Jensen Huang Founder and CEO of Nvidia

0:08Jensen, a big thank you for taking the time today. I have to say I'm normally very excited when I record this podcast, but wow, when I did the preparation for this podcast.

0:30And I'm happy for our company to be able to contribute to so many different areas of science and society. And so it is a thrilling time. Now, you delivered, well, you hand delivered, actually, the first supercomputer to open AI some years ago. Tell us about that. We were just getting into artificial intelligence ourselves and we're working on autonomous vehicles. And so we were imagining what kind of computer has to be built for this new way of doing software. You know, as the audience probably know by now, that artificial intelligence is a computer that works with software programmers to write software that is refined from data.

1:21And that software is impossible for humans to write. It's a gigantic body of code. And it requires a special type of computer. And so when deep learning first came along, we reasoned about how it would fundamentally change computer science. Because the early effectiveness of the first deep learning network that the industry, a lot of people saw, was AlexNet. And AlexNet was designed for computer vision. And its effectiveness was so incredible. It shattered records and shattered the effectiveness of computer scientists of several decades. And it was a piece of software that required NVIDIA's GPUs to produce.

2:14And we were so inspired by that. We reasoned about what kind of computer, what kind of software is this? Where can it go? What kind of problems can it solve? and what is the implications to everything about computer science from chips to systems to operating systems to data centers to networking to the algorithms all the way to applications. And we came to the conclusion that a new type of computer had to be created and we created this new computer we call DGX. It's a deep learning system and it's an AI supercomputer, if you will. And I delivered the world's very first one when I announced it, that we were building it for ourselves.

3:03I thought that some people would be interested, but it turned out a lot of people were interested. And so I delivered the very first one outside of our own company to OpenAI. What were the big breakthroughs that you had to go through to get there? Well, the way that the software is written, it processes a giant amount of data to find patterns and relationships, patterns and relationships in the data. And the data is ones and zeros. and somehow using these architectures of deep learning models. And the way that the deep learning neural network is constructed, it was possible to scale into very large models, very large networks, and process a gigantic amount of data looking for patterns and relationships.

4:01And so the question is, what kind of computer would be good at that? And so we reasoned through the computer architecture and came to the conclusion that every aspect of the computer has to change from the way the processor is designed. Our GPUs, our graphics processors, which simulates the world, was almost a perfect starting point for understanding the world. You know, a chip, a graphics chip that was designed to simulate the virtual world, as it turns out, is fairly good at understanding the world. The mathematics is similar. But yet, the size of the data and the amount of computation necessary to do it, to go through all of that data is enormous.

4:47And so we broke it down chip by chip, and the CPU was no longer the ideal processor. And so we evolved our GPU to become a deep learning, if you will, a artificial intelligence processor. The PC architecture is suited for personal computing, but it's not suited for giant AI systems to learn from data. The IO has to change. The networking has to change. We bought a company called Mellanox so that we can change the way computers are connected and the way computers can work together. Instead of maybe a handful of CPU cores working together, we have millions of GPU cores working together to process the data, to find patterns, and to learn what is called a representation, to learn the language of the subject that you're trying to learn.

5:42You could be learning the language of music, you could be learning the language of language, human language, or the language of the world, the physical world. To learn computer vision, we can learn proteins, we can learn chemicals, we can learn all kinds of things now. And of all these things, what were the biggest challenges? Well, the biggest challenges, and it remains now, is that the problem space is gigantic. Whereas an application for a phone can fit in a few gigabytes, you know, or a PC, a few gigabytes. The software that we're talking about here, you know, ChatGPT is 100 and, well, ChatGPT3 is 175 billion parameters.

6:32And each one of the parameters could be a floating point number. And so 175 billion parameters just for the neural network math, not to mention the applications that sit around it. And so this is a giant application that doesn't fit in one PC. It doesn't fit on any phone. It doesn't fit on one PC. And it takes many computers working together just to run it. It takes a giant data center to learn from it. And it takes a large number of computers just to run it. And so this type of application has never existed before. And so every aspect was changed. We literally re-architected everything of computers that we know of from the ground all the way to the top.

7:18And so now these AI computers are unlike anything that we've ever built before. And because the performance needed is so great, it takes weeks and weeks to process the data so that we can learn from the data and learn its representation, learn the model, if you will, the model of the subjects, the world model that you're trying to learn from the data. It takes so much time to train these models that if you could even shave half of the time off, it's measured in weeks. You mentioned GPT-3. What's the progress from 3 to 4, just in terms of complexity? Well, the complexity is hard to estimate because OpenAI hasn't really described it.

8:03But there's a great number of new breakthroughs. One of the ones is the fact that it can learn from both language and images at the same time. Just as with humans, we learn more about anything if we could see, if we could read the words and see the images. And one good example is if you've only seen horses and you've never seen a zebra, but I told you that a zebra is like a horse, but with black and white stripes. the fact that you have knowledge of both modalities, images and language, allows you to connect the two and learn something in your brain, even though you imagine a zebra in your brain, even though maybe you've never seen one actually.

8:56And so the ability for us to learn from multi modalities is very important and GPT-4 has learned that capability.

9:07You mentioned recently that we have now reached the iPhone moment for artificial intelligence. What do you mean by that? Over the last 40 years or so that I've been in the computer industry, we transitioned. We phase shifted. You know, at first slowly and all of a sudden abruptly, just like water turning to ice or water turning to vapor. At first, the temperature kind of, you know, increases linearly. And all of a sudden at some phase shift, the structure of the molecule changes all of a sudden and instantly something happens. It happened when we went from many computers and workstations, client servers to personal computers.

9:59The very first four or five years, it grew linearly. All of a sudden, Windows 95 came and everything about the personal computer changed. And yet the personal computer was introduced some 10 years before that. And the same thing happened with the Internet. The first five, six, eight years, scientists and researchers were already using the Internet. And then all of a sudden, one day, Mosaic came along, Yahoo came along, and bam, there's a phase shift in the way that the Internet was perceived and used. In every single one of these transitions, then cloud, then the iPhone, the mobile cloud, in each one of these transitions, the computer itself is programmed differently.

10:52It's easier to program. Let me give you one example. The number of mainframe applications in the world is not that many, but the number of iPhone applications is over 5 million. And so the fact that there are so many applications must suggest that it's easier to create amazing applications. And it's absolutely true. The applications we have on iPhones and the mobile devices is surely beyond expectation, beyond imagination just 30 years ago. And yet, you know, people are creating these applications, obviously, very, very quickly. And so the programming model has changed. The application capability has changed.

11:31And the reach of the computing has changed. And so let's apply that to GPT. Let's apply that to artificial intelligence as we know it. The way that you program this is just with human language. This is the only computer, this first computer in the history of humanity, that everyone can program the computer to do something. And there's no programming language. You don't have to use BASIC. You get into Fortran, Pascal, C, C++, Java. You don't have to use any of those programming. Like Python, you don't have to learn anything. Which is really good because I tried to learn Python last year and it didn't quite work out.

12:06To the extent that I don't have to sing at the summer party. The number of programmers in the world just increased from tens of millions to several billion. And so we've narrowed, if you will, we've democratized computers. We've democratized computer programming. And we've closed the gap between the have and the have not access to technology. The technology divide has now been closed. What are the implications of this democratization? When you democratize technology and you put it in the hands of almost everybody, you empower everyone. Look at the number of applications that are coming out now that are based on ChatGPT, where people are connecting it to applications, making the applications better.

12:54People are using it to write stories, create music, write programs. So instead of writing a program, you tell ChatGPT to help you write a program. You tell it the problem you're trying to solve, and it writes you a Python program. It writes you a SQL query, and it even creates a website for you. And so if you want to go into business and you don't know how to create a website, you can now just tell ChatGPT to help you create a website. You describe what you want, and it connects it all up for you. and it's operating. And so here's a computer that can help you write programs, solve problems, empower you.

13:31And so that's one of the greatest things of democratization of technology. We've now put this amazing tool in the hands of everyone. What do you think it would do to society? Well, the first thing that's going to happen is our productivity will go up. Any profession that relies on knowledge and the access of knowledge, the application of knowledge, will now be boosted. And so if you have domain knowledge, and most companies in the world has very, very deep domain knowledge, it's the reason why they're a company. Their domain knowledge can now be put in the hands of their employees and applied and accessed and applied in a much more rapid way.

14:16Of course, there's a lot of mundane information tasks that are now, if you will, commoditized, and it's automated. You know, one of the things that's really incredible about artificial intelligence, and the reason why this is definitely going to be the next industrial revolution, is instead of producing steam to electricity, instead of mass producing physical things, we are now going to be producing the most valuable asset, the most valuable commodity that we know as society, information, knowledge. And so the production of intelligence is going to be what all companies do in the future. NVIDIA has AI factories.

15:03We put data into it and improved software, and the software is artificial intelligence, improved intelligence software comes out every single day. I go to sleep and it keeps producing it. And we keep refining more data. We keep improving the software. The software helps us design chips. It helps us operate robots like self-driving cars. It helps us do computer vision for quality inspection. It helps us develop software that helps us design and manufacture chips better. And so every company will be able to do that for their own particular domain. So I think the next industrial revolution is going to be about the production of intelligence.

15:46And for industries that relies on intelligence, our productivity will be insanely boosted. And of course, there will be some jobs that will be changed. There will be some jobs that will be created. Right now, we're creating a whole lot of jobs for artificial intelligence data scientists and people who understand this field. Of course, some jobs will be displaced. And so we have to make sure that as a society that we understand what this technology is and take advantage of it as fast as we can so that we understand it. And apply it to social benefits. How much do you think productivity could increase on the back of this?

16:28Well, there's a few ways we can measure it. So let me give you a couple of examples. One of the hardest things that we do in our company is designing chips. The chips that we build are the largest chips, the most complex chips the world builds today. No singular entity builds such large, enormous semiconductor chips. And these chips are simply impossible to build anymore without artificial intelligence. And the reason for that is because the number of transistors and the way that we can connect up those transistors, the combinations is just so insanely great. And because so many people work on it, the optimization of the mathematics, the optimization of how to place, you know, it's kind of like imagine in New York City.

17:18but it's a thousand times bigger than New York City. And you're trying to figure out how to organize New York City from the ground up such that it is the most optimal placement of every single building. And then you have to understand where the traffic goes from building to building. You have to understand which buildings are associated with other building and what buildings are necessary to support certain buildings in certain infrastructure. Where do you put the parks? Where do you put the restaurants? It's insanely complicated, as you can imagine. You know, it's the number of combinations is off the charts.

17:57And we can't solve these problems anymore without artificial intelligence. So on the one hand, it lets us do things that we can't otherwise do. Another hand, let me give you another example. We use artificial intelligence right now to try to better understand climate change. And in order to understand climate change, you have to simulate the weather a lot more quickly because you're trying to further extrapolate the implications of climate out in the distance, not just tomorrow, but ideally next month, next year, next 10 years, next 30 years. So in order to do that, we have to do weather simulation a lot, lot faster.

18:40And so we've created artificial intelligence that helps us simulate the multi-physics of weather. And we're already simulating weather now 10 ,000, 50 ,000 times faster than you using numerics. And so that's another way of thinking about productivity. When you can do something 10 ,000 times faster, you're doing it 10 ,000 times faster. One last example, the single greatest expense in our company is software engineers. And now with Microsoft's co-pilot, you could, And they've estimated that some 40%, 50 % of the software that's now written in GitHub is produced by AI. It's a little bit like text completion.

19:22It's a little bit like grammar correction in our word editing documents, except this is for program completion. And so the AI can suggest, based on what you've already written and what you intend to write, it can write the program for you. So if you could imagine, the single most expensive population at NVIDIA is now amplified by a factor of two. That's incredible. And so our estimate is we're going to improve the productivity of our engineers by a factor of 10. When you talked about complexity being a thousand times the complexity of New York City, that you put in on what kind of area? How big is Wollongonger chips?

20:04Rich ships are probably, comparing it to a stamp, it's a couple of inches per side. And what is a couple of inches per side? It's kind of like a, it's smaller than a coffee cup, a coaster, probably. Probably two-thirds the area of a coaster, if you will, just to put it in perspective. The R &D budget for it is probably something like$5 billion. And then it costs more to build one of these generations than, for example, to build Rocket. And the R &D budget is very high. When you put together everything you said about productivity gains, when you look at the whole society, how do you think this could drive productivity gains in the whole society if you were to put a number on it?

20:57I don't know how to do that. But one thing for sure, the countries that don't have the richness of computer scientists and haven't benefited as greatly from the enormous capabilities of computers, this should be a reckoning moment for them. This should be just an extraordinary opportunity for them. The up-and-coming economies, the up-and-coming industries, I think India, Southeast Asia, Africa, these are regions and economies that I think has a real benefit from artificial intelligence, enhancing the capabilities of the entire industry and their economy. and then driving productivity to the limits.

21:54So I think for the rest of the world, for the developed countries, the ability to reduce costs is incredible, not to mention accelerating everything that we do. So when you look at the most important problems that AI will solve over the next five to ten years, what are they? One of the most important ones is digital biology, drug discovery. Just as we've learned the language of humans, we've now learned the language of proteins. And we've learned how to understand proteins. And we've learned how to, from the desired function, a protein is a machine, the biological machine. The way that it's connected, the amino acids, the chain of amino acids, and the way that it comes together, the 3d shape of that protein determines its mechanical functionality if you will it's kind of like you know the difference between the shape of a of a motorcycle in the shape of a car in a shape of a you know unicycle the fact that there are different shapes their functionality is different a plier and hammer the fun the functionality is different because of their shape and so proteins have different shapes and different functions we can now from the desired function of a protein synthesize other proteins, other potential proteins that have properties that are maybe better for temperature or better ability so that we could, it goes into our bloodstream better, or maybe we can use it to synthesize energy from light.

23:26Maybe we can break down plastics. Maybe it could break down oil leaks in the ocean or, you know, whatever the The interesting problem is we can now use protein machinery and protein engineering to go help solve that problem. I think that that's tremendous, incredible potentials. We can understand the language of chemicals so that now that you can understand chemicals and proteins, you can understand their interactions and do a better job discovering drugs. Drug discovery still costs enormous amounts of money. It takes a very long period of time and our success rate is very low. And so now we can improve the odds of that.

24:09And so drug discovery is one. The other one is climate change, both in understanding the impact of human factors to climate change, predicting climate change and the climate effects in regional climate impact, whether it's extreme weather in the Gulf of Mexico or the number of fires that because we just have so many dry days in Northern California, climate change has a different impact in different parts of the world. And, you know, people are interested in average climate change, but not really. People want to know what climate change has an impact on them and their local economy and their agriculture and their water supply and the quality of life and the impact of extreme weather and so on and so forth.

25:04And so we want to have a better understanding of the future of climate. And by doing that, the side effect is that the algorithms, the mathematics that we and the way that we do computational physics can have a tremendous impact in just about all other physical fields in reducing the amount of computation necessary to do it. And so on the one hand, we have the ability to predict climate. On the other hand, we use less energy to predict climate. And so that artificial intelligence makes that possible.

25:41We had Bill Gates on the podcast recently, and he talked about a personal agent in a way like a digital personal assistant. How do you see that? We'll have personal digital assistants. We'll have group digital assistants. Maybe we have a study group and we have a digital assistant to help us. We'll have a company, digital assistants of all kinds, somebody who is a digital assistant for HR, somebody who is a digital assistant for IT, somebody a digital assistant for programming, somebody who just wants to – you want to understand, you want to model our company's business or model the potential effectiveness of a new product or a new service.

26:26And so there'll be digital assistants of all kinds. How close are we now to artificial general intelligence? I think Microsoft said that they have seen some sparks of it. What's going to happen here? Well, I'm anxious to see the paper. And it's 150 pages. And so I'm looking for it. I've downloaded it in one of these weekends. I'm going to go through it. Intelligence is about perception, reasoning, and planning. we have done an extraordinary job with perception and but we still have a long ways to go in order to in order to really have a you know the goal of perception is to create a a model of the world around you a model of the world around you in a static form but also in its dynamic form you know what are if i did this what would happen to that you know we do this all the time.

27:21Today, the conversation is you're imagining if you ask this question, it might lead to this answer, which leads to another question, which leads to another answer. And we do this in human interaction. We do this in company and industrial interaction. We do this all the time. And we have a mental model. Some of it is supported by simulation, a mental model of how the world behaves. We have to go create that model of the world. And there's so many different worlds. There's the world that is the human scale world, but there's the world that is molecular scale. There's the world that's atomic scale.

27:55And then there's, of course, there's the world that's galactic scale. Each one of these worlds are described sometimes by different physics, right? At some level, you have to go to quantum physics. And all of these, understanding these different worlds all matter. And so the first thing is just understanding the world. The second part is how do you reason through problems in a way that achieves the objectives but are supportive and within the realms of your core values, your principles, keeping other people safe, that's explainable, interpretable, and that's in a transparent way. how do you reason through all of this with those things in mind and then how do you come up with a plan that is efficient and cost effective and all of the things that we do we do as humans and as industries and so those three those steps if you will AI is making tremendous progress along that entire arc robotics is making great progress and that's understanding the world and be able to plan your motion.

29:11We're making great progress in autonomous vehicles. We're making great progress. ChatGPT, obviously the fact that it can take a problem that you described and be able to break it down into a computer program, obviously suggests it has the ability to reason through several steps. It might not be able to reason through as many conceptual steps as we can, but it's surely demonstrating the ability to do some early level reasoning. And so the progress is quite fast. Changing your tack a bit and zooming in on the ethical side of this. Now, we just had a letter recently from a thousand really well-respected people who said it's time to slow down, think, and reassess.

30:00What do you think about it? AI is a very powerful technology. And it's a very powerful technology because it can perform tasks and do things that are of great value. And technology that has this level of capability obviously can also be applied to do harm. And so regulation is necessary. We regulate serial, for God's sakes. We should regulate AI.

30:31And this technology, of course, is moving very quickly. And so it's sensible that regulators really have to get engaged and understand technology to the best of their ability. But put some guardrails in, put some regulation in so that the technology can advance in a way that's helpful to society and not hurtful. Have we got any guardrails in place now? No, not really. Not really. Who should put them in place? Well, the same people that governments. There's really no choice but for governments to step in and regulate this. We regulate food. We regulate drugs. We regulate transportation. We regulate industries, the creation of chemicals, the creation of materials that could be toxic.

31:27We regulate just about everything. We regulate electricity. We regulate communications. We regulate the broadcast of television. There are certain things that you can't broadcast. There should be certain things you can't generate. Generative AI generates information. There are certain things you can't generate. On the one hand, it's hard to regulate people expressing themselves because of open speech. However, it is possible to regulate what information you produce. And so the regulation of the production of things, and now we're producing information using computers, you can regulate that. And there are many things that you can regulate.

32:12When you look at this arms race we are having now in AI and, you know, powered by your technology, are you afraid? Whatever power of technology there is, we should try to democratize it the best we can. If it were to land in the hands of one company, it's obviously less good than being available to everybody. it's less likely that in the near term that ai is going to displace our jobs it's more likely that someone uses ai is going to displace our jobs and and the same thing could be taken to all all kinds of extremes and so so when the new technology that comes along that produces so much productivity gains, whether it's the steam engine or heavy machinery, it gave us superpower.

33:06A tractor gave us superpower. Forklift gave us superpower human strength. And now we have this capability to give us amplifier intelligence and help us solve problems a lot more quickly. We've got to find a way to use that technology as soon as we can, but make sure that that technology is available to everybody who would like to use it and regulate it, you know, regulate it as soon as we can. You said it should be democratized, but, and of course, OpenAI was meant to be open, right? Now it's turning into a commercial product. You were more guarded when you talked about the specifications, you know, there is a lot less disclosure about the underpinnings of it.

33:53How do you read that? Well, that's a company choice of theirs, and they have the right to do that. In the meantime, there's a great deal of AI research that's still done in the open. The number of large language models that are available in the open is quite abundant. And so it's not about access to the technology that is keeping anybody back. It's simply the willpower to go and the insight that the technology is at a very close to useful state, that insight was terrific. The insight that between GPT-2 and GPT-3 is a very useful product. The difference between a marginally useful product to an incredible useful product, that was a great insight.

34:38Those are the same insights that led to the iPhone or that led to the PC, that led to the internet. before each one of those that led to Google search. The insight that led to each one of those innovations is really about timing. You know, technology was invented early on and it was even cultivating and brewing in certain circles for quite some time. And yet the innovators are the ones that realize the timing is now and to jump on it and industrialized it. and turned it into a really great product. ChatGPT is unquestionably the single best software product the world's ever made. And by that definition, let me defend that.

35:29A great software product is something that does amazing things and surprisingly amazing things. And a great software product is also easy to use. This is the easiest product to use on the planet. Anybody could use it. Over 100 million people have used it. and there's no instructional manual. You don't read anything. You just start typing into it. And if it's not sure what you meant, it asks you questions back. And it tells you that it's not sure. You just keep talking to it with whatever language you use. And it produces amazing things, surprising things. It is the single most useful, best application the world's ever written.

36:10How is this going to change geopolitics? How is it going to change the relationship between the U.S. and China? Well, hard to say. Hard to say. I think there's a genuine harm that can come from fake news that's being generated, fake information that's being generated. And that could cause real harm. The same harm that's currently happening in social media and fake news, and some of it is generated by human. well most of it is generated by human today and so you could imagine that this AI has a better ability to detect human generated fake news but this technology also has the ability to generate fake news and so both of those both of those possibilities exist in abundance

37:10Changing tack here. Let's talk about the young Jensen. Who were you when you were young? Let's see. I was... I'm not saying you're not young still, but you're really young. Statistically, I'm on the other side of that hill. Let's see. I was focused. I was curious. I was a perfectionist. I wanted to do everything well. I worked hard. I would say that those things characterize me. Do you think it's understood how much hard work that goes into great achievements? Oh, yeah. I mean, the amount of hard work. There's hard work and then there's insanely hard work. In order to be... And where are you on that scale?

38:03I'm on the insanely hard work. What does a day look like? I work every day. There's not a day that goes by I don't work. And if I'm not working, I'm thinking about working. And when do you kick off in the morning? Well, I wake up at 5 o 'clock. And the moment I wake up, I start working. And so I work every single day. There's not a day that goes by I don't work. When do you go to bed? As early as possible. I'm in bed probably. I'm asleep probably by 9.30. and I like my sleep and sleep is really important to me. What do you do to relax? What do you relax? I relax all the time. I enjoy relaxing at work.

38:43Just working is relaxing for me. Solving problems is relaxing for me. Achieving something is relaxing for me. And the most relaxing, just hanging out with my family, doing anything is relaxing for me. Yeah, I relax in a whole bunch of ways. I'm reading about things that's important to me is relaxing to me. So just hanging out with my family is relaxing. I relax in a lot of different ways. I'm pretty relaxed. What do you read? Let's see. I just read Chip Wars. I skimmed through a lot of AI papers. I don't understand all of them, but I try to understand all of them.

39:26I try to read everything that's of curiosity to me. You started NVIDIA in 93. You were 30 years old. If you were to boil down the essence of the success, what type of characteristics is it that the company has that makes it so successful? So your perspective about the future has to be, you know, on a fairly long arc, pretty important. And it has to be somewhat directionally right. And we were, I would say, absolutely directionally right. Now, the question is along that direction, there are a lot of different paths. And some of those paths would have been easier if we had better skills. You know, I didn't know how to be a CEO and nobody in the company knew how to build a company.

40:19And we didn't even know what a PC looked like at the time. I never even used one before. And so there were a lot of things about the company that the skills that we didn't have, that we had to develop those skills. How to raise money, how to organize the company, how to recruit people. You know, those were all skills that we had to develop along the way. I think that those skills are probably skills are learnable. I think the, you know, the attitude of an entrepreneur and the attitude of somebody who does something new is how hard can it be? You know, and my attitude has always been, you know, how hard could it be to learn, learn PC?

41:00How hard could it be to build a company? How hard can it be to hire people? How hard could it be to create an organization? It turns out all of those things were super hard. All of those things were super. It turned out it was super, super hard. But I think you want to go into it with the attitude, how hard can it be? And so when we got into the journey of artificial intelligence, we got into the journey of scientific computing, we got into the journey of autonomous vehicles, we started with the attitude, how hard can it be? And so if it's a solvable problem, how hard can it be? And we reasoned about everything from first principles.

41:37And if anybody could do it, I'm sure we could. We could. And we'll just learn as fast as we can. And so I would say we didn't have any of those skills. But if I had to boil down what led the company to be successful, our vision was right. But the character of the company is probably the most important thing. The character of a company is what makes it ultimately successful. And how resilient is it? How does it deal with adversity? How does it deal with learning? When it's presented with new assumptions, if the conditions change, how agile is the company? The world changed around us continuously.

42:32Those values, the learning, the agility, the ability to change, how do you install those values into the company? You talk about it. You teach it. You live it. NVIDIA is really fortunate. As a long-term successful company, we have excellent chance. And the reason for that is because we suffered so greatly in the beginning. For the first 15 years of our company, it was one adversity after another. And then every, you know, and after that, there were adversity after another. But the company was able to deal with it. The first 15 years, the adversity were incredible. Maybe five, six, seven times it was existential.

43:16What's the key to coping with adversity? I think in the beginning of a company forming the corporate character, the corporate culture, it's people. It's the people's resilience. It's the character of the people there. Unfortunately, a company is made of people. And it's not made of the document that describes the culture. It's not made of the document that the inscription of the core values on the building. And that's not what makes the company's culture is the people and how the company overcame existential crises, how the company overcame the incredible adversity that was presented at the time, both in the agility of the people, the cleverness and the creativity of the people, the ingenuity of the people.

44:11And then also the will of the people, the countless times that our company has been presented with challenges and the willpower, the utter incredible ability to suffer, you know, willpower to be able to do something even in just extraordinary pain. That is corporate character. In 2003 at Stanford, you said, my will to survive exceeds almost everybody else's will to kill me. Yeah, right. Exactly. Where does that come from? Well, you know, I think everybody's everybody's upbringing is is unique to them. I've just always had that. You know, it's pretty hard to discourage me. And if I believe in something.

44:59I'm just going to, if I believe in it, I'm just going to keep on doing it until it's done, until we're great at it. It's hard to deter me. It's hard to distract me. It's hard to discourage me. And in my mind, it's always, how hard can this be? And it turns out, every time I say, how hard can this be? It turns out it's incredibly hard. and I'm surrounded by amazing people helping me. And it remains incredibly hard. And last question, what is your advice to young people? Well, there are a lot of things to learn. I would advise be a learner.

45:44but probably the best advice that that I can imagine is is think from first principles don't worry about anybody else's advice you know I've been given a lot of advice over the years some of it have been very good most of it has been irrelevant and the reason for that is because it was either the advice was either an opinion um it was perspective of the time it was based on on wrong assumptions um and and my advice would be you know think for yourself think from first principles and and and a lot of people say you know find something you love uh i i don't know about that i guess i guess i've fallen and fallen in love um and many things that i do I loved it when I was a dishwasher.

46:39I loved it when I was a busboy. I loved it when I was delivering papers. I loved it when I was waiting tables. I've loved every single job that I've ever had. And I loved every single day at NVIDIA that I've ever had. And I just learned to love what I'm doing. And so I guess it's probably harder to find something that you love. but it's easier to fall in love with what you're doing. And once you fall in love with what you're doing because you just desperately want to do a good job at it, it's easier to do it well and do it hard. Well, I think that's a beautiful place to end. I have to say this has been one of the most intriguing and interesting conversations I ever had in my whole life.

47:27And I think when we look back at this time, 20 years from now, 30 years from now, you could potentially have been the person who's changed the world the most. I enjoyed our conversation, Nikolai. Thank you so much. Thanks for the time. Thanks for the opportunity. Keep it up. Thanks. All right. Take care.