Pratyush Kumar, Co-founder, Sarvam AI | "Sarvam means everybody- AI should be for everyone."| Ep. 24

Pratyush Kumar 0:00 we should build, um, AI for Indian languages, right? Because it was an interesting problem. India is so diverse, very rich culture, and it shows in our languages. So many different scripts, so many different unique ways of speaking, and we said it's a good, uh, challenge, uh, to build AI for Indian languages. You should have the ability to build it yourself, because that gives a very different stance to who we are as a country, right? Um, and I think this is very, very important for strategic, uh, technologies. Uh, and I think it is very true for AI. [upbeat music]

Speaker 0:51 Hi, my name is Amrit. We've heard that IIT Madras is the best place to build. So we've come down to the Sudha and Shankar Innovation Hub. We want to meet some people. These are builders. We want to talk to them about their work, and also ask them, "What makes IIT Madras the best place to build?" Hi, and welcome to The Best Place To Build Podcast. This is Amritash. Today I'm sitting with Pratyush, the co-founder and CEO of Sarvam AI. Hi, Pratyush.

Pratyush Kumar 1:30 Hi, Amritash.

Speaker 1:30 Pratyush, you are, uh... you like to stay away from the limelight, for good reason, so if it's okay, I will run an introduction. Pratyush runs Sarvam, one of India's hottest AI startups. They solve for, uh, AI for India, so Indian languages, Indian use cases, uh, speech to text, text to speech translation, and of course, LLM. Uh, company has raised funding from Peak XV, formerly Sequoia, Khosla Ventures, and Lightspeed. You should know that Khosla Ventures was an early investor in OpenAI also. Uh, recently, it was also announced that, uh, by the Honorable Minister Ashwini Vaishnaw, that Sarvam has been selected by the government of India for- to build India's sovereign LLM, uh, under the India AI mission. So we'll talk a little bit about that. Um, the TLDR version is that Sarvam is one of India's hottest AI companies, uh, and is deeply linked to India's AI objectives, and Pratyush leads Sarvam. Uh, hi, Pratyush.

Pratyush Kumar 2:20 Hi, Amritash.

Speaker 2:21 Um, that's a very heavy introduction, but I want to say that you were into AI long before this LLM craze began, right? So can you give us a sort of an introduction on how your journey has come to be?

Pratyush Kumar 2:34 Uh, certainly. Thanks. Thanks for the invite, and it's great to be sitting here in an Bajaj garage in IIT Madras. Uh, it's definitely a place to build, um, so, so glad to be on the podcast. Um, yeah, my background, um, uh, I studied electrical engineering, um, at, at, uh, IIT Bombay. Uh, I sort of, uh, do believe that electrical engineering is a broad topic that teaches you various things, right? From, you know, understanding linear algebra to, to writing code. Uh, and then went on to do a systems engineering PhD at ETH Zurich,

Pratyush Kumar 3:04 and basically learned more about high-performance computing, uh, reliable computing, and so on. Uh, so not really AI at that point, but then I joined IBM Research. Uh, this was the early, uh, time of deep learning. AlexNet had happened, and so people were aware that you could just make these models larger and get more out of them. But interestingly, the area was taking a turn. Uh, it was going from one where, uh, the algorithms mattered a lot, especially algorithm for computer vision, image, uh, for video, for audio separately, uh, versus it becoming, uh, very compute-driven, where you had to build efficient systems on very large amounts of data, uh, and you did much better. So at IBM Research, I got to see a bit of that and started working in this area, and then decided, uh, that it's so- it's an important area, that it's worth doing fundamental research on. So I joined IIT Madras here, uh, as a faculty member, uh, in the computer science department. Uh, and I met a very willing, uh, colleague in Mitesh Khapra, and we decided to say, "Let's focus on, uh, building AI."

Speaker 4:10 This was when?

Pratyush Kumar 4:11 Uh, I should check my dates, but I think it was 2017. I will try to remember. Um, and, um, and we decided to think about what is the practical thing that we could do, uh, with technology, but with a focus for India, right? Uh, and that's when we started, um, AI4Bharat, right? Um, and the intention was to build, um, build technology that is, uh, using the best-in-class methods that a company like Meta or Google would use, because they have access to compute and, uh, data, but do it in a way that is, uh, locally relevant, right?

Speaker 4:48 Yeah.

Pratyush Kumar 4:48 Um, so as part of it, broadly, we, we ran large courses on deep learning. Uh, we also ran efforts on hackathons, and so on. And then finally, we, uh, converged on the fact that we should build, um, AI for Indian languages, right? Because it was an interesting problem. Um, India is so diverse,

Pratyush Kumar 5:08 um, very rich culture, and it shows in our languages. So many different scripts, so many different unique ways of speaking, and we said it's a good, uh, challenge, uh, to build AI for Indian languages, and that's where the, uh, the journey began.

Speaker 5:21 Nice. So, um, get- just before we get into Sarvam, um, what is AI4Bharat? I understand that you said that it started as a, as something from the offshoot of your course, and I think right from the start it attracted a lot of interest, right? So what, uh, so what was it started as, and what is it now?

Pratyush Kumar 5:40 Uh, yeah. So we, we ran, um, online courses on deep learning, Mitesh and I, with an intent of making it very, um, hands-on. Um, these were the days when we were teaching people how to use PyTorch, et cetera, um, and that was a very popular course. Over 50,000 folks did that course online.

Speaker 5:58 50,000?

Pratyush Kumar 5:58 Yes.

Speaker 5:59 Okay.

Pratyush Kumar 5:59 It was, uh... It- I still, I was, I was just giving a keynote at AWS's event a few days back, and in the backstage, the person who was there said, "I just did your course, and now I'm at AWS building stuff." So it's, it's very, uh, gratifying to see. Um-... and that was a good, um, good experience. We could have done more of that, but then we thought, "Let's leverage the community that we have, the understanding we have, uh, to sort of, like, have a broader hacker motion around building for, uh, AI in India." Uh, but that didn't work very well. We had a decentralized setup where we said people can come sign up for problems, and then solve them, uh, offline. Uh, but, uh, that didn't really work out. So we instead then decided that we'll take one or two problems, which we run out of our lab, uh, and, uh, get it built with the help of students, plus volunteers. Right? And that's when we pivoted towards Indian language AI. So it started as a research lab, so to speak. I remember, uh, about a dozen CFI students were the sort of the founding members of that. Um, and then what we did was, uh, we started building these, uh, components. For example, one of the first MS students I had built something very trivial but very important, which is high-quality web scrapers on top of Indian web. Right? Okay. Uh, because this whole thing starts with the data flywheel, right? Yeah, so, uh,... sort of like a series of things added up, and in about a year, we had built, uh, translation systems, right, uh, that can translate from English to Indian languages and vice versa, which were competitive, uh, with, uh, big tech's offerings. Right? Uh, and that attracted attention. Uh, we managed to receive, uh, support financially from the Government of India as part of the KASHINI project, and also philanthropic money from Nandan Nilekani. Uh, in fact, Nilekani Foundation is now a key sponsor of AI for Bharat. Uh, and AI for Bharat continues as a very successful, uh, center of excellence at, uh, IIT Madras. Uh, it,

Pratyush Kumar 8:00 uh, it's probably over two hundred people now, both the language experts and the researchers working there, and probably is the foremost lab doing open source work on Indian language AI. Right. So it's really matured, uh, and I'm very glad to tell you that Sarvam is now deeply collaborating with AI for Bharat, both from a research perspective and understanding how we take these technologies to the LLM world. So you were, uh, a faculty at IIT Madras in the computer science department. You were, you were, um, collaborating with Professor Mitesh, who is still a faculty at IIT Madras for AI for Bharat, and then you started Sarvam. So maybe you can just give us that segue of why did you start with Sarvam? Who are your main collaborators in Sarvam, and s- build that story a bit for us. As I also spent some time at Microsoft Research, uh, I was seeing what was happening broadly in the AI space. Uh, um, there was the GPT stuff that was happening. Uh, so it was clear that, uh, we need to take a more, uh, uh, foundational view of building this technology. Unlike things like translation systems, which can be built with, you know, about $100,000 of compute, uh, you need a lot more money to be able to build, uh, production-grade large language model, right? So, uh, we decided to, uh, start Sarvam, it was actually a co-founder, Vivek Raghavan, who also has an interesting, uh, background, sort of built successful startups in the US, IIT Delhi guy, uh, CMU, and then, uh, came back to India very early and decided to work towards, uh, digital public infrastructure, as it's called. In fact, he was one of the founding members of Aadhaar. Continues to remain involved with them for years hence, um, and built various parts of it, including the latest face recognition stuff that, uh, received the Prime Minister's Award for Technology. Right. Um, so we decided that, um, uh, we should start something, uh, with venture capital, uh, and, uh, build it out, and so we just- we started Sarvam, and the goal was, let's build from scratch, um, LLMs and whatever is to come in this line of technology in India, for India. So a bit of a contrarian bet to say that let's build for India. Uh, people talk about it being a small market, low ARPU, and so on. But we said it's actually a large market, and it's actually a very early cycle of a long-evolving technology, and so let's focus and build this technology in India. I mean, if you're from an Aadhaar background and you've seen UPI, then you know that Aadhaar and UPI scaled way beyond a- anything anybody imagined, right? So... UPI's role in AI, I think is a bit, it's, it's a bit interesting because scale is one of the key things that distinguishes this technology from others, right? What I mean by that is two things. One, this technology continues to scale- Mm ... uh, on axes such as the size of the model, the amount of data you have, and so on. So you need to do bigger and bigger things to make this technology better, right? Mm. And it's not very clear what those limits are currently. There are some limits, but it's not very clear. So the-- you need to have a path towards scale, and one path towards scale is what the US shows with extremely large, uh, big tech companies, right? Another path to scale is probably what China does, right, with, with a very heavy-handed government. I think India's path to scale, at least in the last decade, that has been successful, uh, is public-private partnership, uh, on top of open platforms like UPI. Mm. Uh, because it's extremely scalable, low-cost utility service that provides open standards, but then people can come and innovate in the private sector and build on top of it. And people have built very successful businesses on top of that, right? I think that's a, that's a model that, uh, we feel AI can evolve towards. Yeah. You still need the scale, but you need it probably in a way in which we can leverage both, uh, the, the population scale deployments, but at the same time, ensure that the models get, uh, better, uh, from these deployments. Pratyush, I want to go deeper into that, but before that, you mentioned the word foundation. Mm. And, um, I'll be honest with you, I don't really know what a foundation model is in the-- I mean, I know that these are foundation models, but I don't know what that means. And I'm curious to know from you, uh, to go from whatever a foundation model is to an application, there must be a lot of steps involved, and what are all these things?

Speaker 12:19 ... um, who are building them separately, and why should we invest in each of these?

Pratyush Kumar 12:25 Good question. So, uh, so of course, a model is used to characterize one of these AI systems. Uh, say they can model something, often it is language or images, et cetera. Foundation models are seen as, uh, models which are so general purpose, and that they don't try to have a particular skill. Let's say, for example, writing code necessarily, or, uh, or let's say analyzing X-ray images, but are more generally capable, right? So foundation models, um, loosely defined, is a general purpose, uh, model that can be of language, and nowadays audio and images and so on, right? And there are many large companies in the world building. Of course, as you know, OpenAI has been building in the, on the first bet that we can keep scaling it and get really good models. Uh, companies like Google have been building it, Meta has been building it. Um, of course, there are companies in China building these. I reduced to, to see very large efforts in, in China as well. So, uh, I- in some sense it is, it is one of those why-- it's also called foundations, because on top of it, we can build various things, like a code assistant, uh, or, or the ability to analyze, uh, legal documents and, and summarize what's happening there. Uh, so I think the reason why different people are building them, uh, is because, um, it is believed that various things we are doing, either in the customer connect world or in the internal productivity world, will have some amount of AI presence in them. And because such a large and general purpose statement, uh, many folks are trying to build it, right? Imagine being, uh, in the 1800s and seeing that you have the, the steam engine-

Speaker 14:08 Yeah

Pratyush Kumar 14:08 ... and figuring out where all can it show up, right? And it's happening significantly faster because, uh, the cycles are down to months instead of decades. Um, and it's digital, so it can be replicated, moved around, and so on. So there's a lot of value to be created in this space, and hence we should, uh, think about building them. The other important point is, um, I think it is one of those technologies that we think of as being in a strategic nature, right?

Speaker 14:37 Mm.

Pratyush Kumar 14:37 We would talk about space or what nuclear, uh, we want up. Uh, we think of these as strategic because the outcomes they deliver, uh, can be extraordinary, right? Uh, I think AI seems to be in that ballpark. Uh,

Pratyush Kumar 14:51 and hence, it is also important for us to, in the country, have the ability to build it. So there's a commercial imperative, uh, because it's actually a super large, uh, market, and the amount of investments that have gone into it is very large, but also there's a strategic element to it.

Speaker 15:06 Um, I, I have, like, two or three follow-up questions on that. Just to, uh, cement my understanding, a foundation model, on top of that, you build applications. So something like a Perplexity is an application layer, right?

Pratyush Kumar 15:17 Uh, so Perplexity is a company, right? Uh, the way I would say it is, web search, uh, on top of LLMs would be an application.

Speaker 15:26 Okay.

Pratyush Kumar 15:26 Uh, or the ability to draw images of a particular style is an application, or the ability to reach out to customer care with voice-to-voice bots is an application. Uh, so these are applications.

Speaker 15:36 And, and in the case of ChatGPT, uh, OpenAI, they are building the application layer and the model themselves?

Pratyush Kumar 15:41 OpenAI is building, uh, of course, the models, right, the GPT models. Uh, and then they are enabling different people to build on top of that, right? They do build some applications, but maybe the ChatGPT website itself is an application-

Speaker 15:54 Mm

Pratyush Kumar 15:55 ... and becoming a popular one. Um, but, uh, they enable others to then build applications on.

Speaker 16:00 Okay. Between the application layer and the foundation layer, are there other things in between?

Pratyush Kumar 16:07 Yeah. So, uh, in fact, we see ourselves as a full stack company. Uh, again, another contrarian bet as opposed to being one layer. I- we see it as four layers.

Speaker 16:15 Okay.

Pratyush Kumar 16:15 At the bottom is, um, uh, is the ability to run these models, right? This is obvious, called inferencing. Uh, which is the, you know-- How do you ensure that the tokens that you produce from the models-

Speaker 16:27 Mm

Pratyush Kumar 16:27 ... uh, are actually, uh, efficiently done, and so that the cost of serving these models can be low.

Speaker 16:33 Okay.

Pratyush Kumar 16:33 Uh, and you need to figure out how to run GPUs well, how to compile your models down, how to optimize it, and so on. The layer above is the model building itself, right? And these could be, of course, they're LMS, but there could be vision models, there could be audio models, and so on. And there could be even a, a multiplicity of models. Uh, above that is something which we call orchestration layer.

Speaker 16:52 Okay.

Pratyush Kumar 16:53 Which is, if you want to build any application on top, typically you'll have to put together these models and other code, um, efficiently, uh, and scalably. Uh, uh, for example, we have at least two orchestrations in the company, one around real-time voice bots, uh, and another around more complex reasoning systems, right? Um, so that's the orchestration layer. Um, and then it's-- this is, this is solid software engineering stuff, right? And the layer above that is applications. Uh, and the reason we separate applications for some orchestration is because, uh, an application to be built by a domain expert.

Speaker 17:31 Mm.

Pratyush Kumar 17:32 Right? If you're building a voice bot to explain to users about your insurance scheme, um, then you can do that without worrying about the layers below, and thinking really about: What is the insurance scheme? What do people care about? Uh, the emotional, emotional design around it, et cetera, right?

Speaker 17:48 Yeah.

Pratyush Kumar 17:49 So that's why we separate these four layers.

Speaker 17:50 Okay. So inference, and then, uh, I, I, I couldn't figure out what this is. Models.

Pratyush Kumar 17:55 Models themselves, and then the orchestration and, uh, application layer.

Speaker 17:59 Inference models, orchestration, application. Interesting. And I think I met-- I heard you talk about it in some- somewhere else, that there's also a lot of challenge in organizing the data to sort of make these models.

Pratyush Kumar 18:11 Yes. Uh, so uh, there are two raw materials, uh, and probably three, four raw materials for model set. One is, of course, you need lots of data, uh, you need lots of compute, right? Um, and you need talent, because, uh, this, this is not, uh-... um, openly discussed, uh, research, uh, methodology around how to train these models. There are still some open-- there's still some unknowns out there in the open domain. So, uh, data is one of the key raw ingredients for training our models, because you're talking about, um, these models requiring trillions of words, uh, to be trained, right? And if it is audio, then it could be hundreds of thousands of hours of audio and stuff like that. So yes, data becomes an important role.

Speaker 18:53 And I guess, um, also, um, again, in the context of Indian languages, Indian languages, the data availability is much less compared to, uh, an, uh, an English language, right?

Pratyush Kumar 19:05 That is true. That is true. So, um, there are multiple aspects to that. One is there are large number of Indian languages, and not all of them have enough data. That is one. Secondly, there is the culture in those languages. They're not fully digitized. And, um, so, uh, we-- first one is more standard language tokens are required. So a token is a, a sub-word or half a word, typically, uh, and that is used to train these models. So there are, uh, language tokens you need for language understanding and fluency, but there's also culture tokens. Uh, tokens that represent the culture, which might be in an old book, uh, even, uh, undigitized, uh, material, which needs to be put out there. And, uh, and, and the third thing is that, uh, there is also new ways of using language that have evolved over time. For example, uh, how do you type Hindi or Bengali, et cetera, on the keyboard, right? Uh, you would use the Romanized keyboard and still type that, right?

Speaker 20:06 Yeah.

Pratyush Kumar 20:07 That is, of course, a new-- uh, and that's how language should be. It should be evolving.

Speaker 20:10 Yeah.

Pratyush Kumar 20:10 So you need to also represent these aspects-

Speaker 20:12 Yeah

Pratyush Kumar 20:13 ... of that.

Speaker 20:13 And in Indian languages, uh, because, um, somebody was telling me about words that have actually never been translated, like secretariat, internet. Most languages, they just use the word internet, uh, and so on.

Pratyush Kumar 20:26 Yeah, and we call that code-mixing. Uh, in fact, it's very interesting. Uh, if you've been tracking how language is being used officially in things like SMS messages, uh, there are English words-

Speaker 20:37 Mm

Pratyush Kumar 20:38 ... in the Roman alphabet, in between an Indic sentence.

Speaker 20:41 Yeah.

Pratyush Kumar 20:41 And that's okay. We are so, we are so, uh, malleable as a culture to, to absorb them, right? So these are the kinds of things you need to map out, right? Can these models be good on Romanized Hindi-

Speaker 20:52 Mm.

Pratyush Kumar 20:52 uh, Tamil plus English in between, et cetera?

Speaker 20:56 Yeah.

Pratyush Kumar 20:56 Yeah.

Speaker 20:57 Very nice. Very interesting. So, um, right, so there are, there are four layers, and in a- addition to the layers, before the layers, there's data, compute, talent to solve for. So it's a large thing. You are in it, in all the layers, uh, but also in such a way that others can build application on top of the Sarvam layers, right?

Pratyush Kumar 21:17 Yeah. So the, the way we think about it is, um... So just to give you a sense, I'll just go a little bit deeper here. Um, what we do on the inferencing layer, for example, so inside, um, the Aadhaar building, there's actually a set of GPU boxes,

Pratyush Kumar 21:33 uh, which contain our models.

Speaker 21:35 What do you mean inside the Aadhaar building? Sorry.

Pratyush Kumar 21:37 So Aadhaar is, I mean, of course, it's, it's as, in some sense, sovereign as it gets. UIDAI is the authority that runs the Aadhaar program.

Speaker 21:45 Yeah.

Pratyush Kumar 21:45 Right? And it's a completely air-gapped system because it's so critical for, uh, as a national infrastructure. And so if they had to do AI, how would they do it, right? Uh, so just to, just to characterize that, in the basement of Aadhaar, we have a set of boxes.

Speaker 22:00 Mm.

Pratyush Kumar 22:00 Um-

Speaker 22:01 You're physically talking about buildings?

Pratyush Kumar 22:02 Physically, we're talking about buildings.

Speaker 22:03 Okay. [chuckles]

Pratyush Kumar 22:03 We're talking basement, the boxes, and so on.

Speaker 22:05 Understood.

Pratyush Kumar 22:05 Like the super real, uh, GPUs-

Speaker 22:08 Okay

Pratyush Kumar 22:08 ... which contain in them several models and the orchestration layer, and bots built on top, uh, to, uh, deal with things like calls that can go out to citizens when things like biometric fails, right? So that is the-- that is the stack. You need to think about the compute. How does the compute actually run all these complex models? What are the models that are required to do a particular task well? It's often just not the model, but everything around it, that is the orchestration, right? How do you keep this thing running? How do you have telemetry? How do you have low latency? How do you have reliability, scalability? And finally, there's application. So in this case, it is a call going out to a senior citizen when biometric fa- fails, right? Uh, or take the example of a large insurance company who we are working with. Uh, when there is a renewal that is due, they would like to inform the customers at scale. We're talking about crores of people-

Speaker 23:02 Oh

Pratyush Kumar 23:02 ... in their language, about what is this insurance claim, what does it provide, right? And what benefits does it provide, how to use it, and so on. So that is the application layer, and, uh, we, uh, we work with partner companies to then build applications for them on top of the stack.

Speaker 23:18 Okay. Okay, I understand. I mean, um, I've seen often when people talk of applications, uh, they give very simple examples, but the examples you're giving are actually large, complicated example. I remember one example from earlier where somebody was talking about Supreme Court judgments being available in multiple languages and things like that, which is, I guess, simpler than what you're saying.

Pratyush Kumar 23:43 Right. That's also an important use case. In fact, we are working closely with a few s- uh, high courts. Uh, there, there's a mandate from the, uh, Supreme Court to make judgments available in local languages.

Speaker 23:53 Mm.

Pratyush Kumar 23:53 But it's not a trivial problem because it's, doesn't, doesn't fall under the usual translate card though, kind of a scheme, right? Because there are le- there's legal language-

Speaker 24:02 Uh, yeah

Pratyush Kumar 24:02 ... there are words that need to read- be written in a particular way. Uh, there's some things that require-

Speaker 24:06 Oh, the-

Pratyush Kumar 24:06 ... clarification

Speaker 24:07 ... the co- the, the cost of an inaccuracy is also pretty high, right?

Pratyush Kumar 24:10 Cost of inaccuracy is high, but, uh, I see it more as, uh, optimizing for accessibility, right?

Speaker 24:16 Yeah.

Pratyush Kumar 24:16 Uh, you want the judgments to be available to people in a way in which they can understand it, where, where it's a common joke that you can't read legal language, right? So that is the key, uh, thing there. But, uh-

Speaker 24:26 Now you can't read legal language in your own language. [chuckles]

Pratyush Kumar 24:28 Yes. No, the intention is to make it simpler. Um, no, but, but the, uh, the use cases are actually very, very diverse. I'll give you another one, just, just for, uh, sake of, uh, uh, difference. Uh, we're working with NITI Aayog, which is the primary think tank body of the Government of India.... to, uh, take data sets available across central ministries and the state governments. There are about five thousand very large tables that MediaIO has right now, and build a layer by which you can ask, ask a question in natural language. And in response to it, AI would first introspect which of these tables should I look at, and it is huge number of tables, right? Uh, identify the right ones. Write code, typically in Python, to access the data, to massage it. Uh, often it'll make mistakes in this writing of code, it self-corrects. Then gets the data, analyzes it, writes a report with the plot-

Speaker 25:20 Nice

Pratyush Kumar 25:20 ... or depending on what is the, what the question requires. And, uh, this has been-- we've been building this for the last quarter, and it's, uh, it's working very well. Uh, what it, uh, what it does is it build this intermediates policy work-

Speaker 25:34 Mm

Pratyush Kumar 25:35 ... from data, right? And often there are many, many steps required to go from policy work and the question that you might have on policy perspective, to actually seeing the data and the results. It completely takes that away. And not just, uh, data in tables, but also things like PDFs. Uh, and you might have economic survey of many, many years, which contain insights in different parts. We're able to also look at that, combine it with the data, and give a result. Uh, that's another use case, and I think this is going to be the other, um,

Pratyush Kumar 26:04 very large mode of consumption of AI, which is doing your work much better, especially in dealing with large amounts of data.

Speaker 26:13 Nice. I think the example you gave, uh, many times, uh, policymakers want to get a policy report on something, say, you know, like where to build a highway or where to invest in a particular crop. And by the time the report is ready, their stint has moved-- they moved on to another location, another stint, or the politicians have changed because this is a democracy, and so on. So I guess the impact of having quicker access to reports will be really large, right?

Pratyush Kumar 26:39 Yeah, I think so. Uh, so it, uh, works across the hierarchy. Even the leaders, I guess, they ought to get real about what it is that, uh, the data says-

Speaker 26:48 Mm

Pratyush Kumar 26:48 ... um, instead of hearing various layers of PP-

Speaker 26:51 Yeah, better.

Pratyush Kumar 26:52 Right. So I think it, it, it works, it works, uh, across the hierarchy, and yeah, uh, data is what we should be driving policy decisions on.

Speaker 26:59 Mm.

Pratyush Kumar 26:59 But that promise has never been, uh, true earlier.

Speaker 27:02 Nice. So cool. I think we are, uh, sort of, um, skirting around the topic, but let's get to it. What is sovereign AI? What does it mean? Uh, what is the context? A lot of countries are getting into it. It's not just India. So what are all the other countries doing? What is India's focus in this?

Pratyush Kumar 27:17 Yeah. Um, no, I just want to start with saying something else, right? So we, we are, we are, uh, shooting this at a time when, um, the armed forces are battling in the borders. And firstly, uh, thanks to all the effort that, uh, the jawans are doing, which ensures that we are sitting here chatting. Uh, the reason I say that is because, um, I don't think everything is out there yet, but, um, a lot of the success that the Indian army had and the armed forces had, uh, was because of indigenously built technology over the last decade. Right? Um, and, um, so in some sense, the way I see it is, you need, you need better tech to have the upper hand.

Speaker 27:54 C- Can I just ask you a question? Did it surprise you that so much Indian tech was used?

Pratyush Kumar 27:59 Yes, yes. I was not, uh, sufficiently educated about what's happening in that space, and I was surprised, yes.

Speaker 28:04 Yeah. So, so was I. I was like, uh, surprised, but also proud in a sense that, okay, we are here, that so much Indian tech could get used.

Pratyush Kumar 28:12 Correct. Uh, and I think we should know more about it, especially in places like here in Madras.

Speaker 28:16 Mm.

Pratyush Kumar 28:16 Because, see, you need superior tech to have the upper hand, right? Uh, but in today's world, I think you need strategic autonomy to have the winning hand, right? And what does that mean, right? You don't want to decouple from the world. India has never wanted to decouple from anybody, but you need to have the ability to build things grounds up, from scratch, as we say, yourself, right? Uh, happy to collaborate with whoever in the world to build

Pratyush Kumar 28:43 whatever, uh, but you should have the ability to build it yourself, because that gives a very different stance to who we are as a country, right? Um, and I think this is very, very important for strategic, uh, technologies. Uh, and I think it is very true for AI.

Speaker 28:57 Okay.

Pratyush Kumar 28:57 Um, and that is the sort of the mood under which the sovereign AI is being discussed, right? So one is that ability to have the talent, the resources to build this tech grounds up yourself. The second one is the ability to, um, use it at scale, right? Um, you need compute, you need power, you need, uh, application developers to then leverage the foundational technology-

Speaker 29:23 Mm

Pratyush Kumar 29:23 ... uh, to create value for citizens, right? Uh, in fact, I think, uh, uh, just like power, electricity consumption was used as a proxy of development for the last couple of decades or even more. I think AI could start looking like that soon.

Speaker 29:38 Mm.

Pratyush Kumar 29:38 It's that kind of a general-purpose technology which helps productivity, it helps, uh, you know, uh, deal with misinformation as strategic sector use cases. It might start looking like your per capita consumption of AI is a decent proxy for how advanced or competitive you are as a country, right? Uh, so in, in that context, right, the idea of sovereign model is to be able to build it in India, grounds up, have the talent that can start building it now, but notice that we are very, very early in what is going to be a long technological cycle.

Speaker 30:11 Mm.

Pratyush Kumar 30:11 And so in 2030, uh, what is the world building in AI, and how close are we to the state-of-the-art, right? Um, that's the kind of thing that we should optimize for, uh, and that requires us to start building now, grounds up. Uh, and that's the mood in which I think of, uh, sovereign AI.

Speaker 30:28 Very interesting. Also interesting that you said long technology cycle, that's in 19-- uh, to 2030, which is just five years ahead. Like, I think when we talk of long technology cycles, in the past, it's been like thirty years, forty years, ten years. Now, it's just like we're talking about five years, and everything will change.

Pratyush Kumar 30:44 The interesting thing is, uh, five years will change-... but there's also a thirty-year window, right? [chuckles] It's, it's not even, uh, clear to most people what it looks-

Speaker 30:52 What it'll look like

Pratyush Kumar 30:53 -data.

Speaker 30:54 Okay, so, um, just a little bit more detail on that. Um, uh, other countries are also building their sovereign AI. So are you saying that they also are thinking of data, compute, talent, models that they own-- that, that the country has access to? Uh, maybe the c- maybe the government doesn't own it, but, uh, uh, a set of companies in the, in that space own it, in that country own it.

Pratyush Kumar 31:18 Yeah. So, um, uh, I think ownership, et cetera, these are details that different countries will, uh, figure out differently. But I know that Japan has a sovereign AI program. South Korea has one. Uh, France, which was the host of the, uh, AI summit, last time we were the co-host, uh, they have a large program. Uh, various other countries like UK have, uh, announced large, uh, investments, right? Uh, because it's, it's, uh, just for you to recognize, it's, it's actually, um, a ball that needs to be set rolling. For example, look at GPUs, because the IndiaAI mission, which is the government's flagship program, uh, invested in building these models, we are going to get large amount of GPU compute coming into the country, right? So in this case, the government is catalyzing it. I don't think the government should be, uh, the only one building it out or supporting it, but a catalysis is required where, for example, can we have compute, uh, of the scale necessary to build out these models, to consume these models, et cetera? And I think we are far from that right now. There's a long way for us to go. Uh, but I see that India being, in the medium term, a very large producer and consumer of this technology, right? Uh, and that, that job starts now.

Speaker 32:31 Oh.

Pratyush Kumar 32:31 And we should then be able to also, um, help the broad global south with this key technology, like we've done in other cases.

Speaker 32:39 Sure, fine. Makes sense. But if you look at the India, uh, compared to the countries you're talking about, South Korea, Japan-

Pratyush Kumar 32:46 Mm

Speaker 32:46 ... France or others, they, they have a single language-

Pratyush Kumar 32:51 Mm

Speaker 32:51 ... uh, in the entire country. That's like, France speaks French. I, I don't know if there are other languages. Maybe if I'm wrong, somebody can correct me. But in India, they're like, I don't know, many, many languages, right? So our problem is uniquely different.

Pratyush Kumar 33:06 Yes. Uh, the problem is, uh, more challenging. Uh, uh, it is not, uh... In fact, people don't recognize it enough, um, which we really, truly want to build AI that works for everybody. In fact, Sarvam in Sanskrit means everybody, everyone, because the intention is that it should be used by everyone. Um, that's a very hard part, right? But can we target that first ten, first fifteen, first twenty-two languages? In fact, what we are going for now is the first twenty-two scheduled languages, as per the Constitution. It's definitely harder problem still, right? Uh, but I think it's something that can be worked at, um, and one can optimize. So language is important.

Speaker 33:44 Can I ask you a question?

Pratyush Kumar 33:44 Yeah.

Speaker 33:45 Will it be possible someday that in, in the upper house or the lower house, somebody speaking in their own language, and everybody has, uh, headphones where they can hear it in their own language?

Pratyush Kumar 33:56 Um, why not? In fact, uh, there was recently a podcast between Lex Fridman and Narendra Modi, right? And, um, uh, we-- the team put out later, um, videos of Narendra Modi's speech in different languages, right, in his own voice. Uh, in fact-

Speaker 34:12 But that was not live, right?

Pratyush Kumar 34:14 Not live. So that, that was the Prime Minister's office themselves tweeting it. Uh, but this can be taken live.

Speaker 34:19 Okay.

Pratyush Kumar 34:20 These, these things are becoming much faster, that they can be taken live. And speech-to-speech translation is a fairly, uh, well-understood problem, and one can look at it. The ability to do it in their own voice, and the, and that, that intonation gives it even more power, right? So I certainly believe we can do this.

Speaker 34:35 Nice. Um, also I, I think,

Speaker 34:40 um, and we sort of discussed it before the podcast a bit, um, from a, uh, Government of India or from a national perspective, there's been a lot of investment in these large population size problems, from Aadhaar to UPI, to now Sarvam AI. Can you give us some... To now LLM, sovereign LLM. So can you give us some, uh, like, context around that?

Pratyush Kumar 35:02 Um, yeah. So I mean, uh, the, um, build-out of, uh, let's say Aadhaar is a good example, where, uh, technology of its-- of the cutting edge was used, um, by the government to build out, uh, a unique example in the world. For example, in the US, there is a little bit of tension between Silicon Valley and Washington, DC, right? Uh, but-

Speaker 35:25 Little bit is, uh, okay.

Pratyush Kumar 35:27 Yeah.

Speaker 35:28 There is tension. [chuckles]

Pratyush Kumar 35:29 There is fair tension at changing the-

Speaker 35:31 There's an X amount of tension, and X changes with time. [chuckles]

Pratyush Kumar 35:35 And X is also a loaded word in that context.

Speaker 35:38 Right.

Pratyush Kumar 35:38 Um, but, um, yeah, I think India has shown that you could sort of, like, bring two, these two things together. Um, get great technocrats to come build technology, which is cutting-edge. Look at BHIM AHAAR, right? Which other country has something that works so seamlessly, and that's all tied to the same identity system that we have, right? Um, and I think the, the core principles there were: build cutting-edge tech, have open standards, right? Allow people to build on top of it. For example, the enrollment on the mobile phones that happened in the last decade, right? Huge part of it. Or online banking accounts, that happened on top of ID systems, which made it very easy to do eKYC, to, uh, to actually for even institutions like banks, to have the confidence that they can transact online with completely new customs.

Speaker 36:27 Right.

Pratyush Kumar 36:28 So I think this principle of thinking of utility, technology providing best-in-class utility, that somehow has the stamp of government, the trust of government, the scale of government, uh, and then people building on top, right? I think that's a good example, and it then followed through with, uh, UPI, uh, payment system, um, with so many transactions right now.... and a few other things. There, there, there is the FastTag thing that happened. Um, then there is, of course, NPCI was created, that does many other things now, including, uh, payments for our utility bills, et cetera. Um, so I guess this is a good template, and it's encouraging to think about-- I don't think in AI we have yet figured out what that model is.

Speaker 37:09 Yeah, actually, I was going to ask, when you say- use the word open, I mean, not in the context of open AI, but open standards. Um, you mentioned open standards, which is, I guess, different from open source.

Pratyush Kumar 37:18 Mm.

Speaker 37:18 And which parts of this, this-- all this, the entire AI setup is going to be open standard, which is going to be open source? And what parts of it will be, um, common to all, and what parts of it will be proprietary?

Pratyush Kumar 37:30 Yeah. Oh, that's a good question. I think, uh, as I, as I was saying, I think the-- in the AI world, this has not been fully understood.

Speaker 37:35 Mm.

Pratyush Kumar 37:35 Uh, but, uh, the-- sort of the metaphor that I carry forward is to think of AI as utility, right?

Speaker 37:42 Okay.

Pratyush Kumar 37:43 Um, just like you access today UPI with, uh, UPI's API or you have ONBC with set of APIs for looking at catalogs and so on, they're utilities. Uh, I think AI should be seen as a utility. Uh, it's, it's serving should be so efficient and available, uh, and cost-effective, uh, that people then are encouraged to build applications on top of it, right? Uh, somebody building an education application, uh, should be able to build it and make available widely, right? Uh, in fact, I, I think... I was talking about tokens, which is what these LLMs work with. Every, about half a word is a token. Um, it's easily possible to im- imagine that in the country in a few years, uh, every person has access to lacks of tokens per day, right? Because whatever you do, whether it's work, whether it's health, whether it's family, whether it's mental well-being, in various aspects of life, you can leverage this technology. And if it is available as a utility and developers build very interesting stuff on top, I think there will be a, uh, a widespread use.

Speaker 38:46 Understood.

Pratyush Kumar 38:46 That's the world we have to build.

Speaker 38:47 And, and you said that, uh, then AI per-- AI use per capita would be something that we

Speaker 38:54 would start thinking about and like-

Pratyush Kumar 38:55 I think it's a good, uh, it's a good metaphor in my mind, because it just tells you the scale we can achieve, right?

Speaker 39:00 Mm.

Pratyush Kumar 39:00 Uh, given the s- the size of our country and the need, um... The size of the country, the tech readiness, uh, we have adopted digital at a huge scale.

Speaker 39:10 Good proxy for what?

Pratyush Kumar 39:12 A good proxy for how big India's use case in AI would be, right? Because today, we don't have a sense of, uh, how many GPUs do we need in the country, right?

Speaker 39:23 Uh.

Pratyush Kumar 39:23 Why should we be building these models? This costs so many dollars. Why should we be building it, right? Because if every citizen was X percentage more productive, was X percentage able to utilize this in their day-to-day lives, the connect between a citizen and the government was X percentage more effective-

Speaker 39:40 Mm.

Pratyush Kumar 39:40 -that outcome is very large, right? And we should be thinking about how we can make that happen with AI as a utility, and then developers building on top of it.

Speaker 39:48 Understood. So AI per, uh, capita would sort of, uh, give a sense of how much of advanced technology the country is using, and therefore is more efficient or more productive or more, uh, uh, able to connect with services around themselves?

Pratyush Kumar 40:07 Right. And also, I, I would say, um-

Speaker 40:10 And, and I guess it's similar to, say, UPI transaction or transactions per capita is a good metaphor for the financial strength of a country or the-

Pratyush Kumar 40:19 Right

Speaker 40:19 ... fluidity of a country as well.

Pratyush Kumar 40:20 And that because the UPI that you and I do versus the UPI that somebody in rural Bihar does, the experience is the same, right? And then that leveling, that flattening is required, and that's what AI will be able to provide. If you had to ask something to the government about, let's say, some particular payment or something, the experience that you have and somebody has in rural Maharashtra should be the same, right? And I think that is possible. That, that is why I think technology has been, um, that flattener. And with AI, we should aim to do that at, at a much larger scale.

Speaker 40:52 Pratyush, I'm curious to know, um, there's a lot of money that's

Speaker 40:57 flowing into this business, so I want to understand, why does it need so much money? Where does it get spent? And, uh, maybe you can take us through all these layers that you spoke about and what takes the kind of costs that we are talking about.

Pratyush Kumar 41:10 Yeah, sure. Um, I think one of the primary costs is the training of the model itself, right? Um, and these are large models that need a large number of GPUs to train, and often there is work to be done before that in terms of, uh, preparing the data. Uh, in fact, uh, the way Jensen Huang from NVIDIA puts it, he thinks of, um, GPU factories.

Speaker 41:31 Right.

Pratyush Kumar 41:32 Uh, these are just like a normal factory. They are churning out stuff. They pre-- They churn out even the data. The datasets used for training the models

Pratyush Kumar 41:40 also come from lots of processing that you can do on GPUs. Um, so there's, there's data preparation, there is the model training itself, and often this is a very, um, competitive space globally, so talent is expensive as well to get the best-in-class people, uh, to do this, right? Um, and then it's not enough to just throw a model over the wall, at least in what we are doing. There's also work to be done on the software engineering layers and also then building applications. Um, I see that this is, this is a phase where, um, the technology is still in that early part of that S-curve, right, in terms of deployments in India. Once, once we hit scale, once we hit critical mass, uh, I see this extremely profitable, right? So that is also true. In fact, that is one of the reasons we should be building this technology in the country, because the value loops should remain in the country, right? Uh, there is value in deploying this technology, using that deployment experience to improve the models further, and this value loop, uh, needs to remain local.

Speaker 42:42 Can, can you just explain that a bit better?

Pratyush Kumar 42:43 Uh, so, uh, unlike other technologies, let's say you built a, you built a camera, right? And you can put it out on the market, and you'll get some feedback, and the next camera that you build can be slightly better, right? And this probably takes, uh, a year, two years to get that slightly....

Pratyush Kumar 42:58 with AI, you build a model and put it out there to see what's happening with the model, and in a matter of months, you can build an improved model that does well on those kinds of tasks, uh, but at a smaller form factor, maybe reduce the cost or maybe increase the accuracy. Right. And this, I think, pops as a value loop that can be created. Um, and this is very fast. It's a matter of months, right? Compound that to a few years, we're talking 2030, uh, we need to have, in the country, the ability to kickstart this value use-

Speaker 43:29 Okay.

Pratyush Kumar 43:29 -to make this work.

Speaker 43:32 Okay, amazing. So, um, so it's, uh, prep- prepping the data, training the data, attracting the talent, the cost of the GPUs itself, um, these are the things that cost that much, right? Will it, will it get lesser? I'm also-- I know you've raised a certain amount of funding. I don't know if it's a low amount or a high amount. Does it feel like it, it, it, it may, like, be a lot more that you'll need, or maybe it'll turn out that the costs have come down and it is sufficient?

Pratyush Kumar 43:59 So firstly, the costs are coming down from what people expected a couple of years back, right? Uh, very large companies in the US have, uh, raised hundreds of billions of dollars, right? Um, and, uh, that, that was exorbitant. I think the costs are coming down quickly because people are understanding how to do it more efficiently.

Speaker 44:15 Mm.

Pratyush Kumar 44:15 But it's still going to be cost, uh, prohibitive for multiple folks to do it. Like, it's not-- there are not gonna be hundreds of foundation model builders, right? But there is room for tens of them, to be fair, right? Um, and the cost will come down, and there will be more value at different layers of the stack. But to be- to keep pace with the foundational work that's happening across the world, you still need to keep pushing, uh, the research thread. So I think many folks could build models at, at lower price points, but you still need to keep pushing if you want to be close to the state-of-the-art in building it from scratch.

Speaker 44:51 Okay. Very interesting. Um, things in AI are moving so fast, I feel like... I mean, I'm, I'm in the outside, so it's-

Pratyush Kumar 44:59 Mm

Speaker 44:59 ... a little like, it's, uh, uh, overwhelming. I don't know if you feel so. In fact, um,

Speaker 45:05 can you give us a sense of how it feels like to be in this business? Does it feel like you're in control?

Pratyush Kumar 45:11 Uh, mmm, yes and no. I, I think definitely we're moving too fast for comfort. Uh, and one of the key things to recognize is to, uh, see the medium-term horizon, uh, strongly enough every day. There might be some fuzzy news, or this happened, that happened. I think one is to see, uh, s- see the medium term and what we are building very clearly. Uh, also, I think it's important, because there's so much money around, there's a lot of, uh, uh... I mean, one needs to be conscious of what the key goals are, right?

Speaker 45:41 Mm.

Pratyush Kumar 45:42 Um, and at least for us at Sarvam, it is about democratizing this, uh, technology, uh, in, in a country like India, which is going to have very crucial three decades ahead, right?

Speaker 45:52 Mm.

Pratyush Kumar 45:52 Um, so that is our, that's our focus. Uh, so with that gives us a little bit more stability. We know what we are here for. We're not for build the next, next week or the next quarter, we are do something, uh, diff-- We have to do the interesting things as well, but I think we have that a little bit more clarity that we are in the long, long haul.

Speaker 46:11 Very interesting. Um, and, and, uh, this ecosystem,

Speaker 46:15 I think, a little bit we wanted to speak about the ecosystem also. I feel-- By the way, I feel so lucky because I get to meet people like you, and I know you're very busy. I think it took us, uh, due to audience, it took us four months to schedule this. So that's how busy, um, Pratyush gets. Um, and I'm very lucky, and I'm very grateful. Um, mm, does the-- Is there ecosystem benefits? Are there, like, benefits to having AI for Bharat here, and access to a student community, and all of those things? Um, from far, we have heard that a lot of Silicon Valley benefits are simply because there's an ecosystem there. Is, is that happening here? Uh, is there a network effect that's building up?

Pratyush Kumar 46:59 Um, there is, there is, uh, and AI for Bharat and Sarvam is a good example of the help, but there's a very, very long way to go, uh, for us to, uh, connect, um, the academic and, uh, student body kind of setup, uh, with what's happening in this technology. And it's so common in the US for sophomores to think that they could go build a startup, right?

Speaker 47:22 Right.

Pratyush Kumar 47:23 Uh, that is building here, and I'm hearing a lot more from students, especially from IIT Madras. Um, but there's a long way to go. And I feel, uh, you need to have, um, uh, more institutional support to, uh, in some sense, cross-fertilize, right? We need to bring in the VC community, that startup community, into the campus much more, um, and expose, uh, students to that, right? And,

Pratyush Kumar 47:47 um, I think with places like Nirman, CFI, and IIT Madras, you're doing that. Uh, but there, there is room to do much more because every, every six months is very different, right?

Speaker 47:58 Is there a s- is there a fear of... Uh, I mean, you mentioned that sophomores in the US have this kind of blind faith, even on their first year, they'll build something that'll, you know, conquer the world kind of thing. Um, and they have access to those kind of professors and those kind of-- that kind of capital, which is what you're saying, that's what's required. Uh, but they have built that over decades, right? And we are building it in, like, five, 10 years, maybe 15 years. Is there a fear that it's going to sort of not build very authentically?

Pratyush Kumar 48:33 Uh, I mean, you have to design systems well, and that, that always remains the case. But I feel that, um, my own personal journey, since I've traveled many of these roads, um, I feel that there's a-- this is the time to cut the crap, right? You should either be builders or sellers, right? Um, and sometimes we remain in abstractions, uh, which are not that helpful, both in academia and in companies.... right? Uh, and I think a startup is a very, a good way to get real, right? If you're really building something with your hands and, and seeing what is coming out, you, you get super fast feedback, right? Um, and I think that's, that's the, the thing you should create more of.

Speaker 49:19 Right. So we were talking about ecosystem building, and uh, uh, w- you were saying that, uh, even though we- I mean, I- in essence, we don't have time, so we have to sort of, like, really put our heads down, use our hands, and build. That's it.

Pratyush Kumar 49:36 Uh, we should, um, not worry about, uh, breaking existing structures. I, I think i- if the feedback is very real and very here and now, I think that's a good system, right? And somebody trying to build and they didn't like it, or they don't have success, they can figure out what to do next, right? But we should, we should throw more, uh, more energy into this. And get, uh, get people who've done it. I mean, there are enough, uh, good startups that have made, uh, made big success happen in the last decade. We need to get them all here, not for talks, right? Talks are fine, uh, on, on, or even podcasts. [chuckles] We should get them-

Speaker 50:14 Thanks for that. [chuckles]

Pratyush Kumar 50:16 [chuckles] No, no, this is a good starting point, but I think we should, we should get them here for actually helping put together structures-

Speaker 50:22 Mm

Pratyush Kumar 50:22 ... uh, that they know from their experience is useful. And I think this mindset of either I'm a builder or I'm a seller, um, and get really real about it, especially when the world is changing every six months, uh, is super important.

Speaker 50:35 Okay, um, we met, uh, Professor Ravindran earlier in the podcast. He's, uh, extremely inspiring, and now he's heading the new AI department. I think, uh, uh, we are one of the first of the original IITs to launch an entire new department. Again, well-funded, uh, fund- alumni funding, and I, I don't know whether that means whether the funding is, um, in terms of compute or in terms of cash, but whatever. Um, they have a bunch of new courses, including BTech in AI. So I think there's a lot of talent coming in, right? Uh, and um, um, what I wanted to ask was, uh... I mean, my original question written down is very different, but now I'm- as you're saying it, I'm thinking, is there enough talent coming in? 'Cause the, the c- the way you have set up this whole problem, it feels like all- it requires like 1,000, 2,000 really smart

Speaker 51:28 folks to come together, right? Or maybe even more.

Pratyush Kumar 51:31 Oh, much more, much more. Uh, a-and, uh, see, uh, I think we have an education system in the country that's producing certain output. You won't go change that all of a sudden. But in higher education, we have a great advantage. Uh, people coming to places like IIT Madras, we should instill in them extreme self-belief, that no matter what, you are, you are fine, right? And this is a time to express yourself, time to build and stuff like that. I think we have a very... I mean, the education problem is much, much larger, but I think we are at a, at a point of that pyramid where we should do much more, uh, to instill that self-belief and make it happen. And by the way, it just doesn't have to be IIT Madras, uh, students. We can be- we can open our doors to other students to come in. Uh, as long as we have an engine that can demonstrate to people that this is how you come and build, and the research park is a great example of it. It's just not IIT Madras folks. People from outside have come to station themselves there, uh, building great companies. Um, I just think we should under access.

Speaker 52:29 Yeah, fair enough. I mean, uh, you studied at IIT Bombay, your, uh, co-founder's from IIT Delhi, and my, my dad went to IIT, but, uh, Kharagpur, so, um, quite a mix. Uh, and, and I definitely don't wanna be beating the IITM drum too much. Uh, can you give us an idea of what else is happening across the country, top labs, top universities?

Pratyush Kumar 52:46 No, certainly. So I think, uh, uh, there's a lot happening across all the strategic sectors. Quantum is something that is really taking off with a lot of effort across cross IITs to do that. Even in AI, there are multiple efforts that are going on, which is, which is very welcome. Um, the um... I, I think whether, whether I look at, uh, strategic sector, for example, I see a lot more willingness from the government and strategic sector to engage, uh, with, um, with researchers and startups. Um, so I, I guess we are in that moment when we realize that for all things we do, whether it is,

Pratyush Kumar 53:22 you know, commercial output, uh, ensuring that we are, uh, taking care of people across the country, or building strategic strength, technology, especially deep tech, is gonna play a key role, right? And I think, uh, gone are the days we believe that you should buy every, uh, deep tech stuff come outside. I think the next decade or so, uh, will be a, would be a sort of like a very important phase for us to build deep tech, right? I don't see we have done everything that we need to succeed, but I see things are moving in the right direction. Um, and we, we just need to talk about this much more, to give examples of this much more, and encourage people to stay back. And this is al- also at a time when the world is shaking a little bit-

Speaker 54:05 Mm

Pratyush Kumar 54:05 ... uh, in terms of its, uh, globalized order, like you had mentioned. So it's, it's a great time to have talent in India and build for India.

Speaker 54:12 Mm-hmm. Yeah. But, you know, I'm just talking to you, and, and we've been doing this podcast for a while now, and, uh, this is, this is one of the more serious podcasts, in the sense that, uh, it feels like half of it went in talking about global security and national security, and so on. Um, the students joining now or who are coming into, uh, this field, uh, are young, so a large component of their life is also just living, right? And, and, uh, does it, does it feel like the getting into AI would mean that you are sacrificing everything and just going for a cause? Is it like that?

Pratyush Kumar 54:50 Uh, of course not. I, I, I mean, you, you can just build fun projects in AI that generate... Uh, example fun project is, uh, can you generate art, uh, based on Indian styles?... and it's a fun project. I don't think you're staking, uh, your career and value

Pratyush Kumar 55:07 system for it, right? Definitely not. Uh, but-

Speaker 55:09 Oh, sure, sure, sure. I, I quickly understood what you mean. Like, instead of Ghibli art form, can I do a Wardley art form of, uh, my family photo?

Pratyush Kumar 55:17 What I'm trying to stress is that, uh, we don't tend to recognize when the forces change a bit, right? And what, uh, we see is AI has that, uh, that... Why, why is there- why is governments talking about AI? There's an AI summit, and, uh, prime ministers and presidents are meeting and talking about it, and we are doing podcasts about item address and AI, right? So this has larger connotations for sure. Uh, but of course, it's a very, very interesting technology, algorithmically, systems engineering-wise, uh, from just building applications, right? Um, and this is like a-- such a rich period to build, right? The serious overtone is because, uh, um, at, at a national level, there are things to do, but at an individual level, this is a great time to study and build this tech. And by the way, it is in fact, not just, uh, those interested in computer science or programming. I think it's becoming something where whether you're a lawyer, whether you're an artist, whether you're a scriptwriter, you're going to use this technology a lot, right?

Speaker 56:15 Right.

Pratyush Kumar 56:15 So it's going to become very interesting from the point of view of combining what you know with what this technology is.

Speaker 56:21 Yeah, fair enough. I mean, in, in-- as a marketing person, in our team, we use a lot of AI in almost every part of our work, and, uh, our workflow has changed so much in the last three years. We actually have a weekly meeting of the entire team to discuss what new AI modes are there for us to use.

Pratyush Kumar 56:41 Just weekly or daily?

Speaker 56:42 No, not daily. Uh, daily. We are not that fast. [chuckles] The community is very fast. Um, yeah, uh, yeah, I think, I think it's, it's, it's quite cool that AI has come into our lives in such a big way. Earlier, it was, you know, under the hood, error correction, KYC, stock prediction. We-- It's something that you would hear other people doing. Face Unlock, uh, I think iPh- iPhone's Face Unlock was a big moment. Uh, maps, navigation, but now it's here, like, in our hands, in our, um, in our, um, palms. It's quite remarkable. And, and

Speaker 57:20 I guess what you're saying is that it's gonna go further, further, further, and it's gonna be a big part of our lives, like UPI and Aadhaar is, or maybe more than what UPI has.

Pratyush Kumar 57:28 Absolutely. I, I would...

Speaker 57:30 Mm. Question for you: Do you think my questions were AI-generated?

Pratyush Kumar 57:35 Um, I would say partially yes.

Speaker 57:38 I, I did have this issue, uh, I think, uh, in our research. So what happens is that we send our questions, uh, and our research work to the guest early on, and you pointed out a bunch of mistakes, uh, in our research work, and all of it was LLM- [chuckles] generated.

Pratyush Kumar 57:53 Oh, yes.

Speaker 57:54 Yes. Uh, in fact, I, I, I was surprised because, uh, we use LLM to sort of do research on our guests, and then we cross-check each source, and I guess the source itself must have been LLM-generated. So, [chuckles] um, there's like some three layers of hallucination that had happened. Uh, but yeah, I think still it improves our workflow a lot. Um, I think, uh, one broad question as we sort of sign out of this, um, the-- our, the human equation with machines is changing very dramatically, right? Um, the way I thought of my phone earlier and the way I think of my phone now is quite different. And, uh, how do you see the future sort of evolving?

Pratyush Kumar 58:40 Uh, I'll take a broader and deeper view at it, right? So I think that... Two, two things very clearly. One is these machines are becoming very good, very fast, and with no clear, uh, boundaries on what they could achieve that we could specify today, right? At the same time, uh, we might increase our dependence on them as individuals, right? As, um, um, whether at work, whether in

Pratyush Kumar 59:06 personal life, we might increase our dependence on them, and like you said, be prone to some of their deficiencies.

Speaker 59:12 Right.

Pratyush Kumar 59:12 Uh, but I do, do, uh, wonder what the gap between human capability and machine capability looks like ten, fifteen years down the line. Uh, and, and that's, that's the scary way of putting it. The positive way of putting it is that we go back to the basics of, uh, what human life is, right? And go back to what, um, what people have discovered since ages, and there are different views of it, and not philosophy, it is about a centralisation. We go back to more real things that we should be doing, uh, and, uh, hopefully, have a very positive and symbiotic relationship with AI. Uh, but there's a quite a complex trajectory from where we are now to get to that positive, uh, point, right? And we'll have to navigate that properly.

Speaker 1:00:02 I mean, as a creator, as a writer, I often wonder... And, and it's, it's a question that I've only been thinking now, because machines have become so good. What does it take to be human? Like, what part of whatever I have written is me, has, uh, or is human? Has become a very difficult question to answer.

Pratyush Kumar 1:00:22 Yeah. Uh, I, I think, I think that going up now, we are getting into philosophy winds, uh, which is, which, which says that we're moving towards the end of the podcast. But I think, uh, I think we are, we are who we are intrinsically, not judged by the outcome. Uh, I mean, your experience is a human experience, is your experience, right? And that, that's as real as it can get, right? Um, we, uh-- And I think-

Speaker 1:00:50 Correct

Pratyush Kumar 1:00:50 ... today's society, we are judged a lot by what we do, the output of that, and so on. And I, I hope the positive place that we can be is to recognize that that's not, that's not what it is. Of course, AI can write a report, AI can analyze data, AI can write code. Uh, that's not what we should be judged by. It's a human experience.... uh, uh, whatever, and the sanctity of that, uh, which we have to recognize.

Speaker 1:01:12 That's right. As you're saying that, I'm reminded, uh, when I was a student here, uh, we had a lecture by, I think, Professor Ramachandran. I think he's, uh, V.S. Ramachandran. He, he was a, a researcher in the brain science space, and he gave this beautiful lecture, and at the end of it, somebody asked him a question. And he said, um, that it's possible in the future we'll know exactly the chemical reactions that cause love, and you can simulate it. But, um, if I told you that, would your experience of feeling love become any less real? And obviously, his point was that just because machines can't do it-- can do it, doesn't mean that the human experience is any less.

Pratyush Kumar 1:01:49 Absolutely. Uh, but what, what we should be, however, careful about is that we live in engineered systems, and machines becoming very powerful can affect what we experience.

Speaker 1:02:01 Right.

Pratyush Kumar 1:02:01 While experience is ours and it's, it's sort of sacred, we live in engineered systems, and hence engineering this whole system to be a positive system-

Speaker 1:02:10 Okay

Pratyush Kumar 1:02:10 ... needs to be a deliberate decision.

Speaker 1:02:12 I, I get what you're saying. I think our entire podcast on that. Um, before I go, I wanna ask you, um, can you just dial back a bit? Um, what does Sarvam do now, and what does Sarvam's next one-year, two-year, five-year pipeline look like? Whatever you can share.

Pratyush Kumar 1:02:27 Yeah. No, absolutely. We, we are in the thick of things. We're gonna build out, um, the sovereign model this year. We're gonna build out... Uh, we, we have been, uh, we have been putting products out there. Our products have scaled up already. They are doing millions of calls now. Uh, we're analyzing data for various companies. Uh, once we have the sovereign model built out, uh, you will see that we'll have a lot more surface area to build for, and we are really looking at that scale-out. Uh, in, in the two to five-year timeline, we want to close the gap to being able to build the state-of-the-art models of that day. What is state-of-the-art today and in two years will be very, very different, and we want to be, uh, uh, uh, able to build the best tech, uh, build it with strategic intent,

Pratyush Kumar 1:03:12 but also build it with the intent of, um, democratizing this in a country like India, where we have a long, long way to go, uh, to providing the per capita resources required for every individual to succeed.

Speaker 1:03:25 Hmm. Nice. Thank you so much.

Pratyush Kumar 1:03:27 Thanks.

Speaker 1:03:28 Let's close it here. Great podcast! Okay, see you next time. Remember to save, share, subscribe, all that stuff. And if you want to connect with, uh, Pratyush, hit his website. Thank you. [outro jingle]