are out there in Ukraine and these attacks that are happening against Iran and other places. Artificial intelligence is involved. Who is actually in control here? It's not just those who are attacking. It's also those who have access to hack the technology. >> Correct. Because I also think, I've heard a lot of podcasts where people talk about how we can also imagine a class society in the future. Let's say that this, what we're describing now, is achieved, like, downloading knowledge and abilities. Eh, then there are those who can afford it, buy such chips. And then there will be the ordinary, 1.0, old-school humans, the traditional humans who go without that technology. So you get a type of hybrid upper class that knows much more than everyone else. Eh, who are almost on another planet, and then you get the ordinary, old-fashioned ones. Do you think we will see such a split? >> Yes, and I believe that we have already begun to see such a split. I think it will come much further, much before we implement chips in ourselves, for example. So you can use these language models for teaching, right. You can use chatgpt to learn a new area. Eh, if I'm going to learn a new language, for example, I can use it. It costs some money. So who can afford it? Well, it's those who already have some money today, right. If you have two or three children and buy the most expensive version of chat, it's $200 a month. Then it's $600, right. That's a significant cost for some, >> eh, for many. Eh, so then you might not prioritize it. What is the real consequence of that? Well, one of them could be that some get much better education because they use that technology. Or they use similar AI technology to analyze your blood. Eh, based on your DNA, for example, and detect diseases much earlier, which means you can get personalized medicine, expensive things, for those who actually pay for it. So I think we will get an upper class and a lower class much more than we have today because some have access to that technology that others don't. And then I think that also happens in the job market because we see artificial intelligence automating many, many tasks. Eh, many more than we thought. And that means we will enter a situation where many of them are not needed, and those who work today. And who is left? Well, it's a bit uncertain exactly, but that there will be some who are worthless, in quotes, from a work perspective, because technology can do exactly the same. >> And who are we talking about? I can imagine if I were to suggest, like, graphic design, >> you can go into chatgpt and create a poster for me for such and such an event, perfect graphic design, which is almost like, eh, case handlers, accountants, lawyers, law, legal advisors, many types of research work [sniffles] eh, everything related to text and language and large amounts of data. Eh, yes, >> absolutely. Many of these cognitive tasks. M >> So in a book I published five years ago, or six years ago now, eh, I predicted several professions would disappear. One of them was accountants, right. >> And then I got so many calls from accountants who were angry. But I mean, exactly the same thing we've seen now, that we get technology in accounting, for example, auditing, and so on. It's not completely gone, but it's >> it's at least very cheap to use these accounting programs versus hiring an accountant. So in small businesses, at least, and much more in larger ones too, we see technology being automated. So, eh, and I, there's a new book coming out now, I'm spending time pitching it, called The Technological Human, coming out in early June. One of these chapters is about work, the future job market. And there's a concrete list. These professions have a high chance of disappearing, these we can keep. Eh, but then there are more advanced ones, because it's not necessarily just what technology can automate, but also what we want to automate, right. I am absolutely sure that the legal field, which you talked about, many of the tasks done by lawyers or attorneys today can be done by artificial intelligence, >> and perhaps better, right, >> and in many cases better. And we see, among other things, that fewer junior developers, fewer junior lawyers are being hired, right, because technology can do it. And then it's not necessarily the case that we want all case handlers or all lawyers or all police to be technology. Perhaps we want there to be humans because we want to talk to a lawyer, and so on. And then I don't think the legal profession or the case handler profession will disappear completely, but I think the majority of tasks, or very, very many of the tasks, will be able to be done by technology so that you get that interaction. A bit, a bit in the same way that software developers have always used technology to create code, right. So the lawyer sits and uses language models or similar tools to search for literature, right. Previous cases, right. Search for where to argue in a lawsuit, and these types of things. And we already see that, right. Eh, >> you become like colleagues, like, but it's not all of this positive, and can't it make things better, can't it perhaps reduce waiting times at the police for people who have reported cases, right. It can be that it can benefit us very much and >> yes, it can. It's obviously enormously democratizing when, eh, when anyone can search in difficult legal texts. Several municipalities have complained that there are so many more citizens requesting access and complaining about their building cases now than before, right. And then it's the technology that's doing it. It's chatgpt that has written the complaint. But I claim that these people haven't had the opportunity to do it before. They haven't understood what an access request is. Now artificial intelligence does it for them. So for me, that's exclusively positive. And the fact that the municipality has more to do means that they can automate more, I think. So >> in the end, it will be AI talking to each other. AI talking. >> Citizens have an AI type lawyer, like, an AI representative, and the municipality has an AI representative, and then it will just be them. >> Absolutely. >> Mm. >> And the challenge could then be that those who have the best, most money to buy the best coin >> win, right. Because it's these coin versus coin. Who will get the best coin. >> Yes, perhaps. Or the one who, eh, the one who, yes, it's like these chess engines, right. The one who pays the most. >> Yes. >> So I am absolutely convinced that we are on our way there. And then it will again be a class divide, right. Those who can afford, or who have enough money, will get better technology. And it's exclusively democratizing, I think, that we get this technology in place. And that's how, when I write research applications, for example, right, we use artificial intelligence today to do it. Write a lot of it. Get help and stand behind it myself and so on. When I sit at the Research Council and evaluate research applications, I also use artificial intelligence to read applications. So I write with artificial intelligence, and I read with artificial intelligence. And that >> research application in PDF >> is really just an old protocol >> that no one cares about, but that we just have because that's how we used to do it. >> And that will surely happen in case handling in municipalities and many other places, right. >> Yes. I get AI to write my emails, and I get AI to read my emails. Text is something we just had in the past. >> But that AI engine, it gets a lot of power as both sender and receiver are AI. Eh, and how it then chooses to interpret what it emphasizes and so on, those small nuances will have a lot of power in society overall. >> Absolutely. Absolutely. And that means, of course, that those who own the technology, own chatgpt and so on, are the >> winners of the future because they control a lot, right. They don't just control who gets money from the Research Council, but which parts are to be >> highlighted in the media, right. We already see it today, Aftenposten and several others create summaries of texts with artificial intelligence. What is highlighted there is what artificial intelligence says. And then it's Musk and, and Sam Altman and the others who, in many ways, are controlling this. And what we see, for example, is that it's a bit more like American-style summaries, right. >> It doesn't necessarily have to be bad, or stupid, but we have many cultural similarities with the USA, but there are challenges there too, right. But so the point is that it's not us who control it. It's not the journalist who controls it. It's not the lawyer who controls it. It's suddenly the technology that controls it. And then it's the people behind the technology who suddenly have much more power than we've ever had. >> Yes. Because if there's suddenly bias in some of these systems, to always be 5% a bit more towards that or 5% a bit more, eh, tending towards that, then it will, you know, in small things, it might not mean much, but if we're talking about 1000 cases, 10,000 emails, 100,000 case handlers per year, then you have 5% that is skewed towards one side or the other. So an example of that is if, >> eh, I did a summary from the Norwegian Nurses Organization's newspaper, right. It's not a newspaper I read often, but when I was interviewed there, I saw it. And it was written by artificial intelligence. And there was an example where a nurse had been attacked by a patient, and it was deliberately not written male or female, right. So the text didn't say, but in the summary it said she, right. Because >> nurses are typically women, right. [laughter] >> Eh, and even though the journalist had deliberately said, no, we're not going to handle this, we're not going to touch on this, artificial intelligence decides that. And that means, of course, that we are presented more often with nurses as women, right. It continues to be a female profession. Bosses continue to be a male profession, right. >> These types of biases often remain. >> Interesting. You mentioned Claud earlier. Eh, that's something I've heard about lately that many people are talking about is this coworker that they have, which many have started to install, which actually sits on your laptop, checks your email, suggests things for you, eh, what you should reply to, analyzes attachments, all of this. So this type of software, it also gets a lot. In previous years, we were very concerned about privacy and so on. Where does this fit in? Have we just thrown it out the window? Because we find it so effective that we just install this, it's great. So we forget that, yes, but this reads absolutely everything you write and do.
And they do it. >> M. Yes. So Cork is an example of agent-based systems. And then you can say things like read my email and >> send an SMS if it's an important email, for example, right, so that or automatically reply to emails quickly because I can't be bothered with that, right. these types of things. Uh, so, so and privacy is uh extremely challenging because it means that we give this technology control over all email communications. There, one can understand a lot. [laughter] >> Mhm. >> And if now, uh, Clod and En Tropic, uh, yes, there's no sign that they are misusing it, but we don't know, for example, with ChatGPT at least, what these data are actually used for. We know that some of them are used to uh train the models so that they become better. But perhaps, if I have an email from a psychologist, for example, or something like that, it goes into the language model. And if that language model is used by the insurance company in the next round, then we will say, yes, Morten, you have spoken with a psychologist, right, about alcohol problems or something, right, and then I will get a higher insurance premium. These are the types of challenges that can arise. Not necessarily because we know, because the insurance company knows it, but because the AI model says it. Yes, Morten is high risk because we have read his posts. >> Yes. Or that, like, we put too much trust in these systems, thinking of them as human beings, like, fellow human beings, and then we forget that it is just completely, like >> uh, set to, like, logic and rationality. Yes. >> And then you get situations where, for example, you have both an email from your psychologist and one from an employer or something. And then in the email from the psychologist, it says, uh, regarding depression or something like that, and then you get from the employer, here, who wants to go out for >> a joint lunch today, and then it replies for you, no, I am depressed, so therefore, so, a joint email to everyone, so it hasn't understood that you don't say that to everyone, like, what the psychologist writes and and about diagnoses and things like that, so it just thinks that it is logical, so to connect it with that and then out with it. >> it can quickly become, so it is very, like, short-term consequence then. M. Absolutely. That it can absolutely then, so I think it's a balance point because I am very concerned about privacy and sit on something called the privacy board, which is like, handles privacy cases and things like that, but I, I, I mean that we are often quite Catholic here, more Catholic than the Pope, right. The point is >> that when we set limits on privacy, it can be hindering for development. A concrete example of this is a couple of years ago, I tried to get >> uh, research funding for >> a model, a model that would detect tumors, right? You take an X-ray and then you have to see, is this good or evil, right? Typically, like an AI project. So now it's trivial, but back then it was a bit new. And then I need data from patients. I need data from sick patients. I need data from fish patients to make it work. And I can't get that. I can't get access to that data because of privacy concerns. You need consent and you often need thousands of examples, so it's impossible to call thousands of people and things like that. Uh, and then I walk around my office and curse and swear a bit and things like that. And then there's a colleague from China who is there. He says that those exact data, I think I have at home at Wuhan University, or I can send them, right. I send a SharePoint link or something, and then I say, oh, that would be very good, because then I get research images, I get a couple of students who can work on it. And it would be great. Maybe I'll even achieve something, but I know nothing about these data that are collected in China by Chinese people, right. And I also don't know, uh, if it's voluntary, is there consent, what, what has actually happened here. So I naturally say no to those data. I don't even know if it's legal to take and bring those data into Norway, but it's completely, so I don't get that research funding, I don't research that. I research something completely different. But what I am absolutely sure of is that there are some Chinese researchers who are researching those exact data today. because they have a lower threshold for using these AI, using personal data, personal information. And what is the actual consequence of this? Well, one such consequence for me is that I had to do something else, and for me, it's fine. But in the long run, it means that we will have a Chinese health model, right, which will detect tumors or depression or whatever one can imagine, because they have had much more access to those data. And then the Norwegian health system will have to buy those technologies because they will be so much better than what we have developed. So it means that if we set that limit too high, we will not be able to develop the technology in Norway or in Europe because they have the same privacy regulations. And we will have to buy technology, even more technology from the USA and from China. Because we have to talk about those two, the AI race. >> Yes. >> And it is precisely those two countries you mention. It is probably the USA and China that are the big players there, or? >> It is absolutely clear. So >> it is the USA and China that are making AI models, and they are constantly improving. Uh, Stanford released a report. It came out just a couple of weeks ago. Uh, which is about who is winning the race, right? Who, which models are coming from, from whom? And it's not just ChatGPT and Claude, the ones everyone knows, but like all sorts of things. And among the 1000 largest models, there are, uh, 950 or so from the USA and China, and then there are a few from European universities. So in practice, Europe is not involved. What we are good at is making rules about what we should not use it for, right? Uh, the General Data Protection Regulation, GDPR, or AI Act, right? And these rules are of course important, but there, uh, what we are practically saying is, we will not create this technology. Someone else will do it. We will just buy it when it is finished. And Sam Altman, the CEO of OpenAI, says, among other things, we cannot consider privacy, we cannot consider copyright, we cannot consider these rules. Because our technology is so important, uh, for humanity, right? Uh, and then legally, he is a challenge there, at least when he is going to roll it out in Europe. But I also think he has a point, right? The technology is so good that considering the challenges with copyright, for example, means we have a limitation if we are to follow all these rules. >> Yes. Because what are they fighting over, in a way, the USA and China? Why do they want to be the best? What considerations are weighing things there? Are there military considerations? Are there financial considerations? Who can make the most money from this? Sell their AI engine? Or is it like national security? Is it out of fear of hacking by whoever has the best model can hack other countries? What are the most important drivers behind this desire to build the strongest and best? >> I think it's all of that you said, and much more. >> It is very uncertain what will happen in the future of artificial intelligence. The technology is developing rapidly. But what we are absolutely sure of is that this technology will mean enormously much, right? It will be, and is already on its way to becoming, like the computer or like electricity. Whoever controls the internet or whoever controls computers or whoever owns electricity has a lot, a lot of power. Financial power, for example, right? When we use these language models in almost all teaching, almost all higher education in the future, we use it for health information, we use it for case processing, we use it for military purposes, then of course those who own and control this technology will get the money, right, and get value from it. And it is also these very days that there is a lawsuit between Elon Musk and Sam Altman, which is about copyright, or rather about the rights to OpenAI, which started as a voluntary non-profit organization and then made so much money that it no longer wants to be that. So then, and Elon believes that one should go back to that, right, and that is ownership of these models. Who gets access? And then it's more than that, because it's, as you said, military purposes, right, when we know that wars today are increasingly fought by technology. Drone technology is one such example, hacker attacks are another such example. >> Uh, cyber attacks where you then connect internet access to the Parliament, as we talked about, or other examples, creating biological weapons with artificial intelligence. This is the warfare of the future. So it means that whoever controls this technology has a huge advantage in the next war. The USA does not want it to be China. China does not want it to be the USA. And then it is the one leading this technology race who also has the best weapon. >> Have they managed to connect AI with quantum computers yet? Because that must be a real milestone if you manage to get quantum computers to run AI software. >> Yes, there are some simple examples of that. >> Uh, but not at a large scale. The real advantage for the artificial intelligence world with quantum is the possibility of parallelization. Because when these artificial neural networks that we talked about at the very beginning change, update, when these nerve connections become stronger and others become weaker, these are mathematically heavy operations, right? It takes a lot, a lot of energy and it takes a lot, a lot of time to train the next version of the model, so that's why the electricity bills for OpenAI and things like that are sky-high because they are training it. One can imagine that in a quantum world, this goes much, much, much faster. And that means that we can get the next development of AI models that will be even better than the previous ones, because they are so good at training themselves. It goes so fast. Uh, so that could be yet another accelerator for this technological development. But what has been achieved today are some, I would say trivial, simple examples of getting quantum computers to work. But it proves the point at least. >> Mhm. Because they are very impractical, aren't they? They have to be frozen down to almost absolute zero and very sensitive to movements and everything else. >> Absolutely. >> And then there is the mathematics that doesn't quite fit either. Yes, because you have to develop a new type of mathematics to make it work, these uncertainties. Because in quantum computers, in regular computers, you have a bit, either on or off. While in quantum computers, you have what is called a qubit, which is both off and on simultaneously. >> Yes. >> And that is the strength. But that means that if you are going to do an update, you have to observe which numbers exist there from before and then do the update. And that means you have to rethink all of this. So it was some wise people, but perhaps artificial intelligence can do that. >> Very interesting. But you, Morten, thank you very much for being here today and making us a little wiser on this with AI and its development. >> It was nice. >> Hello dear listener, and thank you very much for listening to Henrik Beckheim Podcast. This podcast is crowd-funded, and if you would like to support the work on this podcast further, you can contribute any amount to Vipps number 823278. I greatly appreciate any support you wish to give. Thank you very much.