Hello. I'm Caroline Steel. This is the BBC World Service. And welcome to The Engineers. This year we're at the Science Specialist University, Imperial College London. And we're here to focus on the technical revolution defining our era, artificial intelligence. I'm joined by a panel of three world leaders in the field and a large, enthusiastic audience in Imperial's Great Hall. Already a computer can defeat the world's greatest player at our most complex strategy game. The first movie written entirely by AI has just been released, and AI may have discovered our first antibiotic in three decades. Together with
our partners, Royal Commission 1851, we've brought together three engineers at the cutting edge of this field to discuss their work and what it means for us humans. Paolo Pirjanian is Armenian, but he was born in Iran and started his career working on Mars rovers for NASA. He's now founder and CEO of Embodied, which is a company that builds emotionally intelligent robots to help with child development. David Silver is from the UK, where he's principal research scientist at the AI Research Lab. Google DeepMind. He led the team that used AI to defeat the world's best player
at the complicated strategy game Go. And he's working on artificial general intelligence. Regina Barzilay is Israeli-American and a distinguished professor for AI and Health at MIT in the U.S. She created a major breakthrough in detecting early stage breast cancer and also led the team that used AI to discover what is hoped to be a brand new antibiotic. So please do join me in welcoming them all. Regina, let's start with you. So what is it that made you shift your work to oncology? Sadly, you were in the perfect position to do that, weren't you? Yeah. So
actually, I started my work at MIT in 2003 as a faculty and I was working on natural language processing and AI And in 2014, I was diagnosed with breast cancer and I was treated in one of the best hospitals in the United States, Massachusetts General Hospital. And what I discovered going through the treatment that there was really no AI or not even basic information technology as part of the treatment. Neither the diagnostics nor the treatment nor the post-treatment. And, you know, after I was treated, I just was totally confused as to what I want to
do, because it was the first time I realised that, you know, my life is finite and I've seen a lot of very sick people there surrounding me. And I was thinking, What can I do? And MGH - this hospital at MIT - is just one subway stop away, they are separated by a bridge I’m saying, how come we have all this great technology at MIT, but none of it is actually coming to the hospitals and helping patients. So after I finished my treatment, I still didn't have my hair. I started kind of going from doctor
to doctor, asking them, you know, how I can bring AI - I will do it for free. I'm a professor. So there were not many takers, but eventually we found somebody who... it was a doctor called Connie Lehman who had the idea that we can apply AI to do early detection of cancer. Thank you, Regina. David, you came to AI via the games industry and you did a PhD in reinforcement learning. What is reinforcement learning and how did you use it in those early days? Yeah. So I guess I started out in the games industry
before I went back to academia and I was working on building computer games. And as a big part of building computer games is building the AI for those games, that kind of makes all of the characters move around. And I found myself being fundamentally disappointed by the methods that were being used in those games. And it felt like what I really wanted to do was build something that had real AI in it. I discovered this idea of reinforcement learning, which is basically a method very much like those that animals and humans use, where the system
is able to learn for itself from experience, from trial and error, from trying things out and seeing what works and what doesn't. So is it sort of like when we learn to not touch fire because at some point we try it and it really hurts and we learn don't do that in the future because the consequences aren't appealing. Is it sort of like how humans learn reinforcement learning? Yes, it's a lot like that. So in fact humans are believed to have, you know, a major part of the brain which is devoted to providing a signal,
giving feedback that that makes the brain actually learn to do more of the good things and less of the bad things. And so actually, that's inspired a lot of work in machine learning to make machines have that same capability. But a machine doesn't feel heat or isn't rewarded by a cookie. How can you reward a machine? Yeah, so for a machine, it's just a number. So you give it a positive number if it's done something good and a negative number if it's done something bad. And at the end of the day, everything stems from that
one single number. So this one single number, which we call the reward, contains enormous power because it's the signal that drives everything. Paolo, you said your experiences of feeling alienated in foreign countries made you want to create an imaginary friend. And I'm sure much of our audience can relate to having an imaginary friend, but for most of us, they stay imaginary. How did you go about making a real one? So unfortunately, there's a lot of people in need of companionship or therapy, and there's a massive gap of labour force that can provide us, as an
example, to use numbers from the U.S. We know the prevalence of things such as autism is growing rapidly. Ten years ago it was one out of about 200 kids. Today, it's one out of about 30 kids. So the experiences I had, which was leaving my family at a very young age, living abroad in a society that's amazing. I mean, these are amazing people, but yet you are different, so you are not going to be embraced. So this is not too dissimilar from a child on the autism spectrum that has a hard time expressing themselves or
reading emotions from other people. And that was the genesis of creating a robot companion that understands human emotions, can create a deep relationship with a child, and it'll help them exercise and practice social skills such as eye contact, turn-taking, joint attention and so on, so that the child has a chance of being successful in their society. Thank you. Regina, what can AI do when it comes to understanding cancer that humans can't? So I think that in cancer and in many other diseases, a big question is always, how do you deal with uncertainty? And unfortunately, today
we rely on humans who don't have this capacity to make predictions. And as a result, many times people get wrong treatments or they are diagnosed much later. And one question that really troubled me is, you know, how late I was diagnosed and when we already developed a model, I came back to my own mammograms and rediscovered the mammograms two years earlier already had on a tiny small cancer. Now, for human eye, for radiologists, it's impossible to diagnose it because it's so, so confusing. There's so many other spots on your tissue. So what AI can do,
it can do a lot of tasks which humans cannot do. Take all the data that we have and remove the guessing out of diagnosis and treatment. Thank you. David, AI had already defeated the reigning grandmaster, Garry Kasparov, at chess well before you started your project AlphaGo and the rules of Go sound quite simple. Basically, on each turn, a player puts down a counter on the board and you gain territory by connecting your counters and the player with the most territory at the end of the game wins. So why is it harder for a computer to
beat a human at Go than at chess? Which sounds more complicated. So the game of Go is this very beautiful and elegant game where it seems at first glance like the rules are very simple. But once you start to understand it a bit like unpeeling an onion, you discover more and more layers of complexity and what's amazing is that when humans play this game, they basically... If you ask them to describe how they did something, they really don't know. They've used incredible intuition. And so these amazing professional players who've devoted their entire lives to this
game have built this incredible intuition and creativity and intuition and creativity are two traits which were previously considered to be very human and very hard to build into machines. So while chess, it was possible to succeed just with tactical look ahead in the game of Go, that wasn't enough because, you know, early on in the game you just have this handful of stones on the board and you really just have to imagine what the game will pan out like, you know, 300 moves later with this sort of intuitive sense of where it will go. And
that required some some major breakthroughs. Paolo, debatably even more complex than Go is human children. Your human centred robot forms an emotional bond with children. How can a robot do that? Well, first of all, it's important to make clear that the robot is not meant to replace the need for human contact. It's really almost like training wheels to teach children the social skills and then be able to practice those in real life. The way the robot forms bond is that humans are wired to bonding. We create connections with inanimate objects all the time. I mean,
with a robot that has eyes, can make eye contact, can smile back at you and can speak to you and express emotion and empathy. It's actually not that hard to create a bond there. And children open up to to these robots very quickly in ways that they may not even open up to their therapists or parents. Thank you, Paolo. Regina, you not only made an impact on oncology, your team at MIT used AI to discover what could be our first new antibiotic in three decades. It seems it can be E. Coli, MRSA, and strains of
bacteria which are currently resistant to all other antibiotics. So I think we all wish it success. How did you do that? So I should say that, you know, developing antibiotics is not an area with an immense competition, even though their resistance to to antibiotics that we have continues to grow. This happened to be an area where pharmaceutical companies are not very active because economically it doesn't work for them. So in some ways we do need to have alternative approaches. I met a colleague and he was working. He was from biological engineering, he was working on
antibiotics. And he was describing the big problem of finding new molecules which are effective against bacteria, drug resistant bacteria. But at the same time, are not toxic to humans. They have some molecules screened against, I think E.coli. We started with that and then we just gave to the machine, you know, thousands of molecules. And for each molecule you knew whether it kills a bacteria or not. It was kind of the first attempt to learn automatically. How do you look at the structure of the molecule and predict whether it would have a desired effect? We found
a molecule that didn't look like something human created. And it turns out in the lab that it was able to kill using a different mechanism of action, kill it in a different way. And that's what made it so effective against so many different species. David, let's go back to you. So, so far we've been talking about systems designed to perform one task - that's known as narrow AI, but you're working towards artificial general intelligence. Could you explain what artificial general intelligence is? So if you think about humans and human intelligence, it's this wonderful and beautiful
thing where we're able to learn skills which are incredibly diverse, where, you know, one person might choose to specialise in learning how to play tennis and another person might specialise in becoming an amazing chef and another person, a pianist and another person, a scientist and so when we want to build artificial intelligence, we want systems which not only solve a single problem but in a similar manner to humans, are able to approach any number of problems with intelligence, and that's capable of doing amazing things in each of those different areas. And that's what we refer
to as artificial general intelligence or AGI for short. And how far off do you think we are from that being a reality? So I think it's going to be a spectrum over many years. And I also think it's likely or at least plausible that there are many breakthroughs that are still required before we can really crack, you know, the same kind of level of intelligence that humans have. Regina, you've developed AI to better predict cancer, but it's only employed in a tiny number of cases, right? Why is AI not used more widely in medicine? The
problem is that we're creating a lot of great technology, but this technology is not really translated into patient care. And if I would ask the audience, you know, how many of you when you last saw your physician, have you actually seen any AI, and I'm sure not many can really attest to it. So I think the technology for many of the tasks is really mature, but there are many other questions which have nothing to do with AI per se. One is the regulation or regulations, in Europe, UK, US - continue to change. Another big problem
is people don't really know how to bill for AI. And today the American doctor who uses AI loses money when they see a patient. So it's not much of a motivation. So just to clarify that, did you say that using AI for doctors in the US could actually make them lose money because it makes treatment more effective? I'm referring to a specific paper. The way the billing is done it somehow relates to the time the doctor spends with a patient. So if you have something that makes it faster, you're actually losing money. Very depressing. And
Paolo, what about your challenges? The robot you've developed mimics human behaviour. What's next? I'm very hopeful with what we are doing in terms of creating social emotional AI systems that can help humanity become its best. If we can intervene early with children for instance on the autism spectrum, they have a chance of really integrating well with the society and doing really well. When we think about other vulnerable areas of our life is when we age. Social isolation, being lonely, and that leads to mental health issues, that leads to physical health issues and so on. We
can have the same systems become a companion that help you there. Once we figure out the physical task, you can also imagine that they can give you assistive care, meaning that they can be not only a social emotional companion for you, but they can also be a companion and say, let's cook some food together, let's go for a walk together and so on, which is going to help a lot with independent living with dignity when we are at that age. Do you think we'll see robots helping with assistive care in the future any time soon,
or is this way off in the distance? I think it is within the next decade. Oh wow. David, you're working on Gemini, which is Google's answer to ChatGPT, and you aim for it to be able to do both tax returns and write a novel. People would probably be very happy to have their tax returns done or I would definitely would be. But novels, should we really be letting AI sort of take over human culture? So it's a great question. I think I wouldn't see it as taking over human culture. I think what will happen, or
the most likely outcome, is that we'll be providing an incredibly powerful tool to human authors. So we've already seen this in a number of areas where we've developed technologies that enable authors of different kinds of media to basically create things much more powerfully using tools. So, for example, there's a music authoring system called Lyria that was released recently, and there's this wonderful footage of Will.i.am when he's playing with it for the first time. And he's just so excited because he says it can kind of speed up his songwriting process by, you know, 10 to 100
times. So I think, you know what I really hope we get to is a world in which the AI and the humans kind of work together to just make everything better. So, you know, I'm excited to be in a world where we have much more... the most amazing novels that we can imagine, that which go far beyond the ones we have today. Thank you, David. Thank you. This is The Engineers - Intelligent Machines from the BBC World Service. We've discussed medical AI, emotionally intelligent robots, the goal of artificial general intelligence and the threats AI might
pose. Has anyone got a question on anything we've discussed so far? Wow. Okay. Pretty much everyone has a hand up, so this is going to be tricky. Could we start with the man in the red shirt? If you could stand up and say your name and your question. Thank you very much. Hello, my name's Simon. The British government wants to make Britain a leader in AI and sees the way to do this by making it a safe space. They're looking at doing that by making sure you can't develop AI where you don't understand the consequences
before you develop it. Does the panel think that's the right approach? So, yes. David, what do you think about legislation around AI? I think we need some kind of regulation. I think it's, you know, an area which clearly is going to have more and more consequence and impacts on society. So I think regulation is important. I think some of the areas which have been agreed in various summits over the last year - you know, fantastic start. One thing I would say is I think it's hard sometimes to come up with like a one size fits
all recipe for AI because it's so different in different areas. So you know, the regulation that you might need in medicine might look quite different to the kind of regulation you might need for, say, a chatbot. So I think, you know, we have to look at each area separately and make sure that whatever we do, you know, we really fully understand the consequences of what the impact of AI will be in that area. There's a lot of anxiety about the pace AI is moving, right? We have people resigning, leaders in the field to campaign for
more safeguarding against the threat of misinformation, the threat to jobs, and even an existential risk to humanity. Regina? I actually feel quite opposite. People are really suffering. There are lots of incurable diseases. It's hard for patients, it's hard for their families. There is a lot of technologies that is out there and because we cannot get together to put the regulation in place, make, you know, the payers take part of it and find a way to bring it I think we are really making many, many people suffer. Thank you. Paolo, what about robots? Are they going
to take all our jobs? Yes. Yes. Okay, great. We've got it. We finally have an answer. Thank you, Paolo. On a serious note, I feel it's a bit of a conundrum to think about legislation, because on one hand, yes, there are definitely risks and you would like to regulate it so that no one with bad intentions goes wild. On the other hand, if you think about it, it's an extremely potent technology. It could change everything in our lives including being strategically important technology to master from a national perspective. And from that perspective, if you regulate
it, let's say part of the regulation is slowing it down, what are our adversaries going to do? Are they going to then have an advantage over us? So it's a bit of an arms race. And I think for that reason, practically speaking, I think it's going to be very hard to regulate it to that level unless you can have international regulation and agreement across all the powerful nations to say this is how we're going to handle it. And I haven't seen that work out very well in the history of time. Fair enough. Audience, we have
time for a couple more questions. So hands up. If we could go to the man in the white top. Thank you. Hello, panel. My name's Rob. And I'm a business and sports coach. And I'm wondering whether what you've been talking about is possible for helping humans to improve their performance at a sport. Good question. Who's our greatest sport enthusiast? Perhaps this one is for you, David. Can we be using AI to improve sports performance? It's a great question. There's a lot of really amazing research that's happening to try and do just that. You know, one
thing which we've been doing at Google DeepMind is actually a collaboration with Liverpool Football Club to try and help them improve their tactics. So that's one example. I think, you know, the amazing thing about sports is it's become over time so refined in terms of the particular approaches that people take that actually, you know, really they're very open to new discoveries and new ways to do things. And so it's been really fun actually, just watching that kind of thing unfold. Very cool. Thank you, David. I can see we've got lots of young people in the
audience, some teenagers even. Does any of our younger audience have a question? If so stick your hand up. There was a young lady who had a hand up here with a ponytail. Yes. As AI develops, I reckon that humans probably depend more on AI and maybe learn less. Is there anything we could do to maybe make sure that as it develops, humans still develop and learn? Ooh, good question. So will we stop learning, as AI does the learning instead for us? I'm going to put that to all of you if that's all right. What do
you think, David? I'd like to imagine a world where we have an AGI which is like a personal friend, assistant teacher, and everything we do understands what we want to learn, and it knows just how to teach us and help us to learn more and more and more and more. Regina, what do you think? Do you think we're just going to end up relying on AI for everything? As a non-native speaker of English I remember as a young professor I spent a humongous amount of time like reading my papers and making sure, you know, because
in my native language we don't have and so I make a lot of mistakes when I write. I see now how I write papers. Actually, it removes a lot of my pressure in writing and I can really focus on ideas rather than doing the small things. So I hope that we will find a symbiosis. We don't really, you know, remove our basic skills, but we can just do more and focus on the things that we can do better. I would like to imagine if Isaac Newton or Albert Einstein had access to these tools today, I
mean, as prolific as they were at a time where we didn't even have calculators, imagine what they would have done and the impact it would have had on the world if they had access to these tools. In the next five years potentially, you don't need to code. You will just tell your favourite AI system and say I want a code that does x, y, z and you can accomplish what would take years of many people in hours today. So it will make you more prolific. Thank you. I think it's fair to say all three of
you have given us hope for the future. Audience, thank you so much for your questions. I wish we could take more, but I'm afraid we're out of time. That's it for The Engineers - Intelligent Machines at Imperial College, London. I'm Caroline Steel. On behalf of the BBC World Service our partners the Royal Commission for the Exhibition of 1851 and my producer, Charlie Taylor. Please join me in giving a warm round of applause for our brilliant pioneering AI engineers Regina Barzilay, Paolo Pirjanian and David Silver Hello. I'm Caroline Steel. This is the BBC World Service. And
welcome to The Engineers. This year we're at the Science Specialist University, Imperial College London. And we're here to focus on the technical revolution defining our era, artificial intelligence. I'm joined by a panel of three world leaders in the field and a large, enthusiastic audience in Imperial's Great Hall. Already a computer can defeat the world's greatest player at our most complex strategy game. The first movie written entirely by AI has just been released, and AI may have discovered our first antibiotic in three decades. Together with our partners, Royal Commission 1851, we've brought together three engineers at
the cutting edge of this field to discuss their work and what it means for us humans. Paolo Pirjanian is Armenian, but he was born in Iran and started his career working on Mars rovers for NASA. He's now founder and CEO of Embodied, which is a company that builds emotionally intelligent robots to help with child development. David Silver is from the UK, where he's principal research scientist at the AI Research Lab. Google DeepMind. He led the team that used AI to defeat the world's best player at the complicated strategy game Go. And he's working on artificial
general intelligence. Regina Barzilay is Israeli-American and a distinguished professor for AI and Health at MIT in the U.S. She created a major breakthrough in detecting early stage breast cancer and also led the team that used AI to discover what is hoped to be a brand new antibiotic. So please do join me in welcoming them all. Regina, let's start with you. So what is it that made you shift your work to oncology? Sadly, you were in the perfect position to do that, weren't you? Yeah. So actually, I started my work at MIT in 2003 as a
faculty and I was working on natural language processing and AI And in 2014, I was diagnosed with breast cancer and I was treated in one of the best hospitals in the United States, Massachusetts General Hospital. And what I discovered going through the treatment that there was really no AI or not even basic information technology as part of the treatment. Neither the diagnostics nor the treatment nor the post-treatment. And, you know, after I was treated, I just was totally confused as to what I want to do, because it was the first time I realised that, you
know, my life is finite and I've seen a lot of very sick people there surrounding me. And I was thinking, What can I do? And MGH - this hospital at MIT - is just one subway stop away, they are separated by a bridge I’m saying, how come we have all this great technology at MIT, but none of it is actually coming to the hospitals and helping patients. So after I finished my treatment, I still didn't have my hair. I started kind of going from doctor to doctor, asking them, you know, how I can bring AI
- I will do it for free. I'm a professor. So there were not many takers, but eventually we found somebody who... it was a doctor called Connie Lehman who had the idea that we can apply AI to do early detection of cancer. Thank you, Regina. David, you came to AI via the games industry and you did a PhD in reinforcement learning. What is reinforcement learning and how did you use it in those early days? Yeah. So I guess I started out in the games industry before I went back to academia and I was working on
building computer games. And as a big part of building computer games is building the AI for those games, that kind of makes all of the characters move around. And I found myself being fundamentally disappointed by the methods that were being used in those games. And it felt like what I really wanted to do was build something that had real AI in it. I discovered this idea of reinforcement learning, which is basically a method very much like those that animals and humans use, where the system is able to learn for itself from experience, from trial and
error, from trying things out and seeing what works and what doesn't. So is it sort of like when we learn to not touch fire because at some point we try it and it really hurts and we learn don't do that in the future because the consequences aren't appealing. Is it sort of like how humans learn reinforcement learning? Yes, it's a lot like that. So in fact humans are believed to have, you know, a major part of the brain which is devoted to providing a signal, giving feedback that that makes the brain actually learn to do
more of the good things and less of the bad things. And so actually, that's inspired a lot of work in machine learning to make machines have that same capability. But a machine doesn't feel heat or isn't rewarded by a cookie. How can you reward a machine? Yeah, so for a machine, it's just a number. So you give it a positive number if it's done something good and a negative number if it's done something bad. And at the end of the day, everything stems from that one single number. So this one single number, which we call
the reward, contains enormous power because it's the signal that drives everything. Paolo, you said your experiences of feeling alienated in foreign countries made you want to create an imaginary friend. And I'm sure much of our audience can relate to having an imaginary friend, but for most of us, they stay imaginary. How did you go about making a real one? So unfortunately, there's a lot of people in need of companionship or therapy, and there's a massive gap of labour force that can provide us, as an example, to use numbers from the U.S. We know the prevalence
of things such as autism is growing rapidly. Ten years ago it was one out of about 200 kids. Today, it's one out of about 30 kids. So the experiences I had, which was leaving my family at a very young age, living abroad in a society that's amazing. I mean, these are amazing people, but yet you are different, so you are not going to be embraced. So this is not too dissimilar from a child on the autism spectrum that has a hard time expressing themselves or reading emotions from other people. And that was the genesis of
creating a robot companion that understands human emotions, can create a deep relationship with a child, and it'll help them exercise and practice social skills such as eye contact, turn-taking, joint attention and so on, so that the child has a chance of being successful in their society. Thank you. Regina, what can AI do when it comes to understanding cancer that humans can't? So I think that in cancer and in many other diseases, a big question is always, how do you deal with uncertainty? And unfortunately, today we rely on humans who don't have this capacity to make
predictions. And as a result, many times people get wrong treatments or they are diagnosed much later. And one question that really troubled me is, you know, how late I was diagnosed and when we already developed a model, I came back to my own mammograms and rediscovered the mammograms two years earlier already had on a tiny small cancer. Now, for human eye, for radiologists, it's impossible to diagnose it because it's so, so confusing. There's so many other spots on your tissue. So what AI can do, it can do a lot of tasks which humans cannot do.
Take all the data that we have and remove the guessing out of diagnosis and treatment. Thank you. David, AI had already defeated the reigning grandmaster, Garry Kasparov, at chess well before you started your project AlphaGo and the rules of Go sound quite simple. Basically, on each turn, a player puts down a counter on the board and you gain territory by connecting your counters and the player with the most territory at the end of the game wins. So why is it harder for a computer to beat a human at Go than at chess? Which sounds more
complicated. So the game of Go is this very beautiful and elegant game where it seems at first glance like the rules are very simple. But once you start to understand it a bit like unpeeling an onion, you discover more and more layers of complexity and what's amazing is that when humans play this game, they basically... If you ask them to describe how they did something, they really don't know. They've used incredible intuition. And so these amazing professional players who've devoted their entire lives to this game have built this incredible intuition and creativity and intuition and
creativity are two traits which were previously considered to be very human and very hard to build into machines. So while chess, it was possible to succeed just with tactical look ahead in the game of Go, that wasn't enough because, you know, early on in the game you just have this handful of stones on the board and you really just have to imagine what the game will pan out like, you know, 300 moves later with this sort of intuitive sense of where it will go. And that required some some major breakthroughs. Paolo, debatably even more complex
than Go is human children. Your human centred robot forms an emotional bond with children. How can a robot do that? Well, first of all, it's important to make clear that the robot is not meant to replace the need for human contact. It's really almost like training wheels to teach children the social skills and then be able to practice those in real life. The way the robot forms bond is that humans are wired to bonding. We create connections with inanimate objects all the time. I mean, with a robot that has eyes, can make eye contact, can
smile back at you and can speak to you and express emotion and empathy. It's actually not that hard to create a bond there. And children open up to to these robots very quickly in ways that they may not even open up to their therapists or parents. Thank you, Paolo. Regina, you not only made an impact on oncology, your team at MIT used AI to discover what could be our first new antibiotic in three decades. It seems it can be E. Coli, MRSA, and strains of bacteria which are currently resistant to all other antibiotics. So I
think we all wish it success. How did you do that? So I should say that, you know, developing antibiotics is not an area with an immense competition, even though their resistance to to antibiotics that we have continues to grow. This happened to be an area where pharmaceutical companies are not very active because economically it doesn't work for them. So in some ways we do need to have alternative approaches. I met a colleague and he was working. He was from biological engineering, he was working on antibiotics. And he was describing the big problem of finding new
molecules which are effective against bacteria, drug resistant bacteria. But at the same time, are not toxic to humans. They have some molecules screened against, I think E.coli. We started with that and then we just gave to the machine, you know, thousands of molecules. And for each molecule you knew whether it kills a bacteria or not. It was kind of the first attempt to learn automatically. How do you look at the structure of the molecule and predict whether it would have a desired effect? We found a molecule that didn't look like something human created. And it
turns out in the lab that it was able to kill using a different mechanism of action, kill it in a different way. And that's what made it so effective against so many different species. David, let's go back to you. So, so far we've been talking about systems designed to perform one task - that's known as narrow AI, but you're working towards artificial general intelligence. Could you explain what artificial general intelligence is? So if you think about humans and human intelligence, it's this wonderful and beautiful thing where we're able to learn skills which are incredibly diverse,
where, you know, one person might choose to specialise in learning how to play tennis and another person might specialise in becoming an amazing chef and another person, a pianist and another person, a scientist and so when we want to build artificial intelligence, we want systems which not only solve a single problem but in a similar manner to humans, are able to approach any number of problems with intelligence, and that's capable of doing amazing things in each of those different areas. And that's what we refer to as artificial general intelligence or AGI for short. And how
far off do you think we are from that being a reality? So I think it's going to be a spectrum over many years. And I also think it's likely or at least plausible that there are many breakthroughs that are still required before we can really crack, you know, the same kind of level of intelligence that humans have. Regina, you've developed AI to better predict cancer, but it's only employed in a tiny number of cases, right? Why is AI not used more widely in medicine? The problem is that we're creating a lot of great technology, but
this technology is not really translated into patient care. And if I would ask the audience, you know, how many of you when you last saw your physician, have you actually seen any AI, and I'm sure not many can really attest to it. So I think the technology for many of the tasks is really mature, but there are many other questions which have nothing to do with AI per se. One is the regulation or regulations, in Europe, UK, US - continue to change. Another big problem is people don't really know how to bill for AI. And
today the American doctor who uses AI loses money when they see a patient. So it's not much of a motivation. So just to clarify that, did you say that using AI for doctors in the US could actually make them lose money because it makes treatment more effective? I'm referring to a specific paper. The way the billing is done it somehow relates to the time the doctor spends with a patient. So if you have something that makes it faster, you're actually losing money. Very depressing. And Paolo, what about your challenges? The robot you've developed mimics human
behaviour. What's next? I'm very hopeful with what we are doing in terms of creating social emotional AI systems that can help humanity become its best. If we can intervene early with children for instance on the autism spectrum, they have a chance of really integrating well with the society and doing really well. When we think about other vulnerable areas of our life is when we age. Social isolation, being lonely, and that leads to mental health issues, that leads to physical health issues and so on. We can have the same systems become a companion that help you
there. Once we figure out the physical task, you can also imagine that they can give you assistive care, meaning that they can be not only a social emotional companion for you, but they can also be a companion and say, let's cook some food together, let's go for a walk together and so on, which is going to help a lot with independent living with dignity when we are at that age. Do you think we'll see robots helping with assistive care in the future any time soon, or is this way off in the distance? I think it
is within the next decade. Oh wow. David, you're working on Gemini, which is Google's answer to ChatGPT, and you aim for it to be able to do both tax returns and write a novel. People would probably be very happy to have their tax returns done or I would definitely would be. But novels, should we really be letting AI sort of take over human culture? So it's a great question. I think I wouldn't see it as taking over human culture. I think what will happen, or the most likely outcome, is that we'll be providing an incredibly
powerful tool to human authors. So we've already seen this in a number of areas where we've developed technologies that enable authors of different kinds of media to basically create things much more powerfully using tools. So, for example, there's a music authoring system called Lyria that was released recently, and there's this wonderful footage of Will.i.am when he's playing with it for the first time. And he's just so excited because he says it can kind of speed up his songwriting process by, you know, 10 to 100 times. So I think, you know what I really hope we
get to is a world in which the AI and the humans kind of work together to just make everything better. So, you know, I'm excited to be in a world where we have much more... the most amazing novels that we can imagine, that which go far beyond the ones we have today. Thank you, David. Thank you. This is The Engineers - Intelligent Machines from the BBC World Service. We've discussed medical AI, emotionally intelligent robots, the goal of artificial general intelligence and the threats AI might pose. Has anyone got a question on anything we've discussed so
far? Wow. Okay. Pretty much everyone has a hand up, so this is going to be tricky. Could we start with the man in the red shirt? If you could stand up and say your name and your question. Thank you very much. Hello, my name's Simon. The British government wants to make Britain a leader in AI and sees the way to do this by making it a safe space. They're looking at doing that by making sure you can't develop AI where you don't understand the consequences before you develop it. Does the panel think that's the right
approach? So, yes. David, what do you think about legislation around AI? I think we need some kind of regulation. I think it's, you know, an area which clearly is going to have more and more consequence and impacts on society. So I think regulation is important. I think some of the areas which have been agreed in various summits over the last year - you know, fantastic start. One thing I would say is I think it's hard sometimes to come up with like a one size fits all recipe for AI because it's so different in different areas.
So you know, the regulation that you might need in medicine might look quite different to the kind of regulation you might need for, say, a chatbot. So I think, you know, we have to look at each area separately and make sure that whatever we do, you know, we really fully understand the consequences of what the impact of AI will be in that area. There's a lot of anxiety about the pace AI is moving, right? We have people resigning, leaders in the field to campaign for more safeguarding against the threat of misinformation, the threat to jobs,
and even an existential risk to humanity. Regina? I actually feel quite opposite. People are really suffering. There are lots of incurable diseases. It's hard for patients, it's hard for their families. There is a lot of technologies that is out there and because we cannot get together to put the regulation in place, make, you know, the payers take part of it and find a way to bring it I think we are really making many, many people suffer. Thank you. Paolo, what about robots? Are they going to take all our jobs? Yes. Yes. Okay, great. We've got
it. We finally have an answer. Thank you, Paolo. On a serious note, I feel it's a bit of a conundrum to think about legislation, because on one hand, yes, there are definitely risks and you would like to regulate it so that no one with bad intentions goes wild. On the other hand, if you think about it, it's an extremely potent technology. It could change everything in our lives including being strategically important technology to master from a national perspective. And from that perspective, if you regulate it, let's say part of the regulation is slowing it down,
what are our adversaries going to do? Are they going to then have an advantage over us? So it's a bit of an arms race. And I think for that reason, practically speaking, I think it's going to be very hard to regulate it to that level unless you can have international regulation and agreement across all the powerful nations to say this is how we're going to handle it. And I haven't seen that work out very well in the history of time. Fair enough. Audience, we have time for a couple more questions. So hands up. If we
could go to the man in the white top. Thank you. Hello, panel. My name's Rob. And I'm a business and sports coach. And I'm wondering whether what you've been talking about is possible for helping humans to improve their performance at a sport. Good question. Who's our greatest sport enthusiast? Perhaps this one is for you, David. Can we be using AI to improve sports performance? It's a great question. There's a lot of really amazing research that's happening to try and do just that. You know, one thing which we've been doing at Google DeepMind is actually a
collaboration with Liverpool Football Club to try and help them improve their tactics. So that's one example. I think, you know, the amazing thing about sports is it's become over time so refined in terms of the particular approaches that people take that actually, you know, really they're very open to new discoveries and new ways to do things. And so it's been really fun actually, just watching that kind of thing unfold. Very cool. Thank you, David. I can see we've got lots of young people in the audience, some teenagers even. Does any of our younger audience have
a question? If so stick your hand up. There was a young lady who had a hand up here with a ponytail. Yes. As AI develops, I reckon that humans probably depend more on AI and maybe learn less. Is there anything we could do to maybe make sure that as it develops, humans still develop and learn? Ooh, good question. So will we stop learning, as AI does the learning instead for us? I'm going to put that to all of you if that's all right. What do you think, David? I'd like to imagine a world where we
have an AGI which is like a personal friend, assistant teacher, and everything we do understands what we want to learn, and it knows just how to teach us and help us to learn more and more and more and more. Regina, what do you think? Do you think we're just going to end up relying on AI for everything? As a non-native speaker of English I remember as a young professor I spent a humongous amount of time like reading my papers and making sure, you know, because in my native language we don't have and so I make
a lot of mistakes when I write. I see now how I write papers. Actually, it removes a lot of my pressure in writing and I can really focus on ideas rather than doing the small things. So I hope that we will find a symbiosis. We don't really, you know, remove our basic skills, but we can just do more and focus on the things that we can do better. I would like to imagine if Isaac Newton or Albert Einstein had access to these tools today, I mean, as prolific as they were at a time where we
didn't even have calculators, imagine what they would have done and the impact it would have had on the world if they had access to these tools. In the next five years potentially, you don't need to code. You will just tell your favourite AI system and say I want a code that does x, y, z and you can accomplish what would take years of many people in hours today. So it will make you more prolific. Thank you. I think it's fair to say all three of you have given us hope for the future. Audience, thank you
so much for your questions. I wish we could take more, but I'm afraid we're out of time. That's it for The Engineers - Intelligent Machines at Imperial College, London. I'm Caroline Steel. On behalf of the BBC World Service our partners the Royal Commission for the Exhibition of 1851 and my producer, Charlie Taylor. Please join me in giving a warm round of applause for our brilliant pioneering AI engineers Regina Barzilay, Paolo Pirjanian and David Silver
