The Cost of Scaling GenAI - and the Search for Smarter Models

Speaker A: So good evening friends and welcome to this episode uh of the AI ah Rising podcast. We uh, have a very interesting guest with us, uh uh but I will introduce you to him uh in a bit. Um but just to give you a context, you um, know if you have not heard of uh EPAM India or EPAM Systems basically which is an IT services, you can sort of liken it with the Indian IT services providers kind of stuff. Um you know and uh they are a global provider of digital digital transformation services and but specifically product engineering also uh they focus on engineering expertise, design thinking, business consulting and of course the whole idea is to help clients transform uh their businesses. Uh they cater to various sectors as is the one with most uh IT service providers which includes software, high tech, life sciences, healthcare, financial services and a whole spectrum of uh services. India uh incidentally is EPAM's largest location in terms of headcount and delivery uh capability. And Praveen Oja who's here with us today, he's IPAM India's chief technologist. So of course his role uh involves uh collaborating with UH EPAM's practice and delivery teams to drive technology focused initiatives both internally and for external clients. Uh he has worked extensively with fintech startups etc and he's certified in blockchain and the subject matter in cryptocurrency and cyber security. Very, very interesting subjects. Of course Praveen has a master's degree in computer science and an executive degree from the MIT Sloan School of Management. So that's all about uh Praveen I think uh we are going to speak about a range of subjects about Jayant and I are going to grill him about AI Jenny I agentic AI and given that he has done his Ms. In computer science I'm sure this is all going to be child uh you know pretty pudile stuff for him. But uh, I guess the insights that come from a company that is you know dealing with clients in and out on a daily basis, that's what we're looking out for and that's what we're going to talk about. But that in a bit. But first as is you know Jayant and I keep on discussing about the week's developments and uh, uh of course uh, you know I think uh two or three things uh caught my attention uh Jayant and Praveen. One is of course you know how AI powered browsers are changing the way we surf the Internet. So you know you talk about um, uh some examples that I have you know are ah, like um, uh we had the traditional browser like Google Chrome which has now uh, developed an AI mode, um, uh, I think because it's being, I think AI overviews is being subsumed within AI mode. Then you have the Mozilla's Firefox and now both are uh, you know sort of uh, Microsoft's Edge of course is there uh, Bing. Uh so now you have these, you know it's facing competition from AI. First browser like Perplexity's Comet which is newly launched. The Dr. Browser now Dr. Browser comes from the browser company which had the ARC browser if you remember and there's uh, AI powered browser called Brave Leo which is for a Brave software and of course there's OpenAI's planned AI browser. I think the whole idea um, uh, you know is that they're using Chromium based browsers which is um, uh basically you know uh, uh Google's open source, it's all chromium based uh AI browsers for uh audience uh, is built on Google's open source Chromium engine, the same core as uh Chrome. But these browsers you know can do, they can summarize pages, uh web pages, they can answer questions, automate searches using AI models like Chat, GPT, Claude or whatever. So they work like normal browsers. But you know there's that intelligent chatbot, that assistant layer. Uh and that helps uh to take all these actions, it's agent take also because it helps the take uh some of these actions. Yeah, absolutely. I think more interestingly Jayant and Praveen, you know uh, the ad spend in search advertising is almost like 351 uh billion in 2025 as per Statista. Uh, uh I think 75% of Google chromes uh, uh ad stuff comes you know uh, ad revenue comes from search. So this is an interesting uh thing that I'm noticing. Uh, so that's one trend that of course I would love both of you all to comment on. But uh, before that there's another which is this, which you know sort of is a rant from my side. It's more on that AI Appreciation Day. Come on. There's nothing called. You remember last year also we spoke about the same nonsense. So everybody is you know uh, sending us press releases about this AI appreciation day on the 16th of July. Come on. I have clarified it last time also that no expert has called for this day. There's nothing called this day. If ever there should be an AI Appreciation Day it should be somewhere in August when you know John McCarthy, Claude Shannon and you uh, know the uh, founding fathers of AI, they coined the term, uh, you know, uh, artificial intelligence. At least it should have been something to mark that this is a company called AI Heart or something which you know, makes T shirts. So they said, okay, Appreciation day. And this PR community has caught on to this nonsense. And worst part is that every company that calls itself an AI company is sending a specialization how they appreciating aid.

Speaker B: You know what you should, you should send back the PR execs. Uh, you know, I'll take this into serious consideration if you tell me which day of the year is Electricity Appreciation Day or Fire Appreciation Day.

Speaker A: Exactly. If AI is the new electricity, why are we wasting our time on this? But of course it was just a rant. I guess, uh, everybody wants to do some marketing and uh, uh, you know, uh, I mean we have written this book AI Rising. So I think we appreciate AI more than others.

Speaker C: But I'd rather go with, uh, June 12th, you know, when the paper came out which said, you know, actually that should have been in the name of my wife, you know, it said, attention is what you need.

Speaker B: Attention is all you need. Attention is all you need.

Speaker A: Yeah. Some of the, it should mark a particular occasion, not somebody who said or

Speaker B: the back prop paper back in 89. The Jeff. Jeff Hinton and Ilya Sudke.

Speaker A: Exactly. But I think most credit goes back to the founding fathers. So I would always say like John McCarthy, whoever. So he's like credited with having coined the term. So let's give them Claude Shannon, John McCarthy, uh, Marvin Minsky, my favorite. So all these names and of course Alan Turing.

Speaker C: July 16th has nothing.

Speaker B: Right, Exactly.

Speaker A: Just hot air and wasted of time and energy. But your, uh, quick thoughts on the AI powered browsers, uh, how it is changing the Internet. Maybe some interesting thoughts for our uh, viewers and listeners.

Speaker C: I think uh, the way it's happening, the way the shift that I see is happening is that you are moving away from task oriented searches towards really goal oriented searches. Right. So what I mean by that is, you know, we're used to search for a specific task. How to do this or how can I do this? You can actually give a goal that. I'm writing a research paper. I need to first talk. These are the topics that I'm actually looking at. And whatever results that you get, you can actually do a kind of uh, you know, what you call as prompt chaining. Right. Where the output of what the AI ah browser is going to give is actually that becomes the input request of the next search. Right. So with that chaining and now all this thing that you can just key in into the browser and you know, as an AI to start constructing and it'll even ask you, do you want me to actually create a word document with all these uh, you know, headers and all? And after that all you need to do is just start getting into a typical kind of a chatgpt mode like answer yes or no and you know it'll start doing those actions. So I see that, you know, it's um, it's largely shifting away from very, very specific tasks which we used to like. I, when I started off using Google, it was all used to be for coding. Right. So how do I write this code? How what is the best practices, what is the best standard? And here like I'm just giving it a far fetched goal. I'm like, you know, that goal is sometimes so abstract that you know, it uh, it is much more better place to actually give shape to what I want. Right. So it starts entering into the sphere where it starts shaping uh, you know, how I want my prompt to be done. And it's, it's something, Leslie Jain, you know, that you see uh, nowadays that the LLMs are better tuned to writing prompts than the humans. So in fact they're sourcing even the prompt writing to do the LLMs.

Speaker A: Yeah, I was wondering what happens to the prompt engineers?

Speaker B: Oh, I was, I was just uh, looking at YouTube, uh, video yesterday which says prompt engineering is dead.

Speaker A: Yeah. And just a year back they were talking about $350, $350,000 annual salaries if you remember.

Speaker B: Annual salaries. Yeah, yeah. No, I agree with what Praveen is saying. Right. In terms of how the context, uh, and the objective of search itself.

Speaker A: Yeah, I, Sorry to interrupt but I want your thoughts more on. Because you have been you know, with uh, uh, Motorola and all. So I want you more of your thoughts on, you know, how does it impact the search advertising revenue? What? Because you know, the kind of stuff that is happening out here, the uh, you know, how are you, how these guys will have to you know, come up with newer models and stickability, uh, is not there within the AI overviews and the AI modes and all that stuff. So this is going to be a very interesting. Love your thoughts on that.

Speaker B: No, so that's what I mean. What, what I'm, I was saying was I, I, you know, I reiterate, I agree with what Praveen is saying as to how are, uh, the search itself is changing but if you look at it, browsers have in the last 10 years Leslie had already become a layer on top of our uh personal computing usage. Be it on the Internet, be it on the laptops apart from the operating system. If you look at it, browsers are the in and uh, the things that were almost a ah OS layer. Right. Uh, and companies like uh uh uh, uh, Google with Chrome or Apple with Safari tried to keep our Mozilla with Firefox try to keep um, their consumers uh stuck to one of the browsers because if moving browsers had completely uh, become um, extremely difficult. It's like changing from iOS to Android or from uh one operating system windows to um, Apple operating system or something like that. So browsers had already become that layer as far as uh consumer usage and the uh, locking enough value from the tech companies point of view. So I mean this I'd see very strategically as a, as a product in the AI age when the, when the companies, the you know um, older tech giants trying to you know bring in more AI features and AI functionalities into their existing browsers and the uh, AI companies like the perplexities and anthropics coming up with their own AI first browsers. This is an opportunity again uh, uh in the aih, um to get consumers hooked onto the new product and new operating system. How is this going to change the uh, ad revenues anyway? See search engine optimization in the last one year we've been seeing uh, has come radically changed. Yeah, yeah, it's going down because you know people are not um, um now that the bots and AI agents are out there and they are looking for intelligent and they are uh aggregating the knowledge and giving um, LLM based intelligent output. Okay. So people were anyway uh looking at optimizing their digital presence be it in the form of websites or anything. Okay. Beyond search engine because you know people are not going to the Google searches anymore. So they were looking at how can uh, this one come up in ChatGPT outputs or llama sorry or deep SEQ outputs and the Gemini outputs. Right. Still the backend of the entire ad revenue moving on to uh, AI based uh search outputs is still not built but it's very clear that the search based ad revenue is decreasing. Where is it moving? How will it move? Who will lead? That is that infrastructure and that business model uh, uh even in terms of a product it's not in, not in place as yet Leslie, but it's definitely going to take an impact. And this uh, AI based browsers being there is actually one step into these companies trying to get closer to Moving of that revenue.

Speaker A: Yeah, but what I found very interesting is also in both cases, you know, uh, Google uh, came up with the transformer technology. OpenAI capitalized on it or monetized here. Also you have the chromium based browsers which is again a Google product and you know everybody else is uh, capitalizing on it. So you know, others are seeing far more value in Google and Alphabet.

Speaker B: It's like cricket has been uh, invented in uh, Britain and you know India and South Asia rules badminton again inland and England and you know, Asia rules it I think. Yeah, that's, you know, what is it called? Uh, inventor's dilemma or. No, there is a term for it. Right. It happens in history. We've seen that happen. The originators, the inventors of it somehow don't seem to be deriving the most value of uh, you know, of the stuff they've invented.

Speaker C: Uh, that kind of like, you know, because now browsers became a layer like it wasn't uh, became designed because you know, data started growing across. So do you see now there's an introduction of a new layer of agents. Now instead of your browsers where your revenue was essentially getting generated based on your searches, now your agents are essentially going to be driving that, you know, by being the agents that kind of consolidate aggregate filter. Uh, and you know you're basically monetizing those aspects of those services of the agents. Because instead of having these two, three browsers tomorrow you might have just one agent, right? You know, for which you would uh, essentially pay whatever revenues that you're. It becomes your revenue source for that. So do you see that introduction as well? Like you know there's a, there's, we are moving into a direction where. And now it's getting more and more abstract with the introduction of a agent layer.

Speaker A: But these browsers are also agent tech. So all these AI houses, they capitalize on that agentic power of the.

Speaker B: Correct. But, but Leslie and Praveen, today I, and to the foreseeable future. The way I look at it, agents are um, you know, more of um, you know in silos. They are, they are task doers. Okay. Um, they are being designed and developed uh, both in terms of, especially in terms of depth. They're depth oriented right now and for the foreseeable future. Okay. Um, and for them to become a layer, you know, we need to figure out a way to build a connectivity tissue and you know, get them to get them to work. That's where A2A comes into picture. But how will that become a layer? And you know, how can a uh, search based ad revenue kind of a model work on, work on that? I don't see that happening because for the foreseeable future I think agents are going to have a lot of SaaS, uh kind of a business model. You know what I built this thing that can help you solve your problem, automate your entire workflow. Your entire team's work can be done by this tireless guy, um, with highest amount of accuracy. And you were spending X number of dollars per year uh, for this team. You pay me a fraction of that I think for the foreseeable feature, the business model and the pricing model, uh, and the solutioning based on agents is going to be that way before an uh, agentic layer um, you know, can be built. Because when we're talking agentic layer we're talking about multiple agents working in unison. For the foreseeable future I see agents being developed to solve issues M. You know, for um, um instead of a common, common issue.

Speaker A: But could it also be that you know uh, as you said, okay, business models are themselves are changing but see it is still a tree and 150 billion uh you know, search advertising market. So it's not a small market that we're talking about. And again as I said you know 75% of Google's revenue comes ah from that, that's around what 165, $170 billion. So it's not a joke. Uh and even when I was uh, you know uh, chatting with admin SRI now I saw perplexity. I remember telling me that we are looking at that, you know we cannot dislodge Google as far as that position because Google clearly holds a you know 90, 95% market. Yeah, almost more than 90. Now he's saying we are looking at the high quality ad because the margins are there.

Speaker B: Mhm.

Speaker A: And that is what the creamy layer that he is looking uh, at dislodging them. But of course that's something. But my question was more on the thing as the business models are uh, shifting perhaps what people are looking out is using these agentic agents as such. This layer that is there, that is being built within these AI powered browsers to have more commerce based uh transactions. So you know whether it's the payment, whether I can actually buy my good. So I would probably make my money through commerce uh rather than anything transaction revenue share maybe, maybe it goes directly to an Amazon, it goes directly to a uh, whatever the gate. If you remember when we used to develop websites way back in year 2000. We used to have these CC nows and also the, all your payment transactions used to go to cc. Some other, everybody used to have that. So probably, you know, whether we want to call it a democratization of, of uh, commerce once again or level playing fields, who knows what, what these models will delve into. But um, uh, Praveen, what are your clients asking you? Because you're dealing with, you know, digital transformation in the engineering space and across manufacturing industries, I guess, uh, across sectors you're looking at finance, you're looking at plenty of other sectors. Um, typically uh, now there are two or three things of course AI. Then we were talking about gen AI, now we're talking about agent. AI are not giving, you know, ah, clients a breather. Too many things are happening at uh, One Stoke, uh, a lot of people are confused. And then you have reports like Gartner and all that are saying that you know, something like about 40% of the agent AI, you know, projects will be shelled by in the next couple of years. I would love your thoughts on this. What are your clients asking you? And you know, how do you look at this?

Speaker C: Yeah, so let me just break it down like you know, so the fundamentals of what an agent does, Leslie. And that hasn't changed. So agent basically provides a level of autonomicity and being autonomous and being able to reason. So how do we evolve? From the early days of AI we were basically a rule based system. We are completely rule based where if you had a very specific set of instructions to be done, then you can write those specific set of instruction and get it done so that you could maybe you know, as a batch process or as a file process, you could essentially go ahead and execute that. So you save a lot of human time, right? So you are basically making human life easier. All these things are nothing but trying to make human life easier. Where you know, where there is a, there is an ability where you can do things more intelligently that you can outsource to specific agents that rule based systems then moved into AI and ML, right? Where you went into deep learning where you identified some patterns saying that okay, now it's not just rules. Now rules also have some patterns of the data, right? So you understand, you analyze the patterns and then you said okay, if you ah, encounter this specific pattern then this is the behavior you need to follow. So it became more and more intelligent. And that is what you know, as you know from machine learning and deep learning and that's been there in practice for more than 10, 15 years now. Like, you know, and some of the organizations have been doing it long before. And then it was towards generative AI, now agentic AI. I'll tell you what happens is, you know, when, when Gen AI came in, it came up with one big bang foundational model, right? That was like, okay, it can answer a lot of stuff. How does it solve enterprise problems? Like, you know, if I'm a manufacturing company, if I'm doing this, how does it really, you know, what do I generate, what do I create that helps my business processes? Because a business process interacts with systems. Those systems are front ending, right? Which are the, you can call it the front office. Those systems are also managing your integrations or could be part of your middle office layer. Or it could be a back office which is doing your reconciliations, your finance, all the grind, right? The numbers and stuff. All those systems are also. So how do you bring that into. That was not available, that was not known. That's when, you know, we thought that, you know, if we have a specific set of agents which have the relevant center of autonomy, can use the power of LLM. So basically what does that make an agent? Agent is nothing but a wrapper over an LLM. So LLM is the brain, it does the thinking and agent is the wrapper on top of it. But because it was just built for foundational, now you want to use it for the enterprise. You try to bring in some context like what the enterprise want to do. So 2024 till December 2024, I can say that most of our customers were trying to build that context that, uh, how can I use this within my organization to able to solve or automate say one business process? And they were able to do that in POC stages in some form or the other. But there were various questions around security and data privacy and all that came through which made the bridge too far where they could not move from staging environment to production. So they were able to do it from development and move it on to staging, get it, you know, using real time data and even real data of the organization. But they were not able to put. Because regulators got cold feet, you know, because, um, they were not entirely up to speed of it. But now if you see, you know, European Union has come up with an AI regulation, you have the feds coming up, India is working on one. So you see now countries are now laying down the foundations. So far, the entire governance bit. Now when you, when I talk about enterprise adoption, the whole governance bit was left to the enterprise. You build, you make sure that you Know the uh, AI respond, LLM responses, filter biases is able to mitigate those kind of violent responses. You know, um, anything which is, which it should not be doing.

Speaker A: From just one, one clarification out here, Praveen. Uh, basically when we're talking about this hallucination and all, as far as foundational models are concerned, yes, I do agree. But when it comes SLMs, basically the smaller language models which are being more task oriented or the AI, uh, agentic systems, I guess a lot of it is working on localized data and where you can sort of control, you know, the hallucination part of it, if that you can clarify for the, you know, uh, viewers and listeners that will help.

Speaker C: So this is whole, uh, related to the evaluation frameworks, right? So what the evaluation frameworks do, Leslie, they have recently come in like, you know, Azure has come up with the AI foundry. You have LangChain, you have Langraph, Autogen is coming up. So many are over there, you know, which basically say that, you know, they have a framework to evaluate the hallucinations. They essentially be able to ensure that the responses that you are getting, how do you evaluate those responses, you know, how do you trust that? Why is that important? Because you are building autonomy, you are building autonomous actions. So how can you allow something? Suppose if an agent is supposed to do four actions as part of a workflow and as part of that workflow it is then going to hit a transaction which is going to be costing millions, for example, right? So the whole decision making process that includes multiple steps for an enterprise till the time it hits that particular transaction. There has to be a way to evaluate that every step has been covered properly. And that is why this evaluation frameworks between agentic AI becomes very important. And now to augment that. So there's nothing new like as Jayant would conquer, right? You know, on the MCP and A2A. What they have done is, is LLMs have been trained on with a cutoff date, right? So they never work on real time data. LLMs never have real time data. So you will always have some cutoff data on which uh, you are essentially trained in. Then you bring in some organization context, you bring in rag, so your RAG database comes into the picture along with an old trained LLM. But now given a situation that you want real time data validation, okay, for example, I have an API, okay, which can give me whether an organization or, you know, or it can give me real time route information. Suppose I'm booking a ticket, right? And I have an agent which is doing this Booking. And now if I use an LLM to do the entire trip planning, it will give me the route which is, I don't know, 2023 old. Right. So how do I do that? So I create an MCP server. Okay. I use the an MCP tool which is already created.

Speaker A: It will be, it will help if you explain to uh, viewers and listeners the model context protocol.

Speaker C: In a simple terms, protocol is exactly that. It's not a framework, it's not an agent, it's not a code, it is a protocol. It is a protocol in which how your agent can actually use external resources. So for example the example I gave you, I want to use a Google API which will give me the latest map, but my entire work is being done by the LLM, but my LLM itself is dated, right? So, but I want to use, right. So then what I'm going to do is my, I'm going to allow my agent using MCP protocol to actually hit the latest API of Google and fetch me the latest route. So suddenly you have a business case that you've built in from, for building an end to end itinerary with the most latest data with an LLM. So that's the beauty. Now uh, then from an agentic perspective you're building in, right? But like I said, you know, you're doing uh. There could be a lot of hallucinations that can come in, there can be a lot of uh, you know, checks. But you know, uh, you have made the right point. Smaller languages model have been trying to do, you know, uh, get away or they're somehow more accurate, I can say. But then you can never completely do away with it Leslie. Right. You know, because at the end of

Speaker A: the day that's why you have humans in the loop.

Speaker C: Right, exactly. And that is why autonomous agents, completely autonomous agents is still a little far cry from the industry because even regulators will not allow.

Speaker A: He will not allow, RBI will not allow, Sebi will not allow.

Speaker C: So we need a babu sitting there who says that, okay, click this. It looks good to me. It's all done and then it goes through. So I think you know that human in the loop is very, very important. And whether it's from government projects which we see, which we uh, you know, there are some governments which are very advanced in terms of doing this and some of the large enterprises that are, they are essentially using it. So. But what the MCP has done is they have given us a standardized way of how to interact and similarly like MCP gives you a standard way of interacting with multiple tools out there, your A2A protocols helps you to basically talk to other agents. So tomorrow if I am having a company which is, is uh, doing a, uh, uh, travel booking and there is another hotel booking company like Marriott, they may launch their own agent. So how does my agent talk to their agent?

Speaker B: Talk to their agent? Yeah, yeah.

Speaker C: So that is where the whole A

Speaker B: two way protocol, standardization and protocols.

Speaker C: Yeah, absolutely.

Speaker B: Yeah. I mean like how we have HTTP for the early days of web 1.0. We need an agents, AI agents equivalent of that protocol, that standardized protocol where these things can communicate and work together. Uh, I have a question for you. Uh, especially uh, since epam, you guys are in the business of digital transformation of companies across sectors. Uh, what do you make of this news that came out this week that uh, OpenAI is getting into the consulting business for $10 million plus accounts and they are using Palantir's uh, playbook of actually forward deploying the engineers. I mean OpenAI has always been this company which is building horizontal, this one. I mean they never got into verticalized, even business models. Uh, but why do you think they're getting into this enterprise consulting services instead of building AI foundational models?

Speaker C: Yeah, I mean, uh, so that's a good question, uh, to be honest, in terms of some of the business drivers that they're doing. But you have to realize that even consulting from that perspective, um, with the advent and the growth that we see, trajectory that we see of how AI is essentially going to be, uh, heading towards, I can be a consultant who can have three or four or five agents at my back pocket that I will come and release in an ecosystem and one will do the requirements gathering for me. It will ask the right questions better than me as a human. One will basically go and assess the entire landscape of what the application looks like. What are the challenges, what are the applications which have been having the maximum number of downtime, what were the reasons, what were the tickets? What is the SLA of those applications getting all those nitty gritty details. I can have another agent that can actually go down and look into. If I have these sort of data points, what is the treatment or remedy that I should propose? Should I talk about re a platform? Should I talk about completely re architecting or should I say that, you know, just grab this out, uh, demise this thing and get something new. So this is consulting, right? So now when you, so now when you are equipped with this information earlier, all this was with the human and the human experience that it came along with, right?

Speaker B: Correct.

Speaker C: So there is a curve that has been shortened in this whole process. So I won't be surprised that you know, from a consulting. And when you talk about 10 million plus consulting, um, those deals are actually coming more and more rare if you ask me, because nobody's doing really big bank transformations. Right, true. Because uh, people are doing micro transformations, they're identifying the right areas, giving immediate business value. Gone are the days, you know, where you could actually do large scale enterprise planning and then you know, come up with a billion dollar uh, you know, transformation roadmap. Because by the time the roadmap is done the whole business scenario has changed. Right. So uh, there is value uh, in what they're doing. They will be very tough competitors to do that because they own the LLM models, they own the performance, they're not going to incur the token cost. You know, if you look at the token cost for doing the maximum number of uh, searches and you know, getting the content and the uh, respective outputs, they don't have to worry about it. So I think it's a smart business move. Uh, we are not worried because uh, of the uh, the fact that you know, the domain expertise that we bring in right across every domain, uh, it is something that uh, yes, an LLM can do up till a certain extent but every business process, even if I say asset management, but the way asset management is structured for one organization is going to be, is different. The platforms, the business processes, some might be manual, some might be automated, they might already be on gen AI, they might already be already having some remote process automations. So those final nuances uh, is something you know we are yet to see if uh, purely uh, you know, an AI based uh, event can uh, essentially solve it. So it's interesting, OpenAI is going a

Speaker A: little too fast I think so because it's going into getting into consulting. It's creating that OpenAI browser, it's creating a hardware device and you know, we know it's a well funded company for now. Uh, but you know at the same time putting your finger in too many pies also gets you into trouble.

Speaker B: Constantly being in the headlines. Okay, yeah, and, and are constantly looking for monetization models.

Speaker A: Monetization. And at a time when key uh, key people are leaving you. So you know, when key executives, uh, sort of, you know, because they feel the pressure when you are having too many things. Not that the people who are poaching them are going, you know, allowing them to breathe something but uh, everybody is creating the super intelligence labs, the intel AI Intelligence lab. As if, you know, uh, AGI or ASI is around the corner. But, uh, and then Praveen has on his hand poor enterprises that ask you, please tell me what do I do with my business? Because nobody's asking the basic question, what problem do I want to solve and how do I do it more efficiently? These are the main things, whether I'm using an LLM, whether AI agentic system, or whether I'm using AGI or using ASI for that. Am I right, Praveen?

Speaker C: We constantly get that, Leslie. We constantly get that, that, you know, let's move beyond the hype, okay? We have seen, we have seen the tricks that, you know, this can pull off. But tell me how I can save like, you know, $5 million out of a portfolio. I mean, because the cost to achieve is going higher and the returns, which has been promising is something that, you know, it's got a curve, right? There is a, there is a curve to achieve that. So it'll get there. But organizations are struggling to identify the right use cases. Like I said, Leslie, when we had earlier met that uh, the uh, U turns over here can slightly be expensive, right? So you got to identify the business cases where you can actually derive strong value and then you can use that to essentially fund your other programs, uh, around Genai. And having the right use case, you can build your whole ecosystem around it. Because some of the capabilities, while you might be giving a very small capability as a microtransformation, some uh, of the capabilities that you build while doing that are going to be used as horizontals across your organization. So those become foundational pillars for you, like building in your AI hub, your governance model, uh, around it, your security. Because agents are talking like, you know, what if one agent just picks up a document and sends it to some other agent out there on the Internet, right? How do you control all that? You know, what kind of uh, guardrails you have around it. So all that has to be centrally managed. So organizations are now very, you know, uh, they're spending a lot in terms of what I understand, uh, in getting these guardrails in place, uh, because regulations are around the corner so they anyways have to do it and for their own security. And secondly, identification of these right, business use cases which can give them uh, quick value returns. And some of the business cases I can highlight, like, you know, compliance is one of the areas which uh, is very strongly, uh, you know, uh, we are pursuing this opportunity because imagine the amount of surveillance, uh, uh, infrastructure that you had to earlier build, you know, to track that somebody is giving off insider trading information or not. You had to look at every single message and try to build a context of, you know, across all these different, different conversations. And now you give it to an LLM and LLM tell you what, what is going on in that conversation. Right. In a very single NLP kind of a thing. So those use cases are very, very relevant. Uh, summarization of documents. Very relevant. Again. So. So um, I see that you know. Yes, uh, uh, it is true that gartner has mentioned 40% of those uh, you know, agentic AI projects will go, but probably that those will go because you know we haven't spent a lot of time in identifying the right business use cases and business value cases scenarios.

Speaker A: No, just to clarify but Gartner I think was very specific saying that you know the problem is that most people are using AI agentic systems for things that are not needed. That's the. I mean they were. That. That clarity was. Sorry, uh, j.

Speaker B: No, no, what I was saying is. Yeah, I mean uh, first of all people haven't identified the right business use cases to use AI to solve them. And the second is scalability. I mean you know, uh, again, you know certain, certain use cases can be solved and solutioned using AI but then at, at for certain scale beyond that, you know, enterprise level scale or deployment level scale. Again there are places where AI uh driven uh systems will break. They're not ready for that kind of a scale as yet. Be it either from a cost point of view because all gen AI uh today and uh, um reasoning AI as the usage increases, the cost also increases. Unlike the scale of economies which the rest of the world and up until now we are used to. Okay, so that is where the enterprises are feeling and enterprise AI implementations are feeling the pinch and they're not looking at ROI because the cost actually increases uh drastically.

Speaker C: Great point. Because scaling. Um, and just one. Um. I would like to reiterate what Jayanti was saying. That is kind of augmented, uh you know the uh, the process of agent operations, agent ops within organizations and finops for agents. This is exactly what happened. You know, nothing new. What happened with microservices and uh, APIs and the API economy is now happening on the, on the agent. So we will see organizations getting much and more wiser in terms of like you know, the use cases and how they use it and you know, uh, setting those token counts, uh the minimal token rather than just firing a very broad based query, those kind of things, checks and balances would essentially come in. You remember Leslie, when the cloud came in and everyone was given a cloud, uh, playing area and then they just said that we'll just put the credit card and suddenly got a $5,000 bill shocks.

Speaker B: Cloud, bill shocks. I mean was it a week or two weeks back, Ah, cursor, uh, overnight changed their uh, pricing model. Right. And all of a sudden people got bill shocks, uh, on uh, people using cursors code.

Speaker A: Okay.

Speaker B: For wipe coding. Okay. So that's already happening.

Speaker A: It is ironic because we said that the costs are falling, remember?

Speaker B: Yeah, yeah.

Speaker A: Just about six months back, I, I still remember that conversation. Oh, costs are going to fall. But the problem is then suddenly will, you know, one of these tech companies will come up with. Oh no, no. This is a very advanced reasoning model, Leslie.

Speaker B: I, I come from uh, this one where I've launched mobile phones with one MB internal memory. Okay. And then early days of mobile smartphone apps, okay. People used to say anything beyond a few mb, you cannot build the app. Today an average app is a few hundred MB and no one's caring. I mean, you know, processing memory, uh, you know, we will keep using them.

Speaker A: Okay.

Speaker B: And uh, the costs are.

Speaker A: Yeah, yeah, they're ballooning. And we don't, we have not accounted for the energy costs out here, you know, because it's scary sometimes because much of this energy is not being derived from renewables as such. You know, there's hardcore coal based kind of stuff.

Speaker C: Um, you look at the technology proliferation and adaptation for this particular technology, you know, it's not taking a decade. Within a year now you have a GPU that is getting launched literally like a SIM card. Right. You know, Nvidia is coming up uh, with that. And you know, if you look at the uh, you know the processing power is increasing, right. Capacity to process more is essentially increasing and your LLMs are getting more tuned and getting more smaller in nature. So you will see that, you know there is going to be that trade off where today it might look very expensive but eventually this whole infrastructure will become like, you know, the whole cloud and you know, the memory thing, you know that Jayant, uh, just gave an example that you know, where it was so precious earlier, right. And you know it'll, it'll kind of find its way to become much more uh, democratized.

Speaker A: But also the question also is there praveen that basically Whether we require GPUs for everything? You know, because much of the work can be done with a CPU TPO combination or CPU NPO combination. So the whole point Is you know um, I mean I know that I myself have just written a piece on whether you know Nvidia can touch the 6 trillion dollar market cap or 5 trillion by 20. Yeah, I understand it has the Cuda uh architecture. It is doing extremely well. It is locking people uh the way you know the Apple ecosystem does it. But uh, at the same time I think people are also realizing that as they come up with more slms etc they would cpu's and infuse do the job. Um Google's tpus do the job. So you know the point is whether we, these are nobody's building found models at that scale now any longer because we are done with that stage I guess uh at this point in time. So unless we require GPU now for quantum processing etc, that's a very different uh ball game altogether. Or maybe for more Bitcoin mining which is your area of expertise

Speaker C: today I have an uh, LLM deployed on my MacBook and I do my own MCP research and A2 agents and tool creations pretty much locally. I mean you don't need that. That's why I feel that you know today uh these things might be pinching the organizations in terms of like you know how uh the price and the processing uh capacity. But eventually I think it'll taper down, it'll become much more uh cheaper in terms infrastructure will not probably be uh that big a decision making. But I think to your earlier point Leslie, you know it's very important to start showing uh results in terms of how your business processes are getting optimized. Right. You know how you're able to do something more faster and smarter and maybe cheaper. Right. You know at the same time so ability to do more, you know deliver more features, you know uh, from a business standpoint uh I think those uh, are the areas where I find enterprises are now uh, you know moving into and struggling and they're now dealing with some uh really uh complex workflows. In some of the industries like manufacturing and uh banking and even healthcare I see some really really very complex workflows where you have massive amount of data getting churned into a large data lake with ingestion pipelines and then building an LLM infrastructure on top of that and talking to them. So the talk to docs architecture you should rather instead of searching the documents you have to talk to the documents payments. So I think all those concepts now the you know very interesting use cases are coming up. So I think we are probably a quarter or two quarters away. Uh in my personal view that, you know, we'll see some of the success across, ah, very complex use cases coming out. So I think it's, it's a very interesting uh, phase, uh, in my view

Speaker A: before we uh, conclude, uh, I would love you know, your thoughts on um, what do you feel, where does India stand in this entire picture? Because according to the Stanford report, if we look at it, okay, the AI index, etc, uh, I mean India's ranking as far as you know, uh, the skill set is concerned. Yes, maybe we have that kind of skill set, but not from the, you know, we have these semiconductor plans, we have the India AI mission. We have uh, I mean 1 lakh 60 crore, 1.6 lakh crore or something that has been uh, allocated for RND, etc. But you know, India has a long way to go, uh, as far as uh, AI is concerned, especially if it wants to match the prowess of countries like say China, US because comparisons may sound odious, but that's exactly what we are going to do.

Speaker C: Yeah. So yeah, from an ecosystem perspective, Lesley, I think, yes, uh, they have an edge, you know, when it comes to say a country like US or Europe, they have an edge because uh, when it comes to skilling, they have a much faster rate of skilling. The AT scale, you know, they can scale or they can skill the entire organization on a particular technology much quicker. Right. You know, when I'm talking about the entire workforce of India, it's only the people who are involved in the IT side, right, who have got the ability to do that at that scale. But if you look at the larger population. Right. I don't think, think it's, it's having that level of uh, you know, uh, impact, you know, because we are unable to scale at that sense. So in my view, I think we, we still have some latitude, especially when it comes to regulation, to wait and watch and not commit the mistakes that some of the others are essentially committing. Like we are working on a regulation. But let's look at what EU is coming up with, right? Let's look at what US is coming up with because they are also going to iterate through these, all these regulations go through some level of iteration. We have no hurry in that sense. We don't have massive level of investments in terms of product architectures coming up on the AI front where if there are organizations that are doing it, then obviously they need to do it as part of it, but then they are plugged into the larger ecosystem as well. So I think from a uh, country perspective, I think we are well poised because we are gradually building up their whole scale of learning uh as well as far as the core tech like technology is concerned. Uh, from a hardware. I still don't have a lot of point of view in terms of saying that you know, uh, because we are still far you know, a little bit behind the curve when it comes to building the kind of GPUs that would essentially be uh, required. But when it comes to partnerships, I know that there are certain partnerships that the government is essentially signing with external governments and we need to see how that plays out. But in the meanwhile, I mean even if some of these things settle down in some of the architectures, the design patterns, the security patterns, uh when it comes to uh, uh, you know, your governance patterns, let's wait and watch and you know, maybe we can adapt which is a much more a uh, slightly more refined uh outcome of these. But uh, yeah, I mean I think we are, we are rightly poised and as far as our resources are concerned they're top notch because we are working with those enterprises directly. The whole GCC boom as you see. Right. So we are essentially driving, because GCC is massively driving some of these AI adoptions in the parent companies. Right. So we are at front of that particular uh curve as well. So I think we are well uh, pretty comfortably poised there.

Speaker A: Jayant, last quick thoughts.

Speaker B: No, I guess uh, uh, Praveen summarized it uh completely. I think uh, he gave a very well this one of uh where India's positioning as he looked at it from a regulation point of view and I think I agree considering his day in and day out involved in enterprise AI Um I think looking at uh things how regulations uh are evolving and uh following that makes sense because enterprises will follow uh some kind of a ah regulatory framework uh because that needs to give them some kind of a confidence um that the technology evolution, it can be all over the place in all directions. But regulations are what will give us certain uh framework and the direction. Uh, I'm in Denmark right now and just this week you know uh, Denmark passed a bill to be the first country in the world where it gave uh every individual, every citizen the right over their face and voice for uh, AI uh, this one. So you know, just to avoid uh uh deep fakes coming in. So uh, every person has the right on of their face and their voice and if uh any deepfakes or any AI LLMs, uh uses this one to um train their models, these guys are ah eligible for compensation. So I think these kinds of things are still evolving. Um, but Um, I agree with uh, Praveen, uh, saying that India I think is fairly well positioned, uh, hardware wise. We are not very uh, deeply integrated and uh, we don't have the chops there. But software and application and uh, his point about gccs, okay. Driving the AI adoption in the parent companies and thereby, you know, India is becoming a GCC hub as we speak right now. And it's uh, for the next four, five years up until 2030 it's supposed to grow.

Speaker A: Yeah. I think with over 2,000 GCCs in India, I think that's you know, part of the strength of uh, India get

Speaker B: to 16,000 by 2030 or something by NASCOM Projection. Questions? Uh, so I think the enterprise AI work is still going to come significantly to India. Uh, and then just like the entire mobile app space, we are already seeing in the AI app space some of the uh, you know, Indian companies which are bootstrapped, which haven't raised funding, which are at 25 million ARR. Uh already. So.

Speaker A: And we're also building those sovereign LLMs where at least that attempt is going on. So a lot of, I mean four companies have already signed and, and uh, we should hopefully see uh, many more in the coming years. So. Yeah, I guess. And uh, I think to uh, to Praveen's point on the uh, legislation, uh, I think that explains why the Indian uh, digital act is being, you know, it's not on the back burner. It's sort of, you know, being delayed very. It's a very deliberate kind of uh, move and I think it's a very sensible thing.

Speaker B: I mean talent will look at Varun Mohan of Windsor stuff. Okay. They're, they are, you know.

Speaker A: Uh, yeah, but talent pool, I have my reservations on that, you know, because most of the talent pool is silicon based. You know, when we say okay, talent pool. But I don't see that in happening in India. So I'm probably, I'm a bit skeptical about that whole claim on the skill sets because these are very basic kind of skill sets. Uh, these are not really. Uh, I, I think you, you look at the IT services.

Speaker B: It is still the breeding ground. Breeding ground for talent, Leslie. India.

Speaker A: I understand that.

Speaker B: Breeding ground for AI talent. Okay.

Speaker A: I'm also a votary for that.

Speaker B: We might migrate out and that's where we get our value. And I'm seeing that happening with IT employees as well. I mean really good IT employees who were the early adopters and practitioners of AI they're already leaving the country to.

Speaker A: My only fear is that we from an AI, you know, a nation of AI, uh, coders. I mean, for coders, rather. You should not become a nation of AI coder skills. I think the skill set should go much above that.

Speaker B: I completely agree.

Speaker A: Yeah. That is my only concern.

Speaker B: More than me on that, because I, I don't know.

Speaker A: Because the point is, I remember having this chat with, um, I, uh, won't mention the deans of the IITs, but they. And, um, many, many professors whom I know have the privilege of knowing well. Many of them have voiced this concern time and again that we have other disciplines also Say, you know, in India, nobody seems to be, you know, getting. Everybody wants to do now, a computer science program and end up as a coder. Uh, that's not the, you know, India will not be able to, uh, build a great nation with that.

Speaker C: Yeah, we have to continue to invest in science. I think this is how it augments science and, you know, various industries. But the core, uh, this cannot be the core. It is always working on something. Right, Exactly. And I think, you know, uh, in India, the advice I would give to anyone is that like, you know, you'll not lose your job to AI. You will maybe lose your job to someone who knows AI. Right. So you. But you still got to do what you got to do. Right. So do what you need to know whether it's a core engineering. And then this has to be that add on that gives that real buzz to your profile. Right. You know, because then you're fully equipped for it. But, you know, uh, as you rightly said, this cannot be the only driver. Like, you know, we have to be assigned to country to grow. We have to be really investing in science.

Speaker A: Exactly. And also, you know, the AI engines are becoming so smart, the AI tools are becoming smarter day by day. So they might advise you at some point in time as to whether how to use them well. So we, we will end up in that. You know, it's not, uh, a SkyFi scenario. It's already happening in many cases. You just prompt. It tells you, actually, like I am in the media field, in the research field and all those kind of things. It tells you the one. Why don't you look at this angle? Why don't you look at this angle? You know, sometimes very humbling because. But then you also realize that, you know, uh, you, uh, don't have to really, uh, cringe because the point is that it is crowdsourced, uh, knowledge.

Speaker B: It's machine

Speaker A: and the best of minds.

Speaker B: Exactly, exactly.

Speaker A: It's a very nice, uh, experience as I said, I think Praveen rightly man. And we have. I've time and again reiterated to our viewers and listeners that don't worry about your jobs too much. I think it's, uh, more that if you don't know AI, or rather you know how to use AI, that could be a matter of concern. But, uh, viewers and listeners, we had a lovely time. I think Praveen has given us, uh, some lovely insights and, uh, take homes. Um, so, Praveen, thank you very much for your time. Um, uh, we really appreciate it. And Jayant, you have a lovely weekend, uh, in Denmark. Uh, please get us some chocolates when you come back.

Speaker B: Uh, yes, some Danish cookies.

Speaker A: So great. Uh, viewers and listeners also wish you a very lovely weekend.

Speaker C: Thank you.

Speaker B: Thank you.