Persistent Systems Limited (PERSISTENT) Earnings Call Transcript & Summary

Sharing with you. And this is mainly for educational purpose. So I just wanted to let you know about that. The other disclaimer -- this is from my side that this is not a Persistent AI strategy discussion. The purpose of this presentation is to provide an overview of what's going on in the industry and what's happening in AI. Again, this is not my original research, so I'm not presenting personal research here. I'm actually sharing with you ideas that have been collected from the Internet and actually have been curated and enhanced by the AI tools that we are all very familiar with, okay? So let me get to the next set. So we are in March 2025. And as you all know, in AI, things move really, really fast. So I just wanted to make sure I time stamp this. And AI is clearly the future, and over the last 2, 3 years, especially with respect to gen AI, I think we have moved forward quite a bit. That said, we are not at the end as such. There's a long way to go. So that's the context I wanted to set. So for me, right, I've been part of the computer science type industry now for more than, I'd say, 45 years, 50 years. And even when we were in college, we had AI as part of what we were working on. And over these years, I have personally seen AI evolve over the last 40 years. And in the first 10 years or first 30 years, I would say, even 35 years, AI was always -- anything that is not discovered was AI. So AI was something that is always the next step. So that is how AI was thought. But this time, I would say that AI is for real. And right now, we are seeing real commercial implications of AI in the context of what's going on. So that is the first thing. The second, as I mentioned to you, AI has been around for a very long time. One of the earliest examples of AI came in all the way from 1940s and '50s when Turing test was trying to -- was created by Alan Turing. His idea was to create a test that will identify if you have a human in the loop in some sense. But I'll not go through the rest of this here. But right now, as we see AI is AI is really transformational. And in my opinion, it's the next big technological revolution. So we started out is -- as big, in my opinion, as Internet, mobile and gen AI. So Internet allowed everyone to have ubiquitous connectivity. So everybody got connected to each other, and that was the main benefit of the Internet. Mobile allowed us to have compute and apps on our own mobile devices that allowed us to do many more things in addition to the connectivity. And now with AI, I think everyone will have access to intelligence with them at all times. And that's really remarkable and really the reason why this is going to be quite exciting. Now people talk a lot about gen AI, but I'm going to say that gen AI is only part of AI. There are many things involved in AI. And when I'm referring to today, I'm not referring to just gen AI, but I'm referring to AI. And AI is really the tip of the arrow, and there are many other techniques that are following gen AI. So it's a combination of those things, and I'm not going to specifically talk gen AI, but pretty much mention AI as the technology overview for the future. Let me start with this simple thing that I have observed in general. And I find that when you start using AI, you will find that AI starts to feel like it's improving your productivity quite significantly. But then as you get further along, you get -- you'll find that now I have too many options. I don't know how to use AI. So you kind of get into this zone where your productivity comes down. But then you have this concept of factory workers. So copilots, which is what people are using quite a bit right now. In the copilot world, you can actually improve the performance of factory workers significantly by improving their productivity on an ongoing basis. And that's where the copilots happen. But I think what we are looking at and what I'm going to share with you a little bit today is about agentic AI, where you are not just improving the performance of factory workers, but you are building factories. And this is sort of how I see this. So you'll see exponential improvement when you get into agentic AI as compared to copilot zone where you get 30%, 50%, 70%, 100% improvement of factory workers. But the factory workers still exist. But when you build a factory using agentic AI, the way you might think about it could be very different. And we'll talk more about this as we go along. Just one highlight slide I have here is that deploying AI in commercial mission-critical applications is not trivial. It's not like snorkeling that you would just pick up your gear and jump in. It is quite complicated, and you would want to be trained, and you would want to have people who understand what's going on. So it's more like skydiving rather than snorkeling is how I see this. So the point here is that if you are going to skydive, you need an instructor to go along with you. So you would want to work with professionals to make this kind of stuff happen. And you wouldn't pick up AI and do commercial agentic AI or commercial and mission-critical applications you wouldn't work on without any help. Now let me get to some basic bold claims that I thought, before I get further, I should sort of comment on where I think AI is going at a very macro level. And these are sort of the 5 bold claims that I would like to make in the context of what's happening in AI in the next few years. So the first is, of course, I believe that AI has implications and will affect everything that we are familiar with, whether it is health care, education, finance, transactions. Just like the mobile phone affects every single individual in the world, AI will affect every individual and improve their productivity, change how we think, it will change how we interact with people, we'll have multimodality. And all of these fantastic applications that you can imagine, they will all get built using AI. Now the key here to understand is to build all this AI, you are going to need people to build it. It's not going to happen out of the box. And many of this is where I think software engineers have a great opportunity to create transformational products that will transform the world in some sense. And of course, as I said, just like as said in Spider-Man, with every great power comes great responsibility. The second thing that I believe here, and this is something that I would have not said about a year back, but having seen the improvements that we see in various kinds of AI products, it is very clear that AI will write better code than software engineers. So software engineers, per se, cannot compete with AI for writing code. But there is a lot that needs to get done, and we'll talk a little bit more about what should software engineers do to survive in the world [indiscernible]. So that's a good question, and we'll talk about how to live with AI. But if you are trying to compete humans with software code writing, AI will write better code, more reliable code, faster, cleaner and repeatable, all of these kinds of things on a long-term basis, much better than humans do it, okay? If you have anything that you don't agree with me on this, feel free to post a question or put your hand up, and we will take you in as we get along. But let's keep moving. The numbers -- third item that I'm mentioning here is about agentic AI. So I think agentic AI will redefine industries with unprecedented power and risk. So I think agentic AIs are where autonomous and complex tasks will get handled with minimum or minimal human oversight. And we see this in automated cars or self-driven cars or things like that. But I think these kinds of agentic AI solutions are going to come in every single domain that we know of. And I'm going to share a little bit more about how this works and what exactly is happening in the agentic AI world. But I believe that agentic AI is going to be the one that is going to really redefine industries and commercial and enterprises that we are talking about. This is where the future is. And I think we are still early in this game. We have been making some investments in this through SASVA and other agentic AI platforms, but this is where you will see the best return on investment. And once companies are able to deploy agentic AI solutions, I think you will see a transformation on what we can do within the enterprise. Now of course, this is one of the risks that I see. Inequality that can get created between the haves and the have-nots. It's a huge challenge. And this can happen because of access. It can happen because of cost. Energy requirements are different. Geopolitics is an issue. We can see this right now with what's going on in the U.S., where people are cutting off other countries from access to AI and other related things. And in addition to that, concentration. Right now, there are a small number of companies that are dominating AI. So that's a huge challenge. And the divide between the haves and have-nots will deepen, and that is a huge issue that we need to figure out how to deal with. And then finally, the fifth thing I would say is that an unchecked AI would be humanity's great risk. And while I'm not a doomsday advocate here. If you want to see some -- in the last few months, there have been many commentaries by Yuval Harari and also by Geoffrey Hinton, who have taken a very negative approach of what could happen because of the thing and the worst-case scenarios that they have been projecting are fairly scary. So yes, there is a little bit of things that we should be worried about. And I think it's important that all of us, both at enterprise level, at individual level and nationally and globally, take into account what responsible AI may be and ensure that we are operating within the purviews of responsibility. So this is the first set of 5 bold claims that I wanted to make today. And then what I want to spend in the next hour or so is to spend a little bit of time trying to explain to you the technology that is coming into all of this stuff. I'm going to go through this rather fast, but it will give a sense of what exactly is happening, and it will give us a chance to sort of put a common framework for which then we can see if this is what is happening and this is how it's happening, how do we, as individuals, respond to it? So what would -- what is the future of digital engineering as I'm going to share a little bit? And what do software engineers need to do in this whole [indiscernible]? So really, that's what the rest of the talk is. Again, I hope I'm making sense, and if not, please stop me. See, some of you here also can interrupt if you think that there is something that needs to be clarified because otherwise -- okay. So the first thing is, of course, everyone has heard LLM. So what is really different here? I just said, for the last 35, 40, 45, 50 years, we've always used AI. And most of the times, AI that was used was mostly symbolic. So you had preset rules that were created, and they were explicitly proven. And traditional AI required specific instructions for each task. So if I wanted to do certain kinds of things, I had to train a particular program, I would have used supervised learning or any of those kinds of techniques. And I would have had to build separate programs for different kinds of applications because you would train them separately, you might have a different program. You might have to handcraft these rules or whatever. With gen AI, what we have found is that in the last few years, we have seen how deep learning has allowed gen AI to learn from patterns autonomously from vast data sets and some of the tasks that gen AI can do can be done and new tasks can be discovered without any additional team. Now I want to share with you just to give you a context of how AI has moved. This is a little bit of a history of the work that Geoffrey Hinton has done over the years. So if you see he started in 1970, he earned a PhD in artificial intelligence in '78. And if you saw the key thing about neural networks, stochastic recurrent neural networks, he invented them in 1985. So since '85, he had that. And one of the more important algorithms that is used a lot in AI is the backpropagation algorithm, which he invented in 1986, okay? Then despite having done this in 1986, for almost 20 years, nothing much happened. Most people, at that point, believed that deep learning or some of these neural networks don't have any commercial applications. And the main reasons were that people were trying out neural networks that were not as deep as are needed to get good-quality results. So most of the parameters that were used were small number of parameters, partly because of the hardware and various other things. Also, the amount of data that is needed to train these things, which did not exist at that point. But over the years, different kinds of networks started to happen. AlexNet got developed, ImageNet got developed and many other. People started to make data available for people to do analysis and to train their models in some sense. So there was a 20-year, which is called -- you call an AI filter in some sense and almost 30 years between 1986 and 2016 when, all of a sudden, people started to see that deep learning can create amazing kinds of things. So the kind of persistence that Geoffrey Hinton demonstrated in getting this to a point, I think, is a very critical factor. So AI, while it seems like it is doing amazing things and it seems to have been born yesterday, it has been around for a long time and many of the work, the algorithms that Geoffrey Hinton talks about were -- have been invented about 20, 30, 40 years back. But all of a sudden, because of the access to scalable hardware, the ability to do parallelism, ability to have large quantities of data, improved algorithms and the willingness to try out a model that was far larger than what was imagined allowed AI to happen at the right kind of things. And all of a sudden, we realized that we could build billions of parameters in an AI model, something that was not imagined or possible with the kind of hardware and data that we had when we were looking at it 20, 30 years back. And of course, funding that went into this was very much responsible for AI to happen now. Now I don't want to really explain what is happening, but there are many kinds of foundational large language models. And these are the ones which tend to want to learn everything. So they are trained on vast amounts of data. They have billions -- hundreds of billions of parameters that they are trained on, and they are general-purpose models. So many of these models are where they are trying to push AGI where [indiscernible] know everything. But reality is that there is a whole bunch of them that have come up. And the most common ones that you would have seen are things like GPT 5 or 4.5 and then so on, 3.7 so on, and there are many of these models. Some of these are in open source and some are closed source. All the same, these models are getting released at very rapid pace. So these -- if you see in the last 3 months, for example, you will see many new models have been updated, or newer and newer models are coming in. The pace at which these are happening is quite incredible. Now what is a model? Now I'm sure you all know this, but I think it's important to explain what our model is. So unlike a search query where you are answering a deterministic question where you say, okay, give me a result in a particular query. In a model, what you are trying to do is you are trying to infer results on the basis of known information that you might have. So a best example to think about this would be to think like a medical doctor. So a medical doctor, over his practice or over her practice, gets to see many patients. And they figure out that, okay, patients have this and this and this is how they can treat them. So over a period, they start to become an expert at various kinds of symptoms and the kinds of medicines they should prescribe to people with those symptoms. Now if somebody else walks in all of a sudden with completely new symptoms, it's not that the doctor is fox, the doctor thinks about them, and using other patterns that they have seen before, they are able to guess what may be wrong and extrapolate or fill in the gaps in some sense, and they are able to identify what may be missing. And this is what the model does. So model is the sort of representation of knowledge that you would have with you, which can be used for interpreting things that may be seen or even those which we haven't seen before. So when we talk about ChatGPT or others hallucinating, actually, they are responding on the basis of the data and the model that they have built upfront. So they are designed to hallucinate in that sense because they are actually answering questions about information that they may not have seen in the past. Now of course, we've been following models for a very long time. One example, of course, is the Galileo's experiments where he dropped balls from a certain height and saw how long it takes to come down. And over a period, he built an equation. By building an equation, which is really a model, I don't have to test for every single value of X or Y to see the equation. I know what the equation is and I can put the right values and I start to get the results I need. Like this, you can train your model using training data. And once you have the model trained, then when you give it an image in this case, it will be able to tell you with some level of confidence that this is really what it is all about. So models are very useful. And models are built by training on data sets. So over the years, people have -- if you look at any neural network, typically, they have multilayer neural networks. So you start for an image like a dog and you say, okay, let's try to find out various kinds of things and you build a graph of this kind where you put weights on every single node and you are able to then identify if a certain kind of an image is of a dog or otherwise. And also, if you find that you are not able to identify, you get a wrong result, then you have the ability to go through a backpropagation algorithm to fix the errors that may be there. Now of course, in today's world, these graphs are very complicated. They have many, many layers. And each of these signals are essentially encoded into real numbers. And the edges, which is sort of like neurons and every such node has weights which adjust depending on the data sets that they see. Now -- and then this -- once you have the net trained in some sense, then you can ask questions against this net that has been built up. So this neural network is really the model that gets created in most deep learning models. So when we talk about what is involved in a model, typically, these have parameters, embedded layers, transformer blocks and a vocabulary. And once you have trained a model, you can create these things and you can block. You can create this, describe what your model is and you can publish the model. And once you have a model, you can keep building off of that model even further on. But all models that you will get essentially, which are deep learning models or neural networks are essentially made up of these 4 blocks. Now the way these models have been trained, especially in the generative AI context, they have been trained to focus on the next element in substance, the next token at a time. And this is happening at -- so what happens is that you are all familiar with what happens on a mobile phone when you start typing a message, you see the next word, it gives you suggestions on the next word. So by learning through multiple text elements, one word at a time or one token at a time, the system has figured out over the model, how do I add a next token to the existing set of tokens. And this token generation happens at great speed and in an interactive response time. So that's really what is the engineering innovation that has happened in this particular thing. And one of the important things that comes up is that when you're looking at the next token, how far do you go to see, okay, maybe I chose the wrong token. Could I go back and fix it? Can I change this around? So that essentially how far can I go? What is the context window in which you can look at? And that is sort of referred to through this attention mechanism, which allows people to go back to a certain window as such. And this is what has really been very critical. And this is a seminal paper that came out from Google but it is 2017 is when this came out. And this is supposed to have really transformed how deep learning models can do meaningful answers rather than just answer some toy questions or small answers. Now of course, as it is said, when you are looking at models here, the -- these kinds of text that you read are actually converted into tokens. Instead of looking at it as characters or words or whatever else, there is a tokenizer, which is the first part of what you do when you read a model and then you break it down into one to 2. I don't have the time to explain what these tokens look like, but typical -- there are tools where you can go online and put in your strings and you can see what tokens are getting created. But these tokens happen quite naturally when you start to implement it. And these can be tuned on the basis of tokens. And every time you are looking at AI models, people are referring to them as the number of tokens you have in your string, and your cost depends a lot on the tokens. One interesting thing I wanted to mention here is that essentially, what is happening when you build a neural network, you have points in a very sparse graph, which is very large and is multidimensional. But the way you are -- most of these tokenizers are built out, you are able to bring different tokens or different concepts in the vicinity of each other. So when you're looking at foods, you'll find all foods are essentially mapped to areas within the same area, similarly, people, toys, food and all that. What is also interesting is that you can actually map audio, video, text and images, different languages, all in that same context. So the ability that AI has of trying to move between voice, video, text and languages is quite incredible. And a lot of this is happening because of the way these tokens are associated. So a dog has D-O-G, dog, and an image of dog and the bark of a dog, they all need to relate to the same object that happens through some of these kinds of tokens. And this is something that I'm going to come back to you again because I think this particular multimodality of AI, actually, the ability to translate between different kinds of modes is, I think, a very phenomenal opportunity for us to reimagine how we interact with systems in a big way. And I think this is a great opportunity for everyone, especially the way I see tech for the tech and the software community about how do we rethink about user experience in the context of multimodality. And again, I already shared with you this that text, maps, all of them essentially get clubbed into similar areas so you are able to deal with these [ effectively. ] What happens is that after you have done all the tokenizing and everything else, a few of these elements end up being real numbers, large numbers, and they are all managed as vectors. So most of the -- if you think about the computation that happens in most of these -- on these things are all relating to large vectors or matrices where you are trying to do computation on them, whether you do vector mechanisms or linear algebra, probability statistics and basic calculus. So fairly simple mathematics is being used in doing these. As I mentioned, matrix multiplication and matrices of different kinds are very critical, and that's where GPUs come into play. So GPUs have become -- were always there as graphics processing units, the game units or whatever else where you're looking at an entire screen and you want to move screen and objects within the screen very effectively or the central processing unit or the typical CPU that Intel had was not exactly best suited for those kinds of raster graphics. So NVIDIA and other companies built GPUs, which were mainly used for improving the performance of display. All of a sudden, since they do very well on matrices, they started to get used for AI-related objects. So AI computation starting to use GPUs potentially was accidental, but now over the years, GPUs have become the primary processing mechanisms for processing these vectors [indiscernible]. Let me move to the next segment here, which is regarding the model architecture and types. So there are many kinds of models. Not all models are the same, and different models do well for different kinds of things. So GPT excels at general text generation, BERT is good for understanding and analysis, T5 can take text-to-text format, domain-specific models. So more and more models are getting built out, and they serve different purpose. So you have to understand which model works for what kind of application. And there is a little bit of that you can think about. Again, these models decide how you are going to respond to various kinds of queries. And again, this is too detailed. I'm going to skip through this slide. And over the years, you would have observed that there are some models that are open source, which means certain companies have chosen to release their models into the open source. Now when they release a model, what exactly is meant by that? In some cases, they are releasing the sort of -- if you were to imagine this whole vector as a large ZIP file, it's a bunch of numbers and you have it compressed, some people are saying, okay, here's the ZIP file, here's the API, and I have released the model to anybody to use so that you can do various things with it, or you can start from that model and keep building on top of it. There are also open bits and open -- there are many varieties of open source or open models that are there in the context of how models are kept open. Of course, the most recent DeepSeek is considered one of the most open of the open models. And Meta has this product called Llama, which has been around for a while. There's Mistral and then there's some -- but of course, the large ones are from OpenAI and others are definitely closed source models. Now again, this is a very interesting situation. When OpenAI came out with their first version of GPT-3 and 3.5 and it started to become very useful, it was felt that the more data you feed in and the bigger the model you create, the better-quality results you can get. So then the game became just a hardware game, and I'm going to get more and more models, bigger and bigger GPUs, bigger hardware, bigger models. So increase the number of parameters as the only set that started to come. Over a period, people realized that for certain applications, if my application only requires a certain kinds of things and why do I need such a big model? Can I do with a smaller model? So while OpenAI has models which are like 700 billion, 650-plus billion parameters, today, 7 billion parameter models, which are much smaller, much easier to deal with, which Llama and others provide, can be quite effective on their own. So small models have a place to play in this world. And it doesn't have to only go to these large foundational models. And this is an important distinction that I want you to realize because as you start to use large models, they are very expensive to use. They are very expensive to train both in terms of cost of running them, the energy costs, the number of GPUs you need, the infrastructure you need to get started is extremely high. Of course, they can answer questions that may be random, which we have never thought of. But if you are within a domain, a smaller model can be as effective for your application as a large one. And again, I want you to note this because we are seeing a world where people are going to build many small domain-specific models because they can be controlled in terms of where the data resides, they can be controlled in terms of how they are used and also controlled in terms of the cost and performance parameters that are necessary for your application. Keep moving here. Now one of the biggest challenges that AI had in the past was data that is tagged and good for learning. Because when you're doing supervised learning or even unsupervised learning, you need large amounts of data set that is pretagged. So that says, okay, here are the dogs and all the dogs are tagged correctly and the cats are tagged correctly. Then when you train your model, you are able to figure this out using these kinds of things. So of course, that still gives you good results. But over the years, people have figured out that through generated results and by perturbing certain values and all that, with limited amounts of good-quality tagged data, you can still get pretty good results if you have large amounts of data. So a lot of the stuff that has changed which has created this transformation for AI, I would say, has been because of the ability to train models using a few-shot or single-shot learning. So this has been a very key discovery or, I'll say, evolution of AI models in the last few years, where people have realized that I don't need to have all the data pretagged for my training, and I can do with a lot less training in terms of where I want to go. Now of course, what happens is that you train a model, which is a large model or even a small model with data sets that are well known or which are available to you. Now this is generic. But when you start using it, you might find that for certain kinds of queries that I need all the time, I can essentially control that model or add some more parameters or do some essential post training, which is typically referred to as fine-tuning. So you take a model and then you fine-tune it with specific things that are related to your own industry. So for example, let's say, I want to build a medical product. I might take a generic LLM and then say that, okay, I'm going to fine-tune it with information that is very domain-specific. So this extra stuff that you can do, which is fine-tuning, can be done with much smaller data sets on top of an existing base model, and you can get fairly good performance or good accuracy of results by doing this kind of fine-tuning. Another technique that is used a lot is something called RAG, which again basically allows you to add -- so when you get a foundational model that gives you results, sometimes these results are from the Internet. So they may not be relevant to your own particular environment. So one of the important techniques that people use is to take data sets that come back or the results that you get back from a model and then do a lookup or basically retrieve additional information from your own particular data sets. As long as it's organized in the right way, you can do a query against that and basically embellish or enhance your results, which are contained for your own organization. So you might ask for a question that says, okay, give me information about something, instead of getting the results directly from the Internet, you can ask it to give you a query, which then you can apply on your own data sets or whatever else. So in some sense, this allows you to actually work with smaller models and enhance them with your own data set because that way, you can get much greater accuracy. As against this, the other alternative is that if you have your own enterprise data, you push that into the foundational model and you extend your model very significantly. So instead of starting from the billions of parameters, I keep adding a few more billion on to it. But then what happens is, of course, it makes it more expensive, but then you have large context models, as they say. So if you see the foundational models, they have been moving into large context. So they are allowing you to ask longer queries and give you far greater context. However, if you are within the enterprise, you may want to use a smaller model and use RAG to enhance the performance. I'll skip through these slides. They explain a little bit about RAG and other related things. Now I want to explain that really there is a lot of discussion today about training and inference, okay? So training is the first part where typically, you would take a look at -- you start with a blank slate in a sense or you start with an existing model and you build that neural network that I shared with you. So you need to get data from somewhere. You need a lot of data to train your data sets, but then you can get this data first, then you tokenize it, you do pretraining, then you fine-tune it and then you make sure your results are good and you basically use humans to ensure that the results are within the right kinds of context and they're not going completely out of whack. And then you deploy these products. So there is several steps that are involved in when you're training an LLM, and training an LLM is very expensive. It requires you to have all these matrix calculations that I shared with you, have to happen again and again at large numbers. And these have 600 billion, 700 billion. So these models are many, many parameters. So these matrices that you're multiplying or computing with 2 different matrices can be so large that it takes a lot of time and energy and computational power to really set up the inferences. Now there are -- you might say, where is the data coming in from, right? So of course, you can get public data sources. Now the thing is you don't need to go and crawl your -- crawl all the sites everywhere all the time because people figured out that everybody needs that crawl. So you can go online and actually buy this common crawl, or some of the common crawl is available even in open source. So you can go to the Amazon site. They have a program there. And there is a website called Common Crawl, which can share with you the crawl that has been done. Now again, this is, again, a very interesting and another strange problem in some sense. And you would see that there has been a lot of copyright-related discussions happening on this. So this common crawl has actually crawled all kinds of things, everything on Wikipedia, everything -- all the books that are available online, news articles, research papers, pretty much everything has been already crawled and is already used for training these kinds of things. Now if I said that, okay, you cannot use my paper for training, it's very hard to do that because that's already consumed in some kind of common crawl. The common crawl also has specialized open available data on medical information or various candidates. So the amount of data that is available for trainers to train from scratch using common crawl is actually pretty good. Now most recently, there has been the most recent new one that came out was this Grok 3 model where this is from X and Elon trying to do this where he established very quickly in Memphis, a very large number of GPUs into a farm and he built this out. A very nice podcast on this is available with Lex Fridman where he had these 2 guys, Nathan and -- Dylan Patel and Nathan, who spent 5 hours almost in explaining how this whole Grok, media, NVIDIA computing, training cost examples. So very nicely done. They also explained in great detail how DeepSeek was built and what are the pitfalls and what are the trade-offs in various kinds of things. So extremely good podcast. Happy -- I would highly recommend listening to this. It's very long. That's the only thing I would say. But you can get summaries of this all over. But this is fairly recent, and they have explained this in great detail. Of course, DeepSeek has been the most innovative, new thing that has come out in the last few months. And essentially, they use some great innovations to build out smaller models, more efficient models. And have released a lot of their data in open source. They also have very good-quality papers that they have published, and they have actually some good innovations and engineering that has made it possible. Now what changes for the world because of DeepSeek? So of course, DeepSeek data is available so you can start -- in addition to a Llama model, you could start with DeepSeek as your initial model, which is quite richer. It also shows that potentially a smaller or a more compact model can be built, which has equally good-quality results. And that has encouraged people to build their own specialized model rather than just depend on these very large Grok-type models or the LLMs that ChatGPT and OpenAI have been building. So I think when we look at enterprises and you look at the enterprise and deployment of AI in the next few years, the fact that DeepSeek was able to build out -- train a model with limited number of hardware resources by using some new innovations actually has made it possible to think about some new sets of applications that will get built out because of this fact. Let me move on to say that what happens during inference. First is, of course, you train the data set. So you -- as I shared with you that run go examples, you take all the images and you create weights and build that neural network. Once I have that neural network ready, I can give it an input and it gives me an answer. Of course, the cost of running that inference is much lower than the cost of actually building the network. And once you have built the network once, you don't have to keep building it again and again. But even then, the cost of inference is 1 0 and depending on the model. So if the model is very large, then more number of nodes get involved in answering the question. And if that is happening, then to that extent, the inference costs are much larger. Now I do want to mention, and I already mentioned this to you that many times small LLMs -- or small LMs can be seen to be quite effective. Small LMs, again, are not trivially small. They are typically less than 10 billion parameters, 7 billion parameters is sort of what has been pushed out by Llama, which is quite attractive. And you can get pretty good results within your environment. And one can build out these small models at relatively low cost. So it is possible to imagine in your own environment to build your own LLM, which would be a small LLM with small number of parameters, 10 billion parameters instead of having to use a foundational model, which has hundreds of billions of parameters for every single app. And this, again, I want to flag this as an important thing to think about because I do believe as we look ahead and the kind of work that will happen in the enterprise context, small LMs are going to be a big part of what will get built up. Now of course, for enterprise use, again, you will take open source LMs and also customize them quite a bit. You can take existing LLMs, fine-tune them. There are many other things you can do with them to make them enterprise-specific, include your data sets. So with limited number of starting point, which is not from scratch, you can start with an existing model, add to it new things, train it further, do various engineering things on it. And you can build custom LLMs that are very specific to your own environment, your own knowledge setup and various other things, including new languages and various other requirements that you might have. And I think this is where you will see a lot of activity in that because you can control what happens, you can control all the queries, you can control the cost of where these things are going to run, and you are not dependent on this data set going all over the place. And as we discussed already with geopolitics and other reasons, if people shut these large LMs, meaning these foundational models out, you could be left with nothing. So people are going to build their own smaller, fine-tuned LLMs using open source as a starting point for enterprise use. As I said, again, if you look at smaller devices, again, you could potentially run these smaller LLMs, even smaller than the 10 billion ones on your telephones, iPhones or devices or various other places. But the whole concept of writing next-generation programs using LLMs rather than writing rule-based programs is starting to take into -- take boot in some sense and you will see even your toaster might have an LLM running on it or having a toaster that accesses an LLM online is likely to happen. So I think a lot of interesting stuff yet to be done as far as customizing LLMs to your own environment. Let me take a short pause and see if you have any questions at this point. I'm going to now get into a bit of the -- we are about 50 minutes in. So we are -- I'm doing on time. I'm going to spend the next 20 minutes or 30 minutes on this whole concept around what agentic AI is and then I'm going to spend the last 30 minutes in some sense on what exactly should good software companies have to do in the context of AI. I know it might take a while to get the questions assembled, but let me keep moving. If you have a comment or a question, just raise your hand and we can take you in and I'll let you ask the question. Anybody here you think I missed something that I should have shared, we can include that, anything that I should have explained that I didn't explain well. Akshay? Okay. So let me just keep moving. I'm using a 50,000 feet overview of some of the teams and techniques, but my purpose of trying to explain this is not to explain the AI, but to sort of illustrate the fact that there is a lot to the AI world in the enterprise context beyond like a ChatGPT or any of the foundational models. The foundational models are good for doing creative work, and individuals will definitely use it. But when I'm starting to build copilots or agentic AI solutions, I'm not going to be happy with just building or doing calls to OpenAI on the network all the time. It's both expensive, not necessary and might give me results that are not as predictable as I would like in my own enterprise mission-critical applications. That's really the point I'm trying to make. And a consequence of this is that there's a lot of work that needs to be done to put this all together for enterprises in their own points. Now let me get to the next part. And here, I'm going to try to give a differentiation between what a copilot is and what a workflow is and what an agentic AI is. So as you would expect, agents are being used by everyone in all kinds of framework. Over the last, I would say, 4 months, starting with Satya having a conversation with on BG2 podcast to various other people, he started to call out agents and then many just went ballistic on agents. He has been just calling everything agent force and everything is an agent right now. But by and large, when I'm referring to agents here, I'm referring to agentic AI to be autonomous systems that operate independently for extended periods. They are not systems that are just improving the performance of one process, but they are a collection of processes end-to-end or a collection of agents end-to-end that optimize the entire process end-to-end with some human intervention, but they have the ability to -- so there are typically 4 things that agents do. They are able to perceive, so they have ways to perceive what is happening in the world. So they are able to listen to what's happening in the world, meaning they have sensors or they have trackers that allow them to observe what is happening on the ground. They can reason, they can learn, and then take action. So self-driving car is a good example of an agentic AI solution, but one can imagine building these agentic AI solutions to do all kinds of complex processes. Let's say, you want to go to a new country for -- let's say, you want to make a trip to Bali or South Africa, then what do I need to do to make this happen? Like when I'm trying to build a house and then I want to sort of decide the tiles and everything else, I can have agentic AI solutions that will go figure out the prices, different models, decide which one is the right one and do many such things. That's really what agentic AI solutions are about. Again, I don't want to go through this on a per-thing basis, but AI agents are -- have limited autonomy and they do human-defined tasks. Of course, they can do certain specific kinds of things as instructed. But an agentic AI, you can give it a goal and then they can kind of figure out a path to get to that rather than just giving them very precise steps to get things [indiscernible]. Self-driving cars are a good example. On cybersecurity, we see some of these. Supply chain management will create some nice, interesting examples. Health care has some very interesting examples of how we go to doctor and then there are many steps that are taken in, but all of them can be automated using agentic AI. AI agents, typically, you see a lot of them which are now already in deployment, which would be for customer support. Chatbots are AI agents. So they answer very specific questions. They have AI in it, but of course, they are automated for answering very specific kinds of things that you are looking at. What has happened in the last few months or say, last few years and now getting exposed in the last few months is that all of a sudden, all of these things have come together. So the ability to bring in data sets from different sources, the ability to bring small models, the ability to pull down and fine-tuning, the availability of compute power at various places, new architectures that allow you to do this and the fact that you can interface with AI systems with APIs and other ways so that you can ask the right questions in real time has made agentic AI possible. This was not that commonly imagined in the past. But over the last -- the alignment of various things in the last few years have made agentic AI really a possibility and it's quite exciting. As I mentioned already, they do 4 things, they perceive, reason, and act, and they can take and execute complex decisions on their own. Now of course, there is a debate on this topic, and there is a friend of mine, Subbarao Kambhampati, who believes that AI and the way AI and gen AI is built out, you cannot basically use them for planning. But there are discussions on this topic, but the combinations of generative AI models along with traditional rule-based or traditional AI models, combining the 2 together, predefining certain things can give you a very good performance on process-oriented stuff. As I said, agentic AI work -- operate autonomously. They do things on their own. But there are 4 kinds of things that people use when they look at agentic AI [indiscernible]. So one of them is called reflection. The other one is tool, planning and reasoning and then multi-agent collaboration. Now reflection, the easiest way to think about reflection, and you can try this out on your own ChatGPT kind of framework, is that you can ask ChatGPT to give you maybe an essay or a poem. But instead of stopping at that, you can just sort of convert ChatGPT from being a generator to be a verifier. And then you can say, hey, ChatGPT, here's what you have generated, can you verify that it is good? Is it good enough or whatever else. So you can make AI reflect on what has been generated and you might get actually better-quality results because ChatGPT and other AI have this ability to refine their results depending on asking them a very different view on the same question. And this happens to humans as well, meaning if you ask question about what do you think about this and then you ask a further probing question that says, are you really sure? Can you think of other ways why you might say this or otherwise? So instead of asking just one question, you sort of go through a loop, give it another alternative, ask a reverse question or ask it to verify if your question is correct and then you go about building these agents. Now when you're doing these verifiers, these verifiers can actually be not necessarily the same LLM, but it could also be a different LLM. So even if you want to try this out, and you should, try to ask a question to ChatGPT and take the answer, ask Claude the same thing in a different way and then ask the third AI a different way, and you will find that you can enrich your answers and you get much better-quality answers by combining the way of thinking of different AI agents. So this is a big part of what is going to happen as we look at agentic AI patterns. And one of the interesting things to note, and I want to flag this is that typically what will happen is when you build an agentic AI solution, in addition to building the solution that does the work, you will also build agents that are going to verify that the work has been built is actually accurate and correct. Now many of these will require different kinds of systems or you will require another additional step to do this or you might have different agencies that build agents and some who build verifiers. And I do expect that in the future, these agents and verifiers will come in sort of an agent store like an app store, and you may be able to buy agents, both agents and verifier agents. So this is sort of where I think there will be a marketplace for agents as we look out. Now as we have seen in general, when you look at this kind of trained data, trained data was trained at a certain period of time and you need the model to use on an ongoing basis. If you have a database query that you want to find out what is the current views on a particular customer, you cannot get AI in your model to really answer that question. So what people do today is that the LLMs have the ability to make API calls and you set your system up in such a way that the LLM can query the system, which is deterministic through standard API calls through an external data set, ask for information that may be relevant for that context and then embed the results that you get in the LLM that you are working with. So the agent typically has call-out points where you go to a real-world system, ask the API call and then get the result back. Now again, I'm just flagging this again here. When you have to do something like that, one, you need your internal systems to be ready enough to be able to accept an API call from an agent. Most systems that we see are not quite ready for doing that. If I have a particular kind of screen and I look at the interface in a particular way, it may not be well suited for an agentic AI call. So there's a lot of work that needs to be done in trying to make sure that your own internal data systems and other systems that you have are able to answer questions using agentic AI calls. In addition to that, the data that needs to be available to these agents needs to come out at the right format, at the right place, at the right time. That also requires a set of engineering redesign to ensure that your agents can ask the questions at the right place. But any agentic AI solution today will have callouts that it will make to perceive data, what's happening in the real world, what's the weather, temperature. So these are all API calls that you will make at different points in your workflow that you have in an agentic AI solution. The third part, of course, is that typically, what has been found is that if you ask the system to do one point query saying that go do something which is straightforward, closed loop, it's pretty easy. But if you have an open-ended task such as plan a travel itinerary for someone who will visit Europe, there are so many options. You might want to ask what's your budget, what kind of place are you going, how many people are traveling, what are the dates, what are weather. So a lot of these things require multiple steps. Normally, when you are building out an agentic AI solution, you end up building it using -- you have to plan what an agentic AI solution might look like and then you build it together by bringing in many agents together as you go along. And you have to plan this by figuring out, you need an architecture of how these agents interact, you have to build out the architecture and bring in these complex agents to address these complex tasks by simplifying them into smaller and more manageable pieces. When you are doing this design, you also need to think about where the human is required and where the system can make its own call. This is just an example that I tried doing is, okay, if I have to build out travel agency type thing then what are the number of agents that I might need, what kind of things these agents do and what kind of API calls you might have to use for building out these agents. So as you see tasks such as, plan my trip to -- travel agency trip to Europe or whatever else, will involve many different agents doing very different kinds of things. And also, they would want to interact with the airline system to check bookings, availability of flights, all those kinds of things. And those require API calls at the right place, so you can answer those questions as you need them. So overall, building an agentic AI solution is pretty tricky actually. It takes a lot of deferred planning and understanding of [ argument ]. Now if you were to think of agents, this is what Marc Benioff has been talking about, it thinks about agents as people, right? So just like you have an organizational hierarchy, in an agentic world also, you have a multi-agent collaboration that you have to set up. So there can be a pattern where you say that there is a supervisor agent that actually supervises the approvals of many agents who do their own steps and then come back. On the other hand, you might have situations where all of them are decentralized and everyone has its own agentic AI solution and they can be working together. So you can have a traditional company like hierarchy, which is top down, you can have a team that everyone is working together or there are tasks that are generated and any agent can go pick up that task and get it done or there may be a supervised group or various other models are starting to get at that. The summary of this all is that when you want to build out agentic AI solution which is actually going to do your complex tasks end-to-end, and this is when you will see exponential improvement in performance because now these agents are able to actually get the job done with very few humans in some sense because you can do the end-to-end. But stringing these kinds of agentic AI solutions is nontrivial. It requires enough skill and expertise to figure out how do you build an agentic AI architecture? What should each of these agents do? What kind of agents will need verifiers? Which agents will be LLM-based? Which ones will be API calls? Which ones can you buy in the market? Which ones you have to build from scratch? And how do these agents then interact with your own internal business systems that you already have? Again, I'll not -- I'll skip through this. We already talked about RAGs in a different way. RAGs, again, fit into these architectures because you're looking at your existing data systems in some sense to see what is happening. Typically now, when you see these kinds of architectures today, you'll find 3 layers. On the bottom layer or the layer 1, as I call it, is the hardware and software infrastructure. So a lot of this is starting to happen through GPUs or Google have their own tensor processing units. And there are other people who are providing them. And when you do smaller models, you can do with other standard processes of the RAG. You need high bandwidth and all that. So when you are setting up your own AI infrastructure, you need to invest in the hardware and software infrastructure that is needed for making this happen. So it's several racks in a data center or you have to go to a cloud provider like an Amazon who will give you some of this or a Google or a Microsoft who might give you access to their hardware and other complex things that are needed. So it's not trivial to just run your own AI environment as such. After you have the AI environment ready, then you need to worry about models. Do you have them retrained? Where are they fine-tuned? What kind of optimization techniques have you implemented? Are they specific to a particular domain and things of that. Then on top of that, then you have to build out the application interface such as the chatbots or virtual assistants, which are one of the most popular, but one of the early ones that have come in. And we see many more applications that are getting built out on top of that, which will allow you to access APIs. There will be automation, robotics, integration with other external systems. So potentially a robotic engine or in a factory operating on instructions of an LLM is not something that is far away from the future. As I mentioned, beyond the hardware, all of these kind of things, you need a team that understands how to go about building this. So you cannot build these agentic AI systems out of the box, and I don't believe they can ever be done in the next few years because they have to interact with existing systems, which have their own idiosyncrasies. And while you might not need large teams to do that, you will still need enough people who understand the process, the ability to build out these architectures, ensure that they are verified and they are running at optimal [ level. ] I'll skip through this agentic AI stack. I'm going to take a pause in a couple of slides. So if you have questions, you can start to get them ready. The summary here is that implementing agentic AI workflow is tough. It requires technical expertise, system design and focus on resource management. Okay. So let me sort of open the floor now for questions a little bit and then see if you are with me in terms of what I have explained so far and if you have any specific comments or questions because what I've shared so far is very obvious in some sense, it is well known, it's well written up. The point I'm trying to make here is that if you have to implement these solutions in mission-critical applications, you need to have a pretty good handle on many of these techniques. And while the trivial programming can be done without any individual, and programmers are able to do it for doing anything that is complex, you need to bring in other people and other expertise. [indiscernible] You've been quiet. I hope you are able to hear and follow.

Anand, I have a question?

Yes. You may have to be unmuted by us so if we can do that. Nilesh Shah. We'll figure that out. I'm sure there will be some way to make them available. Do you have a question? See what he asked. So Nilesh Shah. Can you unmute his mic? Somebody else also has a mic?

Anand, I think Nilesh's question is whether we can get access to the tech, which I think we can make available later.

Yes, we can answer that.

So Anand...

Yes, Go ahead, Saurabh.

Yes, if I can ask -- so Anand, you did touch upon the subject -- you did point out the fact that it is likely to be nontrivial for a lot of these agentic AI frameworks to be, first of all, built and then configured in the context of enterprises. Do you have thoughts on which industries, which sectors may be able to do that better to begin with over the next couple of years compared to others? And also maybe given that a lot of these things may require significant investments. So larger companies, would they get even higher edge over a lot of their other medium and small competitors? Just some thoughts around that.

Of course, the first question that I would mention regarding which industries will come in first. See, as we discussed here already, right, you need to have good-quality data at the right place to be able to use and leverage AI effectively in your own content. And that's sort of where those industries that have invested in the last few years in building out these kinds of data infrastructure, they are the ones who are going to be the first ones. So for example, the financial industry, for example, they have good-quality data with them already. The health care industry might have some good-quality data with them. Some of the supply chain problems also might have good-quality data with them. So anybody who has specialized data that is good quality is a good target for starting on some of these kinds of examples. That said -- so that's a good place to start. And yes, it is expensive, but it's not for [indiscernible]. Those who don't have data at all, they could directly go to one of these foundational models, and they might find that the results that they get from the foundational models are as good as a good guess anyway. So they may be better than just a good guess that you are taking without data. So I think a little bit of that might also happen. Of course, when you use foundational models, you are able to or open to more creative answers. These answers may not be exactly accurate. They may not be in the context that you are looking for. But they may be still better than no data answers that you might have on your own. So every industry has an opportunity of where this can go. And I think we will see an evolution of how data gets [indiscernible], how AI gets deployed. So first round of things will be on using these foundational LLMs like ChatGPT or Grok or [ broadband. ] Every individual will start to use them for improving their creative writing or various other activities. Students are using them for assignments. So at the individual level of -- this is starting to happen. I think the next big step is going to be where people are going to use copilots, where there will be guardrails for their own system and people will have the ability to ask these questions within the context of their own system. And then finally, the real value addition or the real benefit that you will see will happen when we are able to get into agentic AI systems, but there is a lot of work that you need to do to implement. Okay. Anybody else have any questions? I just want to make sure that I haven't lost [indiscernible]

I have one question.

Yes, go ahead.

So how do you see this agentic AI model -- engagement model evolving? Is it more software led? Or is it more [indiscernible] wherein there will be a few people who will be specializing more and more refined solutions for a use case versus everyone trying to build to a services model like through us companies like...

I think it will be a mix of the 2 where you might end up with people. So someone -- see, so let me explain that a little bit more as we go along. We get to the next part, and I'll explain some of this [indiscernible]. When you look at what is happening, and I'm going to focus a little bit on agentic AI because I think this is where the next action or where the next opportunity is present, and I already mentioned to you that we need significant expertise to put AI to full use. And another way to look at it is to look at it here. So when you have an agentic -- when you do chat or personal assistance that we are looking at, they are about individuals and creativity. They are unconstrained. You can ask a prompt and you'll get results. They are pretty much out of the box. And you can get any kind of request that you can send and you'll get a response. Of course, depending on the model, depending on the kinds of questions, the fact that they have seen it before or otherwise, you get good-quality results [indiscernible] good-quality results. When you get into the copilot zone, typically, they have been trained on specific kinds of problems. So they work within certain guidelines and they are able to answer questions which are based on templates and improve individual [indiscernible]. And when it comes to agentic AI, the way to think about it is to think about an end-to-end process, not think about the workers who are on the project, but to think about the process end-to-end, saying, okay, you want to look at a code to cash or decide whether some loan should be given, I will make a travel to [indiscernible]. It's an end-to-end process. There are many steps involved in it, and every step has certain uncertainty, but you might want to bring them all together, right? So when you are trying to build out solutions that are relevant to an individual process that is highly customized, you will need people who will bring it together. It's a complex engine [indiscernible]. Some of that will be a services-type opportunity, but you may be able to buy agents or get a marketplace where some of the agents can be made available. So it will be a combination of prebuilt solutions integrated by teams that are going to happen. But as you think about it, there are way too many such processes that can be built out, and it may not be possible to get all of those out of the box. I just want to...

I just have one quick question.

Yes, sure.

So do you have any thoughts on possibility of overreliance on the gen AI output in general in various context. And especially when maybe not so tech-savvy folks have no way of verifying what they are getting in return for their -- and yet like blindly believing on that and then going forward with the results. So [indiscernible].

Yes, that is a huge risk, for sure, where people might believe -- see, every time you ask a question to a gen AI solution, you will 100% get an answer. Whether right or wrong, you'll always get an answer. So the reality is that not all these answers are going to be correct and you have to verify them for sure. But -- yes, so after a while when you start getting used to them, you just sort of stop verifying and you assume that whatever is coming back is good. Now if you are looking at questions that are well known and which have been thought of enough, like if Internet has many places where that particular question has already been answered, then gen AI has already learned enough about it, and they can always give you very good-quality results. However, if you ask for something where there isn't enough data, you ask for opinions, for example, then the opinions can be anything and there's always a little danger on when trying to believe something that doesn't exist, right? So yes, I mean that is a definite risk. So this is again a joke that I was very tempted to bring this up because I've used this twice before, and I think this gets used again in the third way. So this was many years back, it's from 2008 when we were first talking about cloud computing. Everyone sort of said, oh, my process is complex. I'm going to move it to the cloud. So there's just one block, which is the cloud and then input output and you have the whole cloud. And it took us about 10, 15 years to migrate data from existing setup to the cloud. It was felt that, oh, I'll just point my data to the cloud and boom, it will be in the cloud. And it will be one step and it will happen. The same thing in '15, '16, '17 when we were talking about APIs, every one of us said the same thing that "Oh, you will have an API to answer that question. Every question you might have, there's an API that answers that question." But who creates those APIs? Where are those APIs coming from? And how long does it take to create the right kinds of APIs? So it has taken many years to build out these APIs. Now I see the same kind of discussion happening on agent. So I have this very complex process. I have one encapsulated agent and the input output and now your agent is working. Unfortunately, while the agents are going to be far more efficient, they are significantly, exponentially more efficient than humans doing it one at a time, the cost and the complexity of building out these agents is [indiscernible]. Now the other thing that I want to point out here is that in general, the Indian software industry has done well on engineering. So typically, when you start, you start with science where you understand principles, then you get into technology where you create frameworks. And then when it comes to engineering, this is when you implement solutions. So I think we are heading into a world right now where now engineering solutions is where the game is going to get played. It's not just about inventing new principles. They are well known. We have good enough frameworks. People have enough open source stuff. Now how do you engineer this into my specific environment, right? That's where the opportunity is, and that's what is going to be the future work that engineering teams will have to do. It's not just coding anymore. Coding -- AI will code better than humans. It's not about writing test scripts or writing test data or any of those things. All of those individual tasks, the AI will do a lot better than human. But when it comes down to stringing these multiple tasks to ensure that you can build a complex system that is reliable, there is a lot of engineering that is required to make that happen. Now let me sort of go into what are the opportunities for the industry and for people like us to think about. One is, of course, if you hear experts talking about it, everyone says that we have 1.4 billion people. There's no way we can have enough doctors. We can now use an AI agent and everyone can have a doctor on call 24/7 and all of those things. Yes, it's all true. But someone will have to build that out. We don't -- general-purpose ChatGPT-type framework and say that, that is going to replace a doctor. I think it's a little far-fetched. It will give you answers, mostly correct, but you cannot run in an environment in a regulated or when you are dealing with humans to be in that [indiscernible]. However, you can build good-quality health care solutions, which target the right audience and the right population with the right language interfaces at the right cost by having to build out these agentic AI solutions for health care, for education, for agriculture, for food and life sciences, for environment and all that. So if you think about the road map of the kinds of applications that need to be built over the next few years, that road map is quite huge, and we can see many different areas where I think new start-ups, new companies or even large companies would look at problem statements that are within the India context and figure out ways to build them out because there is a lot to build to make them reliable and robust and at the right price point that can be used at scale. The other thing I want to mention here is this is how we've been looking at it. So if you look at the kind of sophistication that you can get in the top companies, of course, they know everything and they are hugely sophisticated. But if you go into enterprises, the next few hundred and the next few thousand, you'll find that the kind of technology readiness that companies need for implementing AI is not there. Also, the skilled people that are needed for making these kinds of things happen is also a big challenge. So I think there is a lot of gap in terms of what ChatGPT and AI can do which is phenomenal. But to leverage that and harness it into enterprise organizational situation requires a lot of work which is nontrivial. Very quickly in the context of Persistent, we have been looking at these in the following way. So we have our own agentic AI framework called SASVA. And we also have a framework called [ iAURA, ] which is again, I don't want to spend time explaining what those are here, but [indiscernible] I'm sure our team will be happy to explain. I see a huge opportunity in designing, implementing, deploying, maintaining agentic AI solutions. There's a lot of work that needs to be done to make that happen. That is going to require software engineering skills. To do any good-quality AI, we need good-quality data. There is no doubt about it. And one of the big things that we see right now is many organizations that we work with don't really have their data at the right place at the right kind of way. So they cannot answer these kinds of questions very easily because one, they don't have it in the right place, and they don't have the APIs set up to be able to ask questions to an agentic AI. So there's a lot of work that needs to happen in those [indiscernible]. The third thing I would say -- fourth in this list is multimodality. So I think the opportunity is -- so traditionally, we have worked with environments where we have always thought of everything to do text and numbers. Most of the database [indiscernible], all of them have used mostly text and numbers. So it's very unidimensional in some sense. Now that we can ask queries about video, voice, different languages and all that, I expect the way we look at interfaces will change dramatically over the next few years. There is no reason why I need a report. There is no reason why I would want a chart, and I can always ask a question of what I should do next, right? So the way I might look at these kinds of things would be very different. And I've sort of advocated this in the past that we are heading into a world where we will see low aperture interface in the sense that imagine your interface is just a cellphone or an Amazon Alexa box sitting on the table and then why do I need the whole concept of monitors and keyboards and all of these kinds of things that we are used to, which are a relic of the past, may not remain important as we look at it. So I think there is a room for rethinking how we interact with computers, how we respond to things. And I think this multimodality is a big opportunity that can create many next-generation ways of how we interact and play with them. Enterprise technical debt is something that I think is a huge opportunity. Many years of code that has been built over the years. People don't know what it is. There's an opportunity to document it, clean it up, transform, remove unnecessary or redundant parts and all of that. And then finally, I think AI plus X is going to be a huge opportunity. So as I said, it's not that easy to build out AI systems. So while one can think about, okay, what does it take for having civil engineering AI systems that can design the next-generation bridges or any of those things. Calculations and specific things, maybe you can do, but what would it take to design the next bridge or do many other things. So there are lots of applications that one can imagine that can be built, rethought of and completely reimagined using AI. Let me stop here and see if you have any questions right now.

So Anand, there are a couple of questions. Unfortunately, he's not able to access the Q&A tab. But he has just...

Put his hand up, we'll just let him know.

Yes. But he has sent us an e-mail. I can read...

No just ask him if he can...

Do you know the name?

[ Apurnay. ]

If you have a question, can you please raise your hand and we can just let you [indiscernible].

Do you want to take the question anyway?

Yes, what's the question?

The question is how different is the AI wave for services as compared to cloud wave for services? That's one. Question two is how much of the current scope of work in services can get cannibalized versus monetized? What is the lead time or lag time between both of these? And the third one is what are the evidence of success markers for services companies, for example, revenue productivity, et cetera?

So I think I'm going to let Saurabh answer most of these questions. All I can say here is that the fact is that there is a lot of engineering that is required to build out agentic AI solutions. And I think companies will get to them one by one. They will not start with them on day 1. They'll start to use standard off-the-shelf models that are available such as OpenAI and others, Microsoft already making some of them available. So that's the first round of stuff. But that is not going to change the -- so real cannibalization of engineers will happen only when agentic AI solutions are deployed at speed. There a little bit of improvement that you can get on copilot. It's kind of very hard to squeeze in terms of getting improvement unless we have a large number of people do the same [indiscernible]. That's true in certain kinds of call centers or knowledge processing units. But for services-type jobs, just having a copilot doesn't really improve enough of their performance. However, when you start to build out end-to-end solutions, that's when you will see significant improvement in terms of the performance. And will the -- yes, of course, certain business will get cannibalized with other app. But also it's about getting ready to cannibalize business at large and kind of [indiscernible] to where the world might be.

Can we get to the next part of it here. So okay, so this is going to be transformational. What will software engineers do to orchestrate in the new world of AI? So of course, one interesting example that I want to share with you is how the world embraces AI. And I want to use the example of chess. So chess has always been one of those games that everyone has felt that I would like to solve chess using AI or whatever can be done to use, solving chess was always thought of as AI. So Ken Thompson, who is one of the very well-known programmers or the guy who wrote C and Unix and all that, he started out in 1979 to build a chess machine called Belle. And so '79 onwards, he's been trying to do and mostly it was about trying to teach the computer about playing the rules of the game. A serious effort was made in 1989 through Deep Thought in IBM. Kasparov, who was the champion at that point, actually defeated IBM's Deep Thought in 1989. At that point, it was felt that the computers are good, but never will be good enough. In 1997, IBM's Deep Blue again defeated Kasparov for the first time where in a sort of official match, chess grandmaster or world champion was defeated by a computer as such. And this was remarkable in some sense, but the algorithms or the technology used by Deep Blue was very deterministic, and the main advantage that IBM had was that they were able to think many moves in advance as compared to what human might do. And deterministically, they knew the errors and all that. So basically, brute force was being used for defining how Deep Blue defeated Kasparov. And he had his [indiscernible] that he said that Deep Blue was intelligent way to make your -- the way your programmable alarm clock is intelligent. So he never called it being intelligent. He just said that AI, in the Deep Blue context, was just brute force ability to do a lot of [indiscernible]. And then in 2016, when Alpha Zero came in, which was starting to use next generation of techniques, the rules were invented and very new kinds of games were played. Anyway, the point I'm trying to make is that starting from 2005, '06 onwards, chess engines, as they're called, have been regularly beating human. Today, if you look at Stockfish, their ELO rating is about 3,800, and Magnus, at his highest, is 2,882. So the reality is that human versus the computer or an engine in chess is no contest at all. Engine will defeat the human every single time. And if you talk to the grandmasters, all of them are using engines for building out their own games and all that. Now in mid-2005, '06, I remember there was a lot of discussion about how chess is dead because there's nothing left to play. If a computer is going to defeat the human every single time, what's the fun in playing chess anymore. But if you see the renaissance in chess right now and the number of interest in chess and if you ask chess grandmasters about the moves they've made now, AI has helped them improve the quality of chess, improve the quality of games, improve the new sets of moves, which people in the past would have never played are being played today because people are training with AI and asking AI to do questions in very different moves. So we had Vidit come in here and he explained how he uses AI to generate chess games and how he creates a whole set of data set for his practice using AI. So essentially, what is important is that AI has been embraced by chess players, and they have become better chess players because of AI. So I expect something similar to happen to humans. The bar for perfection will keep going up. Your data is as good as the questions you ask. So the premium will not be on the ability of the craft, but on the creativity and the ability to ask hard questions. So I think there needs to be a lot more thinking in terms of how do I frame the question, what is the domain, what am I expecting as a result? How would a product be positioned and all of those kinds of things? So you will need very different kinds of skills to run and manage AI projects as compared to the skills that we have done -- we have used in the past. Again, you need very different kind of skills to succeed. AI programs will require very different kind of change management. Risks and liabilities are going to be very different. The kinds of infrastructure you will have to manage will be very different. So anyway, we are running into a world which will feel and look very different. You may not need the same number of programmers to do the same kinds of things, but a team of programmers will have to worry about many different kinds of things, and there will be a premium on building out the ability to do design thinking or the ability to think about various kinds of things. In addition to that, this whole question of what is truth and what is correct and all of these kinds of philosophical questions will become far more important than they have been in the past. In the past, we worked with a system that was deterministic. You give it a question, you have the rules, you can verify the answer. Now you are asking an LLM, think in between, there are lots of creativity, but then you don't know if you are on the wrong track, right? So very different kinds of skills will be required for programmers in the future. But I see it's quite exciting, and I believe programmers will adapt to working with AI. So AI will not be a foe but will be an assistant, and you will do better-quality coding, better-quality systems will get built, and new kinds of systems at scale will be built because you have the ability to have a programmer as part of your repertoire and as part of the AI [indiscernible]. So I think there will be an opportunity for engineers to reimagine, implement and operate new generation systems. And this is a very different world that we are looking at. And I -- so again, this is something that I've been thinking a lot about in terms of what should we be teaching engineering students today so that they can live in the world with AI. And it's a very challenging thing to think about, but really what many of the things that we learn in college today, AI will be able to do very effectively. One story that I want to share with you would be during my early years in IIT, I joined Kharagpur in '79. At that point in my first semester, I was sort of in the transition year. So in the first semester, we were not allowed to use calculators. And so everyone had to use log tables and logbooks. And there was a lot of discussion within the IIT professors about how if we allowed students to use calculators their entire skills about calculations and whatever you can do with calculator will go away and kids may not be able to use log tables. But this is not a battle you can win. So by the end of my semester, people had chosen saying, no, everyone has to use the calculator. Everyone now uses a calculator. My first year, calculator was part of every assignment, everything, and you had to carry calculators to your exams. Now yes, I have lost my ability to do log table calculations. But has that really changed my life? Meaning do I really care to know about log table calculations on a daily basis. So like this, I think there are many skills that we have learned the hard way. As we look ahead, maybe they won't be needed because AI can answer those questions. Already my ability to spell without a spell checker has gone by. If you have to write on a board, you can realize that I don't have a spell checker on the board. So like this, things are changing rapidly, and I think we have to rethink about AI in the way we think about the world in the future. Again, I want to close here with this, and we'll take questions for the next minutes or so if you have any; otherwise, I think that is pretty much the last slide [indiscernible]

The question I have is what you talked about the deterministic nature. So far in the programming, most of the emphasis was doing the formal thing [indiscernible]. Suddenly now we are moving to a world which is not deterministic. LLM can give someone -- it gives a different [indiscernible] How to handle that part?

No, I think this is a good question. LLMs are not -- we have traditionally worked with very deterministic systems and now we have LLMs that are very creative and can answer questions that are not exactly deterministic. So how do you deal? So I think you will have to worry about the fact that the context in which you are trying to solve the problem. So for certain parts of the problem, I think having that creativity is huge amount. AI can generate many alternatives for you. And you can look at those alternatives and then decide which ones to keep and which ones to remove, right? So there is an opportunity to relook at some of these things in a big way. So that's the benefit of having an LLM. While if you have a requirement where you want to look at somebody's balance in a banking account, you don't want that to be probabilistic or LLM to tell you how much money I have in my bank account. You would want that to be very deterministic. So I think choosing the solution in the context of the problem is where I think a lot of [indiscernible]. But the fact that you have a creative assistant who can throw out many interesting opportunities is very [indiscernible]

One more question [indiscernible] So far, the LLM are good at the textual representation, but...

No, actually, it's not true, right? I mean, LLMs are actually not -- the reason we think about textual representation is because we are used to thinking in that way. And the data that we have is predominantly textual. LLMs on their own are not constrained by the text part.

Yes. So that part is correct. But what I see is like a lot of things that we now have, we have moved to more of a structured world where [indiscernible]. How do you see the interaction between them happening? More and more digitize things, we are moving to more and more structured things. This LLM is more of a semi structured [indiscernible]

They have to live together. As we just discussed, if I'm looking for bank transactions and bank accounts and money being transferred from one unit to the other, you would not want to keep that data. You still remain some kind of a [indiscernible] database banking system, whatever else. But that's the way it will be. So I think we are looking at a world where we can combine the 2. One of the advantages that we have today is the fact that we are able to live in a world where we can combine what may not be a structured text along with video audio and data services. So like, for example, I'm sure we have seen this as well, is that in the past when we were looking at testing screens to see if they are good or not if they're matching or otherwise, traditionally, we would have tried to look at the [indiscernible] we would look at the metadata about the object. But today, you can take screenshots and compare screenshots very easily. So the way you are looking at things is kind of changing. Today, I can take an image. If I want to find out if there's a skin problem, take a photograph. So the ability to handle different kinds of images or different kinds of modalities, I think, is extremely powerful. I would go on to say that we, in general, have not figured out or haven't pushed the boundaries on that multimodality enough, both in terms of how do you ask the questions, what kind of data can you answer? And can you generate videos? So for example, today, if you have lots of interesting data, you want to publish it, you end up publishing it as a report. It's a 2-dimensional flat, static document, right? Maybe you can create a movie out of it. So instead of looking at the result as a table with a chart, it might be a movie that might give you very different kinds of insights. So I think the way the world is moving, I think there are many different opportunities. I think there are no -- there don't seem to be too many questions. So let's stop here. I'd like to thank you for being here, and I'll hand it back to Saurabh.

Yes. Thank you very much, Anand. This presentation has been a long time in waiting, and we had received a lot of request for this, I would say, this teaching session from your side. We could not get all the investor participants to get registered for this in time. Maybe we'll do another session in person, maybe as part of our Investor Day where we can cover more details. But thank you very much for today, Anand. And to those on the call, if you have any follow-up questions, please do send across them to Alankrita and me, and we can see if we can get you the right responses. This presentation and recording will be uploaded on our website, and you can have access to it going forward. So again, thank you very much, Anand, and everyone on this call and in the room.