Transcript Disclaimer: This transcript has been generated using automated tools and reviewed by a human. However, some errors may still be present. For complete accuracy, please refer to the original audio.
00:00:10 Sangeeta Gupta: Welcome to the second episode of Nasscom Conversations on #AIspeaks. Today we're delighted to have a veteran as our guest today, Dr. Alok Agarwal. He's the founder of Scry AI, a company that performs research and development in data science and related areas. Dr. Agarwal had known him for a long time as the pioneer for the whole knowledge process industry in India, especially when he coined the term KPO because before that it was all about BPO. Yeah, he set up a Value Serve in India, which is the first company that started doing KPO services. And we've come a long, long way since then. He's had a very illustrious career working in IBM and you know, many other organizations and, and I think today we're going to talk to him about what's happening in the AI world. I'm, I'm really what lies ahead. So welcome, Alok, and thank you so much for making the time for us.
00:01:07 Alok Agarwal: Thank you for having me.
00:01:09 Sangeeta Gupta: So, so Alok, I want to start with your book, right? You've written this book on the fourth industrial revolution and 100 years of AII know AI has been around for a long time, but I really did not realize it's 100 plus years that you know, AI has already been in existence, but somehow, you know, with the inventor generative AI and ChatGPT AI feels different. It feels very new now. What do you think? You know how I know a lot of groundwork was laid in these 100 years, but how have you seen the development and what was seen now for it to become AI, to become a part of almost every conversation now?
00:01:45 Alok Agarwal: Yes. So it actually started in 1950 hundred years. I'm actually covering from 1950 to 2050. I bring about 327 more years or 26 more years. In 1950, there was a after the Second World War, there was a research paper written by one of the famous mathematician, some scientists, Alan Turing. Alan was well known already for actually helping the Allies win the war, right? Because he was able to break German cryptographic codes. This was in 1945. In 1950 he wrote a paper: Can a computer imitate Games? Essentially the the interesting experiment he he posed for the following. Suppose there is a computer in one room, there is a human, a man in the other room and a judge in the third room. Can the computer imitate the man so that the judge is confused and that believes that the computer? That is how an in fact there is a biography or a bio picture about Alan Turing, which is called The Imitation Game. People who haven't seen it are welcome to see is he's considered one of the biggest computer scientists and there is a Alan Turing award or literally like a Nobel Prize in computer science that's given out when he wrote this paper in 1952, BBC interviewed him and asked him, so when, when do you think this will really happen? When will computers be able to imitate humans? And his answer was not for another 100 years. So he's already talking 1950 to 2050 or 1952 to 2052. And then they asked him the next question, which was equally interesting. How do you think computer computers will get to be there? He said we'll have to train computers like we train our children. Just like we give our children examples and then we ask questions and validate and test them. We'll have to similarly train this these computers also. He called them child machines and actually that's what most of our AI algorithms or machine learning algorithms are. They're called supervised machine learning because they're being supervised just like we supervise their children. So you've just been around since 1950s. In 1950s, I'll just go back. What was the state of computers? What was the state of electricity and so on? We did not have semiconductors, so we did not have any small chips. Computers were made out of what were called valves and these were essentially like bulbs. If the bulb is on, then it could be A1. If the bulb is off, it's a zero. So that's how they were simulating 1 and zeros. And of course the bulbs fuse very easily. So they would take out to make these valves. They would take out the air inside the bulb so that it's vacuum so that the filament will not get oxidized and it will last longer. So that so this is the period. So semiconductors only came around in 1960s. So this is a period before in fact, when semiconductors were there. So you could not produce very large computers. In fact, that is how bugs will get in. And that's why the word bug came about because bugs will get in these computers. They will touch the valves, which are really hot and they'll die. And I mean, then they will short circuit the computer. And that's how the first word bug came out, thanks to Grace Hopper. I want to remind this history because we have come a long way. In 1965, Intel was formed and Dr. Gordon Moore, who was one of the founders, he actually said a very, very profound statement, which became Moore's Law. He said that he sees that these semiconductors will become smaller and smaller. Every two years. They will reduce in size by two by a factor of 2 and the cost will also reduce by a factor of 2. Therefore almost which implies that in 20 years it will reduce by a factor of 2 to the 10 which is 1000 thousand 28. Indeed 50 years have gone by and it has reduced by A2 to the 25 which is about 3232 million times. So the power of computer has gone up by 32,000,000 times if you can compare to 1965. Now that is the main reason why today we can do most of the things basically because of Moore's Law. The second one is because we have a lot of data, because Moore's Law, a lot of computers are around, they're connected by Internet and therefore there's a lot of data around. And as I mentioned earlier, to train children, we need data, we need to train them on examples here, we need much more data and the data is around. So those are the two main things. Why things began to move very fast in in 2005 after the computers became really very powerful and the data was around thanks to the Internet. There are other small reasons, but these were the two main reasons that propelled propelled this movement that we see today, which essentially started in 2011 with IBM winning the Jeopardy and opening up everyone's mind that I didn't can. I mean that a computer can beat in a game, which is not easy to beat with the three contestants and they're asked questions about trivia, about gender knowledge. Those are the main reasons why we see some of the some of the big changes today. But again, the hype is much more than than actually the changes that are occurring. And those who are deeply into computer science realize that this is the third hype in AI. The first hype was after during unfortunately passed away 1956. There was a conference when the word AI was point, and then people got all researchers got all excited. And Marvin Minsky, who was one of The Pioneers, in fact said that by year 2000 we would have a computer which would pretty much do anything that human can do. And that was the notion of artificial general intelligence. Artificial general intelligence, even though it's not completely defined, implies a computer should be able to do what an average human does. And we'll discuss that more. The hype was so strong that even a movie came out and this is the second movie I would recommend. Although we're talking about AI, I would recommend more about 2001 movies Space Odyssey. In 2001 Space Odyssey, they show this computer Hal 9000 who is an artificially general intelligent computer and one of the discussions that he's having with one of the people who are who's an astronaut is computer Hal 9000 series does not make any mistakes. All mistakes are due to human error. So that was the hype that got created in 1960s. By 19, early 1970s, they realized all these pioneers realized that to create artificial general intelligence will take a long time. Maybe Einstein, maybe Alan Turing was right that it may take 100 years and that was the first hype that died. Then there was a second hype in 1980s and that hype was not as big as this hype or the very first hype. But people thought again that these computers are going to rain the world. And if there was a there was a front page in business week, which says AI, it's finally here. And you see again today people talking about AI, it's finally here. So every generation I would say go through it's AI finally here or AGI, it's finally here. In my view it's still far away.
00:09:29 Sangeeta Gupta: So, so you're really saying that this is the third wave of third hype curve for AI and you every generation has seen it in some form or the other. It feels a little more real to me. I know I've probably not been as involved in the other generations, but but you know, just coming back to that whole discussion around super intelligence, we do see AI becoming smarter. The use cases that are now being talked about every every day there is a new model being launched. There's so much conversation around, you know, this is the potential that AI can do. These are the kind of different use cases. These are the productivity boosters. Just every process can be re imagined with AI. It does. It does make you feel that somewhere in the path of AGI or super intelligence seems closer than it was potentially even in the earlier hype cycles, right?
00:10:19 Alok Agarwal: I think AGI may be closer. We are still far away. So first of all, yes, because of Moore's law, because of the fact that computers can work 32,000,000 times faster than what they were doing in 1960s. There is no doubt that we can use these computers to predict and to understand the data from the past and therefore predict the future. That's how we humans learn too. We humans learn by experience, which is looking at the past. So and then oncologist, a cancer specialist had looked 25,000 cases about cancer, He or she would be much better in predicting what kind of a cancer it is for the new patient that has come in. The same is true with with computers. They, they can now take a lot of data and that's why it's called big data, enormous amount of data and they can learn from that. So therefore, there are, there is no doubt that AI will be very much like motors. In fact, I talk about that in the 1st chapter that just like there are about 35,000 motors right now, there would be about 100,000 use cases of AI by 2050. We already have more than 1000 use cases and they go all over. I mean you pick any department, you pick any industry, there are use cases of AI, some very strong, some not so strong, many boosting basically productivity and many others fundamentally changing the way we live and the way we will, we will behave and will interact in 10 years or 15 years from now. So that having said that with the ATI concept is a much harder concept if you look at it. And if the whole point is literature or language is connected to literature and value system, whatever, it's not that you and I are simply talking. We also come with our baggage, you can see, or with our value system that has been there for since we were born. And the systems, computer systems do not have that particular advantage or they're not learn through the years. You and I could talk about coconuts being thrown out in water and it may, I mean for computer it's coconut is no different than a ball and would not understand it. So the context and the semantics aspect of learning over the years that is missing and we do not know how to how to include that. In fact, we are. So there is one of the senior, I would say one of the current very senior scientists, current pioneer of AI, Jan Lakun, who's a professor at NYU, and also that of Facebook. He actually made a very strong comment about two years ago. And I put it in the book also. But we've built a very large ladder to the moon. It is not there yet to the moon. It's a very long ladder and we can do a lot of things with it. But will it this ladder take us to the moon? And his view was no. And my view is very strongly no, because we just do not know how to how to, but fundamentally solve some of the issues. Because you and I can talk about, for example, how does a mango taste like? Yeah, because we've all both eaten lot of mangoes and and all the good stuff. And we can discuss about various varieties of mangoes. And computer has no such understanding of taste. So you have no understanding of taste, understanding of smell, understanding of all the various other aspects of human senses or even for that matter senses that a rat has, that a mouse has. So the whole notion of creating emotions for exam, we are extremely far away from creating any emotions. Yeah, we can make the computer feel like it has emotions, but whether it has really emotions or it can even get close to what humans have is very far. In other words, the Hal 9000 of 2001 Space Odyssey is still far away, and maybe it will happen by 2052. Maybe it won't.
00:14:33 Sangeeta Gupta: Yes. So I think another comment we keep hearing and some of it may be actually true is that AI is almost like a general purpose technology, how we thought about the Internet, electricity. So the transformative potential of what AI can be more thought of as a general purpose technology. Do you think it's implication will be as wide reaching? I mean today we can't think about life without Internet and electricity. But but do you think AI will penetrate our lives to that same extent?
00:15:02 Alok Agarwal: Absolutely. And it will penetrate even more so. So in my first chapter, as as you rightly pointed out, the title of the book is the Fourth Industrial Revolution and 100 years of AI. And the 1st chapter is actually about the three industrial revolutions and eight characteristics that they've, they had shown in the 8 characteristics which are becoming clear even in this revolution, which I believe started in 2011. The first characteristic is there are a lot of inventions and evolution happens. It's not just one or two inventions and this one has inventions related to data, inventions related to AI, but other inventions like blockchain, like gene editing and many others, IoT, robotics, lot of work is going on climate change inventions. The second is that there is one invention which becomes the which creates the infrastructure. For example, in the first revolution it was the water and steam infrastructure. Canals were created in science like 1760 to 1840. Second, infrastructure was about electricity. Electricity was created. You had electricity generation, electricity transmission, distribution and consumption. The turn is exactly what you talked about, which led to Internet really becoming popular, which was electronic communication. Electronic communication lot of where broadband C cables were put in the sea, they were put in in the earth and a lot of wireless communication started happening so that became pervasive. So but the infrastructure required about $500 billion. This kind will have infrastructure about data because as going back to Alan Turing, you need the data to train these children from child machines and therefore this will be the infrastructure about data. You require a lot of data to be cleansed, harmonised, because your way of looking at revenue sitting in the finance department may be very different than my way of looking at revenue was sitting in the sales department and therefore these things have to be harmonised. We did some study that about if you want to just clean up and make the data useful, about 1% of the data will require about a trillion dollars to cleanse.
00:17:20 Sangeeta Gupta: 1% data.
00:17:22 Alok Agarwal: Yeah. So it's not not that surprising, especially that we spent $500 billion putting tables around the world to make Internet successful. So that's the second characteristic. The third characteristic is there is another invention which feeds on the second invention and becomes pervasive and ubiquitous. Everyone is using it. In the first revolution, it was paper, it was team engines. They used team and we had steam locomotives or steam trains even until about 40 years ago in India and around the world. It was at least until 60 years ago, a steam locomotive, steam rail, boat, boats, ships, all of them were basically a textile. Machinery was running on steam engines. In the second one, it was electric motors. As I mentioned earlier, there are about 3500 different types of electric and if you compare electric motors and what will happen with the AI, there will be at least 100,000 different type of AI use cases and AI systems. In the third revolution, it was the processors, it was CPUs, you have them in your smartphones. We don't even think about a processing unit running in our home, in our, in our pockets and so on and so forth. And, and, and those are the first 3 interesting characteristics of, of the fourth industry revolution. And definitely AI will pervade. It's already pervasive and will become even more so in the near future.
00:18:54 Sangeeta Gupta: Especially the whole focus on data and how you know data will be the new lever for this AI economy. I think the more we talk about it, the more the potential. In fact, you started the whole KPO revolution. Maybe the whole data annotation as a business opportunities is something else that India should focus and build on because there is a lot to be done there and I think India does have the people and the challenge there.
00:19:20 Alok Agarwal: And I've been talking a fair bit about data annotation since the paper. My paper came out and people didn't realise it and they think it's all the lower end. And the interesting part it is not all the data annotation is not at the low end. A lot of, I mean, for example, if you need to annotate data to figure out whether a particular medical X-ray is showing some kind of cancer, then you need people who are in medical domain and they would charge substantially. Similarly, if you're going to annotate data about finance, people better have at least masters in commerce or target accountancy before they can start doing some of these things. So annotation is actually very much like a BPOKPO process. It's spans the entire domain all the way from doing very simple annotation. What's this is a dog. This is a cat. Versus going all the way to being a medical Dr. or being an actually an insurance writer or a finance analyst. And that's something which I personally believe that's something which India is very well suited for. And given it's it's traditions of IT outsourcing BPO, outsourcing BPO and BPO, IT can take on and it's already taking on and people are not even realizing what's happening now, already taking on India quite rapidly.
00:20:47 Sangeeta Gupta: Yeah, I think it's sort of annotation we hear fancy was like data engineering, right. So I think it's probably getting subsumed in some of that.
00:20:55 Alok Agarwal: Everyone wants to everyone wants to add engineering, the word engineering to prompt engineering. Now I would say prompt engineering is not really engineering. Prompt engineering is more of an art rather than engineering because.
00:21:10 Sangeeta Gupta: It's like how we would do search, right? I mean, you had to be smart to do search, so you have to be smart to do prompting. So I wanted to follow up on the whole BPO question and, you know, the impact of AI and what it will do to jobs because clearly we see all these negative articles on how you know, AI will destroy all the customer service jobs because AI can do all of that work. And you know, you will have people who will potentially, you know, struggle with what are the new opportunities and AI can take over everything that is happening. There is the other school of thought which says no AI will be a more augmentation tool. It will actually help people who are not as qualified to address queries that will be structured. While, you know, the more intelligent people are, the more qualified people can do the more higher end jobs. It is not about, you know, we keep getting these arguments. It'll automate tasks. It will not automate jobs. But just want to understand how you're thinking about it. And especially because for somebody who's been actively involved in the BPOKPO industry and now doing AI for your clients, how do you, how do you see this transition on what, what is, what is happening to jobs, right?
00:22:19 Alok Agarwal: So let me again go back a bit in history, not 70 years, but only 10 years. So the first article on this which created a lot of waves in the research community was published by 22 researchers at Oxford Free and Osborne in 2013. They go wrote a very nice article which basically said that 47% of U.S. jobs are at stake. That is roughly half of the jobs in US are stay at stake because they can be automated. So given that US has 116,000,000 jobs, that means about 80 million are at stake. They could be all automated and the analysis was very, very good and it actually followed very similar analysis that was done about BPO and KPO. As to how many jobs will move from US to India or low wage countries in early 2000, similar analysis has had been done. Now the only mistake in my view they made was then this could happen in one to two decades, which means 2013 to 2023 or 2033. Now by 2023 we clearly have the world has not lost any jobs and it's very unlikely that we will lose jobs large number of jobs to AI by 2033. We are only about 7-8 years away and or nine years away and it's not at a stage where it, the, it shows promise, but it's not at a stage where it's mature enough that it can begin to take jobs. So in, in my book, actually in chapter 16, I cover job losses and job gains. And I do say that we analyse the whole thing. There were several other papers written after that. We analyse those also and did our own analysis that yes, jobs will be lost. In fact around the world by 2015. We believe that about 400 million jobs will be lost around the world. That's 40,00,00,000 jobs. That's a huge amount. But there are other aspects that are coming into picture by 2050. First of all, it won't happen in the next 10 years because this particular technology, even though it is very, very interesting, it is going to change the world. Things take time to change. They do not happen overnight and it is at least 10 to 15 years before the jobs, large number of jobs begin to go away. Until then, it will only augment people by being decision support systems that it will help people become better. It'll help a radiologist become better because the computer is saying, look, there is a small dot here, which I think is cancer. You may look at at it. Now this radiologist, she might have overlooked at it. And so this is augmentation, a decision support system that the computer is providing. And I think that will happen for the next 10 to 15 years in a major way. But jobs will not be lost now. They will begin. We will begin to lose jobs because they will become better. Even without AGI. Computers will begin to be better than humans and what they do, and therefore you don't need humans. The interesting part is that this is we should not look at it in just an isolation, but as a whole, as a whole, as human society. The interesting part is that the society, human society will begin to taper off around 10. We'll have we, we have 8 billion people around the world now. It can go to 10 billion in 2050. People will be aging. There would be not enough people who would be providing actually work. So right now there are about 36% of the worldwide population is working. That time there will be 32% working. So we'll have a 4% lost right there in the number of people who are working. So 4% of 10 billion is also interesting before. So when the jobs begin to be lost, we will actually need those computers, almost begging those computers and the robots to help us out because otherwise we will not have enough people to to do the jobs. And the problem that we've seen with Japan, China and not China, but Japan and Italy for sure where there has been a decline in population is that unless you have immigration, especially dependent has no immigration almost if you don't have the working population, your GDP begins to stagnate at best. So, so we will need computers at that time that very strongly to to help us out in 2050. Now there will be other aspects like India, China, other emerging economies will create many more jobs, especially at the lower end like plumbers, electricians, Hwy. builders and so on. And of course there will be more jobs created because the population would be aging. So you need more medical Dr. s and nurses and so on. But the very fact that you don't have enough workers around the world, I think we'll have two or three very major consequences. There'll be much more migration from humans from 1 area to another and we will need robots and AI to help us out. I.
00:27:36 Sangeeta Gupta: Think fascinating how how you think about in the right context, right. But given the children has come and the work we do tends to be very narrowly focused on the technology and the operations business operations world. How do you think if if you're a company that is providing basic customer services and you know you employ a lot of people for that, how do you see that? What kind of focus should they be doing because their outcomes or the challenges for them may be much more short term as opposed to a 2050 kind of a timeline?
00:28:10 Alok Agarwal: Absolutely. So I'm, I'm actually in the process of writing a paper that AI will be transformational to the Indian outsourcing industry, which includes IT outsourcing, business process and knowledge process outsourcing. And of course, customer service is right in the middle of it. And the jobs they will begin to get lost sooner than than many other jobs. For example, that of a Dr. will be lost much, much later than that of a call centre agent or a customer service representative. And I think that that's a place where the companies will have to do 2 things in my view, because it can be transformational. One very fundamental way that all of the Indian companies will have to, to move away from is charging on an FTE model that hey, I charge you $70,000 per person for this kind of work. I mean, for example, I'm sure TCS goes and provides call center customer service agents and they were saying $50,000 per person and they will have to start charging by the amount that the customer is using. So it will be usage basis just like we we pay for electricity on the usage basis, we pay for our Wi-Fi data on the usage basis. And that could be a very fundamental change for, I mean literally transformational for the for the Indian night outsourcing industry. And partly it could be transformational because it's human nature. If we become successful in something, we always say what got me here will get me there also, meaning I can go to the next level also. And exactly the opposite happens. The world has changed and therefore what got me here will definitely not get me there. I mean, if I wanted to start a KPO company today, it'll my likelihood of it be failing completely is probably more than 95% because there are so many companies in the world have moved on. So that's one very important transformation that the IT, VPO and IT outsourcing, all the outsourcing groups should be thinking a lot about, not only in customer service and so on. The second thing specifically about use cases like customer service representatives, I think there is an silver lining because you can create many of these people using AI, again, using AI as an augmentation tool. You can take all the data that of the past customer service data, like what conversations agents had with their clients. You can take all the data, you can take all the manuals and it can be fed into AI systems, live language models to train new AI agents much better, much faster. So India, as we all know it has a very severe problem with attrition that people don't last for more than three years on an average, maximum four years on an average. And this would actually alleviate some of the attrition and the HR issues because you will be able to train people much faster using AI systems which have been used here, They've been used as augmentation systems. So you will be able to train and educate them faster. And secondly, again, there's a newcomer who is a service agent. If he or she is not able to answer a question that the, that the customer is asking, he can ask the the AI system, the large language model to answer it, which may give 3 answers today, but he can pick one of those. So again, it is a decision support system which would help. So I think, again, it will be fairly transformational because we're not thinking in those lines. And some of these jobs will begin to go away in the next 7 or 8 years if they are not modified appropriately. And if the companies, which I don't think most companies are thinking, they're thinking at a cerebral level, they're not thinking at a doing level, at an actionable level at this.
00:32:23 Sangeeta Gupta: Point is if I'm if I'm studying in college today or for the fresher who's in college today, right and may join the workforce five years from now, what would be your recommendation to him to be ready for this AI augmented world?
00:32:40 Alok Agarwal: First recommendation will be be passionate about what you're doing because it's even in 2015, it's not yet clear that we would have AGI where it will actually be creative or can can actually replace computer humans in creative aspects. So be passionate about what you do if you're a writer, right? I mean, be the next DK Rowling and create next Harry Potter series. That's number one. I mean, it would be very hard, at least the way the kind of technology we have, it would be very hard for us to replace humans in the near future. Secondly, don't get worried about AI. It's another tool. I mean it's no different than what automation did in 1980s, what RPA robot process automation did in about a 10-15 years ago. It's just another tool, a very important tool which will be diversified. But think of it as a tool as we use our smartphones to look at the Internet even while we are travelling in a bus or in a car. So think of it as OK, this is part of something which can help me, rather than something this will take my job away and how can I use it to the to the best of my advantage?
00:34:00 Sangeeta Gupta: So working with AI, learning how to leverage AI, I think that's a skill that is that will become even more important.
00:34:08 Alok Agarwal: Right. Exactly. And that's why it will be transformational both for students of today and also for companies of today, especially outsourcing companies, IT BPO outsourcing companies, that they will actually begin to use AI. And the reason I'm saying it's transformational is that because they will have to cannibalize their own revenue. Now going from an FT model to here, I'll charge you by the drink, you are going to cannibalize revenue to a certain extent, at least in the beginning. Now hope is that there will be some companies who will not be able to make the transition. Therefore, hopefully you can be amongst the leaders and take the revenue away from in the long run from those companies. But that will definitely be transformational, both for students and for companies.
00:34:55 Sangeeta Gupta: Thank you so much Alok, this has been a fascinating conversation.
