Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
100 AI Unicorns From India: Rahul Agarwalla's Bold Prediction | SenseAI Ventures
0 Playsin 2 hours
How did a serial entrepreneur who built AI solutions for Toyota and Honda before ChatGPT existed become India's pioneering AI-first venture capitalist?
In this episode, we explore Rahul Agarwalla's bold prediction of 100 AI unicorns emerging from India and his contrarian investment framework that's reshaping the startup ecosystem.
Rahul Agarwalla, Managing Partner at SenseAI Ventures, brings a unique perspective to AI investing - having built and sold three AI/ML companies since 1996, including enterprise solutions used by Fortune 500 giants like Toyota, Honda, and Canon. As India's first AI-focused VC fund, SenseAI has deployed over ₹200 crores across 31 investments, achieving four successful exits while pioneering the shift from AI infrastructure to applications.
In this candid conversation with host Akshay Datt, Rahul reveals his VDAT framework for evaluating AI startups, explains why most "AI wrappers" will fail, and shares controversial takes on outcome-based pricing, founder requirements, and market timing. With AI startups capturing 57.9% of global venture capital in 2025, Rahul's insights on building sustainable AI businesses, navigating the India-US corridor, and identifying the next wave of AI unicorns provide essential guidance for founders and investors riding the AI revolution.
Key Highlights:
👉How Rahul built AI solutions for Toyota and Honda before the ChatGPT era and transitioned to become India's first AI-focused VC
👉The VDAT framework: Why variety, data, architecture, and team are the only metrics that matter when evaluating AI startups
👉Why "AI wrappers" will fail and how to build defensible AI-native applications that OpenAI can't replicate overnight
👉SenseAI's contrarian bet on applications over infrastructure and why outcome-based pricing will replace traditional SaaS models
👉The roadmap to 100 AI unicorns from India and why technical founders are non-negotiable for AI success
👉Market insights on AI tooling opportunities and the shift from foundational models to specialized vertical solutions
Subscribe to the Founder Thesis Podcast for more deep-dive conversations with India's most innovative entrepreneurs and investors. Follow us on LinkedIn for exclusive content, founder insights, and startup ecosystem updates. Visit founderthesis.com to discover more inspiring founder stories and actionable insights from India's leading entrepreneurs.
#AI #venturecapital #AI startups #IndianStartups #StartupEcosystem #AIinIndia #TechEntrepreneur #DeepTech #StartupFunding #AIInvesting #ArtificialIntelligence #MachineLearning #AIVentureCapital #AIApplications #AIInfrastructure #OutcomeBasedPricing #AIWrappers #VDATFramework #DefensibleAI #AIUnicorns
Disclaimer: The views expressed are those of the speaker, not necessarily the channel
Recommended
Transcript
The AI Business Transformation
00:00:00
Speaker
kind of thesis when we started that every business will be an AI business. does better where there is higher context If it's an easy decision to make, the value AI will deliver will be low. What should be my framework that something which cannot be a feature for an open AI to launch tomorrow? Strong belief, especially as an early stage investor, is that it's the people who make the big difference.
00:00:29
Speaker
Only people make
Understanding the AI Ecosystem
00:00:30
Speaker
you money. We strongly suggest that, especially for AI applications,
00:00:35
Speaker
are
00:00:47
Speaker
Rahul, you run an AI first venture capital fund. Let me start by understanding what you mean by AI. What kind of... Because every business today is an AI business. Yeah, that is the kind of thesis when we started that every business will be an AI business.
00:01:06
Speaker
oh So AI is obviously when when you start looking at it from an investment lens, you have to look at not just the overall picture that there is AI. Of course there is AI. But there are multiple layers to it.
00:01:18
Speaker
in From a Sensei perspective, we look at it as a four-layer ecosystem. Right at the bottom, you have the hardware, the data centers. Like the NVIDIAs of the world. NVIDIAs, Microsoft, Google, oh that entire ecosystem of hyperscalers.
00:01:36
Speaker
But the data, just think data center and that will give you one, the first layer. on which AI will run. Then is the core or what people call foundational AI, which is where OpenAI and now of course increasingly many Chinese firms, of course, entropic.
Importance of Application Layer
00:01:52
Speaker
The guys are building the foundational models which run on this hardware. And then above that, right at the top actually, and let me go there first, which is applications, is what you and I will use, right?
00:02:07
Speaker
um It could be a chatbot. It could be a personal finance application. We all use applications and that's they will be AI applications. And to build those applications, even on those foundational models, you need certain...
00:02:21
Speaker
tooling, right? Like for example, data security, data management, you need guardrails, observability, how is your application working, things like that. that So that tooling layer. So that's the four layers.
00:02:34
Speaker
Infrastructure, foundational, tooling, applications. Okay, ah let me just ah ask you a little bit more to make this clear to the audience. So ah obviously, OpenAI is the foundational layer, NVIDIA is the data data center layer. I think these two layers are fairly easy to understand. Anybody who's built a LLM, Gemini, Claude, Grok, all of these come in the foundational layer.
00:02:59
Speaker
um Perplexity, what is that? So perplexity very clearly started in the application layer. It was actually in the beginning more of a wrapper on top of OpenAI and Anthropic and on top of Claude. They are increasingly trying to build their own foundational models, right?
00:03:18
Speaker
So they are backward integrating. But they started, and even today, they are a great example
Security and Observability in AI
00:03:24
Speaker
of a lovely AI native application. right He thought, how do you redesign this entire knowledge discovery and consumption model for the AI era? And that's what he's built.
00:03:38
Speaker
Okay. Okay. Okay. I guess in terms of consumer... facing applications, perplexity would probably be the biggest example, right? I think most of the other applications are either built for coders or enterprise use cases.
00:03:53
Speaker
Well, in some some form, I agree with you, but take another one which I love is lovable. right which actually build applications it's designed you can think it's designed for coders but really it's designed for non-coders to be able to build applications so it's an application right
Venture Focus on India
00:04:12
Speaker
and uh these guys have 800 million with 20 employees in a year yeah it's just phenomenal the way some of these guys have grown the AI ecosystem is just phenomenal uh as far as that is concerned
00:04:24
Speaker
Tell me about the tooling layer. Okay, so I understood the application layer, foundational layer, data center or hardware layer. What is the tooling layer? so ah so So if you go look at building a AI application,
00:04:37
Speaker
You want to build a user-facing application. Now, much like when we were building software, any application historically, you needed tools to build those applications. So the AI tooling layer is just the AI-enabled portions of that tooling, right? Focus around AI. And what are the kind of problems you need to solve? So one, all AI applications will have tons of data around it.
00:05:01
Speaker
right um And of course, to train some of the AI, you need to even do synthetic data. i mean, you may not have enough data right in the beginning. So synthetic data and data management become one kind of tool, which is around the data piece of this problem.
00:05:16
Speaker
Then second key piece is security. right I mean, ah especially when you look at enterprise applications, security becomes a big concern. But not just that. If you are using building, a let's say, a financial application,
00:05:29
Speaker
oh something which robo advisor, how you invest into the markets. You need to have your data very, very deeply protected, right?
00:05:41
Speaker
It's financial information. It's your information. And you want lots of safeguards around it as well as, um so how is that data stored? How is that data accessed?
00:05:53
Speaker
And when it gives you a recommendation, also you need to think about security. So security is not just about your data, but also the data, the application surfaces for you, right? If it starts surfacing, like, for example, if you go and ask a financial advisory application, how to build a bomb or ah how to do fake videos, it can go ahead and answer you if it's trained on LLM, LLM is sitting behind it, but you don't want to be answering those questions. So you want your areas, which you,
00:06:26
Speaker
plane to be clearly guardrails to be put in. So security and guardrails becomes a second very critical pillar in the tooling space. A third big pillar to me is observability. right How is your application working? How how many people are coming there?
00:06:41
Speaker
What kind of questions are they ah asking? Are they getting answered well? Now see, this is now qualitative, not just quantitative. So historically, observability was always just quantitative. How many people came? What did they do?
00:06:55
Speaker
But you want to know whether the application is actually helping the user meet his end goal.
Evolution of Indian Tech Landscape
00:07:00
Speaker
And that is a non quantitative problem. It's as qualitative angles to it.
00:07:06
Speaker
And so observability is getting rebuilt. right As you think of how applications will get built in the future where they we've heard all this about agentic and I'm sure we'll talk more about that in the but as we go along.
00:07:22
Speaker
But you're going to now AI agents on your behalf are going to different websites, using different APIs, different tools, bringing all that together. Now, all of this is very, very different from how software applications used to work.
00:07:37
Speaker
And given that they will have agency, based on your credentials, based on who you are. And because of that, the information they will access on your behalf and get that data back.
00:07:48
Speaker
And then the AI will work on it and use that data to access the next app. So now this becomes a lot more complicated. So you you the tools are in some ways old.
00:08:01
Speaker
i mean, they've always existed, right? yeah When we were doing software applications, but they their facets have completely changed with yeah What were the legacy tools like? Any big names I would have heard of?
00:08:15
Speaker
So if I take, ah since I was talking about observability last, Datadog would be a great example, right? On the observability side. um Data management, I mean, there were two, three really big guys, Databricks and so companies like that. So there's a whole bunch of companies became very, very large in tooling.
00:08:37
Speaker
global and cloudflare for security yeah cloudflare for security and then i mean from authentication to management of authorization lots of other tools also there got it okay ah why is ah security a big deal for example companies use gmail um gmail has access to personal information whatever a lot of proprietary information of the companies, ah but I don't hear ah companies thinking of some additional tool to secure their emails which are running through Gmail. So but why is it any different from Gemini on which I am asking questions and getting answers?
00:09:27
Speaker
Why is it different from Gmail? No, no, no. so Gmail is the application, right? So the guy who built Gmail, so in this case, Google, would have put in the security layers, right? Because how they handled your data.
00:09:40
Speaker
So these are the tools used by application builders to build the applications. Not by the user, but by the builder. Got it. Okay. Understood. Understood. Okay.
00:09:51
Speaker
So the the security behind the scenes ah is ah what you mean by security. Yeah. Is there, I was under the impression that observability means being able to understand the black box of an LLM. LLM is a black box, right? You don't know why it gave an answer that it gave.
00:10:11
Speaker
That's not what observability implies. No, so observability is much larger as a concept. So though what you said would be one portion of it, right? Why did the AI give this answer? So actually, one key piece of observability that you have available today is to understand that given the same prompt, what are the different answers AI gave, right?
00:10:34
Speaker
Which while you may not understand why those answers came, for the same or similar questions, these were the different answers it gave, gives you a sense of whether your AI is performing well or not, right?
00:10:45
Speaker
um And of course, ah variations on that. But the in X. So what you're talking about is actually comes under the term of explainability that you can, yes, you can understand why i did what it did.
00:11:03
Speaker
So the why not the what observability largely is a what question. And not the question. and not the whiteish Okay. Okay. Understood. Okay. So I've understood the tooling layer. Now, given this is the lay of the land, um as a VC fund, where are your binoculars focused on? Where are you looking for ah breakout companies that you want to fund?
00:11:25
Speaker
So we actually focus on the top two layers, which is applications and a majority of our investment will go into applications because we think the applications for the next era yet to be built. Right.
00:11:37
Speaker
I mean, a few have started being built. I gave a couple of examples earlier, but there's a whole host of things. Literally every problem you can think of will be solved with AI or will be attempted to be solved with AI, which means there'll be applications for literally everything which is native.
00:11:54
Speaker
And those companies will be very interesting to us. And of course, because, and also you must understand our lens is as an Indian VC, right?
00:12:05
Speaker
We look at Indian founders mostly. And we have to figure out what are we good at? We as a country, we as a people, what do, what our ecosystem produces well of?
00:12:17
Speaker
So we do a great job of applications, right? We've got, I mean, we started, of course, our industry started with the history of ah you know Y2K and the entire so system integration companies, the Infosys, Wipros, DCSs of the world built our first set of tech.
00:12:35
Speaker
Then the second set of tech happened with the internet boom and you had ne Indian native unicorns like Flipkart, like or your all all the internet unicorns that you can think of is the second generation where we built software for ourselves right like in the first gen it was building software for us companies on their designs for their users then we built software for our users with our own design And then we went to the SaaS, oh third generation would be SaaS in my view, where we were building applications for now global customers, but based on our own design, our own understanding.
00:13:15
Speaker
And this you say is now moving into the AI layer. So we did well. Right. And because we're really good at problem solving. So applications are about solving problems, users, problems.
00:13:28
Speaker
So that's why we do it. The second also tooling is also a problem solving layer. It solves the problem for application builders. So we again and and we have today, i mean, then arguably the largest pool of AI developers in the world, right?
00:13:45
Speaker
ah Stanford index had it at one, then two, and it keeps shifting between those two numbers. So either we are the largest or second largest density of AI talent in the world. And that talent then obviously when it tries to build AI applications faces problems. So these problems are well understood by the developers in India. And thus we do a decent job of tooling as well.
00:14:06
Speaker
So we think those two layers is where we want to focus on. More importantly,
00:14:13
Speaker
And this this, if you go back to any boom in history, see the first, most of the capital in the beginning goes towards building the infrastructure on which everything will run.
00:14:24
Speaker
Right. But all infrastructure doesn't matter which industry we're talking about is a revenue generating layer. Right. It's not a valuation play.
00:14:35
Speaker
It doesn't give you 100x, 10,000x. I mean, if you go look at the internet companies which built at the top of the stack, which is all the trillion dollar companies today, which is Microsoft, Google, Meta.
00:14:48
Speaker
These all built applications. They did not build the foundations of the internet. The foundations of the internet were built by people like Cisco, which built the routers. And of course, its valuation went crazy in the beginning of the internet boom, but collapsed later on.
00:15:02
Speaker
or AOL, America Online, which doesn't even exist today, right? Or the browser companies. So you the infrastructure gets built first. And I think infrastructure is a large, it requires large amounts of capital, like $300 billion dollars went into infrastructure last year globally.
00:15:21
Speaker
This year, more than that will go in. I mean, that number has already got crossed. US, of course, has Stargate program and 500 billion dollar programs and Meta is building a 70 billion dollar data center. So you have massive projects going on in the US.
00:15:40
Speaker
We can't match that. Capital is not available at the same level in India to as a VC, what you're looking is to multiply people's money. I cannot do that if I invest in infrastructure, I invest in data centers. Yeah, it's a great revenue yielding asset, but it is not a value generating value multiplying asset.
00:16:00
Speaker
So we focus only on the top two layers and it's a great thing that Indians are good at it. So it works in every way for us. and Okay, okay. ah I'd argue that most Indian unicorns are still solving roti kapra makan problems, right? the ah Beyond that, it's mostly like American tech, like say ah Instagram or whatever that people use or Gmail for email, e etc. There are no Indian unicorns in the B2C space beyond that roti kapra makan layer.
00:16:30
Speaker
ah So is that why you're not that bullish in the AI B2C space? No, I would not say... Okay, firstly, that focus comes from what we are good at, right?
00:16:42
Speaker
If you look at the founders of the fund, which is Raja and myself, we're the two partners, right? We will be the ones ah investing the money and spending our time helping our portfolio companies grow.
00:16:55
Speaker
That piece, we both come from enterprise background. So we understand that piece really well. Second is also a question of,
00:17:05
Speaker
I think the consumer piece, we still are solving so many, there's still so many unsolved problems just generically. I mean, distribution. Distribution is not an AI problem. It's purely a last mile and all those kind of, these are physical problems, right?
00:17:21
Speaker
We still have so many physical problems to deal with in this country um that I think there's a huge amount of play for B2C businesses built without AI itself. of And So our expertise lies more in enterprise than in consumer. So thus natural affinity towards enterprise.
00:17:38
Speaker
And two is the Indian opportunities in B2C are not very AI specific always. Got it, got it. Okay, okay. Understood. Okay.
00:17:49
Speaker
um Okay. So now within the ah tooling space, what are the types of markets there?
00:17:59
Speaker
so um So, most of the markets we play with are global, right? And I think we should look at it. When we come to B2B, the markets always are global.
00:18:12
Speaker
Now, US firms are focused largely on the US market, so they not may not be competitors for you today, but at some point they will be. Now, Adobe is there in every market. Microsoft is there in every market.
00:18:24
Speaker
So much like that, all those global AI plays will come to this country also at some point. It's a question of priority. As a startup, you have to choose your battles very carefully. So today, most of them are battling for the US market than the Indian market.
00:18:37
Speaker
But ah we will not bet on Indian founders who are only building for India. Built in India for the world. So now take Turing, that's a classic example. So while we want to see founders who understand the problems Indian developers have and have learned from that, those are the same problems. The tools actually most people use are the same globally.
00:18:58
Speaker
I mean, the infrastructure layer is the same, the same data centers, the same hyperscalers everybody depends on, the same foundational layer, which is the same models. So the problems are the same for all developers of AI across the world.
00:19:12
Speaker
So we are looking at guys who can solve problems for them globally. So yeah and I'll give you examples from my portfolio. So one is like a company called PipeShift. Now PipeShift does, the it builds a full orchestration. So today, if you want to use open source AI, which is getting better and better, and I strongly, i mean, I think that'll be a large adoption of open source AI versus proprietary models.
00:19:35
Speaker
I mean, while proprietary models take the news today because, that's how this entire boom started you'll see the shift happen I mean this is classic iPhone of course ruled the roost ah in terms of number of items shipped all of that in the beginning but Android took over units very very soon so same will happen in the AI world and pipe shift enables that shift to happen right you can go build the entire your entire application with your you take open source model train it with your own data and then deploy it in production and it works forex more efficiently than deploying it on servers through amazon or google or wherever because you use something like pipe shift now this kind of efficiency this kind of manageability of your entire yeah infrastructure simply in the end ai application guys why are they using tools to simplify their lives it's really a simplification
00:20:30
Speaker
a problem because what they want to be really good at is solving the user's problem. So why take on all this headache of doing hard engineering work at the lower end of the stack to, you know, that's what's called pipe shit. i mean, make all your pipes work, all the plumbing work.
00:20:47
Speaker
It does the plumbing for you. It takes care of all of that. And that's what we love about that. So what are things which an OpenAI i will not build? that you know like if we talk of this meeting note taker, yes, it's a feature that OpenAI can offer, or Microsoft Co-Pilot is offering, or Google's Gemini is also integrated into Google Meets now.
00:21:09
Speaker
So if I'm a builder thinking about building something in the AI domain, ah what should be my framework that ah something which ah cannot be a feature for an open AI to launch tomorrow, something which is ah scalable, where there is some way of distribution, you know, what are some ah suggestions you have for builders?
00:21:30
Speaker
I think that's a great question, right? And we should dig deep into this. So, See, start from the first principles. The first principle is that these large guys can only go after large, very, very large markets, right?
00:21:44
Speaker
A billion dollar market is great for a startup. It does not move the needle for Google. Google, you cannot get even two minutes during the board meeting to discuss a market which is one, two, three billion dollars.
00:21:57
Speaker
and Nobody's interested. Nobody's going to agree to enter that market. So the largest markets are typically out, right? Do not go after that.
00:22:07
Speaker
As a startup, it's going to be very hard. And second piece is look at where domain knowledge starts coming in. Like Google is known for its search, right?
00:22:19
Speaker
It also got into the enterprise search space. But enterprise search is not as simple as web search. i mean, web search has its own complications, but they are standardized. Right?
00:22:30
Speaker
Once you solve it, you've solved it. Enterprises don't work like that. Every enterprise has a different in set of systems and different set of security rules and different way the content is consumed.
00:22:42
Speaker
And you need to literally cater for all of those variations to really build an enterprise search business. So Google never did well at it. So wherever you start seeing this specificity, so any domain, as you go deeper into, and I'll give you examples from my portfolio, right?
00:23:00
Speaker
Best way for you to understand what I invest in is to look at my portfolio. So for example, auditing, right? Internal audits. Each internal audit is performed for each company.
00:23:12
Speaker
Each company has its own systems from which you have to pick up the data. Two, you have to understand the industry and company specific terms. Both are important, right? And now once you understand it, you should get better at understanding that for that industry.
00:23:28
Speaker
And of course, when you do the audit next time for the same company, right? Now this creates a different memory, different set of knowledge base for each customer.
00:23:44
Speaker
Third is that there is ah domain expertise around auditing. How is auditing done? How do you measure risk? how What is the data you look at? What are the rules you run? And while that may be slightly standardized, what when you combine all these pieces, it no longer looks like a business a big guy will do well at.
00:24:06
Speaker
Right?
00:24:09
Speaker
Second example from, let's say, the construction industry. So I'll do a vertical now. Auditing works across its functional specific. If you look at construction, it is vertical.
00:24:20
Speaker
Now, in the vertical piece, construction as such is a massive industry. It's a very, very large industry. But how is that tooling getting deployed? How is that application getting deployed?
00:24:32
Speaker
It depends on each side. So and here there's a company called continue.ai in my portfolio, which helps you manage the construction process. Why hasn't it? Construction is an industry which uses very little of technology. So it's project management ah tool.
00:24:48
Speaker
Yeah. Continue. Okay. ah to and actual construction project management. What I mean, the difference is you could do project management using Microsoft Project where you enter the data, it gives you pretty chance.
00:25:01
Speaker
But how does the data come about? Because it's all blue collar workers. Who's going go enter the data as to how much work was done today? You literally have zero control. The reason you have a lot of delays in construction projects is largely because you have no visibility no transparency in actually what's happening on at the site.
00:25:19
Speaker
The planning process is very digital and very you can do 3D walkthroughs, all that latest technology you can put. But as soon as construction starts, everything disappears. It's WhatsApp and, you know, calls happening, site visits.
00:25:32
Speaker
That's how you get actual visibility on what's happening. Here the AI does all of that automatically. It just takes a, you just have to put a 360 degree camera on the hard hat of one of your employees there, typically the site manager.
00:25:46
Speaker
And when he walks through the site, the data is automatically captured. He doesn't have to do anything specifically to capture data, not even walk a specific path. He can just walk the way he does. Data gets captured, gets translated into a 3D model by AI and progress judged, defects detected, workflows done.
00:26:04
Speaker
Now, that's a massively Massive leave for the industry. But look at the pieces you're solving. Now, this is not just a LLM solving it, right?
00:26:15
Speaker
There is... Yeah, need computer vision and... Yeah, so there are multiple pieces to this workflow. There multiple complexities. The data you need is... So each site, I mean, what is the under... Actually, you'll find that under construction site photos are much fewer online in the public domain than completed building pictures.
00:26:35
Speaker
Right? Yes, right. So you see, data also becomes a domain. So is the data easily available? If the data is easily available, it becomes easier for the big guys to learn from it.
00:26:47
Speaker
As the data is harder to get to, what is unique in the data, then you get closer and closer to what we love to find. Okay, interesting. interesting interesting ah and Talk about your framework for evaluating opportunities. We've kind of gotten into data as a mode but uh you know just help me understand the full framework in which you evaluate opportunities so uh you know uh our framework is the called vdat the v stands for variety d for data a for architecture and t for t right uh let me go dig deeper into each of those layers so data piece partly you understood right it is
00:27:31
Speaker
If every data that you are using is public data, then everybody has access to it. It's going to be hard. But sometimes even there, you can go build different combinations and take out different insights from other people.
00:27:43
Speaker
And we look at that. But without great data, there's no great AI. That's fundamentally where we start from. So we're looking at what data you have. And the more proprietary or hard to get data.
00:27:54
Speaker
I mean, construction data may not be proprietary, but that it is hard to get. As soon as you start doing that, it becomes harder and harder for other guys to compete with you. So that's why we look at it.
00:28:09
Speaker
Architecture, people say I'm using the latest LLM models in the world. It's the fancier, it's the most scalable, all of that. Yeah, those are important pieces to it.
00:28:20
Speaker
But fundamentally, you want to understand why you chose a piece of technology to solve a specific problem. See, each The why is the real key piece for us as a fund.
00:28:33
Speaker
Why did you do this? Because the same problem can be solved in six different ways, at least, if not more. How you chose to slice that cake, how did you choose to build that combination makes me understand how you look at technology. What are your strengths in technology?
00:28:50
Speaker
And if yours is a well-considered design, it's a good design and it's well-considered. You came to the, you asked the right why's, you can answer your logic to reach these conclusions really well, that becomes very important for us as a investor because we are not investing for what you've already built. We're investing for the future.
00:29:09
Speaker
We're going to give you money. You're going to hire a bunch of engineers. You're going build a lot more. So to me, the capability of the the architecture, what capability we can understand, the technical capability that lies behind that architecture is what I'm trying to get at.
00:29:24
Speaker
Not just the architecture. I don't fund architecture. I fund the people who built that architecture. And that brings me to team, right? I mean, so team one is this technical capability. We really want deep tech because we we're looking for people who are AI first, AI builders, as we call them, right? Not just people who are AI users.
00:29:44
Speaker
So that's why we don't fund wrappers. So you want people who can build. So that's one part of it, right? But then there's the other pieces, right? it so It's a journey where you have to build a team.
00:29:57
Speaker
Can you build a team? How persistent are you? You're going to face difficulty. It's a given. mean, there'll be a few startups that don't. And I wish every one of my startups is like that, but the reality is probably not like that.
00:30:11
Speaker
And which means you need founders who have that. who have the ability to raise money, sell their product, you look at that entire ability in the team. So that's where the team comes. I left out variety, right?
00:30:24
Speaker
Variety sounds like a weird metric to have, right? But actually it's the easiest way me metric, which you can measure, you can literally count it to understand complexity.
00:30:37
Speaker
AI does better where there is higher complexity. If it's an easy decision to make, the value he will deliver will be low. Now, how do you, ah you know, given this VDAP model of evaluating a startup, um typically, you know, I hear from VCs that the jockey versus horse kind of a framework of evaluating startups, the jockey is the founder, like, are you betting on a founder or
00:31:07
Speaker
Horse is the TAM. Like how big is the addressable market? Which I didn't hear in this framework at all. ah ah Do those factor into your decision making? Of course.
00:31:18
Speaker
i mean, see the jockey part has already come in, right? So this filtering is what... So today we see about 100 AI startups a month. Now, that number has grown since we started over the, I mean, it was about, if I go back about 15 months, it would be 50 startups. Today, it's 100.
00:31:36
Speaker
Of course, not all 100 are really AI first startups, but all 100 are claiming to be AI startups. So, this framework helps us filter on those. So, typically today, out of the 100, about 15 will come through.
00:31:52
Speaker
15 to 20 will come through. Right. So about 80 percent, 85 percent are eliminated with this framework that allows us to scale the number of startups we look at and focus on the ones where we truly add value.
00:32:07
Speaker
Right. In the end, it's a ah match between the founder and the fund. Right. We should be able to that symbiosis is very important. It's a five, seven, ten year relationship ah which you're building.
00:32:18
Speaker
So once we've narrowed it down to this lot, then you start looking at other factors. But primarily, this, once you have the right team in the end, my strong belief, especially as an early stage investor, is that it's the people who make the big difference.
00:32:35
Speaker
Only people make you money. People are the guys who build that business. I mean, the same business handled by different people has very different results. I mean, I'm talking about the same business, not even two similar companies.
00:32:51
Speaker
There been enough business families in India which where once the business changes hands, it does very, very different results. So people, and especially at the early stage, see, you can change everything about a startup.
00:33:07
Speaker
And it still stays the same startup, but not the founders. As soon as you change the founders, it's not the same startup at all. right So You have to bet on the founders.
00:33:17
Speaker
Of course, each of the, like the TAM helps me understand whether the founders have business sense. If they've gone after the market, which is really small, then there's a problem with the founders. I mean, it's not a TAM problem because of which I won't invest.
00:33:30
Speaker
I'm not investing because the founders are not smart enough. So you see, actually it all boils down to even everything I gave you in the VDAT framework. if So there are founders who come in and say that we are a startup, so obviously we don't have data.
00:33:45
Speaker
Seems obvious, seems like a very fair argument to make. But every guy we've invested in has had the data. It's not an unsolvable problem. It's a hard problem to solve for a startup, but definitely a solvable problem.
00:33:58
Speaker
So a smart AI founder knows he will live or die by data.
00:34:05
Speaker
Right. So you would have got, you would have made the effort to get the data and build ah ah great, some visibility into your product, make it actually work with real data before you come to a person like me.
00:34:21
Speaker
If you're not making the effort, how do I know you'll make it in the future? Right. So see, we in the end are in the business of choosing people.
00:34:32
Speaker
Primarily. Though I invest in AI, I invest in human on intelligence. First and foremost. Okay.
00:34:41
Speaker
A lot of ah VCs very clearly say that if the market is not large enough, no matter how good the team, you'll not get a great outcome.
00:34:53
Speaker
So you don't agree with that? No, I don't totally disagree with it. actually so Actually, I disagree with the market size thing. See, if you think that Facebook thought its market size is what it is today, and there is no way, right?
00:35:07
Speaker
Businesses evolve. I actually like sharply focused businesses which are going after a niche because that niche, as long as the founder has the vision to expand the niche as he builds the business.
00:35:18
Speaker
Got it. Okay. yeah But however, on the other side, There are industries and markets which are bad. Bad industries. Like where the margins are very competitive, very bad. Like I would not fund an airline.
00:35:33
Speaker
um Now there are new airlines still being born in this country. So there are investors who like that. But to me, the industry structure is against you. So even given a great team to go after ah market which I don't believe in would be a harder problem for me.
00:35:50
Speaker
My advice to the founders will be to find a different marketer. And many times they do.
00:35:57
Speaker
Yeah, understood. Actually, in fact, one of my team's company called Lumo, they went after that same otter market that we spoke about, right? The note-taking. I rejected them. They were called Instaminates at that time. I gave them a note.
00:36:11
Speaker
But I loved the founders. So kept in touch with the team. Then they went ahead and built something called Lumo. And that's when I invested them. And what does Lumo do? So Lumo is in the tooling space again.
00:36:21
Speaker
They sit on both sides of Gen AI, right? From prompt. So they help you in prompt engineering engineering to give great prompts because prompts also determine cost.
00:36:33
Speaker
Cost in AI is measured in input and output tokens. So of course, you prompt determines the input tokens, but... Based on the design of the input tokens, out number of output tokens also have a relation, which LUMO understands and thus is able to guide you towards ah getting the same kind of output with lower number of tokens, so which is lower cost.
00:36:55
Speaker
and Two, then measuring the output, whether it meets your business outcomes, right? Your business, what you want to be able to do. What are you trying to say? Are you really answering the user's question? Are you really meeting, delivering outcome, measurable outcome?
00:37:11
Speaker
So they measure both the outcome side as well as, so they control the input and outputs. They don't touch the, between it is a black box. Okay, understood.
00:37:22
Speaker
So, like, a good analogy for that founder versus market question is, like, if you were to define Amazon as a bookseller, so that would be the wrong... um You could look at an early investor on Amazon could have said that, how big is the market for books?
00:37:39
Speaker
and So, yeah, got it. Okay, you... I mean, good founders make their own markets. Understood.
00:37:47
Speaker
but But, you know, if you get stuck in the wrong market, then it's tough because see, if you can build success, if you even if the market is small, but you can build success there, then you can leverage that success to build beyond that market.
00:38:03
Speaker
But if you enter a market which itself will not give you any revenues or margins, then it's going to be tough for you to be able to do anything beyond it at all. You'll have to pivot out.
00:38:14
Speaker
And then that becomes a completely different story.
00:38:18
Speaker
what kind of That also works in some cases, but then it's a very different story. What kind of founders do you like? Are you looking for grit and hustle and somebody who's able to like show sales achievement like you know that I have paying customers? or like what What do you look for in a founder?
00:38:39
Speaker
So the first thing we look for is the technical capability, right? And the VDAT framework kind of covers that, right? Because today I think building in AI itself is hard. It gets easier. I mean, LLMs become more capable by the day.
00:38:53
Speaker
Gen AI keeps changing, advancing. So it will become easier and easier to build, but still it is hard. It's not knowledge which is generally available.
00:39:04
Speaker
And how deeply you understand AI will determine the kind of products you will be able to build. So the first thing we're looking for is that technical ability in the team.
00:39:16
Speaker
However, we come in early. So mostly at the time we come in, you may not understand. i mean, you may not have much revenue. Mostly our companies have zero revenue. They have early pilots, POCs going on, but they do not have revenue proof.
00:39:32
Speaker
That comes later on from our perspective. We are a post product pre revenue company. So you must build great product. So that's technical capability plus the ability to wrap it into a user usable product. right In the end, a product solves a problem. A technology is just capability, but a product solves a problem.
00:39:51
Speaker
So one, have to see that problem solving ability very, very clearly. That's solution orientation, customer orientation. see When you build technology, you don't care how it's used. You just say, oh, I solve this technical hard challenge.
00:40:06
Speaker
That's not enough. So that will give you the technical capability, but not enough for us to invest. For that, you have to be able to solve a real problem, real business problem. So that's the product piece. And that's fundamentally what we're looking for.
00:40:20
Speaker
Early stage, we're not looking for revenue, but how you think about the market matters. Like, see, almost for anything, anything consumer for sure.
00:40:31
Speaker
India has a huge market, right? The numbers we have, one and a half billion plus people that allows us to say every market is large. But distribution is highly expensive, really, really expensive.
00:40:45
Speaker
So just saying large market doesn't do much, right? How am I getting to that market? So those numbers of thinking from sales cycle to number of people I need, how many meetings do I need to get to a sale?
00:40:57
Speaker
You need to think of your business, your numbers behind your business very, very clearly. Else if the math doesn't work, it will never work. So while I'm not looking for proof, I want the thought process to be clear.
00:41:10
Speaker
Something I should be able to say, yeah, I feel comfortable saying, yeah, this looks like a sensible and something that will work.
00:41:21
Speaker
In the end, there are always changes once you hit the, you know, the rubber meets the road. That's market always teaches you 10 things which you did not know before you went into them. But if your approach is ah from the starting, not you know based on reality, then it's going to be a very hard, the learning cycles are too long.
00:41:41
Speaker
You don't have that much time available today. Business is happening at the speed of light. It's just crazy the way AI businesses are growing. So you you can't be taking months and months to just figure out all those pieces.
00:41:58
Speaker
What are the typical mistakes of distribution slash sales slash GTM? like Like what are some pitfalls to avoid? So I think founders um
00:42:13
Speaker
underestimate the difficulty of getting into the US market. For us, that makes a difference because we do this India-US corridor a lot.
00:42:21
Speaker
US markets are far more, far harder to enter than a lot of people imagine. At the same time, and they take typically more capital than most founders estimate. So those are one or two of the simple mistakes you made and they are easy to recover from, by the way, also.
00:42:40
Speaker
The other is in the beginning thinking you can get a sales guy and build sales. right I am a technical founder. I need to just hire a great head of sales who will, you know, he has this magic wand which he will do like this and we our sales will just skyrocket.
00:42:57
Speaker
is not going to happen. Founders have to, in the early phases, until you have demonstratable PMF, you know, and product market fit. And it's a term which is explained by many people differently.
00:43:11
Speaker
And there are some great videos available online. I'll just say this. You as a founder always know when you're PMF because you see your curve sharpening, right? Your growth curve just sharpens.
00:43:25
Speaker
sales cycles become shorter, number of deals and in every that entire sales funnel, all the metrics around it fundamentally shift. When you sense that happening, you know that you've hit PMF.
00:43:38
Speaker
right Hard to define exactly that moment, but as a founder, you know it. And I'm saying that until you reach PMF, you cannot scale sales.
00:43:48
Speaker
Really speaking, it's still artisans work. You still need to you're going to make mistakes in how you pitch what the real problem is. See, enterprises sometimes buy for completely different reasons than we thought of.
00:44:04
Speaker
buy thereby. Like that internal audit company startup that was talking about, it's a company called Iramay. One of the highly used use cases of theirs, which was actually asked for by a customer. They did not have that in their toolbox when they went to that customer, was moonlighting to figure out how many of our employees are moonlighting.
00:44:26
Speaker
Now, this is a really popular use case in India. Now you would not guess it looking from outside. mean, it doesn't have any compliance, any regulatory overheads, nothing.
00:44:40
Speaker
But post COVID, this is a key concern for a lot of businesses in India. And these guys can address it. So they're selling also like hotcakes because they can just do this one problem.
00:44:51
Speaker
How do you figure this out? sitting Your sales guy won't help you figure it out. See, fundamentally sales is about So when we think of scaling a business, there are two pieces. There's a product piece, which you want to be largely like a cookie, right? So once you make the cookie, it's just a cookie cutter, same shape. You just keep beating it out more of the same.
00:45:14
Speaker
The same goes for sales. People don't get this. That sales needs to be cookie cutter also before you can scale You cannot scale consulting based sales.
00:45:26
Speaker
People will look at large, the world's large consulting businesses and say, Hey, those guys do consulting led sales. But those organizations took years to build.
00:45:38
Speaker
You don't have that sort of time. You need something which is, you cannot hire that ideal sales guy who understands your industry, understands your product and understands the customer's problem.
00:45:51
Speaker
Then can put all three together and make it work. Right? That's the founder. You fundamentally describe the founder when you describe those three capabilities.
00:46:03
Speaker
so So you can't hire for that. It's much harder to hire for that. So you need to be able to bring it down to a cookie, which then any guy who has, if you look at sales as persistence and discipline and going after customers, the how many calls you make, you look at ah that funnel perspective, running the funnel in a disciplined manner, then there's a lot of sales guys you can hire.
00:46:28
Speaker
yeah Tell me about pricing. You said pricing is... still something people are figuring out for AI applications. I've seen a bunch of models. I've seen, of course, the standard monthly subscription model, but then There are also models where the tokens are priced, like you pay for X number of tokens, basically consumption based pricing.
00:46:48
Speaker
um I guess these are the two models I've seen so far. Which model do you like better? Are there other models out there? What's your recommendation? So actually, we are actually, we strongly suggest that, especially for AI application companies, not true for tooling, but for AI applications, but price outcomes.
00:47:08
Speaker
See, AI applications are delivering outcomes almost from day one. like Definitely within two weeks of you buying the AI product, you have outcomes, measurable outcomes in your business.
00:47:23
Speaker
Unless that's happening, i mean, we've invested in the wrong startup. right So if if you are able to deliver outcomes, then you should be able to price those outcomes. Now, what is an outcome and what you should price is thus the debate.
00:47:37
Speaker
And that is not clear fully. don oh Let me give you an example from again from my portfolio. so there's a company called Flowworks. Now Flowworks basically does what? You describe ideal customer for you.
00:47:50
Speaker
It then goes out, looks at multiple say databases, builds a list of potential customers for you. Then it goes through your website, takes content from there, understands the customer's profile, builds a personalized campaign just for that one customer, email, LinkedIn, multiple communication channels, communicates with you and actually does answer, reply, I mean, that entire cycle, you know including objection handling until it gets you to a meeting.
00:48:22
Speaker
yeah Now, there are two outcomes here. One first outcome, which is the easier to measure outcome and probably what most traditional businesses will focus on is the number of emails sent out or number of contacts reached out to.
00:48:38
Speaker
And if and so, contacts is first. Second is number of because the number of communications you send to each customer will be multiple. So you could price on that. And third is you can price on every meeting you get.
00:48:50
Speaker
Now, ultimately business is going to measure ROI based on item three. It does not measure a ROI based on how many people, um many emails you send out to how many people you contact.
00:49:04
Speaker
And that's the difference between ai and old software pricing. So old software like email software was priced on the number of software the emails sent out. Right.
00:49:17
Speaker
So you're moving two layers away from that. Not even the number of contacts you contacted, but actually meetings you got set. And once you start moving there, the alignment with business is very high.
00:49:31
Speaker
However, no software has ever been sold like that. How much do I pay a meeting? Now, i you need to understand the customer's business to price it correctly. A meeting for you versus a meeting with somebody else means different things.
00:49:46
Speaker
I'm an investment banker. I'm doing a hundred million dollar deals. If you get me a meeting with the right company, that meeting has a lot more value than I'm a guy selling water purifiers. 20,000 or 10,000 would be product.
00:50:03
Speaker
Very, very different outcomes for us. so So now the pricing has become much more complicated. But of course, the key point is direction of the pricing is clear.
00:50:15
Speaker
It will move away from input and more towards outcome. It has always fundamentally been priced on input, on effort, on seats, all subscriptions, on time.
00:50:29
Speaker
Why is time a great metric or number of seats a great metric? Those are cost-based metrics. But If you go look at any business, it has to justify its cost based on the ROI.
00:50:44
Speaker
So now AI allows you to jump straight to ROI rather than do all the intermediary steps, which is why the J-curves happen faster because you you are able to justify the ROI. ROI is clear to you. Your product delivers ROI.
00:50:59
Speaker
It delivers outcome. And then then businesses feel much more comfortable spending more on it.
00:51:06
Speaker
ah Will more pricing become concierge pricing then? Because like you said, the outcome-based pricing needs you to evaluate the value of that outcome to a business.
00:51:20
Speaker
So it's hard. i think it will move towards that, but it's a hard journey, right? I mean, if the same outcome means different things for different businesses, right? It becomes harder to price.
00:51:33
Speaker
Two, what are the data points you need to be able to price well? Not yet clear. Will businesses protest too loudly? Because see, it's much easier to do from a budgeting perspective.
00:51:47
Speaker
Number of seats is very easy. You know the number of seats. Now, every time you hire a guy, you just assign this is the software budget. So from a budgeting cycle perspective, cost-based pricing was easier to absorb. So so I'm saying we we probably will see a mix of pricing models and where will the dust settle?
00:52:11
Speaker
So my view or my advice to founders is experiment with it here. Go do multiple things. Look at covering cost in one manner, but really earning money by delivering outcome.
00:52:24
Speaker
I mean, to my mind, it's very simple. You must earn money when I as a fund only earn money when I give you superior returns to the market. That's so only way. ah we We run on something called carry, which is basically a share of profit, which only kicks in when I beat the market.
00:52:44
Speaker
So performance-based pay, and you know you want you keep advising HR wants to keep building performance-based pay. Why not for software? And if AI is virtual employees, and lots of AI is being sold as virtual employees, and instead of having a fixed salary, they just have performance-based pay.
00:53:06
Speaker
so so So we'll see a ah lot is to evolve here. I think, Akshay, we are still in the early days of this evolution.
00:53:17
Speaker
Fascinating. um What is the amount of money you give to a startup that you invest in, an average check size? For us, the average first check we write is about $750,000 and is will be we We can go a little higher than that, we can go a little lower than that, but that's the sweet spot for us.
00:53:37
Speaker
And you're on the fund one, fund two, with what is the current fund? So I had a fund alpha, which was a syndicate, right? So that's why we call it fund alpha, not fund one. And we call my current fund fund one, but technically it's the second fund.
00:53:55
Speaker
I had investors and so I did 12 startups out of Fund Alpha, which are all that fund is fully invested, deployed. We've sold four of those 12 businesses, eight are still in the portfolio.
00:54:07
Speaker
And we'll look at exiting those and over the next two years. While Fund One, which we launched in end of 2023, we've is we've made nine investments out of it, out of the 18 we plan to make. So we're only halfway through discovering the portfolio. And of course, you keep money back to double down on your start some of the startups.
00:54:32
Speaker
So that all capital is still in the bag. So yeah, we'll be investing out of this for quite some time. And what is the size of this fund? So this is a 200 crore or $25 million dollars fund.
00:54:47
Speaker
And which is why that average check size for the first check, for the first 18 companies is that. And then the winners, of course, will invest more capital behind. And are these ah family offices, your LPs or who are they?
00:55:01
Speaker
A lot of family offices. Our anchor LP is SIDBI. oh and then a lot of family offices both raja and i the two gps have put in a significant chunk of money also we we feel more comfortable that way i mean we should risk enough of our own capital to give confidence to our lps that actually will be larger than most lps i mean not as large as sidby but larger than most almost every other lp Okay, okay.
00:55:32
Speaker
ah What is the difference between a syndicate manager and a VC fund manager? So you said you were managing a syndicate earlier. What does that mean? So see, syndicate, firstly, each investment... one happens, you know, it's a standalone investment because you use a the angel fund structure, right? So each investment is a separate risk pool kind of thing.
00:55:57
Speaker
And typically most syndicates don't have teams running them, right? So when I ran the syndicate, I did not have a large team. As a VC, what happens, it just gets more structure. The capital is raised upfront and the portfolio is discovered later.
00:56:15
Speaker
That's fundamentally, would say, is you know in practice the difference. The other piece is, of course, the amounts of capital are different.
00:56:27
Speaker
Syndicates will typically be smaller. no Once you like we are a 200 crore fund. Today, I will only take a check of at least two crores from any investor, which limits the pool of people who can participate in my fund.
00:56:47
Speaker
ah The syndicate is like a bunch of friends co-investing. Like one of them would take the lead, present an opportunity and out of the other, some would say, yes, I want to invest along with you.
00:57:01
Speaker
So it is very flexible. It could be that friends coming together and doing it, right? Or like in my case, I had it a little more structured, right? I mean, I had investors who had committed to X amount of money to my to to the syndicate and I could take it over five investments, six investments.
00:57:19
Speaker
i mean, they said, this much money is there, you take it when you need it and invest it on my behalf. right So it was a blind pool per se. They were not looking at the startup. They were not individually assessing. yeah okay But a lot of syndicates will work the other way, which is that each investment, they'll go talk to their friends about it. They'll hold one meeting and people will then say, okay, i i also like this startup and they'll put in money.
00:57:43
Speaker
So if you it it is a flexible structure because each investment stands alone. It's very flexible how you raise that money, how do you define all of the pieces around it is flexible.
00:57:55
Speaker
So I think it's a great innovation in the ecosystem. From ecosystem perspective, I think it was great. It's also a great learning ground. See, I've been an entrepreneur operator for most of my life.
00:58:06
Speaker
I built multiple businesses, sold multiple businesses. To become an investor, to I mean, i I had to prove to myself also that I can do it. It's very easy to, you know, say you can do it.
00:58:18
Speaker
Yeah, you understand. Yeah, you understand. But to be actually able to do it. So to me, the syndicate helped me establish all the pieces, understand the entire ecosystem, build the infrastructure, build the relationships into the startups, into investors, into the other investors, other funds, etc. Because ultimately,
00:58:39
Speaker
So, for example, the fund ecosystem works only, it's like a, take a startup journey from being a startup to a public company is a long marathon, right? Which is run in multiple legs.
00:58:52
Speaker
And as an early stage investor, I'm kind of running the first leg. I'm carrying the baton in the first leg. The second leg, I'm probably running alongside the guy who's holding the baton.
00:59:03
Speaker
And I'm probably out by the third leg. So, ah Is it harder to, as a VC fund, is it harder to find deals or to raise money?
00:59:19
Speaker
I think both are hard.
00:59:24
Speaker
That's a good question. I'm trying to think which one is harder. Probably raising money because we see a lot of startups maybe today. But that that may be unique to us.
00:59:36
Speaker
oh Actually, I find the startup piece, I've just been around. We are the only ones doing really yeah yeah first fund. We were first in that. But AI has a lot of funds now, right? like And large. literally i mean, everything is AI. Everything. yeah um but but But then like this Girish has this very large fund together fund.
00:59:59
Speaker
ah And so that has that celebrity tag of being backed by Girish. ah so So I mean, there are ah alternatives. ah Why is it that you don't face any problem in attracting startups?
01:00:13
Speaker
So two parts to it. One is that yeah having been around ah long time, we've built a decent brand recognition, right? And this focus around AI, like, of course, yeah, together, Girish is doing that with that focus today.
01:00:28
Speaker
But historically, that's not where they were. I mean, historically, there were nobody. we we We were there first. were pioneers in the space. You were AI first right from the beginning.
01:00:39
Speaker
Correct. From 2017, been doing only ei okay which gives you a certain and you are able to build the relationships into the ecosystem and thus the inbound flow for us is pretty decent.
01:00:51
Speaker
Today we get about 50% of our deals through the inbound funnel. But the rest 50% we work really hard for and we in fact use some portions of AI for doing that, which is hunting for AI startups, right? Because that allows us to build relationships early on.
01:01:10
Speaker
Sometimes we'll find these founders long before they start raising money. They'll only take the check from us when they're raising money or when they're ready for money. But having reached them first, having had the conversations, and i'm I'm happy to help, right? it's I didn't get to where I was without advice and without help and support from a lot of people.
01:01:31
Speaker
And um as part of that, that outreach program that we run, we go out, we find founders, some are very early, they're still figuring things out and we help them, we happy to share our thoughts so that they can figure it out faster.
01:01:47
Speaker
And that works. That builds that ability to get into the deals at the right time into a lot more. So, so to me,
01:02:00
Speaker
It's not as hard to find investors, so like family offices, you talked about so multiple family offices today are investors thankfully in Sensei are very thankful for all the support we've received from multiple families here, but The pool of capital in India is very different from the pool of capital available in the US, let's say, ecosystem.
01:02:24
Speaker
So I think the venture ecosystem is still young also, right? I mean, I think the first funds would have come around 1999. And that time, maybe the first, I'm talking literally the first fund.
01:02:37
Speaker
So the history of venture in India is less than, is 25 years old. And if you see the amount of capital deployed, all of those pieces, we are in the early stage of the venture ecosystem developing, investors understanding the asset class, really being able to understand which managers to support or not, what thesis to go up. So I think there's a...
01:03:06
Speaker
That part of the industry is still very much um being growing and maturing as we speak. Okay. and You said you do outbound. as So how do you decide that I want to like, what are your filters? It's a very large space out there, right? So what signals do you look for?
01:03:31
Speaker
a great It's a great question and a little bit of our secret sauce. So I'm going to try and answer it, but ah I won't be able to be fully transparent here. right I talked about the technical founder. right So now you can build that profile of that technical founder, kind of.
01:03:49
Speaker
He would have gone to a great technical school, college in India or wherever, but a great technical college. to work for a great tech company.
01:03:59
Speaker
What do I mean by a great tech company is where they build tech at scale. I mean, not just innovation, but also scale, because without scale, without scale gives you data, data gives you AI. So that entire journey, you would have seen some portions of that journey also happen, which means, let's say, the of course, all the global big tech, um Indian big tech, all the unicorns, sunicorns maybe.
01:04:23
Speaker
So that, that, that background itself allows you to filter through a lot of the clutter but not all of these guys are a into ai b ah building ah entrepreneurs so then you add additional pieces around signals to figure out that he is a founder and he wants he's building something in now actually the filters are less important because in our business in this in this outreach Let's say i reached out to you, you know, Akshay, because you let's say you did this podcast.
01:04:59
Speaker
And after that podcast, somehow or the other, we figured out that, hey, he's doing something in AI. Let's talk to him. What do you do? You will politely say, no, I'm not building an AI startup. yeah okay and No harm done. You're not going to get irritated that sense that I reached out to you.
01:05:15
Speaker
So even if our, you know, precision is low, But doesn't it dilute the brand? as I wouldn't imagine Sequoia reaching out ever.
01:05:28
Speaker
I think, see, the there are two pieces, different, one, different eras and two, different focus areas, areas right? Sequoia is not really your seed stage investor.
01:05:41
Speaker
Right. It's Series A investor. It fundamentally is a Series A investor. What it does... It's a smaller pool... Yeah. so they we are the we are the discovery layer in the ecosystem.
01:05:54
Speaker
There is no list to work off. There is no list of startups. Sequoia is looking at our list, people we funded. Yeah, yeah, yeah. True, true. That piece. And two, I think also the era has changed, right? I mean, um ah ah the amount of data available...
01:06:13
Speaker
is very, very different from what it used to be. What I can do today using AI and using digital tools and be able to build outreach program wasn't possible. So Sequoia had to do things like the Scout program, where it had Scouts which literally used to go out to look for startups.
01:06:33
Speaker
and And I'm just giving you, ah don't want to give you names, but it would be one of the funds we've talked about, which has reached out to me, said, can you introduce me to this founder? and Not even in my portfolio, but saw that I am connected to him on LinkedIn. So ask me for an introduction. So it's not that the these funds, the ones we think don't reach out, don't reach out. They do reach out.
01:06:58
Speaker
Yeah, yeah, yeah. Okay. Understood. It's a much smaller program than ours is. Hmm. Hmm. so you You said you've been an operator, entrepreneur, exited ah multiple ventures. Just give me like a summary of your journey.
01:07:13
Speaker
ah what What led up to you becoming an AI first investor 2017 when most of the world was not even talking about AI? Okay, so I'll give you the the quick and dirty.
01:07:26
Speaker
So the first venture I built was, you know, when I was coming out of business school, had a choice of taking my day one job. But there was this thing called the internet, which was kind of just coming around. I'm talking 96, 97, right?
01:07:39
Speaker
right So 1996-97 is when the Netscape IPO had happened. had started using the internet. I was excited by it. So I thought this is rather than do the traditional path, which is after MBA do a day, you know, join a campus job.
01:07:55
Speaker
i went and did an internet startup. I was lucky to find the Kodak Mender group. They were the initial backers for that internet startup. There were no VCs actually to approach.
01:08:08
Speaker
So built that business, sold that in 2001. This was digital content aggregator. We aggregated content literally from everybody who had digital data, right?
01:08:18
Speaker
From newspapers to people like CMI and Crystal who did corporate data to then later on, we even spoke to people who had music and any kind of magazines, you name the data, literally we had 200 different providers in by 2000, right?
01:08:36
Speaker
right And who were the buyers? Sorry? who were Who were the buyers? Who was buying this content? This was being sold ah to… So we had two basic communities. One was corporates who were using it for competitor research so that there it was more the chrysal and news and competitor information kind of data industry analysis.
01:08:57
Speaker
the But the bigger business for us was portals. So all the dot-coms which came around the from 1999 onwards and got funded, literally a series A would happen and the founder would call me, Rahul, data laga.
01:09:13
Speaker
yes You see, funding was that time based on page views. Right. And the more content you publish, the more page views you get.
01:09:24
Speaker
So you can't you can't set up a content operation overnight and have pages, right? You can have an idea for setting up a portal, but setting up the operations for it is very tough, much tougher.
01:09:35
Speaker
And so it was much easier to syndicate content. So 80, 90% of the content they just take from people like me, and then add 20% of very specific content, which they thought they needed to add for the audience.
01:09:46
Speaker
And that became the initial set for multiple startups. And you were not an engineer, right? so So you were more on the sales and business side? No, no. So I was a product guy.
01:09:57
Speaker
So I'm not really a software, software guy, but I know enough software. it I look at architecture, but I'm a product guy. So I helped take that as the bridge between engineering and business.
01:10:10
Speaker
mean, I was one of the guys who ran the business, but point making, I'm a product guy. So I helped build that product. Then the next business, so and after selling this, what I've done is taken lots of data.
01:10:22
Speaker
So when I said digital, so imagine what but digital meant in 1999 also was very different from what it means today. So newspapers would publish basically, you know, the pre pre-press files because the presses had become digital.
01:10:39
Speaker
Yeah, e-paper it used to be called. Yeah, so what looks like an e-paper today, I'd get that.
01:10:47
Speaker
It was digital, but it was not really... You yeah did not have... You're getting a PDF file with everything laid out and the ads also laid out and... So sometimes a little better than that, but that's fundamentally it. So we knew we learned to work with lots of data, which was not well tagged, well organized, well structured, and then structure it, tag it, metadata, manage all of that.
01:11:12
Speaker
So that I'd set up a production system for that. So basically took that and converted that into a new business where we did data management for global companies.
01:11:24
Speaker
and which Which is where your familiarity with enterprise search comes in. You were talking of how Google couldn't crack enterprise search. also So enterprise search was also that familiarity happened actually even in the first business because all this data, we had these 200 different suppliers.
01:11:38
Speaker
We were used to use a search engine called Verity, which nobody even remembers today, but that was literally the first engine that came out. Right. And so we used Verity. And so I've been doing search all my life.
01:11:52
Speaker
I'm very familiar with that set of technologies. And which is which brought me to my third business, which was a big data business out of Japan.
01:12:02
Speaker
And here again, so but I'm talking 2002 and 2003 beginning. beginning At that time, there's no term called big data or no SQL. These terms did not exist in the world.
01:12:17
Speaker
But what was clear to me is that when you have large enterprises and who have large quantums of data, doing any kind of analytics using... databases wasn't working.
01:12:31
Speaker
So you had these, you know, BI, i business intelligence tools, data warehouse, these are all terms which existed in the industry. high Very expensive, very inefficient. For example, one of our first customers, um actually the first customer was Honda, the second customer was a company called Toyota.
01:12:49
Speaker
And Toyota's ah data center had
01:12:56
Speaker
37 servers which we brought down to 4 servers.
01:13:01
Speaker
Because of redundancy of data. Yeah, no, and the database would not otherwise be able to perform. It was not performant, right? so you needed multiple copies of the database to serve the queries they had.
01:13:17
Speaker
We just replaced it with search, which is no SQL. And we had a very unique design of the index and things like that. And that kind of answer. But the other problem we solved during this piece was, so Japanese data, Japanese text has no spaces.
01:13:33
Speaker
And, you know, it it uses kanji, which is symbolic. And so a single kanji can be a word. Of course, combinations are also words. So if you have a sentence which has 30 kanji on it,
01:13:46
Speaker
It could logically be 30 words. Of course, it's not going to be 30. But is 7, 8, 12, 15?
01:13:55
Speaker
It was very hard for computers to figure out then. Dictionaries and rules did not work. I mean, you just had really bad accuracy. So that problem we solved using and NLP. And there's this Norwegian engineer who actually came up with that algorithm to use NLP and solve it. And he became the prototypical technical founder in my head.
01:14:16
Speaker
i I mean, he's the Christian that you were speaking about. Okay. Yes. Yes. Yes. well So before that, um so so I'm talking 2003, 2004, we were using and NLP to solve a real business problem and it developed value. I mean, this is classical machine learning NLP, natural language programming and not AI as we know it, Gen AI as we know it today.
01:14:42
Speaker
But actually that still falls forms the base layer even for large language models. Essentially, it's language, right? like yeah zero so yeah No, and the first thing you do, so this breaking up that Japanese sentence into constituent words is called tokenization.
01:14:59
Speaker
ah Which is, yeah, even today. And now you see, now you know why you use the word tokens all the time. Yeah, yeah, yeah. And why a token is not equal, exactly equivalent to a word. So 75 words is 100 tokens. i mean that's the thumb rule.
01:15:16
Speaker
So that's what we solved using machine learning right in 2003. So that business built up. I got Honda, Toyota, Canon, a whole bunch of Japanese companies, then US, s then India, and then sold that business in 2016.
01:15:30
Speaker
So when that business got sold, had capital. That business was helping clients to... make sense of the data, search for data, find relevant data, stuff like that. Absolutely. And so all big data analytics or reports and then visualize it and um it had a monitoring pieces inside it as well. So all the nice, lovely things you want to do with lots of data, we allowed you to do as a business.
01:15:56
Speaker
And and ah to automotive was a huge segment for us and electronics was the second. So these were the two big. And these are industries where Japan does well. So they ended up being our focus areas.
01:16:09
Speaker
So when I sold, after that business got sold, I had enough capital to now decide to do whatever I wanted to do. yeah Who bought it? Like a Japanese company? Yeah, it stayed in Japanese ownership.
01:16:21
Speaker
ah Still is. So then I, ah what had become clear to me is that yeah is the future. I had made money using AI.
01:16:31
Speaker
And so to me, it was very, very clear, crystal clear. This is the future. Of course, in 2017, 18, 19, 20 even, when I to tell people, yeah, it's the future. People used to look at me, and yeah what they yeah, you know, like there would be, um you know, all kinds of skepticism around it. It's too small a world.
01:16:51
Speaker
Today, everything is here. But, you know, I mean, both even three years back, ah by you two you Were you surprised by ChatGPT 3.5 or 3?
01:17:02
Speaker
but Like, did that surprise you? Because you had a framework of this is AI, these are the use cases for AI, this is how businesses can use AI. Did ChatGPT ask, like, kind of make you question that framework and, you know?
01:17:18
Speaker
Not really totally, I would say. ah the The key piece about... So one, OpenAI had been around for a few years and that they'd been hiring ah researchers at crazy salaries and they're trying to really...
01:17:31
Speaker
You know, AI on steroids is one way to think of what was there before ah large language models and ChatGPT. um They're just taking all of that and putting it on steroids. And that thought process was clear to us.
01:17:45
Speaker
And that we could see that those things will happen, right? I mean, ultimately, you had to... You could not keep you could envisage that something like ChatGPT would exist. Right. Yeah, time frame, I would not say that.
01:17:59
Speaker
or Where we are today, if you asked me to even four years back to predict this point in time and this kind of capability, would have got it wrong. would have underestimated for sure.
01:18:11
Speaker
So, the steep curve which we are seeing, the amount of capital we are seeing coming into AI would not have been able to predict in 2021 also. But that this is the future and this will be the future and that, you know, intelligence has always been so expensive and so so critical to success, right?
01:18:34
Speaker
It's one of the scarce resources.
01:18:38
Speaker
People keep thinking of job loss, but you don't understand that the a lot of people just support a few people who are really doing the decisioning, who are really moving the needle, right?
01:18:53
Speaker
There's one guy who rented the car. You know, Ford, literally it was one man's vision, which everybody else in Ford supported in such a large company today. So once once you start thinking like that, you realize that shortage of being able to think through things and execute them has... If we could do multiple...
01:19:14
Speaker
eras of industrialization compressed into 10 years, you'd see that the number of jobs would multiply. Lots of things will change. So if you start thinking of intelligence that way, you realize its power is immense.
01:19:29
Speaker
It's not just about, you have to think about productivity, you have to think about access. Like so many people, in India we are underserved literally on every front. If you look at our population, whether it's education, whether it's access to healthcare,
01:19:44
Speaker
You go, we are, you know, I'm lucky enough to stay in a large city, in a metro. I have access to the best healthcare, the best legal advice, the best education. But you go 200 kilometers away from where I stay and it suddenly starts disappearing.
01:20:01
Speaker
And how are we going to ever deliver high quality education, high quality health care if everything is going to have to, I mean, imagine the number of schools and colleges we need to keep up with our growing population and thus the number of teachers.
01:20:18
Speaker
We cannot have that. You know, government has this vision of having 35% GER ratio, which is gross enrollment ratio in higher education. If you do the math in terms of number of college seats you need and how many teachers you need, you'll find those numbers become very hard to imagine.
01:20:40
Speaker
So to solve these larger problems, you need AI. You will need. And once you start, once you see that, then it becomes easy to see that there's going to be a huge amount of education around.
01:20:54
Speaker
So that's fundamentally some of the pieces we saw, I saw at least early on. And what convinced me this is, this is the, I mean, it's necessary.
01:21:05
Speaker
know, there used to be talk that Earth cannot support so many humans.
01:21:13
Speaker
We are supposed to peak oil. We're supposed to run out of oil. We've, yeah, so these are all issues which have at some point or the other come up. And the limitation has always been our imagination.
01:21:28
Speaker
Fascinating. ah You know, there are a lot of these ah pet phrases of people in the AI world, like scaling law, and the bitter truth and ah attention is all you need. Context is all you need. Do you have any of these favorite pet phrases which you can explain to my audience?
01:21:48
Speaker
I happen to explain some of those. But so, see, you want to make it easy for people to understand. So you need those pet phrases. mean, so ah attention is all you need comes out straight. It's the headline of the title of the paper, which set off this entire large language model boom here.
01:22:05
Speaker
It was a seminal paper, which was actually came out of Google. And OpenAI beat Google to the race to get the world's attention, right? Yeah, yeah. the Yeah, to me, to mean the what I love about this intelligence era is that our imagination literally will become unbounded.
01:22:28
Speaker
We'll be able to experiment with and understand things. big We will live longer... healthier lives. AI will be critical in solving those problems around humanity, around ah diseases, aging, all of these problems.
01:22:45
Speaker
There's so many of these, which we have not really conquered. ah And ah not our lifespans, but our health spans will fundamentally increase. And, you know, we'll we'll be thanking AI for all of that.
01:22:58
Speaker
Good that it will do humanity. So ah but the timeline of that was 10 years. um and And that remains that. I mean, maybe nine years left now.
01:23:11
Speaker
But if you go look at the world's fastest growing AI companies, they're all application companies and not foundation AI companies. There are at least 20 AI application companies that have hit 100 million in revenues, which means their valuations are well north of a billion dollars.
01:23:27
Speaker
Well north of a billion dollars. And these guys have hit those numbers. Of course, there the US markets are larger. Easier to hit those revenue numbers there. um But this has happened in three years, two years, even one With a very small headcount.
01:23:47
Speaker
Yes, with very low headcounts. So in India, we may not match the revenue per employee. ah That metric is going to be different for us than it is in the US. But the fact of the matter is it can be done in a small amount of time, short amount of time with a limited amount of capital.
01:24:05
Speaker
If you go look at these, all these hundred million dollar companies, I mean, Harvey, Lovable, Together AI, I mean, I can just reel off names.
01:24:15
Speaker
curor I mean, these are not all of them are not super funded.
Rapid Growth of Indian Startups
01:24:21
Speaker
They just wrapped very, very quickly because they built really, they solved real problems. Well, And I think that's really the challenge for us. as i'm I'm seeing I'm very, very bullish about this because of the kind of innovation I'm seeing.
01:24:37
Speaker
Even within our portfolio, there are multiple companies that have gone from zero to close to a million dollars in revenue in less than nine months, six months. These are the kind of timelines. And these are Indian companies, some of them, of course, selling into the US.
01:24:50
Speaker
Those are actually growing faster. So that then the example would be from 200,000 to million in six months. so Those kind of numbers you should take two years to achieve.
01:25:05
Speaker
So we are seeing a 3x to 4x acceleration. And I'm talking about Indian startups. Today, given the kind of multiples we're seeing in the AI era, we are talking about a 25x of revenue at least.
01:25:19
Speaker
Which means to hit a billion dollars in valuation, you need to hit 40 million dollars in revenue. Actually, if the growth curve is steep enough, I mean, 30-40x is not a problem at all.
01:25:36
Speaker
So how many companies can we build in the next 10 years or 9 years, which will do at least 25-30 million in revenue? Starting today, let's even forget the last one here that's gone by.
01:25:51
Speaker
um my bet will be at least 100. If not, no, I mean, you'll you'll be, a few years later, you'll be saying, Raoul, you're underestimated. Yeah, yeah, twenty thirty and we it already god um yeah. What are you about?
01:26:08
Speaker
Though the assumption there is distribution, right? I mean, tracking distribution might not be as easy, right?
01:26:22
Speaker
So for consumer products, distribution becomes critical. For B2B, it's the enterprise sales piece and or so maybe, you know, product-led growth, PLG, that motion, whichever path these businesses follow, I'm trying to say 30 million in, let's say, a seven-year timeline is not a hard problem to crack, right?
Demand for AI Innovations
01:26:47
Speaker
If you start looking at it like that, how many companies can hit 30 million in revenue in seven years' time, then the number of the hundred becomes a much easier to understand problem. Yeah, yeah. True, true, true, true. Yeah, yeah. So, and see, distribution...
01:27:03
Speaker
If the product is good enough, distribution follows.
01:27:08
Speaker
The number of people are willing... So, again, going back to my portfolio, Irami, which does internal audit, signed a distribution agreement with PricewaterhouseCoopers, one of the big four in audit.
01:27:22
Speaker
Why? Because the product is that good. Distribution follows. People are looking for, people want AI innovation. There is ah unmet demand in the market. I mean, in literally every boardroom I have walked into, every C-level guy I've spoken to, keep asking you for yeah AI solutions.
01:27:43
Speaker
What do you think? Which is the best AI marketing tool? Now, it's a hard question for a VC to answer. it may not be in portfolio. So it's in my portfolio, easy question answer. And B, I'm not in the business of saying which is the best product. I mean, more in the thing I'm saying, like, ah you know, which is the best business to invest in rather than which product to buy. But I get asked that question a lot. I mean, now we built answers for that because simply because I get asked so often.
01:28:09
Speaker
But you see the point I'm trying to make. There's an unmet demand which is clearly visible and the budgets are moved towards the eye. People are, and the if you also look at build versus buy.
Buy vs. Build: AI Solutions
01:28:22
Speaker
So there was a debate whether people build or whether they'll buy. Right now, buy is winning. I think build will catch up over time. But buy is happening across the board.
01:28:34
Speaker
People are experimenting across the board. Some of that revenue will disappear because people will run, know, after the experiment, choose one tool to buy or no tool to buy and things like that. But the point I'm trying to make, the momentum, and this is, the momentum is driven by what?
01:28:50
Speaker
Not buy. Revenue momentum is never driven by Revenue momentum is driven by products that deliver value. If AI keeps on delivering value at the pace it is delivering value, building the 100 AI unicorns, I'm telling you, is going to be underestimated.
01:29:10
Speaker
Should India try and build its own foundational model? Of course we should. Should I fund it is a very different question. yeah No, we have so many languages to protect.
01:29:23
Speaker
We have so much rich history. Our people are different. our questions we ask are different. Our culture is different. All of that, AI at one level subsumes all of that, right?
01:29:34
Speaker
If you go look at a large language model, it has all that built in. And the way it answers has a cultural basis to answer, ah language basis to answer those questions.
01:29:48
Speaker
So like, for example, if I'm explaining a math problem to an engineering student in India, in let's say in Patna or any other a smaller city or town in India versus I'm doing this sitting in New York City, should i the explanation be the same?
01:30:08
Speaker
No. Yeah, the examples need to be looked like. How I take him through the problem, what problem do I take him, what examples do I give? What names do I use? Yeah, yeah, yeah, yeah. So when you start thinking cultural context and language are integral to any country, I mean, forms the basis of our culture.
01:30:30
Speaker
So that we should have a large language model is a necessity in my opinion. I mean, I'm highly supportive of what the government is trying to do here. And hopefully we will have those great models, much like China has been able to do a good job of it. yeah China has done an outstanding job of it, honestly speaking.
01:30:50
Speaker
ah China's approach has been open source models,
Open Source vs. Proprietary Models
01:30:53
Speaker
right? What is the difference between open source and proprietary model? so in an open source model fully open source model even the way it's the entire you can just download it and run it on your machine and you can also make changes to it because it is open source you literally have the diagram and you can say no no move it here do this do that and you can change the behavior of the product and you can run it anywhere without paying any license downloading a trained
01:31:24
Speaker
brain in a way, like a very layman understanding, which already can read, write, speak, whatever, certain language or based on whatever has been the training data and the focus. So you can download that brain and then you can modify it on your local hardware.
01:31:41
Speaker
And that modified version is only with you. Nobody else ah has that modified version. Absolutely. So that entire piece is what the open source piece is.
01:31:53
Speaker
And by the way, PipeShift, that's what PipeShift enables, right? That exact process you described. On the other side, you have OpenAI, which has, today OpenAI has also finally lived up to its name and released some open source models, two of them.
01:32:09
Speaker
Till now, it was very closed AI, not OpenAI, but Anthropic. There are multiple guys who, Gemini, will build models which, they own the source code. So you don't have access to the brain.
01:32:23
Speaker
You can ask the brain questions and the brain will answer your questions. So you access to its outputs, but not its internal design.
01:32:34
Speaker
You can't run it for free on any server you want. It is hosted and controlled by Google or Meta or OpenAI or Anthropic. Somebody owns the model,
01:32:47
Speaker
and controls the model and only they know how they turn internally the model looks like. You have no access to it. It's a complete black box as far as you're concerned.
01:32:59
Speaker
It's like the Google search engine. You don't know what's inside the search engine or how it was built. You can just ask Google a query and it will give you an answer. So when you talk of proprietary data and training a model on proprietary data, like say the ah the construction example which we discussed. And so if you have proprietary data of construction site videos and you want to train a model on that, do you typically train like one of these ah proprietary models or do you train an open source model in these kinds of scenarios where you want to train a model?
01:33:36
Speaker
What does that mean to train a model? So one, you can build a model from scratch. That's a fully proprietary model. And that is what China has done, and which is why it's phenomenal. And when you say build a foundation model, you say start from scratch.
01:33:52
Speaker
Literally, you just have the main base algorithm and it builds the entire thing. And you can you've designed that algorithm. So you start from scratch.
01:34:03
Speaker
On the other hand, oh
01:34:07
Speaker
where you take an existing model and then you modify it. You you know post train it, you fine tune it.
01:34:18
Speaker
And fine tuning is a technical word which means to make these changes in a large language model or a Gen AI model. That's where you take something which came out of the box, but you say, no, this will change my face.
01:34:31
Speaker
Like, my pink. You know, change the color. If you think of other things, that's what you would do, right? And that can, obviously, the cost of doing that is much lower than building from scratch.
01:34:51
Speaker
And the third model is, of course, to just use something which is ready-made and just so like OpenAI and you you have no control on it. right So you have three models, starting from zero, starting with something which is pre-built, but then changing it to your liking.
01:35:06
Speaker
And three is using something which and somebody just gives you and you consume as is, where is. So those are the three different approaches, I'd say.
Custom AI Models for Industries
01:35:16
Speaker
Now, when you come to something like Continuo.ai and their construction piece, that is not a large language model.
01:35:26
Speaker
That is a computer vision model. So in their case, they had to do a lot of work, which is very, very proprietary to them. And it's 3D.
01:35:39
Speaker
It's three dimensional. It is not even a two dimensional model. So they built a lot of proprietary stuff. On the other hand, if you go look at companies which are building in the language space, right? Like the internal audit one.
01:35:55
Speaker
The IRAME, which is the internal audit tool or Code Karma, which is doing helping you optimize code or... um Confido, which acts as a nursing assistant for US hospitals and consumes a certain amount of medical data, but a lot of patient data and hospital records.
01:36:15
Speaker
Each of them are working in different domains. And none of these architectures is only just a JNI model or a large language model. But they built some custom models, they built some proprietary ah ah pieces around it.
01:36:30
Speaker
And so that's typically how AI has been done in practice. So, when you say build a custom model, that's like a small language model, an SLM that they would have built.
01:36:41
Speaker
So, small language model would be a good example. like So, Flowworks, for example, built multiple small language models because they were doing that sales piece. So, they were building small language models.
01:36:54
Speaker
Some guys need to build, sometimes only work with a larger model because they need larger context. They need... more diversity of knowledge available to them.
01:37:07
Speaker
So depending on the use case, what is your starting point and what is your end point has to be very, very clear. And you can fine tune a proprietary model also, like you could take a Gemini, fine tune it, feed it your data so that it is... No, it's not fine tuning. That is not fine tuning.
01:37:25
Speaker
though the The model can consume your data along with it, but typically that sits outside the model. Okay, so that's not the prompt stage. that Your prompt can include. Okay. Yes.
01:37:37
Speaker
So you can give it context. So that's where the context window and size of the context window starts to matter. Okay, okay, okay. So the reason to build your own small language model is to reduce your token costs.
01:37:51
Speaker
but Otherwise, you will consume a lot of tokens because you'll have to give context to every prompt. Absolutely. So, token cost would be one simple example, one very clear reason, but more also domain specific.
01:38:04
Speaker
Like each industry has very technical terms in inside it, right? Like take prompt engineering. I mean, prompt and engineer, both English words, they mean very different things once you put them together and they're very AI specific.
01:38:20
Speaker
They're not going to bang into it in literature, for example. If you are studying English literature, would you have it learn prompt engineering? No, probably not. But then you would have all of Shakespeare that you'd want to learn in the turn of phrase and what that means.
01:38:37
Speaker
Right? Or idioms, for example. so so So depending on your use case, what is the syntax or the structures you wanted to learn change. And you want it to do better. It's not that you need it to only know that, but it needs to really know your domain very, very well.
01:38:56
Speaker
When an internal auditor talks about revenue, what he means by revenue is not the same as what you and I may mean with revenue. Because revenue annual recurring revenue can mean actually, annual recurring revenue itself can have five different definitions.
01:39:15
Speaker
But which one does it this business mean when it says ARR? It means this specific way to recognize this revenue. And that definition is specific to that maybe that industry or that business in particular.
01:39:29
Speaker
And you have to thus understand what it means because you're going to use it for audit and not just for general purpose to understand what is revenue.
01:39:39
Speaker
Awesome. Thank you so much, Rahul, for the lovely chat. i learned so much. We could keep going on, but I guess it's time for me to say bye to you now. Thank you, Akshay, for having me here. It was a great conversation.
01:39:54
Speaker
Lovely set of questions. Enjoyed… really enjoyed the conversation. Thank you so much.