Current AI innovations reshaping how people interact with language, says Rohan Murty

Thank you so much. We request everybody to please be seated for the next session. Up next, we are honoured to welcome two distinguished guests today, Mr. Rohan Murthy and Professor Chaz Bontra, who will be engaging in a fireside chat on the topic Harnessing AI for Sustainable Development in India. Mr. Rohan Murthy is the founder and CEO of Sorocco, a company specialising in process discovery and automation. Mr. Murthy holds a PhD from Harvard, ABS from Cornell and was a Fellow at MIT. Remarkably, during his time at Harvard he was selected as a Junior Fellow, becoming only the 2nd computer scientist in the program's 80 year history, following the legendary Marvin Minsky, the father of Artificial Intelligence. Moderating this session is Professor Chaz Bontra, the Pro Vice Chancellor for Innovation at Oxford. Professor Chaz is a leading figure in drug discovery and innovation, recognised as one of the top innovators in industry and honoured with the Order of the British Empire. Please join me in giving a very warm welcome to Mr. Rohan Murthy and Professor Chaz Bontra to the stage. Where do you want me to sit? Or Rohan, why don't you tell us a bit about yourself? I heard a little bit about your background, but tell us a bit about it. That almost feels like an into your question. OK, I'm I'm an academic turned entrepreneur. I come from a family where we are teachers for seven generations. And in my generation, it was the last hope, the last Jedi, and everyone thought I'd be a professor. I gave up academia to start a tech company. We are based primarily out of Boston, but also between India and Singapore and now London. And so I was in the UK and Siddharth invited me saying come to Oxford, to the Business School and speak. And so here I am. So tell me what? What do you worry about in terms of AI sustainability, India? What keeps you awake at night? So actually before that, maybe if if it's OK, let me take a slight detour in answering that right. Why is there so much excitement around AI this time around? Because if you sort of see trends in technology, it's sort of always in waves. And the 1st wave may fail, the 2nd wave may fail, finally in the 3rd wave really lands. In the 80s, there was the great hope of AI and everybody thought that was the time when AI was really going to deliver on all the promises that Alan during and many others had dreamt of for computing, and it did not. And so when I was an undergraduate student, which was early 2000s, it was not very fashionable to be working on AI those days. And the difference this time around perhaps is that 10 years ago, AI was very much used, but it was used in ways that maybe, perhaps many of us were unaware of. So if you were transacting with your bank, it was being used in the back end to find anomalies, to detect malicious, seemingly malicious transactions and so on. So it doesn't, it didn't really directly affect you on a regular basis, but it was there somewhere lurking in the background. But this time around, in this current wave, I think it fundamentally changes the interface between humans and machines. And that's why everybody's grandmother here will know about what or would have heard of GPT may not know what it is, it doesn't matter, but would have heard of it because for the first time it it's doing something that fundamentally alters how humans and machines have a relationship. And I think that's, that's very important to understand why it's much more real this time around than in any of its previous avatars. And so having said that, you know, that's sort of the first thing. And the second thing is how does it change this relationship between people and machines? Fundamentally, it changes this, it, it's some of these advances that have happened in the recent past. Have, you know, made it? I'll use some words fairly loosely, but they are, I don't mean to be precise, sort of, if we now have these algorithms can that can understand relationships between large pieces of text, can read Faisal Devji's book and pretend to summarise it as if an intelligent person has read the book. But it's not really the case. There's some very complex relationships between words that have been deciphered and so on. But to cut a much longer story short, I think what the current avatar of what AI does is it redefines humanity's relationship with words. And that is actually an extremely fundamental point. We have not had an invention or an innovation in computing that is being this transformative in terms of changing how we think of words or our connection to words. Now, why is that important? Because all of recorded, well, most of recorded history, not all is in words, right? Our histories are written in words. Our literature is in words. Our communication is big part of it is in words and so on and so forth. And so if that, if that is the case, then then I think it the future sort of that the current wave portends is one where you start to ask things like what is real, what is fake? Turns out you can extend this argument to not just words, but to audio, to video. And so then the questions become even more urgent and pressing. What is real? What is not real? What is synthetic? What does it mean to be synthetic? I don't know. How do you take exams? What is original piece of work? I don't know. I I feel like at least in my own case, I don't have a very good answer for these kinds of questions. Seeing all of the recent advances. I thought I knew, but I don't anymore. In the earlier conversation, Miles was saying in talking about carbon dioxide capture and, you know, and I agree with him. I mean, I think when we have an emergency, like the climate emergency, we need to throw everything at it. You know, the kitchen sink. We need to try everything. Green energy, biodiversity, carbon capture, everything. We should be doing it just like in the pandemic, You know, nobody knew which vaccine was going to work. Many of my colleagues said we won't get a vaccine for COVID-19. And fortunately, we tried lots of things and we had a number of successes, but it was we were lucky. So I think we need to do exactly the same. I think we're not good at predicting the future. You know, I mentioned earlier these trillion dollar companies. I don't think there's anybody in this room who could have predicted NVIDIA would today be worth more than 3 trillion or Apple would be, or a book, a company that started off delivering books through letterboxes. Amazon would be worth nearly 2 trillion. We're not good at it. So the tough question for you Rohan is what's this world going to look like even in 10 years, 20 years? Because we all talk about AI, it's going to impact everything. And we we can talk about the sustainable development goals, etcetera. You know, we can talk about the impact of AI and each of those etcetera. But what's your where do you think we're going to be in 1020 years time? I have no idea. In fact, I'll, I'll tell you in the confines of my own, our own company, OK, forget the world. I mean, this is a company that I started. I continue to be shocked by what I see that our teams are doing because we've built a foundation model of a certain kind. Our model basically today behaves like it's a McKinsey consultant. Any of you are going to be analysts or associate McKinsey, our model will do whatever you're doing. I didn't think that would happen three years ago. So I, I mean, if I can't foresee what we're doing within the confines of a young organization, I don't know where we'll be 10 or 20 years from now. It's exactly because of what you said. In fact, I'm sure we have lots of theories and hypotheses, but the pace of development, as an example, two years ago, I think the world had at least publicly released open source models, something like about 6 or 7000. Today there are over 50,000 and some of these models can do really, really dramatic things. I listened to a so my uncle is a professor of astrophysics at Caltech. I listened to his talk recently and he was talking about how his group are using AI models to completely automate discovery of the night sky without a human in the loop. And this AI model is even right, is doing the discovery and then writing these papers. And if you really think about it, the one of the oldest activities of humankind is to look up at the stars and wonder about our place in the universe. And we're, and we now say that we have models that can pretty much do most of these activities. So it's, I think the pace of change, particularly in this, this area is so high. At least I, I may not have the intelligence to say what will, what will it be five years or 10 years from now? Rohan, I don't want to embarrass you. You're one of the smartest people in this room and on the planet. But sort of, you know, if you can't give us advice on this because the purpose of this discussion today is, you know, I could make a case for the importance of AI in healthcare, energy, climate, food production, water, reducing pollution in air, land, sea, everything. The things I worry about with respect to AI is the energy consumption of these data centres and where that energy is going to come from. And I also worry about increased inequality. You know, at the moment we've got massive educational inequality, income inequality, gender inequality, but there's a third of people on the planet who do not know how to use the Internet, etcetera. So potentially the inequalities. So what do you worry about? I mean, I, I'm afraid my brain works. I always think about problems and then I think about solutions. So give me some problems to think about South, a couple of things. See, see for, for all the things that you just mentioned, right? Energy is a very legitimate 1 legitimate concern because so far, the only the large successes we have seen have been models that require tremendous amounts of compute power to train in particular. And inference is also expensive, but there's still hope because you're seeing, for example, a very young company with only 40 people, I think as a few days ago valued at 6 billion sitting in Paris, who have developed mitral, for those of you who don't know, developed models that are quite competitive with these large models, and yet their energy footprint is much smaller even in their training. In fact, that has been the entire value proposition to say, look, you don't need to be energy quote UN quote inefficient. You can actually do it with far less resources. You can be far better. So then there may be some hope when it comes to inequality of people not having asymmetric access to the Internet and so on. Well, if on device AI models do succeed, which I think they will. So you'll not only going to have these large models in the cloud consuming tremendous amounts of power, you'll have models on phones and so on which don't require Internet connectivity and which can be multilingual. I actually think perhaps our best chance of levelling the, the the inequality will be these very helpful agents on these devices, these AI agents. So I tend to look at all of these and think I don't see the future as being full of problems. I see it as being full of possibilities for each one of these things. And that doesn't mean that the concerns today are not legitimate, but I also see green shoots that suggest that maybe the future will be different. So I don't worry about it in that sense about some of these because to me, these are a matter of time before they get resolved. I mean, I do agree with that. I mean, I, I think where I get my optimism from is sort of, you know, I'm fortunate in this great university. I'm surrounded by many geniuses. Many of them are very young and some of them are sitting in this room and some of them are even younger than yourself, Rowan. And they are so smart and passionate and creative and entrepreneurial, etcetera, etcetera. And that's where I get my hope. I completely find that. So let me try and prod this another way. So one of the other reasons for this meeting is sort of, you know, I always think working in Oxford is an absolute privilege. We've got brilliant people, we've got great brand name, we've got alumni all over the world, we've got wonderful networks, we've got geniuses like Miles working on all sorts of problems, etcetera, etcetera. We've got a brilliant Business School. There's nothing we cannot do. All the problems I talked about at the start, we've got colleagues in this university working in all those areas, etcetera. Now the purpose of this meeting is to think about sort of how can we in Oxford work with our friends in India to do something that's going to be step changing, transformative, etcetera. You know, I would like out of this meeting, I don't just want us to go home and you know, say, oh, we had a great time chatting with Rohan, etcetera. I want us to go home and think about there's one or two things we're going to follow up on. So what are those one or two things we could do? So let me first say something about India and then maybe tie it back to Oxford. In my limited estimation, I don't yet know if we have we have the talent in India, but we don't have the training in India to be able to innovate on models. At least this is what I find when I recruit people in India. America has the largest talent pool. I mean, this is nascent talent, but it's it's exceedingly complex, this work. It's not easy. It's not for the faint hearted. I suspect the Uki mean, I'm not, I suspect I know for a fact the UK does, some parts of Europe do, but I don't think we have the requisite training to be able to, you know, say, OK, we'll we'll iterate and build the next version of the model, the next more energy efficient model as an example in India. Now, of course there's a separate debate. Do you need to do this or not? I mean, we'll put that aside just for a moment, OK. So if you can't do that, then what can you do? What you can do is you can take any of these open source models and you can train it on a whole bunch of data. And that's how maybe you can build some use cases, you can build some competitive advantage etcetera and so on. And so you either need trained talent as your secret sauce and or certainly you need data. Now data we have plenty of in India. And I know that India also has data protection laws and so on, which in this case work to it's favour. We have data from landed records to health records, of which we don't have a standardization, but nonetheless, it's a great opportunity to standardize to financial data to pretty much everything else. We have data at large scale and large volumes. And it turns out that is exactly of the kind or nature of what you need for these models to do quite well. They need a lot of data and so on. And so sort of my perspective on this has been to build any kind of a play or advantage or of any kind in India with respect to AI, It has to revolve around data. It's not yet a talent play, not in India, not yet. Again, it doesn't mean I'm not saying I don't want someone to think, hey, Indians can't do it. Obviously, that's not what I'm saying. It's quite to the contrary. If you actually see the revolutionary paper in 2017 published from Google Brain, I think the two lead authors were from India, but sitting in London. So I think what India can do, at least in the meantime, is to invest very heavily in curating these data sets, cleaning these data sets, training models on these data sets, and then discovering use cases of what does a train model on these data sets, What can it do? Let me try and have another go at this. I mean, I, I, I feel like I, you have, I'm resisting something that you wish to get to it. This is good. And then I want to open it up to questions, etcetera. So our Vice chancellor is a lady named Irene Tracy Praveen knows him well and Miles, you know her a little bit well, I think sort of absolutely brilliant lady. She did this review for our Chancellor, Jeremy Hunt, which published at the end of December. It was on sort of innovation, entrepreneurship in universities and spin outs, etcetera, etcetera. And then a couple of months ago there was a a House of Lords Select Committee and unfortunately Irene couldn't go so Muggins had to go. But sort of I was asked the following question. What is the difference between the US and the UK? And I said the difference is money and culture, you know, and, and by let me just explore that culture bit. So in my sense in the US they have more of a moon shot culture. We're going to land somebody on the moon, we're going to cure cancer, etcetera, etcetera. We don't have that in UK, Europe, you know, somebody comes up with a big problem here, immediately 10 people jump on that individual and say, Oh no, no, you won't do that. It's too risky, you won't get the funding for it, etcetera. But I, I think the, the other thing I like about the US is they're risk taking, they're confident, they're ambitious. And I said earlier that the reason I'm optimistic about these problems I worry about is young entrepreneurs like the people sitting in this room and say they're going to solve them. People like Rishi over there, he's going to solve them, etcetera. People like you. That's why I'm optimistic. But sort of, so if there was one ambitious goal that we could do together, what would it be right when you see we could do together? You mean Oxford in India or the UK with well, UK and India, I don't mind. I don't care. I mean, I just, I just care about bringing scale and urgency to some of these problems and bringing the right people to those problems. So I actually think the UK has lots of right people, trained talent in AI and India has all the data. And I think the mix of the two is, is where the intersection really lies. We have to find use cases now. OK, so you have talent, you have data, OK, what can you do with the two? And you have to find use cases for these. And it can be across the board. It can be in, I mean, pretty much for example, I know there are lots of folks in the Valley who work on on health training, health training models on health data. But we have far more health data in India except you know it's not yet aggregated, not yet cleaned etcetera and so on. But with the right kind of talent, partnering with the right kind of data curation in India, I think something meaningful can be done when it comes to health, it comes to detection of of various problems. I mean that's just one of far too many examples like this. Whenever you go to register a house in India, it's, it's not a straightforward process in terms of it's not digitised and so on. And so we have lots of land records, historical archives and land records, none of which or rather all of which are prime fodder for AI models because you often have to find the lineage for is this property really owned by the person selling it? And you often don't know. And it turns out after you buy years later, you're stuck in lawsuits and so on and so forth, all of which can actually be really dealt with with, with the appropriate data being trained as the right kinds of models or the, or the right kinds of iterations of models. So I always think of India as large data source and you bring the right set of people in and you can, you can solve some problems. So Praveen AI Healthcare data India, that's your next job. OK, so let's open it up to questions, please. Oh my goodness, over there, that young lady there. Could we get a microphone please? And can we make the questions brief? Arun, this is what you said about talent and training. And my question is how easy it is to commoditize anything in India, especially with regards to templatization, education, everything else. And that's that's makes it very, very easy for AI. But you come to leadership and management, what we study here and which is more about EI, right, the emotional quotient and emotional intelligence. Could you tell us a little more about your journey from AI to EI? Well, according to my wife, have not achieved that journey. So I've been married a lot longer and I can tell you you never achieved that journey. So I'm not sure I have a very intelligent response to your rather thoughtful question. So yeah, I don't really know what to say. I think that's a conversation over lunch. So can we have a microphone here, please? 2nd row behind Sumitra. Thank you. Thank you, Professor. Hi, Rohan. I'm Jagushiri. I'm a defilant medical history and I know about the translation work that is done in the Murthy Classical Library. I like what you said about AI redefining the relationship of human beings and language. What I want to know is how can AI redefine and perhaps improve those the understanding of the history of language literature and perhaps what is lost in translation. That's a fantastic question. So I'll just add a little bit of context to the question that's asked. So many years ago I started a non profit, which is perhaps amongst the largest translation projects in the world. It's called the Murthy Classical Library of India. It's based at Harvard. What we do is we attempt to broaden the definition of classics. So now as a grad student, I happen to be sitting next to a classics professor at dinner somewhere. And then he said, you know, classics is he's talking about Greek and Latin is, and my mother tongue is about roughly 2000 years old, a language called Kannada spoken by 60 million people in India. So I asked him, why is my mother tongue not a classic? He said, no, it wouldn't. It actually is. It's possibly it is. We just don't have have access to the text. So one thing led to another and ended up starting this project we have. What we do is we take texts from about 14 different classical Indian languages, Sanskrit, old Tamil, old Kannada, medieval Marathi, medieval Bangla, etcetera, etcetera and so on and we translate them to English. So if you open our books on the left hand side of the original script, the right hand side of the translation, and we do this for literature in these languages, in the classical forms of these languages. And the project is now roughly actually it's about 10 years since 10th year of its iteration. And we want to do this for the next 90 years or so, so about 100 year project. The reason your question is so interesting is that, you know, at its launch, we had launches around the world. And we had one launch here in London at the Indian High Commission in 2014. And we had this fantastic Punjabi text translated by a professor, I think at UCL, one of the schools in London. And his name is Christopher Shackle. And so Chris Shackle spent 50 odd years of his life being an expert on Punjabi literature, which is fascinating for me. I'd never met somebody like him. And so somebody at the time said, hey, you're a computer scientist. Why do you need scholars in 40 different countries doing these translations? And I remember saying there's no, there's no technology I can think of that'll be as good as Chris Shackle because, I mean, he's a remarkable scholar, 52 years of his life just studying Punjabi and doing these translations. Now, having said that, today, if the same question were asked, I'd probably think of it a little differently because I don't know if these models are at the maturity of a Chris Shackle, but they can begin to emulate a Kirk Shackle in limited ways. And that's far more than what we could ever do earlier on. And so there are debates that we're having, for example, at the Classical Library project where we're wondering, saying, because we're doing these translations and we have these expert translators doing these things across the globe, and the left hand side is the original script, the right hand side is English. What happens if we feed all of this into a model? Do we in some sense preserve some kind of essence of these languages in these models? Because there are languages, by the way, where we are not able to find translators because the next generation of all people, like most people in this room and myself, you know, glasses, engineering, computer science, don't really study the classics, right? So may not know the classical forms. And so it's a, if you want to do this for the next 90 or 100 years on some languages, should we quote, UN quote, give up and just train models? And yes, our translations will not be as good, but maybe that's the best we can do. And so there is this active debate going on at the board of trustees right now. In fact, they've brought in computer scientists and classes, and they're all debating. I happen to be sitting in through one of these meetings listening in. And the classicists were, of course, saying to the computer scientists saying, oh, but this translation is nothing like the scholarly translations we do. And they're right, but we may have no choice. And and that is, you know, five years ago, 10 years ago, that was not even a considered option, which it is now, right? And so we'll have to see the Ranjit Hakari from India. So when we talk about sustainable development in India and especially especially about the grassroots development. So it's all depends on the critical demographics, geography of India, which you, you know, so how the AI can help in this because it depends on the sampling surveys. Even if we do it manually, we find errors. How AI can help in it and what are the gaps, challenges and solutions in this? The most important thing it can do at any grassroot anywhere is level education more than anything else. Perhaps health too, but I've not thought about health enough. But certainly on education it you can bring uniformity to understanding or to it explanation of information in a way that we previously just could not see today. You have Wikipedia, right? And so if you have, so if it really depends on the reader and the context of the reader to be able to understand it well enough, that's what information is useful, but not enough. Whereas if you want to make any change in grassroots, my first bet would be on education. Last question. Last question, Rohan, I know Wasim, you can ask your question offline. Wasim, you don't even need to ask the question here. There is something which you don't answer sometimes. This is public, you're supposed to answer. Keep it brief please. Simple question, why did you choose what you chose to do? You could choose anything in the world when you passed out of college. So why did you choose to be doing what exactly you're doing in Soho? Secondly, how you're applying or harnessing AI because you're supposed to be a very big advocate of AI. So how these two things you thought about and you're actually doing it? Great question. So very specifically, just to give some context what, what I found, I'm very inspired by manufacturing. I think all those folks who do anything in computing should look towards manufacturing for more lessons. We tend to have this artificial distinction that somehow computing is the future and a lot of these industries with their hands of the past. I, I, I don't quite buy that. What I find fascinating about manufacturing is it's taken us 100 years to develop a disciplined way of building things, right? You can take any large complex object, whether you're sending into space or a car, breaking down a million little parts, assembly line, etcetera. And you have a detailed discipline for doing this. In fact, there are departments in universities, operations research, industrial engineering that study all of these things. But when it comes to office work, which is what a majority of people here in this room do, where we use some computing device for doing our work, whether you're in a bank or retail company, whatever it is, we tend to think that's very different. And there's no rigor or scientific way for understanding office work, for understanding what's wrong with it. And that always struck me as rather remarkable. The best guess that we had, we have had for 40 years, is to hire consultants. They'll come, you have an instinct, they'll ask a few people a few questions, and then they tell you what's wrong and what to fix. And So what I found very interesting was how do I create a science of work, how we work? I felt as I do not, not that I felt, I know for a fact there is no science for office work. And so I saw an opportunity to take some of the work that I was I had done in academia with a little bit more effort to be able to create a fundamental scientific basis for understanding and improving office work. And I thought, hey, this is great. I get to do some academic work and build a product out of it and maybe have some impact in the world with it. And so that's why I decided to leave academia to start a company. So I think we need to finish there. The time is up. I'll just make a couple of comments. Firstly, I mentioned earlier some of you may have heard me that sort of my mother had a stroke 3 weeks ago and so she's now stopped talking and she's hopefully starting to talk. And you know for somebody who's got to had a mother for 63 years and she's nagged me every day for 63 years and then she stops talking, it's quite hard. So when I go and see her in the evening, you know, I often think sort of what could I do because she is trying to talk and what could I do to sort of help facilitate, etcetera, etcetera. So it's just another example of how this sort of technology skill set I think is going to completely change all walks of our life. The final comment is, look, you know, I absolutely love my job and I love my job because I meet so many truly brilliant people. And increasingly these truly brilliant people are very, very young. And Rohan, you're another one on the list. So thank you. Please can I ask you all to applaud them?