Detecting cancer at the DNA level with AI

Detecting cancer at the DNA level with AI
New research shows that cancer could be detected in the very building blocks of life – our DNA, possibly leading to a diagnosis when the disease is in its infancy.  Dr. Shamith Samara-jiwa from Imperial College London is on the show to discuss how AI can be used to detect tiny changes to our DNA called methylation patterns.  Genetic factors play a significant role in the development of cancer, making it essential to analyse disruptions in our DNA for accurate diagnosis. However, identifying specific genes affected by cancer is not a straightforward process. The impact of cancer on our DNA may appear random initially, but by studying numerous human genomes and disease cases, researchers can start to identify patterns. This requires analysing billions of individual data points to determine any significant findings.  

Could AI help tackle the loneliness epidemic
Being lonely doesn't necessarily mean you are truly alone. We are currently facing what some refer to as a 'loneliness pandemic'. According to a new book, artificial intelligence is becoming more adept at providing social support and helping people overcome the feelings of low self-esteem and social isolation that often accompany loneliness. “The Psychology of Artificial Intelligence” has just been published and author Tony Prescott, a professor of cognitive robotics at the University of Sheffield in England joins us on the podcast.   

The programme is presented by Gareth Mitchell and the studio expert is Ghislaine Boddington.  

More on this week's stories:
Early detection and diagnosis of cancer with interpretable machine learning to uncover cancer-specific DNA methylation patterns
The Psychology of Artificial Intelligence

Chapters

• 3:24 Detecting Cancer at the DNA level
• 14:55 AI and loneliness

Transcript

00:00:00 Gareth Mitchell 

Hello this is the Somewhere on Earth podcast. I'm Gareth Mitchell and it is Tuesday the 25th of June 2024 and we're here in London. All our voices are from London today. In fact, we're all around the table, which is rather nice. So none of these problems with lines going down or anything like that. We are right here to bring you this podcast. 

00:00:27 Gareth Mitchell 

And one of us around the table is expert commentator Ghislaine Boddington. Hello, Ghislaine. Again two weeks running. Lucky me. 

00:00:34 Ghislaine Boddington 

I know Gareth lovely. Yes. To see you again. And Yep, I'm with you today to speak to these two amazing live interviews. You're right. It is really lovely. All four of us at the table, yeah. 

00:00:43 Gareth Mitchell 

Yeah. And we're really teasing that, aren't we? There are these two people there, kind of these ghostly presences that we've not introduced yet, but you will hear from them. And they're not ghosts. Don't worry, folks. So, Ghislaine, what? What? What do you have any gossip for us this week? 

00:00:55 Ghislaine Boddington 

Well, I want to do a little push today, a little plug, that's it. A little plug. 

00:00:59 Gareth Mitchell 

All right. 

00:01:00 Ghislaine Boddington 

It's a link to the Stemettes where I'm a trustee for the Stemettes, which is an organisation london-based charity that works all over Britain and sometimes internationally, helping girls and non binary women stay within the science, technology, engineering, maths sector and keeping making it fun making it accessible, making it, you know, the STEM, 

00:01:21 Ghislaine Boddington 

the STEM push so the Stemettes where we've just produced a white paper and this White paper is online, its Stemettes.org White Paper. You can download it and read it and you can add a quote to it and pass it on to people. And we do need help to do this,  

00:01:35 Ghislaine Boddington 

because it's about getting a more diverse female and non binary representation in the GCSE and A level curriculum. It's UK but this is probably all over the world where at the moment these young women, when they're doing science, technology, engineering or maths, yeah, they only learn about one female scientist. 

00:01:56 Ghislaine Boddington 

And they learn about 14 men, male scientists. They have none mathematical, three men. No women in engineering, two men. And none in computer science, and one man. I think it's 20 to one. 

00:02:09 Ghislaine Boddington 

And there's only one woman who is mentioned in the GCSE actually curriculum and that's Marie Curie. Fair enough, but we're trying to get that changed to make a more equitable place for for women to be referenced properly, historically and present day scientists, technologists, engineers and mathematicians. 

00:02:19 Gareth Mitchell 

Right. 

00:02:26 Gareth Mitchell 

Hence the White paper, yeah. 

00:02:28 Ghislaine Boddington 

Hence the White Paper, because we need these role models, we need this in the statutory curriculum, and this probably is the same in many countries. Can't believe this in 2024, can you? 

00:02:39 Gareth Mitchell 

Awesome. Ghislaine, thank you very much for that. Let's jump in. 

00:02:44 Gareth Mitchell 

And coming up today. 

00:02:48 Gareth Mitchell 

There's a definite AI theme this week. Well, when isn't there an AI theme, I guess. But anyway, I have two remarkable stories for you this week about machine learning's different facets. In one case, potentially helping to diagnose cancer, giving doctors better early warnings in this most time critical group of diseases. In the other case, how AI can help tackle loneliness, as much of the world finds itself in a so-called loneliness epidemic. It's all right here on the Somewhere on Earth podcast. 

00:03:24 Gareth Mitchell 

First, then AI and cancer. Now, for one thing, cancer isn't just a single disease. It's a whole family of diseases. And to a greater or lesser extent, cancers have a genetic component. So looking for disruptions to our DNA, our genetic code, well, that seems a logical idea when  we're trying to diagnose cancer. 

00:03:42 Gareth Mitchell 

The trouble is, it's not quite as easy as saying, oh, look, this gene is affected. That must mean it's this kind of a disease. At first sight, then, the way that cancers affect our DNA, it can seem almost random actually. It's only when you look at enough human genomes and enough cases of disease that you begin to notice patterns. And when I say enough, I mean literally billions of individual data points, sifting through them to see which, if any, might be significant. 

00:04:10 Gareth Mitchell 

Now that's an area in which many researchers are rolling out the AI. And Doctor Shamith Samara-jiwa of Imperial College London, is no different. He's the lead author on a new paper on the early detection and diagnosis of cancer through disease specific DNA methylation patterns. Don't panic, folks. We'll explain that to you. What it all means. And it's great to have you in the studio. Welcome along Shamith. 

00:04:31 Shamith Samara-jiwa 

Hi. 

00:04:31 Gareth Mitchell 

So just getting some of the genetics sorted first. In my very simplistic way. I often think of cancer as being to do with mutations. You know, where there's some kind of, you know, you think about the genetic code, those four letters, the the DNA, the A, the G, the C and the T and there's a typo. A mutation, in other words, but when you talk about methylation that's something different, isn't it? 

00:04:54 Shamith Samara-jiwa 

Yeah. So normally people look at genetic mutations like aberrations where you have damage to the chromosomes. But epigenetic changes, where environmental information gets temporarily activate like that, can imprint the cell in a way that modifies gene function so that these enzymes that can methylate the C letter C, the cytosine molecules in DNA,  

00:05:24 Shamith Samara-jiwa 

and it adds a little methyl group onto the molecules. It normally does these additions to regions that have CG pairs and about 80% of the genome can be methylated this way.  

00:05:37 Gareth Mitchell 

OK, OK. Yeah. So I suppose if I can simplify it very wildly, cos I'm gonna move on to the AI in a minute. It's it's almost if you think about mutations as being like a typo, if you like, like letters missing. Can I just put it to you that this methylation effect is not so much a typo, but as if the text has been smudged slightly. 

00:05:58 Shamith Samara-jiwa 

Yeah, it's kind of like temporary white out of letters so that it it's almost like a mutation, but it can, it's reversible so when the environmental messages stop or change, they can come off. 

00:06:01 Gareth Mitchell 

Yeah. OK. Right. Right. So it doesn't change the initial letters in the same way that the mutation might. OK, I I get that. 

00:06:19 Shamith Samara-jiwa 

They get chemically modified. 

00:06:19 Gareth Mitchell 

Yeah. And you can say it can be temporary and when you talk about environmental factors, you just mean what if somebody's smoking for instance, or if it's it's exposure to the sun or any of the other things that we know are cancer risks and this effect can happen. 

00:06:31 Shamith Samara-jiwa 

Yeah, there's lots of these things that can affect the epigenome. 

00:06:35 Gareth Mitchell 

And it's hugely complicated, so let's move swiftly onto the AI, which gets me slightly back into my comfort zone as well. As you can tell folks, I'm no expert in the genetic side of things here, but what even I can understand is that you're embarking on a massive pattern recognition exercise, aren't you? Because these methylations, these disruptions in a way, to the way that the DNA presents itself, they come in patterns, and if you look at enough of them, you can begin to associate a particular pattern, like a signature, as it were, maybe with a particular kind of cancer. Can I go that far? Thank goodness. right. 

00:07:13 Shamith Samara-jiwa 

Yes. 

00:07:13 Ghislaine Boddington 

We're getting. We're getting there. We're getting there, all of us. 

00:07:13 Gareth Mitchell 

We're getting there. And this pattern recognition is incredibly complicated. I don't know if you can, but is there any way you can quantify it? In my introduction, I just fudged it and said there are literally billions of possible patterns. But is that the kind of thing we’re getting into here? 

00:07:28 Shamith Samara-jiwa 

So there are sequencing methods that detect these methylation changes in the antigenome and if you sequence enough tumors, you'll find the patterns that are specific to each cancer type. And the important thing is these changes happen really early in cancer before you see symptoms. And so if you can detect these most cancers can be treated to have a positive outcome. 

00:07:57 Gareth Mitchell 

So the changes are there, and I suppose the the good news story here is in fact it's it's a completely good news story because the the good news initially is that these changes are there. We have the technology to detect them. Maybe the not so good news is it's very difficult to pull insight out from these these changes. But now we're back on to the good news again, with the kind of technology the AI that we have, the power that it has, you're now able to get meaningful insights from these kinds of disruptions to the DNA and cancers. Yeah, early on as well, crucially. 

00:08:28 Shamith Samara-jiwa 

Yeah. That's important, I think, because the earlier you detect these things, the better the outcome for patients. 

00:08:36 Gareth Mitchell 

Of course, it's crucial. It's such a time critical disease or set of diseases. So Ghislaine, what do you make of this, you know this, especially that early detection, it really is essential, isn't it. 

00:08:46 Ghislaine Boddington 

Now that feels incredible. The possibility for early detection. And the possibility, therefore, for people to possibly change the environmental issue and and move backwards like you say for the for it to recover or but probably the wrong word recover. But so the complexity of this means that actually at this point you are working with a large scale computing, huge computing setup. I was wondering about where you can. I know this isn't right now, but where can you see this going in the future into a more mass use? 

00:09:17 Shamith Samara-jiwa 

So once the models are trained, they don't require that much computing. The patient sample, like the [?] stage to prepare and sequence DNA. You require some computation, but it's nothing massive like it's more sequencing centers can handle that kind of, modern hospitals can handle that kind of throughput.  

00:09:43 Ghislaine Boddington 

I was thinking about this in relationship to biomarkers. Is it, does it come under this, this head of biomarker, if for example we were at a stage where I could go to hospital and your system could say yes, you've got the beginnings of your very early prediction. Does that become a biomarker for me that can be recognized and put into my record? 

00:10:05 Shamith Samara-jiwa 

Eventually it can be used in that way, but you'd need to do a lot more work to get to that stage, like explore these methylation changes a lot more. People already use a small number of ways in sequencing panels to diagnose cancer. What we've done is we've cataloged the entire, all the methylation changes genome wide, not just single chromosome, not single marker or single gene, but all the changes in each cancer type genome wide. So I think that gives us more information in the future to come up with like proper biomarkers. 

00:10:46 Gareth Mitchell 

Yeah. So and and that's crucial, isn't it? Because with the AI, I suppose it's tempting to focus on the algorithm, but what you're saying is yes, but we need the data as well. And the big part of what you've done here is getting the data. 

00:10:54 Shamith Samara-jiwa 

We didn't generate all the data sets. We used a large cancer genome studies from various consortia and we we generate from other studies. But the majority of data was out there. 

00:11:09 Gareth Mitchell 

Also, a really big important thing that you're doing here that I was struck by in your paper is that you're looking at the inner workings of the algorithm itself. Because we so often hear that AI is great, but it's so often a black box, isn't it? We don't really have any idea of how it's going from A and ending up with B or C or goodness knows where it's going to end up. Tell me a bit about then, this opening up the algorithm so you can see what it’s doing. 

00:11:32 Shamith Samara-jiwa 

We haven't just any specialist techniques, right? We are keen on using methods that were interpretable and explainable. So that the features or the methylation positions it figures out can be used to understand the biology and so the method we used is called XGBoost model like gradient boost, boosted tree classification model and it identified all the important methylation features in each cancer type and then we fed that to a deep neural network, which improved the classification. So it's very simple. It's not, nothing complicated. 

00:12:15 Gareth Mitchell 

Yeah, I I love the way you said. Oh, it's really simple. And and I'm. I'm I'm sure it is the the the essence of it. But you know, just the way you're putting it there as well. We'll we'll stay with you, Shamith. But I do want to bring in Tony Prescott. He's our next guest. He's going to come in and talk about his book, but rather handily, Tony, you're a professor of cognitive robotics at the University of Sheffield in England. 

00:12:35 Gareth Mitchell 

And even more handily, you've just written a book about the psychology of AI. So you know a bit about this stuff and I, look, I know you haven't read Shamith's paper and you're not across this research, but what are you picking up in general just about this, the whole conversation, for instance around the algorithms being able to look inside the black box and understand more about it and all the other stuff you've been hearing. 

00:12:53 Tony Prescott 

Yeah. Good evening, everybody. And there's a fantastic study and really quite exciting and I think the the power of AI to unlock this kind of data. So you've got this huge mountain of medical data of which this is just a tiny fraction and just by using AI we can extract signal out of what's probably an extremely noisy data set. So you know in the old days the PhD student might have poured over this data for three years and come across one or two things. And with this technology you can now you know in an afternoon, find out you know, once you've set it up, you run the algorithm and it, you know, it goes across the whole genome. It's super powerful. So it's really gonna transform how we do medicine and how we do diagnosis. And it just really shows the power for good of artificial intelligence. Shamith. You were nodding there. 

00:13:51 Gareth Mitchell  

Sharmith, you were nodding there. 

00:13:51 Shamith Samara-jiwa 

No, I agree. 

Yeah, OK, Ghislaine, a final thought for you on this. 

00:13:54 Ghislaine Boddington 

Yeah. No, I I agree too. And I think what I'm fascinated by, what you said Shamith about working with the consortium. Yeah. And everybody pooling. And it's like a crowd science scenario. 

00:14:06 Ghislaine Boddington 

I think that I've only just begun to realize the last couple of years how much that's happening with biomarkers, with the biobanks across the world where you know there's a lot of them in many different countries and they're and I realise now that they are sharing and working together, and sharing of other scientists and it's all been poured over in this  kind of crazy huge huge pool where actually yes, things are moving very fast and that's for the good of us all. It really is AI for good, yes. 

00:14:34 Gareth Mitchell 

Well, there, we'll leave it at Shamith for now, but hopefully you will stay back for the the the podcast extra for our subscribers. So Shamith can can sit and relax now and listen to us talk about about Tony’s books, but thank you very much for coming in and telling us what you have done so far. And so that's Doctor Shamith Samara-jiwa of Imperial College.  

00:14:52 Gareth Mitchell 

So we're going to talk a little bit now about loneliness with Tony because I suppose ironically, in a way, one thing you can say for yourself if you are lonely, is that you're actually far from alone. In fact, we are in the midst of a loneliness pandemic, as some people have put it, but Tony’s argues in his new book that artificial intelligence is increasingly capable of supporting us with social interaction and breaking that cycle of low self esteem and social isolation that often comes with loneliness. So those revelations and more come in the forthcoming book The Psychology of Artificial Intelligence. 

00:15:30 Gareth Mitchell 

So shall we pick up on that, Tony, just on loneliness and what you have to say about it in the book and indeed, the way that AI can help that, that cycle that does have for so many people, that it drags them down into loneliness. 

00:15:44 Tony Prescott 

Yeah. 

00:15:45 Tony Prescott 

I mean the the strange thing about loneliness is we think of it as a psychological phenomenon, but it's also a real health issue it's, it's you know, we are mammals and we are social animals and to be isolated actually impacts our health. So I think the World Health Organization has described loneliness as being as bad for your health as smoking 15 cigarettes a day. 

00:16:09 Tony Prescott 

And if you look in the UK, nearly half of us describe ourselves as sometimes lonely, and I think 7%, which is nearly 4 million people, say they're lonely all the time. So for those people, there's a really negative health impact of this social isolation. So it's not like being hungry or thirsty and that you know, if you're hungry, you eat. And if you don't eat, you would starve to death.  With loneliness there is a kind of biological need for a company. If you don't have it, you won't thrive, but you'll you'll get by and the question is, are there things that we can be doing? Obviously we could do what we should do and can do to scaffold people with human companionship. So what extra could we do with AI? And that's what I talk about in the book. 

00:16:57 Shamith Samara-jiwa 

Hmm. 

00:17:01 Gareth Mitchell 

Yeah. Would this be chat bots for instance, or are there lots of other ways that AI can help? 

00:17:07 Tony Prescott 

Yeah, chat bots is a good example, and people, and obviously with the latest large language models, chat bots have really come a long way because it used to be that they would just parrot back what you just said to them and they wouldn't have any memory. But the chat bots now can remember the whole conversation or back to a previous conversation and they can offer advice. 

00:17:26 Tony Prescott 

And these large language models have read hundreds of thousands of books, including quite a lot of books about loneliness and therapy and help, so they actually could give you some psychological support, but they don't have to be doing that necessarily to be helpful,  because it's just a lot of people just the social companionship of something that listens and responds. 

00:17:49 Gareth Mitchell 

Yeah, but my mum has a virtual assistant. In fact, I'm not on the BBC anymore, so I can say it's Alexa, but other brands are available, and she she loves Alexa, youunow, she really does. And I I get a bit creeped out sometimes by all these voice assistants and about, you know, the data and what have you. 

00:18:04 Gareth Mitchell 

But Mum's not gonna have a thing, you know, she lives on her own and she says, you know, if say, you know, she's unplugged Alexa to take her upstairs, which I think is also rather endearing because sometimes she prefers Alexa to be in the bedroom to wake her up in the morning. Sometimes she wants her downstairs. So anyway. But, you know, if she's in a room where Alexa isn't, she kind of misses her, you know? So there's there's something in it you know. 

00:18:25 Tony Prescott 

There is this myth that older people don't like technology, but I think it's completely untrue. Yeah, the number of old people that we talked to are really excited about this technology. And we might say a bit about robots, but I think you know there is a real interest from people that live alone in having social robots, maybe ones that are more pet-like, but could have, you could have a conversation with. 

00:18:47 Gareth Mitchell 

Yeah. There you go, Ania says in my headphones, says her mum loves Alexa, switches the lights on and off. My mom likes that, lounge off. Yes. 

00:18:53 Ghislaine Boddington 

I think it's seriously well, you know, these voice, voice editors are very helpful in senior citizens homes and I know definitely and also obviously Facebook and you know and video conferencing, FaceTime or video catch ups on families has been really, really important. 

00:19:11 Ghislaine Boddington 

Yeah. I think one of the things that worries me though is that you were mentioning about the feedback from the chat bots around say, more therapy based feedback. I do know a number of colleagues who are therapists in various different ways who are absolutely freaking out about this harvesting of multiple. 

00:19:26 Gareth Mitchell 

Yeah, but isn't that cause they're worried about their jobs, though? 

00:19:28 Ghislaine Boddington 

No, I think it's about the the the mass harvesting of multiple articles and books, say on loneliness. Yeah. To give this kind of slightly, no, what is it? AI feedback, which isn't necessarily as sound as being with another human. Yeah, so. 

00:19:41 Tony Prescott 

Yeah. I mean, there's there's benefits and risks and there's obviously things like intellectual property rights. I'm kind of more on the side of benefits, I understand the risks. 

00:19:52 Tony Prescott 

But another way to look at it, not to be critical of your therapy friends, is that this is in a way democratizing access to that knowledge, a lot of it, which was produced by universities. And there are some books published, you know, that were copyright, but there's a lot of stuff that scientists like me put out into the public domain and these have been harvested by AI and so yes, I think maybe we should be more careful about creating the data sets that are used to train the large language models, but that doesn't mean to say that we should throw the baby out with the bath water and say the the idea is not good. 

00:20:25 Ghislaine Boddington 

No, not at all. I just think it's not so much about the harvesting I'm talking, it's more about the kind of way it's put together and comes out and whether it's the right advice, yes, so. 

00:20:33 Tony Prescott 

Well, I think it's about doing things which are complementary. So I think that what you don't get is access to your therapist at 1 o’clock in the morning. They probably won't be very happy if you phone them up and say, you know, I'm desperate for a chat, whereas the AI is always there, happy to listen to your story, even if you've told them the same story multiple times and so I think there are different uses. 

00:20:56 Tony Prescott 

And I I wouldn't want to replace the human therapist and I think absolutely human contact is super important and and one of the ideas is that with this artificial companion, you get the social scaffolding. So the loop you were talking about, the spiral downwards of self esteem, which causes people that are feeling lonely and social isolated,  

00:21:17 Tony Prescott 

it makes them feel I can't talk to anybody. And so when there's an opportunity for social contact, they let it pass by because they've got zero self-confidence. So perhaps with the AI chat bot you could practice your social skills, scaffold them a bit. It could give you some positive feedback, help you feel better about yourself, and that might help you then build those human contacts. 

00:21:39 Ghislaine Boddington 

Although we do have now AI girlfriends where millions of men just have these virtual Asian girlfriends. 

00:21:45 Tony Prescott 

And that's another of the risks. You know that we 

00:21:47 Ghislaine Boddington 

So that's a de-socialisation. Yeah, because they they can talk to AI girlfriend, but they can't talk to a real girl. 

00:21:54 Tony Prescott 

I I agree, no, it's there's . There's definitely challenges. I'm not sort of saying let’s go down the route of 

00:21:47 Ghislaine Boddington 

No, no, no, that's 

00:21:57 Gareth Mitchell 

In some cases. 

00:22:01 Gareth Mitchell 

Exactly. But well, you’ve brought Gordon with you haven’t you.  

00:22:03 Tony Prescott 

Yeah. So Gordon is an animal-like robot. He’s part of a a species of robot if you like, which we call Miro. And it looks a bit like, people say it's a puppy or it's a rabbit, somewhere in between the two. It's got a tail, it's got ears, it's got eyes which are cameras, ears which are microphones. It can move around and look at you. This version, with its current software doesn't talk, but it does make animal-like noises. 

00:22:30 Gareth Mitchell 

Oh well, this is radio. Can we give that a try. Miro? 

00:22:32 Tony Prescott 

And we can turn our Gordon on and see what he does.  

00:22:32 Gareth Mitchell 

Gordon, sit. 

00:22:32 Tony Prescott 

So he is fully autonomous. So I'm going to be starting him via Bluetooth controller on my phone. But then what he does after that is up to him or up to it. I should say, it's not gendered. So hopefully, here we go. So you probably can hear the motors. 

00:23:00 Tony Prescott 

So you could probably hit Gordon making animal like sounds. He has a sound generator which is modelled on the mammalian larynx. So he makes the right sounds, kinds of sounds for an animal of this size. So this was built by a colleague of mine at Shepherd called Roger Moore. It's a simulation of the larynx which is generating that sound. 

00:23:15 Ghislaine Boddington 

Fantastic. Yeah. 

00:23:23 Gareth Mitchell 

OK. Yeah, I was interested in the sounds. Yeah, because it's it sounds almost human, like or childlike. The sound rather than. And I'm not dissing it, you know, but it's. But you're saying the point is that the the range of the sound is what you'd expect for, like, physiological entity of this size. 

00:23:40 Tony Prescott 

Yeah. We wanted to make a robot that was animal like but wasn't any specific animal because people have expectations about puppies and rabbits and so on. 

00:23:46 

Ah yeah. 

00:23:48 Tony Prescott 

So we wanted people to respond to it and Ghislaine is. 

00:23:51 Ghislaine Boddington 

Come here then. 

00:23:52 Tony Prescott 

So he's actually able to move, but I've turned off his motors so he won't drive off the table. 

00:23:54 Ghislaine Boddington 

Yeah, you're right. But he's looking at me and he's he knows I'm here. He's come round to look at me. 

00:23:59 Gareth Mitchell 

Yeah, that is. It was really quite striking. Yes, Ghislaine talks to Gordon. Ohh. Now I'm talking now. Gordon's listening to me and actually pivoting his ears up towards me. And now looking a bit taken back probably cause I'm kind of talking about Gordon, but not really to him and. 

00:24:14 Tony Prescott 

He’s totally not listening to what you're saying.  

00:24:17 Gareth Mitchell 

Well, welcome to my life. 

00:24:19 Tony Prescott 

He's paying attention to anything that makes sound or moves visually, so within his visual field. So if he was able to spin around the table, but I've disabled that so he can't, then he would now look at me. But if you start talking again, he'll look at you again.  

00:24:31 Gareth Mitchell 

OK, right. So if I start talking. Yeah, sure. And if you can see. Yeah. He's peering at me through one of his eyes. And now the other, how endearing. 

00:24:37 Tony Prescott 

He has tactile sensors.   

00:24:45 Gareth Mitchell 

Ohh right. So you're kind of stroking Gordon now, yeah. 

00:24:47 Tony Prescott 

Yes, I’m stroking him. And you can see he's got lights in his body. 

00:24:49 Gareth Mitchell 

Yes, there are. There's he's got a kind of translucent body, hasn't he? And then through that body we can see lights. 

00:24:52 Tony Prescott 

He has and you can see you can see these stripes, which are kind of like ribs. Yes. And those are the touch sensors. So he's detecting, and Ghislaine's also stroking him. 

00:25:00 Gareth Mitchell 

Oh. 

00:25:02 Gareth Mitchell 

He's detecting you and and the point being as as you as a I know you're a roboticist scientist, but as a user, say a child for instance, they they're getting feedback as well because they can see Gordon responding not just to his movements and his eye gestures, but also what's going on in his body. 

00:25:16 Ghislaine Boddington  

He's wiggling his tail now. 

00:25:19 Tony Prescott 

And his lights have gone green and he's wiggling his tail. So. So you get positive feedback so. 

00:25:22 Ghislaine Boddington 

Yeah, he's very happy. 

00:25:28 Tony Prescott 

We've done a lot of outreach with children and there have been some studies with children and it has an effect of calming people. 

00:25:38 Tony Prescott 

As do some other robots like this. So you spend a few minutes interacting with with Gordon and children. When we've done this at public events, children that are really quite hyper because you know they're an exciting place have come and they've often sat down and and cuddled the robot for 10, 15 minutes. And it's had a calming effect which has really pleased the parents. 

00:25:57 Gareth Mitchell 

Yeah. Right. 

00:25:59 Tony Prescott 

And so that's the basis for thinking it might be useful for a kind of animal assisted therapy. At the university, where we have a lot of students with social anxiety, one of the ideas is that they would come to the counseling services for therapy and they'd meet a therapist. But also this would be complementary, you know, and perhaps if they were too nervous and shy to talk to a therapist, they could still get a positive experience by meeting something very non-threatening like this robot. 

00:26:27 Ghislaine Boddington 

You just wanna take it? Take it home, though, wouldn't you? Yeah. Can I have it please, to take home so that I don't have any, so much loneliness. 

00:26:36 Gareth Mitchell 

No, I I can see, I can see the function of it and the the serious research purpose behind it. So one level. Yeah, really cute. You know, good fun in a podcast studio, but really the basis of some important research. So we've only got about 3 or 4 minutes to do the whole of your book and it's quite a long book and very deep it goes too. But it is The Psychology of AI. So let's do the basic one first, then. Why did you want to, cause it's such a massive topic to take on, the psychology of AI. So why did you want to write the book? 

00:27:04 Tony Prescott 

My background is I'm a psychologist trained in psychology originally. Then I did AI, and then in the last 20 years I've done robots, but really using robots to understand minds. So I write programs that go inside the robot and control it. So this one is controlled by a very simple model of the brain, the core theme of the book is what is intelligence and what is human intelligence. And what would it mean to have a machine that was intelligent and have we got there yet? So that's what the book is about for most of most of the seven chapters. 

00:27:36 Gareth Mitchell 

Yeah, because you finish with artificial general intelligence, don't you? Which is this move to something that we might think is well, you tell me. Is it intelligent? It's I don't to spoil the book too much, but that is 

00:27:48 Tony Prescott 

My answer is that AI is intelligent, but it's not intelligent in all the ways that humans are, and there are aspects of intelligence that AI hasn't captured yet. 

00:27:59 Gareth Mitchell 

Because one brilliant example that you give I think really helped make me understand that argument. As you say, that LLM for instance, you can have a conversation with a large language model chat bot about the Arctic Monkeys, for instance, and this thing, and it will be actually a meaningful intelligent in inverted commas conversation with a chat bot AI about this band. And yet it's never listened to the Arctic Monkeys. It's incapable of engaging with it musically. And I thought that was a very good example to say what it can do, but what it can’t do. 

00:28:29 Tony Prescott 

And there's actually 2 two sides to understanding, and I think this is where people get confused. So when we understand language, we understand the relationship between the words. And that's what large language models do, because they've absorbed all this literature and, you know, millions of web pages. Ii’s pattern recognition again. It is, yeah, but, and and what's remarkable and surprised many people, including me, is how effectively these large language models have built a model for the world, if you like, in order to understand language because you don't understand language 

00:29:00 Tony Prescott 

the way that they do without understanding something about the world. So they've had to build a model and your models are the same. 

00:29:07 Gareth Mitchell 

And you're talking back to Shamith, by the way, not my models. Yeah. 

00:29:13 Tony Prescott 

In order to explain the patterns of cancer, they've had to understand something about DNA and cells and when you dig into that neural network, you will find there's some understanding in there. When you say pattern recognition that's kind of making it sound quite simple. There's a really sophisticated analysis that is going on inside that deep network and your own brain is a whole bunch of deep networks that are networked together.  

00:29:36 Tony Prescott 

So the book is also about what a brain is and a brain is, you know, probably 200,000 deep networks. That's what your cortex is that are all talking together and networking. So we are getting close in our technology to building something that's very brain like. 

00:29:50 Gareth Mitchell 

Wow, Ghislaine, you have loads of thoughts on this, don't you? This whole area. 

00:29:54 Ghislaine Boddington 

No, I think it's completely fascinating. And I'm really interested in how you've been working. I know you're working a number of years, yeah, with the robotic body. The robot. Yeah, as a as a, as a representation of, in a way, our brains and ourselves. 

00:30:10 Ghislaine Boddington 

And as you know, my work is more with the Avatar body and you know, but they're both digital interfaces of different types, so where we can reflect ourselves and work with and get mirrors back. So we are at a very interesting point with the AI assistant, with the robotic assistant, etcetera. 

00:30:28 Ghislaine Boddington 

Coming into our mundane lives. Yeah. Making maybe a bit more room for creative thinking for us, getting rid of some of the more robotic things that we have to do ourselves. Yeah. And opening up space I hope for a bit more critical reflection and and solution creation. 

00:30:47 Ghislaine Boddington 

By working alongside and with with robots and with avatars etcetera and AI.  The plus of it to enable our lives to be better, to actually sort out some of those issues like the loneliness issue that you're you're dealing with, so so clearly on this, yeah. 

00:31:04 Tony Prescott 

Yeah, I think, I mean, I absolutely agree with you about using these technologies as a kind of mirror to ourselves. So they kind of reflect back at us something about ourselves. So  saying about what large language models understand. They understand the relationship between words but don't understand the relationship of words to the world. 

00:31:21 Tony Prescott 

So these are all just patterns. With a robot you can start to relate words to the world so you can start, the robot can start to understand what a chair is, what a table is. You know what, what sort of things you can do with the ball. You could push it, you could move it. So AI's don't yet have that knowledge, but they could do through robots. 

00:31:43 Tony Prescott 

So it's that grounding of your language in the world. That's the difference between a large language model and and even a small child. They might, a child might have a vocabulary of of 200 words, but they know what those words mean. You know, they know what? You know what cup is. Yeah. And they know you can grasp a cup you can drink from it. And so on. A large language model can talk all about cups and grasping and drinking, but it has no knowledge of what that means? 

00:32:07 Ghislaine Boddington 

I think that’s really a great explanation actually.  

00:32:11 Gareth Mitchell 

Is it basically impossible for an AI to have anything like what we might consider to be intelligence as long as it is unembodied? If it doesn't have sensors and other ways of interacting with the world, and I love the way you mentioned the child at the end, he might only have a vocabulary of 200 words, but that child exists in the world, and that's the difference, is it? 

00:32:31 Tony Prescott 

Yeah. It's difficult to know with AI because of course you can pass knowledge from one AI to another effortlessly and seamlessly, whereas we all have to discover the world ourselves for the first time. And until we get language, you know, we're we're on our own in terms of understanding the world. And then our parents can start to help us by teaching us, but with AI's, you could pass the knowledge on. So it may not be that you need that much robotic embodied embodiment before the AI's will have this kind of understanding of the physical world. We’ve still to find out. 

00:32:59 Ghislaine Boddington 

Yes, that's to come? Yeah, it's exciting time, actually. 

00:33:04 Gareth Mitchell 

And also to come for those of you who subscribe, we will carry on talking about this in the podcast subscription extra. If you're not a subscriber, I know it's really annoying that you think that you're being left out and so on, but don't worry, you've got a fair crack of it in this version of the podcast. It's just that we give you just that little bit more in the subscription extra as well. But thank you very much to you, Tony for coming in and to you Shamith. We’ll talk to you on the other side, as we say here in podcasting, and you, Ghislaine as well. Two weeks running, lucky us. Good on you. 

00:33:30 Ghislaine Boddington 

Yes, see you soon. 

00:33:32 Gareth Mitchell 

See you soon. Thank you very much. The sound and audio today is from Dylan Burton here at Lanson's Team Farner. The production manager is Liz Tuohy, the producer and editor is Ania Lichtarowicz. And I'm some kind of embodied humanoid, not particularly intelligent at the best of times, called Gareth. Thanks for listening. Bye, bye.