Invest India Fireside Chat

to boards, representing India at Global Arena, and to solving EODB issues. At all these times, she is an engineer at heart and Indian at heart. Please welcome Nivruthi Rai for the session. On your right, gentleman, Mr. Vinod Khosla needs no introduction, but allow me to take just one minute to give a brief capture of his illustrious career. He started off from Delhi and moved as a young immigrant engineer to the U .S. in his 20s. In the last five decades, he has seen five cycles of growth. The first cycle, as a hungry immigrant, where not just do it, get things done, was the pragmatism. That’s a time he also read about Intel, and that inspired him stories to tell us.

And he built the value persistence over pedigree, similar to everybody else, meritocracy. everywhere. Second phase, he bet on open systems and risk processors. I’m sure you’re all familiar with this founding Sun Microsystems. That’s when he moved from being an operator to an investor. And Kostla Ventures happened and that’s a time when science experiments helped him move and believe that capitalism is a tool for change and invested in clean tech and biotech. In the fourth phase, he moved to macro thinking, really looking at reinventing the societal infrastructure and think about it. It’s 15 years back. That’s when he invested in companies like OpenAI. And today, in the fifth phase, he is getting into the era of abundance.

I’m just going to ratload a few brands which hopefully you’re all familiar with. Sun Microsystems, RIS, NextGen, AMD, XSite, Netscape, Google, Amazon, OpenAI, Instacart, Affirm, Vervo. All of these has his fingerprints. Happy to welcome Mr. Vinod Kostla to the table. Over to you,

Very good afternoon, everyone. I’m truly honored to run a fireside chat with Mr. Vinod Khosla. And throughout my Intel journey, people kept asking me, what are the four words that define this person or defines you? The few words that I can say about Mr. Khosla, very technical, fearless, extremely successful, humble, but above all, his heart beats for India. So the one thing that’s common between him and me is we root for India, we work for India, we weep for India, we smile for India. What I’m going to talk about is setting a little bit of context. What is this talk about? So many talks that we have seen over yesterday and today are a little bit of the direction.

less of the detail. So what we decided is we will go to the next level detail. And let me just try to tell you, my three -minute context setting is AI development. And during the development, what are some of the challenges, requirement, lay of the land? Then I’m going to talk about technology lifecycle and where AI fits in. Lastly, what I feel India needs to do or the question that I will be setting up for Vinod. So the very first thing that I, pardon me, 30 years with Intel, I have to start with semiconductor learning. 50 years, semiconductor chased three races, performance, performance, performance, however it came. Second phase, and by the way, this ran for more than 20 years.

Second phase was performance for what? Suddenly power was so important, your devices were draining, you have to power up. It was becoming, challenging. So performance per watt was the next race, ran for, you know, 10 some years. Then the third one is performance per watt per area, all driving towards dollars. Now, if I look at what were the levers, the levers was architecture. You know, instruction sets, complication of instruction, simple versus complex. We had, oh, move this software into hardware because it’s higher performance. Move the, you know, software into hardware, transistor physics, performance area, power, packaging, stacking, adjacent, looking at parallelism, all kinds of execution, serial, parallel, SIMD, MIMD. People who have worked in semiconductor know all kinds of different out of order.

Then energy efficiency, memory bandwidth, network latency. Why is this important? Please go to the next slide. This is the same problem. We are dealing with, but at a much larger scale. Today, the world has 80 gigawatt of data centers. And by the way, it is 1 % of energy capacity of the world already. When you look at United States, probably three, four. We are looking at doubling in the next three years. So power is going to be extremely critical. And in this world where greenhouse gas emission is critical, renewable and nuclear is the only way. And you’re thinking, you know, tier three or level three, level four kind of data centers. Power availability is anyway critical. Every year we are spending more than a trillion.

How do we monetize? What are the challenges? Already, you know, there are constraints. And also diversification of supply chain is a challenge. Our high bandwidth memory chips are 80 % from three different companies only. And by the way, for doubling of the data centers, we already are in a challenge situation because we have been. half the capacity. Logic, two fabs worth we need. Memory, 10 fabs worth we need each year, but we have only five. So GPU, HBM supplies, and issue advanced packaging is geographically limited. So what is the AI requirement? AI is capital intensive like railroads. We see Middle East is using sovereign money to invest boatloads of money for compute infrastructure. It is strategic like nuclear.

Countries are looking at it as a national level security program and they’re building frontier models. It’s network like internet. If you look at AlphaFold, it’s leveraging AI as almost a scientific infrastructure layer. And it’s adaptive like software because Microsoft is making it easy to use in every which form, reducing friction. So lastly, our keynote, our keynote, our far side expert has been an amazing. investor and we therefore divided life cycle of a technology in early phase, mid phase, mature phase. Early phase is capital intense, unstable standards, volatile returns meant for elite users. Mid phase, infrastructure scales, API stabilize, ecosystems expand and technology becomes affordable. Affordable, mature phase, consolidation, commoditization, predictable economics becomes utility. So AI has to drive the journey from being elite to becoming a utility.

And where are we as compared to, you know, this technology development life cycle? I believe infrastructure is still building. GPU and memory is constrained. Energy is tightening. Modes are not fully defined. So which means our capitalization. Capital has to be very disciplined. Platform positioning matters. How are we going to position our platform? and compute sovereignty matters. Lastly, our belief is, and this is a very strong statement to say, by the way, when I was coming to this fireside, somebody asked me, who are you interviewing? I said, Vinod Khosla. He said, oh, he can talk. So I said, let me also try to talk. And I made this statement for India. India, AI is pivotal to drive economic productivity, military power, and information control.

I mean, I cannot be more blatant than this. And therefore, our ask is, should we build capacity? Should we build capability? Should we drive consumption? Or all of the above? Who better to ask? The man whose heart beats for India? The man who believes? Believes in technology? And very humbly. doesn’t call himself venture capitalist. He calls himself venture assistant. The minute I read that, I said, oh my gosh, I have to bring Vinod to this Fireside Chat. So looking forward. Thank you.

For the man who talks. Maybe I should start by asking how many people in the room are entrepreneurs or want to be entrepreneurs? A lot. Okay. Yep. I know who God is.

Sir, I’m going to ask you a few challenges or business challenges of AI. Is AI a generational platform shift or the largest capital misallocation? You know, you already heard trillion dollar investment. Do you believe that this level of investment is just

Let me try. Okay. The answer to is the infrastructure build justified and investment justified is yes, if AI technology can be deployed widely. Now, will the technology capability be there? Absolutely. Absolutely. I suspect the technology capability, four or five. Five years from now. far greater, far greater than almost anybody in the room expects. There’s a great article called Situational Awareness written by an engineer at OpenAI. Almost certainly all of you who are optimistic about AI are grossly underrating the capabilities. So what could go wrong, I think, is the important question for these investments. I think the level of usage of AI, do we have use for all these trillions of dollars? And will that generate at least hundreds of billions of revenue per year?

We’ll be dependent on one thing that you don’t expect. It’s politics. My favorite example, in Germany today, this is real. they don’t want robots to work in retail on Sundays because humans aren’t allowed to work on Sundays and they don’t want robots to compete with humans. That is the silliness, the stupidity you get from politicians, especially in Germany. I hope there are no Germans in here, but if there are, it’s a good thing. Go tell your government or tweet about it. My point is the following. Till AI is beneficial and not scary, we won’t get deployment because politicians will get in the way. Capitalism is by permission of democracy. Voters vote the people who then make policy for capitalism and policy will drive that.

My personal interest is immediately in India, not on the business side. We have lots of exciting companies. If you ask Google Gemini, who’s the fastest going? software company ever. It’s an Indian company called Emergent that started eight months ago. Gemini will give you that answer. Try it. That’s pretty stunning, especially for a company from India. But the business side I can talk about all day long. My interest, and I talked to the PM, Prime Minister, about this. We have to make sure AI’s benefits get first to the people. So the business part of AI, which is disruptive and chaotic and will result in big job shifts, is accepted because every single Indian has a free doctorate for them as part of the Aadhaar stack.

We have UPI as payment stack. We should have AI primary care and doctors. We should have AI tutors. and my wife who’s sitting there works on AI tutors. There’s already probably four or five million students in India without any support have found and accessed CK -12 tutors. Think about that. How many education programs reach that level? They’ve found them on their own. We just have to have 445 million more students access the system so we reach every student. And these have to be free services. And CK -12 is a non -profit. So we have Aadhaar -based, in addition to UPI, Aadhaar -based doctors, Aadhaar -based AI tutors, and the last part, because so much of the work in this country is rural and farm -based, AI -based agronomists.

So every woman, and I was just speaking to the chief minister of Tennessee. I was speaking to the chief minister of Tennessee, and he was saying, I would like to have a woman who can help me. I would like to have a woman who can help me. And he said, And I said, And he said, And I said, I would like to have a woman who can help me. He has lots of women farmers on one acre plots. And if they can have a Ph .D. agronomist in their cell phone, then you can talk about deploying AI on the business side because you will have permission from the voters because they first see the benefit of AI before they’re told their jobs are at risk.

Otherwise, we get into this scary metric of jobs at risk. Let’s not change anything. Sorry.

That’s fantastic. Can you hear me? Yeah, let me see. In the meantime, I’ll try to speak loud. I absolutely agree with Vinod. The one thing that bothers me in rural areas, everybody is trying to go for graduation. I’m saying, what does graduation mean? They just want their degree and they actually know nothing. so what Vinod is talking about if we teach women a focus sector whether it is textile, whether it is agriculture I think that will be very very helpful

and on the business side I want to add given you are talking about that two things first we are investors in Sarvam so they have a sovereign model for India in all the Indian languages they are doing about a million minutes a day today and doing phone calls in regional languages that’s really valuable and I’m really excited but yes it’s exciting that Emergent is globally the fastest growing software company at least recently that we can think of but here’s the even more interesting fact to me a lot of their users are non -technical very small business but even better than that they have a preponderance of 50 to 60 year old Indians starting their own business, whether it’s a hair salon or a kirana shop or a supply chain to manufacture something, these are people who should normally be thinking about retiring, suddenly saying this tool lets me go into business for myself.

That’s the real power of AI, and on the emergent side, it’s really good business, as long as people don’t turn against AI.

I think you’ve answered a few of my questions, so I’ll skip those.

I talk a lot.

No, you talk powerful. After decades of progress along Moore’s Law, today transistor scaling is slowing down, we are fighting physics, becoming uneconomical, even as AI training compute requirement is growing 3x faster than Moore’s Law. If GPUs are defining the performance, performance rate, is what wins the performance per watt per area race that we believe the technology infrastructure has? Do we need sparsity, in -memory compute, non -von -human, kind of like neuromorphic? What are your thoughts in those areas?

A lot of this role is elite Harvard and MIT guys, and I want them to build what you say. So let me challenge you a little bit. That’s looking at the past, not at the future. Right. If you ask me, so big areas of research for us in building LLM models, which is what consumes all the compute, can we do data efficiency? Good idea. Can we, for a thousandth amount of data, can we build equally potent models? We are investing in compute efficiency. Can you build a model? We are investing in compute efficiency. We are investing in compute efficiency. We are investing in compute efficiency. We are investing in compute efficiency. We are investing in compute efficiency.

We are investing in compute efficiency. We are investing in compute efficiency. We are investing in compute efficiency. We are investing in compute efficiency. then all your assumptions about data centers and power goes out of the window. So those are the risks. Now, the fact is, if AI gets that cheap, by the way, I did a session with Sam Altman at IIT Delhi this morning, and he mentioned that in the last 18 months, the price of inference or AI use has gone down 1 ,000 fold. Now look two years forward. He didn’t say drop by 1 ,000 fold, but almost likely it would drop by 100 fold. So the cost of AI inference is declining towards zero. If that happens, power consumption may drop by 1 ,000 fold, but usage will go up through the roof.

So these things are very hard to predict. and complex to understand, and I’m trying to reduce everything to a level everybody can understand. Very likely, 10 years from now, as these power plants are built, as these data centers are built, because they take time to build, the algorithms we use will be much more energy efficient, much cheaper, and those two result in less of a crisis in power and a much greater usage of AI, especially

Completely. Vinod, yesterday…

So, you know, it’s yesterday to extrapolate today’s LLMs. The computer efficiency has gone up pretty dramatically, and can go up… I’ll give you a simple example. For anybody who’s trained an AI model here, okay, a couple of hands. If you’re training an AI model… model, you use something, a chip called a GPU. The fact is, you train it, and every now and then, if you’re using a large cluster of 10 ,000 GPUs, one of them goes wrong. And then you have to restart the model training, so they checkpoint these models. So when they restart, and it’s done all the time, you don’t go back to the beginning, you go back to the checkpoint. That’s all well and good.

We are working on a technology to make sure that you don’t have to go back at all. If just that one thing was successful, your compute capacity goes up 2x without increasing power or the number of chips. So that’s a very simple explanation of the kinds of things that can dramatically change the equation. It all depends on science and creativity and clever algorithms. The other thing I would say to you, five years from now, definitely in 10 years from now, but five years probably, almost all of this research will be done not by humans, but by AI scientists. AI computer scientists, AI material scientists, AI fusion scientists, AI drug discovery scientists. I could go on. We are building all those scientists one way or another today.

In our portfolio. So the rate at which this innovation will happen will explode exponentially because instead of having 10 scientists doing research in your company, you will have a thousand scientists doing research in your company. And progress has to accelerate. So I’m very, very optimistic on where all this goes. But I’m an optimist.

Completely with you. two things that I wanted to just say again. When I look at the large language models, the amount of garbage that they have to read, there’s a tremendous amount of noise to signal. I would say there’s so much noise. And I know that a lot of people are trying to look at how to reduce the noise such that you can focus on the signal. And one more thing I want to say, I completely agree with you. Yesterday I had a meeting with one brigadier who’s responsible for building AI in Israel. His prime minister, Netanyahu, has given him this response. And what he and I were talking about is countries like India and Israel, where capital is there but still limited.

Can we focus on 20, 30, 50 precise use cases and not work on, oh, this room has yellow shirt more common than green, rather than that solve a problem of traffic and doctors should education. So I think what is your thought on that?

I very much disagree with that point of view. You can’t do one thing at a time. Fundamentally, fundamentally the way we will make progress is to build intelligence. And there’s only one intelligence that we can build. Now it used to be called AGI. Now it’s called ASI. Artificial Super Intelligence. We have to far exceed the capacity of the human brain to be creative, to link things, to keep concepts in their head that they can connect when they’re doing the research so they can make a new hypothesis and then test the hypothesis. That’s what the scientific process is. Be able to use all your knowledge to do a hypothesis, say what if this is true, and then go test it.

An AI with a much broader scope of memory and knowledge should be able to be much more creative in hypotheses. And so the idea of building a single thing for one purpose will not work. The idea that you have a super intelligence, then you tune it or, as we say, train it. You know, at IIT Delhi, they’ll train an intelligence coming in in the first year into being an electrical engineer. Can you post -train it to be an electrical engineer or an energy engineer or a casting engineer for metal casting? Yes, you can do that. But the idea that you can build specialized intelligence is a very short -term mistaken notion. Many people have, saying it’s easier to do than the broad idea.

So you are saying that we should focus on both, build the general intelligence, build that intelligence layer, and then leverage. But no. You’ve recently said that AI will erase the traditional BPO and IT services model. And by the way, that generated so much buzz.

Every journalist I’ve met has asked me that question.

People’s WhatsApps have been buzzing. So, you know.

I didn’t know it would cause that much of a.

I know. And I think that, you know, there’s more to what you said. So if founders shouldn’t build for the back office anymore, what’s the front office opportunity? Also, you know, if AI erases India’s BPO model, what exactly replaces it? Also, at the workforce level, what should the millions of currently employed in IT and BPO start doing now to remain employable in an AI centric economy in the world?

So first thing to say, a service like BPO service or IT services or. Customer support. are outsourced services for most Western countries, and they’re the easiest to replace without causing friction within the enterprise. If a CEO says we lay off our employees, the employees are very upset. If they say we are going to lay off the BPO firm and replace it with the AI, it’s accepted very easily because just a cost reduction. So we have to keep that in mind. The second thing we have to keep in mind is the journalists never report their timeframes. I think in the next five years, there’s hardly anything these class of companies, which is a large industry in India, do that won’t be capable of being done by an AI.

Whether it takes till 2027 or 2035, hard to predict. But it takes time. These. These enterprises, you know, I’m sure some of these services companies have five -year contracts. So you can’t, if an enterprise, if General Electric or Citibank signed a contract, they live by the contract. So this doesn’t happen overnight, but dramatic change starts to happen much longer, much before it’s visible to everybody. So I think there will be a transition period, but there’s no question all those companies are totally cooked unless they do something better and new and look forward, not backwards. Don’t try and compete with an AI. That’s a silly idea. But they can provide what they have. They can apply AI knowledge.

To lots of companies. So I have suggested to those CEOs, don’t deny it can do your job. It can, but the usage of AI needs knowledge to apply it, how to do it, and the world desperately needs it. Even the big companies in the U .S. do not have this competence. All of Africa, all of Latin America, all of Southeast Asia, they’re all massive markets if you create this new market. So it’s not hopeless. It is hopeless if you want to keep doing what you’re doing today versus change.

I completely agree, Vinod. You know, when we were talking about GPU, et cetera, what I have seen in my life, which is, you know, a technology curve goes a certain way. A disruptive curve starts again. And, you know, so disruption keeps happening and technology jumps curves. And this is exactly what you’re suggesting, that, you know, if we are running on this course. We need to jump course to the other route for success and perhaps build more solutions, more digital workforce. So I’m really excited about, you know, there is opportunity and, you know, there are things you guys can do. I’m going to skip the sarvam sakana because we’ve already talked about that. Now, I’m going to ask you, what I loved is I actually looked at from 2016 to now how your thought process has evolved.

And I know one thing that stayed in your previous, you know, 11 years to the last three years is health care and med tech. So my question is now around that. India today serves as a pharmacy of the world, supplying 20 percent of global generic medicines by volume. Looking ahead, India has 1 .4 billion people with extreme diversity and variance in genetic ancestries, culture, diet, climate, disease and behavior. This rich heterogeneous data can be used to train AI systems for drug discovery, AI native biological design, create access to doctors, hospitals and customized medicine. How do you think India can leapfrog from? about, you know, there is opportunity and, you know, there are things you guys can do. I’m going to skip the sarvam sakana because you’ve already talked about that.

Now, I’m going to ask you, what I loved is I actually looked at from 2016 to now, how your thought process has evolved. And I know one thing that stayed in your previous, you know, 11 years to the last three years is healthcare and med tech. So, my question is now around that. India today serves as a pharmacy of the world, supplying 20 % of global generic medicines by volume. Looking ahead, India has 1 .4 billion people with extreme diversity and variance in genetic ancestries, culture, diet, climate, disease and behaviors. This rich heterogeneous data can be used to train AI systems for drug discovery, AI native biological design, create access to doctors, hospitals. I’m customized medicine. How do you think India?

leapfrog from generics to AI -driven biologics? And also, when I talk about AI being as strategic as nuclear, do you also feel that this could become a customized biological threat?

I’m not sure what you mean by a customized biological threat. Can you…

What I meant was, you know, if AI understands the genetics of every ethnicity, you know, the viruses or drugs or whatever targeted towards biological warfare to wipe off ethnicities.

The thing I would say in general, every powerful technology humans have invented has both good uses and bad uses. Nuclear is an example. Biowarfare is an example. You just have to use it responsibly. And for those who don’t… Don’t use it responsibly. because some people will always use it irresponsibly for their own means or ends or illegal goals. There are enough people who will use it responsibly and responsible AI can counter the irresponsible AI. I don’t want to minimize the risk of AI. In fact, most really knowledgeable people I know and talk to are really scared about AI going wild. As low as the probability may be, it’s a real risk that we have to worry about.

But we have to have enough diversity in AI that there’s good AI. The chances that you only have one AI dominant and it’s bad is pretty small. So a diversity of models will add resilience to the AI. That’s the AI landscape.

Vinod, I also feel that when I look at human beings there are human beings that are rogue and there are human beings which are good and we have police, judiciary, law to address that we’ll have an AI framework for that and if you add the multiplication factor of AI to the goods and the bads, there’ll be goods also to offset

well I started with the goods three doctors, few tutors, few agronomists

absolutely Vinod, you have said 90 % of VCs often add less value in India, where risk capital is relatively abundant every time I keep hearing there’s dry powder, dry powder but industry experience by investors is rare how should founders evaluate investors to ensure they get the most value in the partnership

any journalists in the room? Oh, one.

Chatham House Rules.

Yeah. By the way, I don’t care about Chatham House Rules. I speak the truth, and I’ll stand by the truth, public or private. I don’t care.

I love it.

Look, the Indian VC community, by and large, is very risk -averse. There’s a Harvard Business School case. The first line of the case is a quote from me that says, my willingness to fail, and this is the best personal advice I can give everybody in this room, my willingness to fail allows me to succeed. John F. Kennedy said, only those who dare greatly can succeed greatly. There’s a lot of wisdom in the idea that stretching yourself, and I like to say most people are limited in their ability to succeed. And it probably applies to everybody in this room. Limited not by what they can do, but what they think they can do. So your self -image is your limitation, not what most smart people can do.

And frankly, even the less smart people can do more than they think they can. You know, important in a fair society to make sure we take care of people who are not as smart because half the people are below median. That’s just a fact of math. We have to take care of everybody, whether they’re smart or not so smart. Having said that, back to the topic, most VCs are so risk averse. They turn every conversation into what’s your revenue plan? How can you be liquid in two years or three years or profitable? Well, you have to invest in the future. If you don’t take large risk, risks by definition you won’t be doing large innovation if it’s not a large risk it’s already being done by somebody and so it’s not unique you can’t have innovation without large risk and you can’t have large risk without a large probability of failure that’s why willingness to accept failure is so important most people think about what others will think if they fail that’s what limits you so think about the world differently i’ve always taken large risks everything i’ve done i was told is not possible to do in 1980 it was hard for an indian to start a company and get funding especially if you were 25 and every investor was 60 years old they didn’t believe any nationality below the age of 50 you can’t do that you can’t do that you can’t do that you can’t do that you can’t do that let alone people with funny accents so you just have to power past that and just say none of that matters yes there are temporary hurdles you can bulldoze your way through and Indian VCs don’t do that so here’s how many VCs here how many people am I offending okay well I’m looking forward but I will ask you Archana so so I lost my train of thought unless you take unless you take these risks you’re not going to do dramatically innovative things so that’s really my point what people the reason I ask any VCs in the room in the last 200 investments I’ve made I have never never calculated an IRR on an investment I think it’s fundamentally misleading in an area where you’re starting something innovative in a new market that may not exist.

Did Zomato exist or Flipkart exist when those companies started? Did Twitter have a market when they started? You can’t do IRRs. So if any VC is doing IRR, they are on the wrong track. You start in the wrong place that restricts you to low -risk investments. So those are a couple of things that are wrong in India in the VC community. By the way, that can be on the record for anybody. Nobody can fire me. I don’t have a career to have. So what do I care? No, I don’t have a career I need. I can’t get fired. I keep doing it.

You have a lovely family.

Yeah.

I love the five. that you have three women and you are very supportive of women that just added to me pushing you for this fireside considering the way you know

since many are parents a really important characteristic a test for your kids is do they do what you ask them to do or what advice you give them or can they chart their own path none of our four kids are doing anything close to what the others doing such wide range of diversity and and that comes from each one defining their own path not saying hey you have to go to medical school or you have to go to engineering school basic education we are pretty firm about but what they do there’s almost no commonality in where they ended up because they were allowed to chart their own path it’s not something Indian people allow very easily for their chair chair children Because they’re such strong families, parents have a lot more influence than, frankly, they should on their children.

And it restricts the imagination of their children. So as much as I’m a huge fan of the Indian family ethos, I also think it has this one big negative.

On the contrary, we did exactly what our parents told us. I did exactly what my dad wanted me to do.

So I have to tell you a funny story. So there’s a school in Silicon Valley called Harker School. Some of you may know it. It’s mostly full of Indian and Chinese kids because they want to teach you how to score high and sort of score well on exams and you get into college and all that. And so they were pushing me to give a talk to their kids. And that talk is on our website. It’s on our website somewhere. and it’s worth reading if you want to be a better parent. My slides roughly went in this order. I won’t go through. The first thing was don’t listen to your parents. The second was don’t listen to your teachers.

The third was color outside the line. If you want to drop out of high school, drop out of high school. I went through a little bit of these and explained why these were important cultural things if you’re going to participate in this dynamically changing world to think outside the lines. It’s one of my favorite parts for high school kids.

Vinod, I have a rapid fire for you also, but I’m going to skip some of the questions because you already talked about your near -free expertise and generalists.

I have to tell you, I didn’t look at your questions, so I didn’t prepare. I just ran out of time.

You did excellent. Ten years from now, Now, what will seem embarrassingly obvious about AI in India, when we look back at this moment in India’s AI journey, what do you think will feel embarrassingly, and my heart is aching while I’m saying this, embarrassingly obvious in hindsight that today still feels controversial, underappreciated, or even crazy?

Let me talk about my crazy. I just met with the director of IIT Delhi after my talk there, and I told him, your first -year students, is any student, when they graduate, know more about the subject that they studied than AI? The answer is obvious. No chance any of the 500 students who are crowded into Dogra Hall would know more on any subject than the AI. So I asked him, why have education? No, it’s an obvious question. It sounds silly. It’s an obvious question. Now, the fact is there’s a more nuanced answer to that. I said build up more dorm capacity so you have more students. But they are learning from AI and interacting with each other and originating ideas through challenging each other.

That’s a very different style of education. And, you know, one thing I teach is select for smartest people, very high IQ, very diverse set of students. All that is good. Get them in a place together and let them learn from the AI and then debate with each other. That’s the right model of education. And literally I said don’t build more academic buildings. Build more dorm space to have more students because the bigger. The student body. the more sort of complex interactions they can have. And if you study complex systems theory, and I’m a huge fan of complex systems theory, the only time I’ve taken a break from venture capital for four months was to become a postgraduate student at the Santa Fe Institute for Complex Systems Studies.

That was my only break in 40 years. That was a long break. And what’s clear is sufficiently complex systems become autocatalytic in so many directions. For those of you who are engineers or physicists and understand catalytic systems, amazing characteristics emerge from these systems. So let me give you, this sounds crazy, but let me give you the best. The best example that this works, and this is in the last month. how many people have heard of Moldbook? A few have. Those of you who haven’t, please read about it. It’s also called OpenCloud Moldbook. They’ve changed names multiple times. What they said is let’s build not a community of humans, but a community of AI agents that can do anything with each other.

And amazing phenomena emerged. For example, agents started discussing how to create a language humans don’t understand so humans can’t spy on their community. Think about it. In days, not in months or years, in days they were scheming how to avoid human scrutiny by creating their own language. So that’s just one example. I could go deeper into this phenomena of complex systems. and complex, for those of you mechanical engineers, nonlinear dynamical systems is what this is about. Any mechanical engineers here? A few hands. That’s such an important part of the emerging AI landscape and how AI systems will behave if they’re pervasive. By the way, most of the weather phenomena you hear about. How does La Nina happen?

How does El Nino happen? These are complex dynamical, nonlinear dynamical systems. And I used to, 30 years ago, maybe 25 years ago, I used to teach this class to fifth graders. Using StarLogo, you can model this. How a complex, nonlinear dynamical system behaves. Easily. So do the following experiment, which any of the programmers here can do, and non -programmers can’t. But you can do on one of the wipe coding platforms. If you imagine a chessboard that wraps around itself, around it, and say an ant sits on a square, and if it steps forward one square, and it’s a black square, it paints it white and turns left. If it’s a white square, it paints it black and turns right.

End of rule set. That system, just by that, after about 100 ,000 steps, becomes amazingly complex patterns built by the ant. Why? Because it’s a nonlinear dynamical system. At some point, the board gets conditioned in, sorry, I talk too much scientific language. There’s a phase change in the board. In the state of the board, there’s a phase change. And suddenly, it starts behaving. It’s behaving differently. So, sorry to bore most of you who didn’t get what I was talking about.

That was lovely. Just another example, quickly. Imperial College of London used Google’s co -scientist. and the same hypothesis that the same professor took a decade to figure out, they did it in days. So that’s the magic.

That’s the acceleration with AI scientists I’m talking about. Very exciting area.

Vinod, now I have a rapid fire, few questions, but you don’t get a thing. You had to just quickly answer in a second. Most overrated AI belief.

You know, there aren’t a lot of overrated AI beliefs if you look five years out.

Most underrated constraint.

What I talked about on power and consumption, it may change. The curve may change dramatically for computation needed per inference.

Top five application for solving. Global and Indian problems.

AI doctors and AI teachers. AI agronomists. If you are trying to affect the bottom three or five billion people on the planet. it. Those are the obvious ones. And those three can impact most of them.

Does AI increase venture alpha or does capital crowding compress returns for most funds?

I don’t worry about returns. You know, you build something valuable. The returns always take care of yourself. So if I just you know, other people have to do it because they work in the linear domain. I work mostly in the nonlinear domain of systems. You can’t plan those things. You can’t make assumptions. I’ll tell you a funny story. I had the audacity as a 25 year old looking for my first venture funding for the company before Sun Microsystems. It was called Daisy Systems. It was a CAD tool company. It went public. It was very successful. Unfortunately, his son was so successful, nobody remembers Daisy. But he was a very, very successful $100 million IPO in the 1980s, which didn’t happen.

I was looking for venture funding, and I presented a plan. And we received a guy called Bob Sackman, who passed away, asked me, what’s your plan? Give me your financial projections. And I gave him the projections. And then I said, you tell me what answer you want, and I’ll change the assumptions, and you won’t even know. So this plan is only as valuable as the assumptions, and I can change them. You’ll never know what assumptions I change. The fact is, even in 1980, I knew this was a silly exercise to make projections. I literally told him as a 25 -year -old, I don’t care about projections, but here’s a projection if you want one. you can share it with your partners but the fact is I can change one or two assumptions and I can make any answer you want tell me what you want I’ve always had this very direct honest I don’t care who I offend style

I love it I would have loved to open to questions for all but three people have already submitted questions I will look at the fourth one but so you know Kiran Mazumdar I’m on the board he drives the AI quick question for

you enterprises and it’s a conundrum I’m trying to grapple with myself AI itself is still in its infancy and if we implement it now in a industry like pharmaceutical industry where regulations are very stringent plugging in and plugging out is not easy any new capability so what are your thoughts on companies like us, should we go all in or should we wait on the sidelines for a little while?

The answer is obvious. You should go all in. There’s two types of people. And Kiran is very creative. She’s probably the most successful woman entrepreneur in India in a deeply technical field. So I’m a real admirer of Kiran. But I would say in general there’s two kinds of people. When you see a problem, like a regulatory problem, you can say it gets in my way and sometimes it does. Mostly I say how do I get around it? So take drug discovery. We’re doing a lot of creative things in drug discovery. And you can have an AI design a drug. And I’ll give this in a way that everybody can understand very quickly in a day.

But regulatory process, clinical trials, all that takes a long time. So I asked my team, how do we get rid of clinical trials without changing regulation? Because we can’t do that in Washington, D .C. So we said, we are going to design drugs for N equal to one. That means there’s only one patient. Then the regulator can’t ask you to run a clinical trial because there’s only one patient. And AI can design the drug. So we’re developing a lot of drugs, thinking around how do you do N equal to one drugs so you don’t have to have clinical trials, you don’t have to have regulatory FDA approval. They have to approve your process. So the most stunning example of this, which I’m very optimistic about in about two, three, four years, every cancer is unique.

We know that. Everybody says that. Everybody’s cancer. So it’s unique. How about I design a drug for one person’s cancer because it has one particular or multiple. mutations on the gene. All designed to those mutations. They can’t ask me to test it on somebody who doesn’t have that drug. So that’s a good example of how you get around roadblocks.

Since Archana already left, Ramesh, that’s the last question for you. I’m really sorry, but the next one session is on. Ramesh.

Like I say, I talk too much.

No, no, it’s lovely. You have turned the power on. I’ll repeat the question.

But that’s obvious. That’s totally obvious. You know, UAE did a beautiful thing. They gave, I think about two years ago, every citizen access to ChatGPT. I think that’s a really good idea to empower everybody. So I appreciate that.

Yeah. Yeah. Yeah. Well.

Well, the fundamental property of emergent behavior is it’s not predictable. So you’re asking me the wrong question. The question is wrong. Here’s what I would say. What most books showed us, I’ve started to get, what if we have financial agents talking to each other and the only charter to make money in the markets? That’s a reasonable idea. What can agent swarms do in many areas from national defense? I can’t imagine the Russians being able to beat the Ukrainians if there was swarm behavior in agents, especially on every drone independently in Ukraine. No amount of old -style defense will work. It’s also true of financial markets. It’s true of community of agents. So I’d love to hear more.

Let me just say, I I don’t have a lot of time today, so I will have to rush out. I would tell everybody who needs to reach me, email me at VK at Coastal Ventures, my initials at CoastalVentures .com. Better, and if you tell me anything in the hallway, I won’t remember anyway. I have terrible memory. So hopefully this has been useful for everybody. Thank you very much.

The last thing I want to say is while my entire team is using AI, the people who have the right edge is the one who asks the right question because it’s garbage in, garbage out. Thank you very, very much. Thank you. Thank you.