Jensen Huang's AI-Era Insights and Leadership Vision

At an AI summit interview in late May, NVIDIA's Jensen Huang shared his deep insights on the AI era and his unique perspective on company leadership. Below is the full video and transcript of the conversation. Enjoy previously believed that there were two tech leaders on this planet whose interviews were most worth following — one obviously being Elon Musk, and the other Sam Altman. Now, that TOP list can expand by one more: Jensen Huang.

Quote Commentary

"AI is deep learning, and it's also an algorithm for solving problems that are hard to specify. It's also a new way of developing software. Imagine you have a universal function approximator of arbitrary dimensionality."

Lord of the City in the Sky: "Universal function approximator" — what a brilliant metaphor for neural network models with hundreds of billions or trillions of parameters, both in implementation and logic. Generative models approximate all kinds of real data, knowledge, and reasoning through innumerable parameters.

"There are many aspects of our intelligence that are about applying rules, either learned rules or expressed rules. Many of the principles and rules in our companies, in our particular societies, in our particular cultures are simply expressed through words."

Lord of the City in the Sky: The observation that "many aspects of intelligence are expressed through verbal rules" also explains why LLMs can achieve such intelligence simply by learning from text — because human cognition and wisdom are largely manifested through words rather than through direct personal experience.

"We use AI to help us design, to explore the design space... AI helps us create more energy-efficient processors, exceeding anything that Moore's Law would have predicted."

Lord of the City in the Sky: NVIDIA using AI to design chips appears to be a done deal. Compared to using AI to assist with software coding (Copilot), using AI to design chips represents a far more fundamental transformation, because it means humanity's computing power will explode further. When Jensen says "exceeding anything Moore's Law would have predicted," that should be credible.

"You have to study hard. People love teaching great students, so I invested a lot of time to become a good student."

Lord of the City in the Sky: In simple terms, Jensen explained how he's kept up with cutting-edge technical knowledge over decades of managing the company — nothing but diligent study. The Lord expressed the same idea in the article When Transformer Arrives: How Ordinary People Can Embrace the AI Tsunami Era. To survive in the AI era, there's no substitute for good study — lifelong learning.

"The technology is so deep that your understanding of it, your interest in it and your curiosity about it, is essential to running a technology-driven company."

Lord of the City in the Sky: Another statement that looks simple but is actually profound. Without genuine curiosity and love for science and technology, it's impossible to run a technology-driven company well. This comes through strongly in conversations with both Musk and Sam Altman.

"So I have a lot of people reporting to me because I don't do one-on-ones... Having a lot of direct reports, not doing one-on-ones, flattens the company, information propagates quickly, people are empowered, and that makes it possible for me to not do one-on-ones."

Lord of the City in the Sky: Jensen's management philosophy of not doing 1:1s shocked and refreshed my understanding.

But thinking about it carefully, this does in some sense align more with the philosophy of flat, fully open communication; and the key here is finding A-level talent that doesn't require the CEO to spend energy on career coaching. Jobs saw through it, and the Old Master of the Huang Family inherited it.

Full Interview

Host: I'm so glad to be with you, Jensen. It's such an honor to learn from you. I want to go back to the moment you became convinced about deep learning. Do you remember when that was?

Jensen: Very good. I learned about deep learning about the same time as everybody else. Maybe slightly earlier. And the reason for that was several research groups were all simultaneously trying to compete in the 2012 ImageNet competition. I think there were two things — one was using the latest GPU, the GeForce GTX 580. It had just come out. And learning how to program deep neural networks. And so we were fortunate to learn about them slightly earlier than others.

At that time it was called neural networks. As you know, during that period, artificial intelligence wasn't that popular. Neural networks was a suspect area of research. But nonetheless, we offered to help. We did our best.

What was really striking to me was its effectiveness. And of course, when you see amazingly effective techniques, the next question is how does that technology scale? What other problems can it solve? Because of the nature of neural networks and the fact that every layer is isolated from each other. The backpropagation method is so effective, you could imagine that it would be scaled massively. And it turns out we were right.

Our observation was that AI is deep learning, and it's also an algorithm for solving problems that are hard to specify. It's also a new way of developing software. Imagine you have a universal function approximator of arbitrary dimensionality. No matter what the dimension and size of the problem, as long as you have a large enough model, you could backpropagate it and learn it. And I think inferring that insight was really important to us, and we were deeply convicted of its potential because we realized that this is going to be a new way of developing software.

From that time on, we called it Software 2.0. But we observed that, first of all, it's a new way of developing software, and we realized that it's likely going to require reinventing computing entirely. And from that point on, you could look back and realize that for probably 60 years since the IBM 360 system, for the first time, accelerated computing with GPUs and deep learning truly reinvented the computer. If you will, we may have invented quantum computing before quantum computing.

Host: I think that was a really important moment. And we were fortunate to connect all the dots. I remember reading the research team's papers and how they started to deploy it. It was incredible, impressive, because no one had really used GPUs for this purpose before.

But since model architectures started to evolve, where do you think they're heading? At the time it was CNNs. We now see more and more support for multimodality. Does that excite you about this? Where do you think model architectures are heading?

Jensen: Let's see. The work that researchers do is to restate almost every problem, every type of data into something that a Transformer could learn. You could create vision Transformers, you could create audio Transformers, you could create text Transformers. It looks like you can create a Transformer out of almost anything.

Multimodality is really important for very clear reasons. They perform better. For example, if we said, you're trying to train a vision neural network, and you've seen nothing but horses, you've never seen a zebra. But if you add another modality, like words, you understand that a zebra is a horse with black and white stripes. And somehow the image of the horse and the words' understanding of black and white stripes allows you to imagine what a zebra looks like without ever having seen one.

And so you expand your representation. You could use multimodality to enhance the stability of perception. And so we do that for cameras and radar and lidar. We combine multiple sensor modalities so that we can extend your perception capability to encompass the superset of all the sensor modalities.

I suppose you could also enhance it with ambiguous things, for example, if I were to say, what is that? The combination of my speech and my gesture could help you understand what I mean. It's hard to do that with pronouns. And so multimodality is really important. And now that we have Transformer models, we expressed the Transformer understanding way to understand multimodality.

And so I think the next generation of AI models will have higher performance, be safer, be more stable, can do more things. And I think that's going to be a big advance. I think what's really fascinating is how much we've been able to infer from language. We basically inferred the world model from language that could be applied in very broad contexts.

Host: But you've also talked about teaching these models physics. Do you think that's necessary? Or can we infer it when these models become multimodal? For example, when they're trained in video training, can they learn that from the training data?

Jensen: Yeah, I think so. Let's see, almost everything in the world has already been described in words. Physics has been described in words, Newton's laws have been described in words. And so you could imagine that we could learn almost all of the physical effects from the word corpus of the world.

If you've never seen red, it's really hard to imagine what red means. But if there's enough poetry describing the beauty of a red apple and how it compares to other red apples, I wouldn't be surprised. And heart red, emojis and things like that. Through all of these words, you've associated, this must be red. Through comparison and contrast, without having seen it, you must be able to understand that.

But if you didn't have the ability to combine all these different things together, you could never understand what something feels like and its nuances and subtleties. I think it's possible to teach people physical effects. But if you want to predict the laws of physics, that is, things that have physical phenomena, you need to ground AI on top of physics. It's no different than what we do today using reinforcement learning and human feedback to shape large language models.

In the future, you're going to use reinforcement learning with physics feedback. And that physics feedback, not human, is feedback expressed by a robot through some kind of physics simulation. We created a system called Omniverse that obeys the laws of physics, and so we can have Omniverse be essentially the digital twin of a robot.

That embodied language model would then get reinforcement learning physics feedback, simulated physics feedback, digital twin of Omniverse feedback. And so I do believe that for certain types of robots, you want to ground it on physical truth. You want to ground it on ethical truth. That's what humans are, that's alignment. And it's one of the ways that helps us create safer chatbots.

So I think these two ideas are, in some ways, parallel and reasonable. I think it would be helpful. That might be why we end up in this simulation, why you start doing these experiments.

Host: Where do you think the ceiling is? Do you think we can keep running with the existing model architectures, evolve them a bit, scale them, compute them? Or do we need a fundamental breakthrough, and are we going to hit diminishing returns soon?

Jensen: First of all, I don't know the science behind it. But intuition would suggest that there are many aspects of our intelligence that are simply the application of rules — either learned rules or expressed rules. Many of the principles and rules in our companies, in particular societies, in particular cultures, are simply expressed through words. You don't learn them through any other means.

"Thou shalt not kill" — that's not something you learn by trying it until you discover you shouldn't kill. We just learn it. We're given that rule. So there are many rules that can be expressed without needing to be learned. I think this symbolic reasoning can augment these exhaustively learned models.

So, I don't know the science behind it, and whether it indicates one way or the other, but intuitively, we augment our intelligence through reinforcement learning, experiential learning, and things we learn simply because that's just how things are.

One of the most exciting things for us right now is that it's also possible to augment human intelligence. It would be wonderful if we could provide researchers with tools to run thousands of parallel experiments, rather than replacing those tools.

Host: Over time, how do you think about the role of humans in this? Once we reach intelligence that's several orders of magnitude beyond human capability in this area, what role do I have left?

Jensen: I don't know. But I'm surrounded by people who are several orders of magnitude better than me at certain skills. I have no problem coexisting with them. So I already live in an environment surrounded by superintelligence. Relative to myself, I feel they can do things I can't even imagine. Somehow, I exist quite harmoniously within it. I think we'll raise the baseline of intelligence.

We've raised many things that have, over time, mass-produced human resources. We democratized the hunting of food through agriculture. We no longer have to chase or be chased by our food. We certainly democratized the production of energy. So physically smaller and weaker people can live in a world of heavy objects. You don't have to be large and muscular to survive in such an environment.

We democratized access to energy. And so society has advanced in its capabilities. I think we are now democratizing the production of intelligence. But I think value will still exist for those who have deep domain expertise and certain incredible passions. As you mentioned, through the ability to trial all these future scenarios, we can amplify our own passions and expertise.

Another aspect I'm hopeful about: in my 40-year career, the vast majority of the population never learned how to use this instrument we call a computer. Only a few people — I think you started learning how to program computers as a teenager. For the vast majority, they don't know BASIC. They don't know Python. They don't know Pascal or Fortran or C or C++ or Java. They don't know how to program these computers.

Even though it made sense for a small number of people in the world — about a million people — billions of people still don't know how to do this to this day. Now, with ChatGPT, the preferred programming language is human. You can program in Swedish, I can program in English, you can program to write a program to program another computer. So, for the first time in history, we have democratized programming. I can't help but believe this will empower the billions of people who now see this value-creation tool and productivity tool.

This incredible machine that travels at the speed of light, which we call a computer — for the first time, everyone can use it. So I hope it bridges the digital divide. I think you can make that case. You can see it happening in real time right now — kids online prompting these machines, writing amazing programs, generating beautiful images. They're just kids. They don't know how to write programs, but they can do this.

Host: I think what NVIDIA has done in accelerated computing is also fascinating, because you're using AI yourselves to develop more advanced computing. Can you tell me what that loop looks like?

Jensen: Our current generation of chips is so large, so complex, that it's impossible for all the employees in our company to design it. So we use AI to help us design, to explore the design space. As you said, we try hundreds of thousands of combinations of design options and find the one with the best trade-offs.

The exact, optimal trade-off is the trade-off we determine. There's no such thing as perfect optimization, only optimal optimization under certain trade-offs. Sometimes, we might want to optimize for speed at all costs, on the grounds that it's on the critical path. You want the critical path to be as fast as possible.

Maybe because it represents something you have to instantiate thousands of times. Now, energy efficiency is really important. So we optimize for energy efficiency. We make that judgment, and then we unleash AI to discover all the corner cases and all the different combinations of options, and the design it brings back is not something any human could make.

What's incredible is, we're looking at these designs, and we've incorporated them into our latest generation Hopper — designs that no one has ever designed before, which is truly incredible. Or, it's because we have to connect thousands of modules on a single chip, and the combinatorial explosion of thousands of combinations exceeds all the atoms in the universe.

So only AI can figure out the optimal optimization in that situation.

Host: Another question I'm really curious about your perspective on is, how should we think about defensibility? Last time we talked about the value of domain expertise and deep integration into workflows. But for companies implementing at the application layer, as these models become more commoditized, as they can learn from less and less data, how should they think about defensibility? Many historical moats seem no longer applicable. How should we think about building a defensible business?

Jensen: I don't know of any evidence that understanding a problem domain or understanding a segment of customers has ever lost value. Our technology, the world's computing technology, and the number of highly educated people are obviously growing.

New college graduates today are truly remarkable. Every single one of them was smarter than us when we first graduated. The types of problems they can solve, right from graduation, are the types that entire companies used to have to do. And yet, all evidence suggests that deeply understanding customer challenges is really valuable.

So, I don't think that's going away. We're going to have incredible baseline AI capabilities that can do amazing things. It can solve all the most difficult math problems, it will pass qualification exams.

However, we know very well that people who pass qualification exams, when brought into a particular problem domain, still have to learn the fundamentals of that industry. What is the problem that needs to be solved? What are the complex relationships between people? Customers are often people, trying to solve their own complex challenges. So understanding people, understanding context — these things are often poorly articulated. These are social science problems.

Host: What are you most excited about at the application layer?

Jensen: For us, if I can categorize it into three areas, it would be:

What can AI do to augment, to revolutionize the way we make products? That's one area. We just discussed one aspect. The way we design chips has completely changed.

Second, the way we design software has changed. Of course, the second area is: what can AI now enable us to do that makes our products different?

Not just revolutionizing the way we design products, but revolutionizing the products we build. For example, there are many gamers in the world, and the way we used to design graphics cards was to design programmable shaders, and of course compilers and so on. But that was it. We released it. We shipped excellent processors with excellent compilers and integrated them into games and so on.

But now, you can't even ship a GeForce card independently, because there's a supercomputer behind it learning how to predict missing pixels, because we have to denoise and infer pixels. We infer approximately every pixel we render — we now infer between 8 and 16 pixels for every one. I mean, it's like being given a puzzle piece, giving you one piece, and guessing the other 16 pieces.

So we've taught a class in AI, and there's a supercomputer in the back just learning how to do this and improving the algorithm, and then we download the algorithm every time we improve it (to the local machine). So now we're using AI not only to revolutionize the way we design GPUs, but also to revolutionize the way GPUs generate images. So, it helps us create far more energy-efficient processors than anything Moore's Law would have predicted.

Then, the third category, I would say, is transforming the entire company into an AI company.

In this way, all our employees can be augmented by this system, which is constantly running. So we don't search for information we can't find, so we might be able to connect dots and predict market opportunities, or that the supply chain has changed, or market demand has changed — we couldn't possibly see all the signals, but it's not impossible for AI to see all the signals.

So over time, from the way employees collaborate with each other, to the way we access and use information, to the way we predict demand and work with supply chains — all of this will be revolutionized by AI.

Host: What do you think is your most contrarian view on AI right now?

Jensen: I don't know that I have any particularly contrarian views, because if you look at the world today, most discussions about AI are either enthusiasm beyond its promise, or concern beyond its harm. So the place where real-time discussion happens, between the two extremes of promise and harm, is probably where the truth lies.

There's no question that this technological capability, like any technology in history, brings tremendous social and economic transformation and disruption. So we have to consider considerable harm.

Who would have thought there would be a large group of people doing web design? That's a profession that didn't exist before, or programmatic advertising. These professions didn't exist 40 years ago when I graduated. But today entire industries exist.

So somehow, this enabling technology called the internet had recommendation systems. Early versions of AI enabled us to create this new industry. I think we're going to have to retrain and reskill. Jobs that are displaced can be transformed. But I'm certain new industries will be created that we never thought of.

Right now, we're seeing one area that's already becoming active, which is prompt engineering. Prompt engineering is going to be a real thing, it's going to be a massive industry, it might be the most important programming profession. So how many people will become prompt engineers? It's truly incredible.

You see AI helping write prompts, right? Prompting other AI.

(Host: That's very meta.)

Indeed, very meta. So I think we'll see all of this.

I think there are a lot of good things being discussed. I love the discussion about safety. We have to invest in AI safety commensurate with our investment in AI capability. We happen to work in autonomous vehicles, and we've probably invested as much in making the AI safe and the car safe as we have in making the car drive. So I think the same has to be true for large language models.

The technology for building guardrails around AI, to keep it within its operating domain, the technology related to alignment, reinforcement learning, human feedback, the reinforcement learning related to physical feedback that we talked about earlier, vector database technology, the ability to reduce fallacies or reduce hallucinations, to augment it with facts.

I think we'll see an explosion of ideas around all of these capabilities to transform that core large language model with surrounding other AI technologies, approaches, and best practices. All of that will turn that large language model into a usable chatbot.

Host: One thing I deeply admire about you is how current you stay on technology. I imagine you must spend a lot of time reading or something. You're so unusually thoughtful and knowledgeable about every detail happening in the industry. How do you keep up while running one of the largest companies in the world?

Jensen: Wow. Let's see. What's the answer to this?

First of all, I have a lot of amazing people around me. When I go visit you, you have a lot of amazing people around you too.

They enjoy teaching me, and you have to work hard at learning. People love teaching great students, so I invest a lot of time in being a good student. Of course, the breadth of areas we touch, from autonomous vehicles to climate research to digital biology, the breadth and scope of where we can have impact in the world is large, but we also have to learn.

So, your industry is a technology-driven industry, solving problems and creating solutions for companies. My industry is also technology-driven. For us, understanding the foundations of technology is essential so that you have intuition. How to change an industry. You have an intuition that one technology is a bit of a left turn, and one technology is foundational.

Recognizing from the generative adversarial models we worked on early, to variational autoencoders, to diffusion models — they're all somewhat related. Recognizing that one insight could lead to another breakthrough that opened up the vision for diffusion models now, which are just incredible.

So, I think having intuition about technology allows you to extrapolate better. Our ability to extrapolate and foresee the future is really essential because, my goodness, technology changes so fast.

But we also need several years to build a great solution. So, on the one hand, how do you focus on building something that takes years to build, building it on top of technology that changes 1000x every few years. How do you do that if you don't have intuition?

So, I think the fact that technology is so deep that you understand it, and are interested and curious about it, is essential to running a technology-driven company. So, I think I love that part of my job, I have people around me who enjoy teaching me, and I have to invest fully in being a good student.

Host: One area I'm also very interested in is how you manage the company. I understand you don't do one-on-ones. Can you tell me about some of the classic management playbooks you've challenged and evolved?

Jensen: First, for building a company, you have to solve all the problems first, Joe, which you do very naturally, thinking about it from the very beginning.

What is the machine we're trying to create? What is the output? What is the input? Under what conditions is it operating? What is this industry like? Is it a fast-moving industry? Is it a bureaucratic industry? Is it a highly regulated industry? What kind of industry is it? What are you trying to build?

So, I think looking at it from that perspective, there are a few things I want to do at the company.

I want to create a company that naturally attracts amazing people. The reason is because we're working on problems. Our company's mission is to solve nearly impossible computing problems. If a problem can be solved by ordinary computers, we don't work on it. So, we have to find problems that ordinary computers can't solve or are nearly impossible to solve. So, you want to attract people who want to invent this new form of computing and apply it to solving some very difficult problems. So, I want amazing people.

Second, I want a company that's smaller rather than larger, as small as possible rather than as large as possible. It needs to be right-sized for the work, but as small as possible. So, naturally, you want to empower people. If you want a command-and-control organization, then build it as a pyramid, like ancient armies going all the way back to the Roman Empire. But if you want to empower people, then you want to be as flat as possible so that information travels fast.

For it to be as flat as possible, the first layer has to be very thoughtful. The first layer happens to be the most senior people, and you would think they need the least management. No one comes to me, no one on my management team comes to me for career counseling. They're doing great, they're doing wonderfully.

So I have a lot of people reporting to me because I don't need to have one-on-ones. I don't need to do career coaching. They're all fine, they know what they're doing. They're experts in their domains. So, those one-on-one meetings are actually not necessary.

If there's a strategic direction, why tell just one person? You tell everybody. So, after we've immersed ourselves in the chaos of strategy discussions and charting our future course, when it's time, I'll send it to everybody at the same time or tell everybody at the same time. People give me feedback, and we refine it.

Because the company is so flat, you've already empowered the organization tremendously through company knowledge and access to information, and the company is also agile. The result is, having many direct reports, not doing one-on-ones, makes the company flat, information travels fast, people are empowered, which makes it possible for me to not do one-on-ones.

This algorithm is well-designed, the architecture is well-implemented.

We also don't have business units, we don't have divisions. Everybody works as one whole. The company is formed in a way that we can best build accelerated computing. If you asked me to do fried chicken, we'd be terrible at fried chicken; Swedish meatballs, no chance. But accelerated computing, very good.

Host: I think you have about 40 direct reports, right?

Jensen: Something like that. The challenge is getting everybody together. I want to get everybody together, but either somebody's out, or somebody's on vacation, or somebody's doing something else. The probability that everybody's sitting in the office is about 0%.

Host: How has your leadership style changed over time? You've been at this for decades now, how has it evolved as you've learned?

Jensen: I actually don't have a style, it's just me. There are a lot of things I wish I did better.

If something's happening at work and I don't like the direction it's going, I'll just say it, I don't pull anybody aside for one-on-one coaching.

If something's wrong, I'll just say it; if I have a different opinion, I'll just say it.

It can be a bit direct. But people recognize that my intent is to be direct, and I spend a lot of time reasoning through my decisions.

Empowering employees by letting them learn how the leader thinks about problems. Through every meeting I attend, I'm explaining how I think about this, let me reason through this, let me explain why I'm doing this, how do we compare and contrast these ideas? That management process is really empowering.

We also don't just have VP meetings or director and board meetings, I attend meetings where there are newly minted college graduates, people from different organizations are all there, we just all sit there. It's a bit like your office, where everybody just sits there.

**Host: That's actually something I find very interesting, because that's one of the rules. There's a very clear leadership team, there are leadership team meetings and so on, this is something I've always struggled with, because you'll have a lot of best individual contributors. They should be in these meetings, it shouldn't just be VPs who don't know the technical details. Fascinating that you can.

Jensen: Exactly. You want the most informed, the most skilled, or the most experienced people, who actually made the mess, or actually faced the situation.

You want ground truth. You want the best ground truth and expertise you can possibly get.

Host: You have some system for people to communicate their top priorities, I heard it's through sending emails. How does this work?

Jensen: We don't do status reports. I don't read any status reports.

The reason is that status reports are meta-information by the time you get them, they're barely informative. They've been distilled and refined, bias has been inserted, opinions have been added, you no longer see ground truth.

I tend to appreciate information from anybody. If you send an email, titled Top 5 Things, and it's just your top five things, whatever you're observing or what you've done or what you've learned or what's going on.

Top five things, whatever they are. You just went to a great restaurant, who doesn't want to hear about that? That's important information. I just had a baby. That's important information.

Whatever those things are, top five things, if you send it out, I'll read it. I read it every morning, probably 100-plus emails, I read this every day.

Host: This is a massive thread where everybody in the company sends this to you, right?

Jensen: Everybody has their own version of top five things, they just send it out. If you send it, I'll read it.

Host: What are your top five things?

Jensen: Top five things doesn't mean from the center outward, think of it as the internet of things.

If I send out my top five things, I'm actually polluting the system, which is why I don't do it.

I have my own top five things, I keep them to myself.

Host: How do you strike a balance in planning? The bottom-up idea is letting the best engineers on your team decide what to do, and sometimes you also have to execute on plans. How do you balance these two things?

Jensen: First, strategy isn't words, strategy is actions.

If the company has a set of strategies, but people's actions, their top five things aren't that, then clearly they're not executing the strategy.

It turns out strategy isn't what I say, it's what they do. It's really important for me to understand what everybody's doing.

I do that by having some awareness of everybody's top five things, not reading all of it every week. It's random and a random sampling system. You can sense whether the company is moving in the direction that you wanted, that everybody agreed to.

That's point one.

Second, planning.

We don't have a regular planning system, because the world is a living, breathing thing.

We just plan continuously, there's no five-year plan, no one-year plan, no plan. Just what we're doing.

Host: It's really exciting to hear this. I think, while executing on first principles and generating some ideas, if you're doing things opposite to the playbook, it can also be hard to trust your intuition. What gives you conviction in some of these things intuitively?

Jensen: Most things you commit to pursuing should first be reasoned through first principles.

There's a foundation, an assumption, that leads you to believe that computers have to change, or chip architecture has to change, or the way software is developed, or the way data centers are transitioned.

Data centers used to be places where we stored and retrieved files. But in the future, every company will have more than two data centers — yet one of them won't be a data center at all. It will be a factory. A factory that produces intelligence. Data goes in, gets refined by computers, and what comes out is the most valuable thing in the world: intangible intelligence.

This building will continuously drive this forward. You and I, we will all have factories.

How do you reason through this? You go back and trace it. Before you know it, you've formed a worldview based on first-principles thinking, and the next step is to pursue it wholeheartedly to make it real. That's usually really hard.

But if you're wrong, you change your mind. That's what makes modern leadership truly great.

If I'm wrong about something, I'll admit it. That was wrong, that was really bad. Then you say, I've changed my mind. Because you're constantly adapting and actually replanning. The interesting thing is, over time, people might not even notice that you've adapted 17 times last year, that you may have changed your mind 35 times.

We don't do these massive five-year plans. I think five-year plans are simply terrible for technology, simply absurd. These continuous planning systems can actually make leadership easier.

Host: Just as we're obsessed with the products we're going to build, we're equally obsessed with the companies we're going to build. I think very few companies are truly focused on empowering employees to pursue lifelong careers, and this has always been a key passion of yours. How do you achieve this? What measures have you implemented at NVIDIA to empower employees to pursue lifelong careers?

Jensen: A leader's job is to create the conditions for others, to empower others to pursue lifelong careers.

There are several ways to fulfill this mission.

The most important way to fulfill this mission is to not have people doing commodity work.

For example, our company never talks about market share. The reason is, why talk about how I have 23% market share and they have 27% market share? Why compete with others for market share?

Because the whole concept of market share suggests there are many other people doing the exact same thing. If they're doing the same thing, why should we do it?

Why would I waste the lives of these brilliant people on something that's already been done? Unless we just love competition — I tend to not enjoy competing with people for already-commoditized market share, competing for already-commoditized markets.

It's a mindset of going and doing something that's never been done before.

Another way is to prove it by exiting commoditized businesses. Whether through our own initiative or otherwise, we've exited many businesses in the past, and this very clearly signals to your employees that we won't do commodity work. The combination of choosing the right work and leaving the wrong work — this is the best way to create the conditions.

The rest is what we've already talked about, which is empowering people through information.

While some companies are very closed, where information doesn't flow outside the organization, I encourage our company to be more transparent. If you ask me about our company's secrets, there actually aren't that many secrets, and this empowers people.

The rest is how you conduct yourself at work. If there's hierarchy in the company, then that's clearly less empowering. Anyone in our company can walk into a meeting and contribute, including new college graduates — that's very empowering.

I think empowerment is a big deal.

Oasis Capital is a new-generation venture capital firm in China, dedicated to discovering the most vital entrepreneurs of the next decade and growing alongside them to create long-term value. "Championing Vitality" is Oasis's vision and mission. This vitality is both the direction of structural transformation in the era and the resilience and evolutionary force of entrepreneurs.

Oasis Capital focuses on early and growth-stage investments, with individual checks ranging from $3 million to $30 million. It concentrates on robotics, artificial intelligence, technology services, and other fields, supporting China's technology-driven upgrade of new services.