In a recent podcast interview with Dwarkesh Patel, Microsoft CEO Satya Nadella discussed Microsoft's latest breakthroughs in Artificial Intelligence (AI) and Quantum Computing, as well as his insights into the future of the industry. In the interview, Nadella not only shared Microsoft's significant progress in topological quantum bits and game world modeling, but also emphasized that the goal of driving global economic growth 10% is much more important than achieving General Artificial Intelligence (AGI).
timestamp
- (0:00:00) - Intro
- (0:05:04) - AI won't be winner-take-all.
- (0:15:18) - World Economic Growth 10%
- (0:21:39) - The lowering of smart prices
- (0:30:19) - Breakthroughs in Quantum Computing
- (0:42:51) - How Muse Changed the Gaming Industry
- (0:49:51) - Legal Barriers to AI
- (0:55:46) - Proper understanding of AGI security
- (1:04:59) - 34 years at Microsoft
- (1:10:46) - Does Satya Nadella believe in AGI?
transcript of an interview
0:00:00 - Opening introduction
Dwarkesh Patel
Satya, thanks so much for coming on the podcast. In a moment we're going to dive into two major breakthroughs that Microsoft has made recently, and congratulations, that made it into the journal Nature on the same day: the Majorana Zero chip, which we have in front of us right now, and the World Model of Human Behavior. But can we first continue this conversation? You mentioned that some of the phenomena that we saw in the 1980s and 1990s have resurfaced today.
Satya Nadella
The most exciting thing for me is ...... Dwarkesh, first of all, it's a great honor to be on your podcast. I'm a huge listener and really enjoy your interview style and the wide range of topics you explore.
What's exciting to me is that this scene reminds me of my first years in the tech industry, probably in the '90s, when there was a lot of debate about whether RISC (Reduced Instruction Set Computer) or CISC (Complex Instruction Set Computer) was better or worse and the question, "Can we really build servers on the x86 architecture? and "Can we really build servers on x86?
When I joined Microsoft, I was in the early days of Windows NT. Every aspect of the technology stack, from the core silicon platform to the operating system to the application layer, was full of variables.
It's fair to say that cloud computing has reshaped all of this to some extent, and distributed computing and cloud computing have certainly changed the client-server model, and the Web has changed dramatically. But this wave feels more full-stack disruptive than anything I've experienced in the past.
Dwarkesh Patel
When you look back at the 80's and 90's, what decisions turned out to be long term winners and which ones fell by the wayside? Especially considering that you worked at Sun Microsystems, who lived through the Internet bubble of the '90s. It's often said that building data centers was a bubble, but at the same time, the Internet today is built on the infrastructure of that time.
What are the lessons we can learn about what will stand the test of time, what are the inherent long-term trends, and what are the flash-in-the-pan? What lessons can we learn from this?
Satya Nadella
If you look back at the four major transitions I've been through, the first was the rise of the client and client-server model. That was the era of graphical user interfaces and the x86 architecture, which basically allowed us to build servers.
The trend was already very clear to me. I remember attending the 1991 PDC (Professional Developers Conference), and in fact, I was still working at Sun at the time. At the Moscone Center in 1991, Microsoft made the Win32 interface public for the first time, and it was clear to me what was going to happen, that servers were going to the x86 architecture as well. You have to bet on this long-term trend when the scale advantage starts to build up. What happened on the client side will happen on the server side, and then you can actually build client-server applications. As a result, the application model becomes clear.
Then the Web became a major challenge for us, and we had to deal with it. In fact, not long after I joined Microsoft, the Netscape or Mosaic browser came out, probably in November or December of '93, and Andreessen and his team launched them.
That's a huge change. It's interesting that the Web wave came just as we were starting the Client-Server wave. But it was clear that we were going to win the Client-Server wave as well. We had to adjust to the Web wave, and we did a pretty good job of it, because the browser is a completely new application model.
We embraced it wholesale, whether it was supporting HTML in Word, building new browsers ourselves and competing, or building Web servers on our server stack and aggressively expanding the market. Of course, we missed out on the biggest business model on the Web because we all thought that the core of the Web was distributed, and who would have thought that search would be the biggest winner in organizing Web content? Obviously, we didn't see this coming, and Google did and executed brilliantly.
This is a lesson for me: you have to not only capitalize on the right technology trends, but also figure out where the value will be generated. Shifts in business models can be more challenging than changes in technology trends.
0:05:04 - AI won't be winner-take-all.
Dwarkesh Patel
So where will the value in the AI space be generated?
Satya Nadella
That is a very good question. I think there are two areas that I can predict with some confidence. The first is the hyperscalers that are outperforming, because fundamentally, even looking back at what Sam (probably Sam Altman, OpenAI CEO) and others have said about AI, if intelligence is a logarithmic function of computational power, then whoever can compute at scale is the big winner.
Another interesting observation is that even when you look at the underpinnings of any AI workload, such as ChatGPT, it's not just the advances in GPUs that people are excited about, although that's really great. In fact, I even think of my compute cluster as a ratio between AI gas pedals, storage, and compute. And scale-wise, you have to keep scaling it.
Dwarkesh Patel
Yes.
Satya Nadella
So the emergence of AI workloads has been a blessing, because you know what? They demand more compute power, not only in the training phase, but also in the inference time. When you think of an AI agent, you see that the AI agent is going to exponentially increase compute usage because its calls are no longer limited to the human user, but rather the program that the human user calls is going to call more programs. This will create huge demand and scale for computing infrastructure. So our hyperscale business, the Azure business, and other hyperscale cloud providers, I think are going to have a huge opportunity for growth.
After the mega-business, things get a little murkier. You might say that there is a winner-take-all model, but I personally don't think so. That's another important lesson I've learned, by the way: understanding very clearly which markets are winner-take-all markets and which are not determines everything to some extent. I remember in the early days when I first got into Azure, Amazon had taken a significant lead and people, including investors, were saying to me, "It's game over, you're never going to make it, Amazon's going to be a winner-take-all."
But from my experience competing with Oracle and IBM in the client-server space, I know that buyers will never tolerate a winner-take-all situation. Structurally, hyperscale cloud services will never be a winner-take-all market because buyers are smart.
The consumer market can sometimes be a winner-take-all situation, but in the enterprise market, any buyer, whether it's a company, corporation, or IT department, wants multiple vendors. Therefore, you must be one of many suppliers.
I think something similar will happen in the modeling space. There will be both open source and private models. As with Windows, one of the key lessons I've learned is that if you have a closed-source operating system, there's bound to be a complementary open-source alternative.
So, to some extent, this is going to create a kind of check and balance on the market landscape. I think in the modeling space, there may be some closed-source models, but there will certainly be open-source alternatives, and the open-source alternatives will actually make sure that the closed-source winner-takes-all situation doesn't happen.
This is my take on the modeling field. Also, if AI is really as powerful as people say it is, then countries won't stand by and allow private companies to ...... run amok on a global scale. So I don't think AI will be a winner-take-all market.
Taking it to the next level, I think the situation is the same as it has been in the past, and in the consumer space, there may be some winner-take-all network effects in certain categories. After all, ChatGPT is a great example of this.
It's a massive consumer app that has gained real escape velocity. I see it consistently in the top 5 on the App Store, which is incredible.
They are able to capitalize on early advantages and turn them into application advantages. In the consumer space, this could happen. But in the enterprise market, I think there will be different winners in different categories. At least that's how I analyze it.
Dwarkesh Patel
I have a lot of follow-up questions. We'll have to talk about quantum computing later, but on the idea that models might be commoditized: perhaps someone made a similar argument about cloud computing decades ago - isn't cloud computing fundamentally about chips and boxes?
But in the end, you and many others have found a way to make amazing profit margins in the cloud. You found ways to realize economies of scale and add other value. Fundamentally, even putting aside the terminology, if you have AGI and it helps you create better AI - right now, we're seeing synthetic data and reinforcement learning (RL); maybe in the future, there will be automated AI researchers - this seems like a great way to build on your strengths. I'm curious what you think about this, especially about the importance of staying ahead of the curve in AI.
Satya Nadella
After scale, nothing is a commodity. It's like when you mentioned cloud computing, everybody says, "Oh, cloud computing is a commodity." But when you scale ...... that's the trick to running a hyperscale cloud provider ...... You might say, "What's so hard about that? Don't I just stack servers?"
Dwarkesh Patel
That's right.
Satya Nadella
In fact, in the early days of hyperscale cloud computing, most people thought, "There are a lot of hosting providers out there, but they're not a good business. Is there a future for hyperscale cloud computing? Does it even make business sense?" But it turns out that hyperscale cloud computing is a real business, purely because of the operational know-how, such as Azure, which operates a world-class compute infrastructure across more than 60 regions. That's hard to replicate.
So what I'd like to emphasize more than anything else is whether the market landscape is "one winner" or "winner takes all?" Because you have to figure that out. I like to get into areas where there's a lot of market volume (TAM), where you don't have to take the risk of a winner-take-all situation. Ideally, you're in a big market that has room for multiple winners, and you're one of them.
This is what I call the hyperscale layer. At the model level, the models ultimately need to run on some kind of hyperscale computing infrastructure. So I feel like that connection will always be there. It's not just the model itself; the model needs state, which means it needs storage; it needs regular computation to run these agents and their agent environments.
That's why I'm thinking about why the "one person with one model can do it all" scenario probably won't happen.
Dwarkesh Patel
In terms of hyperscale cloud providers, by the way, it's interesting that as a hyperscale cloud provider, you have the advantage, especially in the context of inference time scaling, that you can amortize the cost of the datacenter and the GPUs not only for training, but also for inference again, if that involves training future models. The
I'm curious what type of hyperscale cloud provider you think Microsoft and Azure belong to? Is it on the pre-training side? Is it providing an O3 (OpenAI o3) type of inference service? Or are you just going to host and deploy whatever models are available in the market and be neutral about it?
Satya Nadella
That's a good question. The way we're building our compute clusters is in part to follow Moore's law. I think it's going to be like everything we've done in the past: renewing compute clusters every year, depreciating them over their lifetime value, and then getting very good at deploying the clusters to be able to run different jobs at high utilization. Sometimes large training jobs require a high concentration of peak floating-point computing power (flops) and need to work together. This is good. We should have a sufficient data center footprint to meet these needs.
But at the end of the day, all of these scales are getting so large that even the pre-training scale, if it needs to continue to grow, the pre-training scale is going to have to cross data center boundaries at some point. All of this is more or less a reality.
So once you start crossing the pre-training data center boundary, what makes it any different from anything else? My thinking is that distributed computing is still going to be distributed, so build your compute clusters to be able to handle large training jobs, inference time calculations, and the demands that RL (reinforcement learning) might bring. To me, it's more like more training floating-point operations, because you want to create these highly specialized, lean models for different tasks.
So you need that kind of compute cluster, as well as service demand . At the end of the day, "the speed of light is the speed of light," and you can't expect to build a data center in Texas and say, "I'm going to serve the world from there."
You have to deploy inference computing clusters around the globe to be able to serve the world. That's my understanding of what it takes to build a true hyperscale computing cluster.
Oh, and by the way, I also want my storage and computing power to be close to all of these facilities as well, because it's not just the AI gas pedal that's stateless. My training data itself needs to be stored, and I want to be able to multiplex multiple training jobs, and I want to have memory, and I want to have an environment where these agents can execute programs. That's my understanding of hyperscale computing clusters.
0:15:18 - World Economic Growth 10%
Dwarkesh Patel
Microsoft recently reported $13 billion in annual revenue from AI. But at current annual growth rates, that number will increase tenfold in about four years. If the trend continues, AI revenues will reach $130 billion. If that's the case, how do you expect to capitalize on intelligence on such a large scale, with such industrial-scale applications?
Is it done through Office? Or by deploying to others for hosting? Do you need to have AGI to realize $130 billion in AI revenue? What would that look like?
Satya Nadella
Dwarkesh, you raise a very good question because in a way, if you are going to have this kind of explosive growth of intelligence, abundance of supply, and even commoditization, the first thing we have to look at is GDP growth.
Before I talk about Microsoft's revenue prospects, we need to be clear about one constraint. This is why we often run too fast on AGI hype. Remember, what is the economic growth rate in the developed world?2%?Nearly zero if you deduct inflation.
Looking ahead to 2025, although I am not an economist, what I see is that we face a real growth challenge. So the first thing we have to do, when we say that this is comparable to the industrial revolution, is to achieve industrial revolution-style growth.
To me, that means 10% growth rate, 7% growth rate for developed countries, and still 5% growth after inflation. That's the real measure. You can't just stop at the supply side.
Indeed, I am heartened by the fact that many people are writing articles on the subject. The article points out that the biggest winners will not be the tech companies, but the broader industry that will capitalize on this ample supply of commodities. Suddenly, productivity has increased and economic growth has accelerated. When that happens, we in the tech industry will benefit as well.
But this is the moment of truth. We boast of having achieved the AGI milestone, which to me is a meaningless benchmarking exercise . The real benchmark is: world economic growth 10%.
Dwarkesh Patel
Okay, if the world economy grows by 10%, then the size of the world economy is about 100 trillion dollars Around. If the world economy grows by 10%, that's an additional $10 trillion in value creation per year. If that's the case, your $80 billion in revenue as a hyperscale cloud provider still seems too small. Shouldn't it be $800 billion?
If you really think that we can actually drive world economic growth at this rate in the next few years, does the key bottleneck become: do you have the necessary computing power to deploy all this AI to do all the work?
Satya Nadella
You're right. But by the way, the classic supply-side thinking is, "Hey, let me build it first and they'll come naturally." It's an argument, and we've done it, we've taken enough risks to practice it.
But ultimately, supply and demand must match. That's why I focus on both the supply side and the demand side. If you focus on the supply side without really understanding how to translate supply into real value for your customers, you may be completely out of touch with reality.
This is one of the reasons I focus on inferred income. That's one of the reasons I publicize Reasoning Income ...... Interestingly, not many people seem to be talking about their actual income, but to me it's important to use as a check and balance on how you think about things .
You can't expect supply and demand to be perfectly symmetrical at any point in time, but you need to have proof of existence , proof that you can convert yesterday's inputs (capital) into today's demand so that you can invest again, perhaps even exponentially, and be confident that you won't be completely out of balance.
Dwarkesh Patel
I'm not sure if there is a contradiction between these two views, because one of the things you have done very well is to dare to make early bets. You guys Invested in OpenAI in 2019Even in the Copilot and before any apps come along.
Recall that during the Industrial Revolution, the construction of railroads and other infrastructure led to economic growth ranging from 6% to 10%, but many of these projects did not have immediate revenues in the early years, e.g., "We've got revenues from ticket sales, now we're going to ...... "
Satya Nadella
Lost a lot of money at the time.
Dwarkesh Patel
Exactly. So if you really think that there is potential here to increase the world's economic growth rate by a factor of 10 or even a factor of 5, but then focus too much on "What's the GPT-4 revenue?" , that seems a bit contradictory.
If you really believe that such tremendous growth is possible in the next phase, shouldn't you adopt a more aggressive strategy, such as "Let's go crazy and invest in building hundreds of billions of dollars worth of computing infrastructure?" After all, there's a huge opportunity here, isn't there?
Satya Nadella
Here's an interesting point, right? That's why even a balanced strategy for compute clusters is critical, at least for me. The key is not to just build compute power, but to build compute power that can actually help me train the next big model and serve the next big model. Until you do both, you can't really capitalize on your investment.
So it's not a race to build models, it's a race to create a commodity that can be used by the world to drive economic growth ...... You have to have a complete framework for thinking, not just focus on one aspect.
One of the phenomena, by the way, is that there's going to be overbuilding . Just as you mentioned in the Internet bubble era, the signal has now been sent: hey, you need more energy, you need more computing power. And thank God everyone will flock to it.
In fact, it's not just companies that are deploying it, but governments are putting capital into it, and it's definitely coming up ...... I'm happy to be a leaser , because, by the way, I build a lot of my own and I lease a lot of it. I'm happy to be leasing a lot of computing capacity in 2027, 2028, because I see the scale of these build-outs and I say, "This is great." The only result of all the compute capacity buildout is that prices go down.
0:21:39 - The lowering of smart prices
Dwarkesh Patel
Speaking of falling prices, you've been on a recent DeepSeek After the model is released tweet Talked about the Jevons' Paradox. I'm curious if you can elaborate. The Jevons' Paradox refers to the fact that when the elasticity of demand for something is very high, total consumption increases even if efficiency gains lead to a decrease in price. Is the demand for smarts so sensitive to price reductions?
Because when I think about it, at least from my perspective as a consumer, smart is already very cheap. Per million token It's only 2 cents. Do I really need it down to .02 cents? My bottleneck right now is getting smarter. If you need to charge me 100 times the price to train at 100 times the scale, I'm happy to do it.
But perhaps you see a different picture in the enterprise market or elsewhere. What are the key use cases for smart? Does it really need to go down to 0.002 cents per million tokens?
Satya Nadella
I think what really matters is the effectiveness of the token. Both need to be taken into account: for one thing, intelligence needs to get better and cheaper. Any time there's a breakthrough, like DeepSeek did, when the efficient frontier of token performance changes, the curve gets bent, the frontier shifts, there's more demand. This is the case with cloud computing.
Here's what's interesting: we used to think, "Gosh, we've sold out of all servers in the client-server era." But once we started putting servers into the cloud , all of a sudden people started consuming more because they could buy it cheaper and it was elastic and they could pay for it on demand instead of buying licenses, which completely expanded the market.
I remember going to countries like India to market SQL Server and we sold some, but you know what? Cloud computing is much bigger in India than anything we were able to do in the server era. I think it's going to happen in the AI space as well.
Imagine if you really wanted to provide really cheap healthcare tokens at very low prices in the South, in the developing world, that would be unprecedentedly transformative.
Dwarkesh Patel
I think it makes sense for someone like me in San Francisco to hear people say, "They're kind of stupid, they don't know what it's like to deploy things in the real world."
As someone who has worked with Fortune 500 companies and helped them deploy technology to hundreds of millions, if not billions, of people, how fast do you think these capabilities will be deployed?
Even if you have agents available, even if you have something that can work remotely for you, is this going to be a huge bottleneck given all the compliance requirements and the inherent bottlenecks? Or will it be overcome quickly?
Satya Nadella
This will be a real challenge because the real issue is change management or process change. Here's an interesting analogy: Imagine how a multinational company like ours does forecasting without PCs, e-mail and spreadsheets. Faxes fly around, someone receives the fax, then writes an internal memo to circulate around, people fill in the numbers on the memo, and the final result is a forecast report, probably just in time for the next quarter.
Then someone said, "Hey, I can just use an Excel spreadsheet, email it out, people can edit on it, and then I'll get a forecast report." The whole forecasting business process changed because the work product and the workflow changed.
This is what needs to happen when AI is introduced into knowledge work. In fact, when we think about all of these agents, the bottom line is that there will be a new kind of work and workflow.
For example, even when preparing for our podcast, I use my Copilot and say to it, "Hey, I'm going to talk to Dwarkesh about our quantum computing announcement and the new model we're building for game generation. Give me a summary and tell me what material I should read before I go." It knew about two Nature papers and extracted the information. I even said, "Hey, present it to me in podcast form." It even modeled the format of a conversation between the two of us very well.
This became my new workflow - in fact, I shared it with my team afterward. I copied it into Pages (Microsoft's document editing software) as my work product and shared it. So my new workflow is to think with AI and work with my colleagues.
This is a fundamental change management issue for all those in knowledge work who need to figure out "How do I do my knowledge work in a new way?" This will take time. It will permeate all areas of sales, finance and supply chain.
For traditional businesses, I think it's going to be a change similar to the adoption of Lean manufacturing by manufacturers. I like the Lean manufacturing analogy because in a way, if you look closely, you see that Lean manufacturing becomes a methodology that companies can use to optimize their end-to-end manufacturing processes and improve efficiency. It emphasizes continuous improvement, which means reducing waste and adding value.
That's what's going to happen in the area of knowledge work. It's like Lean Manufacturing in the Knowledge Work world, especially Knowledge Work. And the management team and the individuals doing the knowledge work need to work on this transformation, which will take time.
Dwarkesh Patel
Can I ask a brief question about the Lean Manufacturing analogy? One of the things that Lean Manufacturing has done is that it has changed the factory floor on a physical level. It reveals bottlenecks that people didn't realize were there until they really focused on processes and workflows.
You just briefly mentioned your own workflow - how your workflow has changed because of AI. I'm curious what it's like to run a large company when you have these AI agents that are getting smarter and smarter over time.
Satya Nadella
It's interesting that you ask. For example, I was just thinking today about how if we look at the way we work now, we rely so much on email. I come into work in the morning and I find my inbox stuffed with emails that I need to reply to. I can't wait to use the Copilot agents and have them automatically populate my draft emails so I can just review and send them.
But I'm already using at least ten agents in Copilot, asking them different queries for different tasks. I have a feeling that there will be a new inbox for the millions of agents I'm working with that will send me exceptions, notifications, and requests for instructions.
So at least in my opinion, a new scaffolding needs to be built, the Agent Manager. It's not just a chat interface. I needed something smarter than a chat interface to manage all the agents and their conversations.
That's why I think Copilot, as the UI (user interface) for AI, is significant. All of us will have it. Basically, you can understand that there is knowledge work and there are knowledge workers. Knowledge work may be done by many agents, but there is still a need for a knowledge worker to handle all those knowledge workers. And that's where we need to build the interface.
0:30:19 - Breakthroughs in quantum computing
Dwarkesh Patel
You are one of the few people in the world who can say "I have 200,000 ...... I'm surrounded by an intelligent cluster of Microsoft and all its employees". You have to manage it and interact with it to get the most out of it. Hopefully more people in the world will have this experience in the future.
I'm curious, if your inbox turned out that way, does that mean everyone's inbox will look like yours in the morning?
Well, before we dive into that, I'd like to go ahead and ask you more about AI, but I'd really like to ask you about the major breakthroughs that Microsoft Research has made in the area of quantum computing. Can you explain a little bit about what's going on so far?
Satya Nadella
That's another 30 years of our long journey. It's incredible. I'm the third Microsoft CEO to get excited about quantum computing.
Our vision has always been that you need a breakthrough in physics to build quantum computers that can be used on a practical scale. We chose a path where we thought that one way to get lower noise or more reliable quantum bits (qubits) was to bet on a physical property that was inherently more reliable, and that's why we chose the Majorana zero modes, which were proposed in the 1930s. The question is, can we really make these things physically? Can we really build them?
So the real breakthrough is that we now finally have proof of existence and a breakthrough in physics for the Majorana zero mode, which exists in a new form of matter. That's why we like to compare this to the transistor moment of quantum computing, where we actually have a new form, the topological phase, which means that we can now even reliably hide quantum information, measure it, and can make it. Now that we have it, we feel that with the breakthroughs in the core, fundamental manufacturing technologies, we can start to build the Majorana chip.
We're calling it Majorana One, and I think it's going to be the first chip that can hold a million quantum bits (physical quantum bits). And then on top of that, thousands of logical quantum bits to enable error correction. And then the game is on. Suddenly you have the ability to build really practical-scale quantum computers, and I feel like that's becoming much more feasible now. Without breakthroughs like this, you can still make some milestones, but you'll never be able to build a practical-scale computer. That's what we're excited about.
Dwarkesh Patel
Awesome. By the way, I believe this is it, right?
Satya Nadella
That's right, that's it.
Dwarkesh Patel
Yes.
Satya Nadella
I can't remember now. Did we call it Majorana? - Yes, that's right. Yes, that's right, Majorana One. In his name. Name.
It's incredible to think that we can build a quantum computer with a million quantum bits on such a small chip. And here's the key: unless we can do that, you can't dream of building a practical-scale quantum computer.
Dwarkesh Patel
You're saying that eventually a million quantum bits will be integrated into a chip of this size? That's amazing.
Other companies have announced quantum computers with 100 physical quantum bits, such as Google, IBM, and others. When you say you've announced a breakthrough, you're also saying that your technology has an advantage in terms of scale.
Satya Nadella
Yes. The other thing we've done is we've taken a hardware/software approach. We're building our software stack, and we're now working with experts in neutral atom quantum computers, experts in ion trap quantum computers, and we're also working with other people who have great approaches in areas like photonics, which means that there will be different types of quantum computers. In fact, our most recent announcement was 24 logic quantum bits. So we've also had some remarkable breakthroughs in error correction, which has allowed us to build systems with more than 20 logical quantum bits, even on neutral atom and ion-well quantum computers, and I think that's going to continue to be the case this year; you're going to see us go up and up and up in this area.
But we also said, "Let's go back to first principles and build our own quantum computers, betting on topological qubits." That's what this breakthrough is all about.
Dwarkesh Patel
That's great. Millions of topological quantum bits, thousands of logical quantum bits, what's the projected timetable for scaling up to that scale? If you've got the first transistor, what does Moore's Law look like here?
Satya Nadella
We have obviously been working in this field for 30 years. I'm delighted that we're now making breakthroughs in physics and manufacturing technology.
I wish we already had quantum computers because, by the way, the first thing that quantum computers would allow us to do is to build quantum computers, because it would be much easier to build new quantum gates that simulate the atomic level.
But in any case, the next real thing to do is, now that we have the manufacturing technology, let's build the first fault-tolerant quantum computer. computer). This would be the logical next step.
So, I think I can now say, "Maybe in 2027, 2028, 2029, we'll be able to actually build it." Now that we have this quantum gate, can I put it into integrated circuits and then put those integrated circuits into a real computer? That's the next logical step.
Dwarkesh Patel
What do you think it's going to look like in 2027, 2028, when you've successfully built a quantum computer? Is it something that's accessible through an API? Or is it something that you use internally for materials and chemistry research?
Satya Nadella
That's a great question. One of the things I've been excited about is that even in today's world ...... we've got quantum computing programs and we've added some APIs to them. The breakthrough that we made about two years ago was to think about the High Performance Computing (HPC) stack, the AI stack, and quantum computing together.
In fact, if you think about it, you'll see that AI is like the simulator of simulators. Quantum computing is like an analog of nature. What will quantum computing do? By the way, quantum computing is not going to replace classical computing. Quantum computing will be very good at what it does well, and classical computing will continue to play a role. ......
Quantum computing would be well suited for tasks that are not data-intensive, but are demanding in terms of state space exploration. It should be data-light, but you need to explore exponential levels of state. Simulations are a good example of this: chemical physics, biology, and so on.
One of the things we've started doing is really using AI as a simulation engine . But you can train it. So the way I understand it is that if you combine AI with quantum computing, you might be able to use quantum computing to generate synthetic data that AI can then use to train better models to understand how to simulate chemistry or physics or other disciplines. The two would be used synergistically.
Even today, we're actually using high-performance computing in conjunction with AI. I would like to replace some of the high-performance computing components with quantum computers.
Dwarkesh Patel
Can you tell me how you make these research decisions? Are these decisions going to really pay off in 20, 30 years, especially in a company the size of Microsoft? Obviously, you are very knowledgeable about the technical details of this program. Is it possible for you to do that with everything that Microsoft Research does?
How do you know that the bets you make today will pay off in 20 years? Does it just need to grow organically within the organization, or how do you track it all?
Satya Nadella
One of the things I think is great is that Bill Gates founded Microsoft Research (MSR) in 1995. I think that in the long history of all curiosity-driven research organizations, Microsoft Research has built up this institutional advantage over the years of having a research organization just to do basic research. So when I think about capital allocation or budgets, we first put money in and say, "This is the Microsoft Research budget." And we do that every year, knowing full well that most of those bets are not going to pay off in a limited time frame. Maybe Microsoft's sixth CEO will benefit from it. And in the tech industry, I take that for granted.
I think what's really important is, as technologies like quantum computing or new models mature, are you able to capitalize on the opportunity? So, as a traditional company, you look back at the history of the tech industry and you realize that the problem is not that people aren't investing, it's that you need to have a culture that knows how to scale innovation.
This is the real problem for CEOs and management teams, and it is very challenging. It requires both good judgment and a great culture. Sometimes we do well, and sometimes we make mistakes. I can tell you that there are thousands of projects at Microsoft Research that we should be leading, but we're not. I always ask myself why. It's because we don't have the confidence or the framework of thought to know how to not only turn innovations into useful products, but also find viable business models and bring them to market.
That's the job of the CEO and the management team: not just to get excited about something, but to be able to actually execute a complete program. Easier said than done.
Dwarkesh Patel
When you refer to the last three CEOs of Microsoft, if each of them raised the market capitalization by an order of magnitude, then Microsoft's market capitalization could be equivalent to the total world economy by the time your next breakthrough arrives.
Satya Nadella
Or remember, the world economy is going to grow at 10%, so we'll be fine.
0:42:51 - How Muse changed the gaming industry
Dwarkesh Patel
Let's dive into another major breakthrough you guys just made. Amazingly, you released two major achievements on the same day, and that's your Game World Model . I'd love to hear from you about that.
Satya Nadella
We named it Muse. It will be a model of world behavior or a model of human behavior.
That's pretty cool. We all know that Dall-E and Sora What has been accomplished in terms of generating models is incredible. One direction we want to explore is utilizing game data. Can you generate games that are coherent and show the diversity of the game? And can you sustain user modifications?
That's the goal of Muse. Researchers at Microsoft Research collaborated with one of our game studios to achieve this result, which is the subject of another paper published in the journal Nature.
What I'm excited about is that we're going to be launching a series of games very soon and start using these models, or we're going to train these models to generate games and then start playing them.
In fact, when Phil Spencer (CEO of Microsoft Games) first showed me Muse, he was holding an Xbox controller, a model that basically takes input from the controller and generates output based on that input. And the output was consistent with the game content. It was a "wow" moment for me. It was kind of like the first time we saw ChatGPT Finish sentences, Dall-E drawings or Sora generate videos. This is a milestone moment.
Dwarkesh Patel
This morning I had the opportunity to watch the live demo video with Katja, one of your lead researchers. It was only after talking to her that I really realized how incredible this is. We used to use AI to model the agent, and now we're able to generate a coherent game world in real time just by using the same technology to model the world around the agent - we'll be overlaying the demo video on the podcast so you can see it for yourself. I think the video should be posted by then, and you can watch it there as well.
That in itself is incredible. During your tenure as CEO, you've invested tens of billions of dollars in building Microsoft's gaming business and acquiring IP (intellectual property).
In retrospect, if you can integrate all of this data into a large model that provides you with the experience of accessing and experiencing multiple worlds at once, and if that's where the gaming industry is headed, then it seems like a very wise investment. Do you have any premonitions about this?
Satya Nadella
I wouldn't say we invest in games to build models. We invest in games, frankly, because - here's an interesting point about our history: we developed our first game before we built Windows. Microsoft Flight Simulator was a Microsoft product long before Windows was built.
So gaming has a long history at Microsoft, and we wanted to get into gaming for the sake of gaming itself. I've always said that I hate to get into businesses that are "a means to an end". They have to be ends in themselves.
And then, yes, we are not a conglomerate . We are a company, and we have to bring all these assets together and be a better owner by adding value. For example, cloud gaming is a natural investment for us because it will expand the capacity of the market (TAM) and expand the ability for people to play games anytime, anywhere.
The same goes for AI and gaming: we firmly believe that it could help change the gaming industry - it's kind of like the CGI (computer-generated imagery) moment of the gaming industry. That's great. As the world's largest game publisher, it will help us. But at the same time, we have to make quality games. I mean, you can't be a game publisher if you don't focus on that first.
But the fact that these data assets are going to be very interesting, not only in the gaming space, but that it's going to be a universal model of behavior and a model of the world, that's fantastic. I mean, I think game data could be to Microsoft what YouTube is to Google. So I'm very excited about that.
Dwarkesh Patel
Yeah, that's what I meant, in the sense that you can have a unified experience across many different types of games. How does this relate to other work Microsoft has done in the past outside of AI (e.g. mixed reality)? Maybe it gives smaller game studios the opportunity to build AAA action games? Looking five to ten years into the future, what application scenarios can you envision?
Satya Nadella
Five, six, seven years ago, I saw these three things as the cornerstones. I said at the time that the three big bets we wanted to make were AI, quantum computing, and mixed reality. I still believe in them because in a sense, what are the big questions we need to solve?
Presence. This is the dream of mixed reality. Can you guys create a real sense of presence? Like you and I are doing right now with this podcast.
Frankly, I think this remains one of the more difficult challenges to solve. I thought it would be easier to solve. But it's probably more challenging, probably because of the social aspect of it: wearing the device and so on.
We are excited about our upcoming project with Anduril and Palmer Luckey, who will drive the IVAS (Integrated Visual Augmentation System) program forward as it is a great use case. We will continue to work on this.
But the 2D interface is still important. It turns out that tools like Teams, thanks to the epidemic, we've really gained the ability to create a sense of presence through the 2D interface. I think that's going to continue to evolve. It's a long-term trend.
We've already discussed quantum computing , and AI is another key area. So, I'm looking at these three areas and thinking about how do we bring them together? Ultimately, it's not about technology for technology's sake, but it's about addressing some of the basic needs that we aspire to as human beings in our lives, and more importantly, we need them in our economy to be more productive. If we can somehow do that well, then I think we've really made progress.
Dwarkesh Patel
When you write your next book, you'll have to explain why these three fields came together at almost the same time, right? For example, you might think that quantum computing and AI should make breakthroughs in 2028 and 2025, respectively, but that doesn't seem inherently necessary.
Satya Nadella
That's right. In a way, my understanding is: Is there a systemic breakthrough? To me, quantum computing is a systemic breakthrough.
Is there a business logic breakthrough? For me, AI is a business logic breakthrough, i.e., can the logic layer reason in a fundamentally different way? Can you replace imperative coding with learning systems? That's what AI is all about.
On the UI side, the key is the sense of presence.
0:49:51 - Legal barriers to AI
Dwarkesh Patel
Let's get back to AI. In your 2017 Publications...... 2019 you invested in OpenAI, very early on, in 2017 or even earlier, and you write in your book, "One might also say that we are birthing a new species, one whose intelligence may have no upper limit."
Of course, it's too early to talk about this in 2017. We've been talking at length about proxies, Office Copilot, capex, and so on. But if you take a more macro view of what you're saying and consider your role as a hyperscale cloud provider, as a modeling researcher, and as someone who provides training, inference, and research support for building new species, how do you see it all coming together?
Do you think we'll move toward superhuman intelligence during your tenure as CEO?
Satya Nadella
I think that even Mustafa (probably referring to Mustafa Suleyman, the Inflection AI CEO) also uses the term "new species." In fact, he's been using it more frequently lately.
My understanding is that you absolutely need trust. Before we claim that it is as important as "species", the first thing we need to make sure of is to build real trust, both at the individual level and at the societal level, and to integrate it. That is the real dilemma.
I think the biggest factor limiting the power of AI will be how our legal ...... infrastructure evolves to cope with AI.Our current world order is based on the notion that humans own property, have rights, and take responsibility. The first thing we need to think about is what this means for the tools humans are using. If humans are going to delegate more power to these tools, how will this structure evolve? Until these questions are actually addressed, I don't think it makes sense to just talk about technological capabilities.
Dwarkesh Patel
Are you saying that we won't be able to deploy these intelligences until we figure out how to ......?
Satya Nadella
That is exactly right. Because at the end of the day, there is no way to get around the legal issues. Today, you can't deploy these intelligences unless someone guarantees them in the name of humanity.
And as you say, I think that's why I think even the most powerful AI is essentially working within the limits of human delegated authority. You might say, oh, that's alignment and all sorts of other issues. That's why I think you have to actually make these alignment jobs work and make them verifiable in some way, but I just don't think you can deploy intelligence that's out of control. For example, the AI takeoff problem may be a real problem, but the real problem will be in court before it becomes a real problem. No society is going to allow someone to say "that's what AI does."
Dwarkesh Patel
Yes. Well, there are a lot of societies in the world, and I was wondering if there would be any that might have a more lenient legal system. If you can't stop takeoffs, then you might be concerned. The takeoff doesn't have to happen in the U.S., does it?
Satya Nadella
We don't think there's a society that doesn't care, right? There may be rogue actors, I'm not saying there won't be rogue actors; there are cybercriminals and rogue states; they've always been around.
But it's also unrealistic to think that human society as a whole doesn't care about it. I think we all do. We know today how to deal with rogue states and rogue actors. The world is not going to stand by and say, "We can tolerate it." That is why I am glad that we have a world order in which there are consequences for any rogue actor or rogue state.
Dwarkesh Patel
That's right. But if you envision a scenario where the world economy grows by 10%, I think it really depends on AGI being realized because trillions of dollars worth, which sounds closer to the sum of human wages, is about 60% of the economy. to get to that scale, you have to automate the labor or the labor assistance in a very significant way.
If this is possible, and once we figure out the legal issues involved, it seems likely that we will figure them out even during your term. Are you thinking about superhuman intelligence? Is that, for example, the most important accomplishment of your career?
Satya Nadella
You bring up another point. I know there's been a lot of discussion on this, David Autor and others, about labor ...... of 60% Another issue that I think needs to be addressed is, at least in our democracies. I think that in order to have a stable social structure and to have a functioning democracy, you can't have just a return on capital and not a return on labor. We can talk about that, but that 60% labor has to be revalued.
In my own way of understanding, perhaps you could call it naivety, we will begin to value different types of human labor. Human labor that is considered high value today may become commodities. There may be new things that we will value.
That includes the people who come to help me with physical therapy, or whatever it is that we're going to value, but at the end of the day, if we don't have a return on our labor and there's a lack of meaning and dignity in our work, then that's going to be another constraint on deploying any of these technologies.
0:55:46 - Proper understanding of AGI security
Dwarkesh Patel
In terms of alignment, two years ago you released Sydney Bing. To be clear, I think it's a charming, lovely, somewhat comical example of unalignment given the state of the art at the time.
But that's because it was the age of chatbots. They could think for 30 seconds and give you some funny or inappropriate response. But if you think about that kind of system - I think a reporter for the New York Times, trying to get him to leave his wife - if you think about the future, where you have these agents that can run for hours and weeks and months on end, like autonomous AGI clusters, and they could be unaligned and screwing things up in similar ways, maybe even coordinating with each other, what are your plans for the future? So that when you have a really strong AGI, you can make sure it's secure?
Satya Nadella
You're right. That's one of the reasons why when we allocate compute resources, we usually allocate them for alignment challenges.
More importantly, what kind of runtime environment are you going to use to actually monitor this stuff? What's the observability around it? We're dealing with a lot of similar issues today in classic computing, such as cybersecurity. We don't just write software and then leave it alone. You own the software and you monitor it. You monitor it for cyberattacks, for injections, and so on.
So I think we're going to have to build enough software engineering around the deployment aspects of these technologies and then, within the model, achieve alignment. Some of these problems are really scientific problems, and some of them are really engineering problems, and we're going to have to solve them one by one.
It also means that we have to take responsibility for ourselves in all these areas. So I would prefer to deploy these technologies in areas where you can really control the scope and scale of them. You can't release something into the world that is going to cause harm, because society won't allow it.
Dwarkesh Patel
When you have an agent that can actually do weeks of work for you, what are the minimum assurances you want before you allow it to run for a Fortune 500 company?
Satya Nadella
I think when I use a tool like Deep Research, even now, the minimum assurance that we want to get is that before we have any physical entities, and I think that's the threshold, when you cross that threshold, things may change. That could be a key point.
Another key point is, for example, the permissions of the runtime environment in which the agent runs. You might want to make sure that it's sandboxed, that it doesn't go outside the bounds of the sandbox.
Dwarkesh Patel
I mean, we already have Web search, and it's beyond the sandbox.
Satya Nadella
But even what it does with Web search and what it writes - for example, as you said, if it's just writing a bunch of code to do some calculations, where is that code deployed? Is that code temporary and only used to create output, or does it get released to the world?
These are things that you can actually control in the action space.
Dwarkesh Patel
Beyond the security issue, when you think about your own suite of products and consider that if you do have such powerful AI, then at some point it becomes more than just Copilot - you mentioned an example of how you prepared for this podcast - -it's more akin to how you delegate work to your colleagues.
Considering your current suite of products, what would it look like to add this capability? I mean, one question is whether LLM will be commoditized by other things .
I'm wondering, if LLM is your main entry point to all of these things, these databases or canvases or Excel sheets and so on - is it possible for LLM to commoditize Office?
Satya Nadella
This is an interesting question. My understanding, at least in the first phase, is that LLM can help me use all these tools or canvases more effectively to accomplish my intellectual work?
One of the best demos I've seen was a doctor preparing for a tumor board workflow. She's going to a Tumor Board meeting, and the first thing she does with Copilot is create an agenda for the meeting because LLM helps her analyze all the cases, which are all stored on some SharePoint site. It says, "Hey, these cases-obviously, a tumor board meeting is a high-stakes meeting, and you need to be aware of the differences between cases so that you can allocate your time wisely."
Even for the reasoning task of creating an agenda , it's great that it knows how to allocate time. So I use LLM for this task. Then I attend meetings and I talk to all my colleagues in Teams calls. I focus on the actual case rather than taking notes because you now have AI Copilot doing the full transcription. It's not just a transcription, it's a database entry of the content of the meeting that can be saved permanently and accessed at any time.
Then she walks out of the conference room, having discussed the case and not being distracted by taking notes. She is a teaching physician; she wants to go and prepare for her course. So, she goes into Copilot and says, "Make PowerPoint slides of what I did at the tumor board meeting so I can explain it to my students."
This is the scenario I'm talking about. The UI and the scaffolding that I have is the canvas that is now being populated with LLM. And the workflow itself is being reinvented; the knowledge work is being done.
Here's an interesting phenomenon: If someone had told me in the late '80s, "You're going to have millions of documents on your desk." I would have said, "What the hell does that mean?" I really would have thought that my desk would be filled with millions of paper documents. But in fact, we do have millions of spreadsheets and millions of electronic documents.
Dwarkesh Patel
I don't, you do.
Satya Nadella
They're all there. The same thing will happen with proxies. There will be a UI layer. To me, Office is not just Office today; it's the UI layer of knowledge work. It's going to evolve as the workflow evolves. That's what we want to build.
I do think that the SaaS (Software as a Service) applications that exist today, these CRUD (Create, Read, Update, and Delete) applications, are going to change fundamentally because the business logic is going to move more to the agent layer. In fact, another cool feature in my Copilot experience today is that when I say, "Hey, I'm getting ready for a meeting with a client," all I have to do is say, "Give me all the meeting notes I should know about." It'll pull information from my CRM database, it'll pull information from my Microsoft Graph, it'll create a composite work product, and then it'll even apply logic to it. I think this is going to revolutionize SaaS applications as we know them today.
Dwarkesh Patel
SaaS as an industry could be worth hundreds of billions or even trillions of dollars per year, depending on how you count. If AI really can turn the SaaS industry on its head, is it possible that you guys could once again increase Microsoft's market capitalization by 10x in the next decade? Because you're talking trillions of dollars ......
Satya Nadella
This will also create tremendous value in the SaaS space. One of the things we may not be paying enough attention to is the massive IT backlog that exists in the world.
These code generation tools, coupled with the ability for me to use agents to interrogate all of your SaaS apps and get more utility out of them, are going to bring about the biggest explosion of apps, and they're going to be called agents, and so all of a sudden we're going to have access to services in every vertical, in every industry, in every category.
So there is tremendous value here. You can't stand still. You can't just settle for the old "I схематизировал up some narrow business process, I have a UI in the browser, and this is my product" model. That's not going to be the case anymore. You have to move up and think "What are the tasks I have to be involved in?"
You'll want to be able to turn your SaaS application into a great agent that can participate in a multi-agent world. If you can do that, I think you can even increase the value.
1:04:59 - 34 years at Microsoft
Dwarkesh Patel
Can I ask you some questions about your experience at Microsoft?
Satya Nadella
Sure.
Dwarkesh Patel
Is being a "company guy" underrated? You've spent most of your career at Microsoft, and it's safe to say that one of the reasons you've been able to add so much value to the company is because you've experienced its culture, history, and technological evolution firsthand. You have accumulated all of this background knowledge by advancing within the company. Should more people with this background knowledge be running the company?
Satya Nadella
That's a good question. I hadn't thought of it that way before.
Every year of my 34 years with Microsoft, I've gotten more excited than thinking, "Oh, I'm just a company guy," or whatever. I take the status of "company person" seriously, even for anyone who joins Microsoft. They join Microsoft as long as they think they can use Microsoft as a platform to realize their financial rewards as well as their goals and sense of purpose. That's the contract between us.
So I think that yes, companies have to create a culture where people can fit in and become "company people" like me. At least in my case, Microsoft has done a pretty good job of that, and I hope it continues.
Dwarkesh Patel
You talk about a sixth CEO who will have the opportunity to capitalize on the research you're starting now. What are you doing to retain future Satya Nadella so they can become future leaders?
Satya Nadella
That's interesting. I've been thinking a lot about Microsoft's 50th anniversary this year. I think the key to thinking about this is that longevity is not the goal, relevance is.
What I and all 200,000 Microsoft employees must do every day is this: is what we do useful and relevant to the trends we see in the world, not only today, but tomorrow?
We're in an industry where there is no franchise value, and that's another problem. If you take our R&D budget for this year, it's a complete guess as to what's going to happen five years from now. You have to take that attitude and say to yourself, "We're doing things that we think are going to be relevant."
So you have to focus on that. And then you have to recognize that there's a batting average issue and you can't succeed every time - you have to have a high tolerance for failure. You have to make enough attempts to be able to say to yourself, "Okay, we're going to make it through this successfully as a company." That's what makes this industry so tricky.
Dwarkesh Patel
Speaking of - you mentioned earlier that Microsoft's 50th anniversary is two months away. If you look at the top 10 or top 5 companies in terms of market capitalization, basically all of them except Microsoft are younger than Microsoft. It's an interesting observation why the most successful companies tend to be young. The average lifespan of a Fortune 500 company is only 10 to 15 years.
What has Microsoft done to stay relevant over the years? How have you continued to reinvent yourselves?
Satya Nadella
I like Reed Hoffman's use of the term "refounding". It's a mindset. People talk about founder mode, but for us mortal CEOs, it's more like refounding mode.
Being able to see things in a new light is key. To your question: Can we culturally create an environment where reinventing ourselves becomes a habit? We go to work every day and tell ourselves, "We have shares in this place that can change the core assumptions of what we do and how we relate to the world around us. Are we allowing ourselves to do that?" I think a lot of times companies can feel overly bound by their business model or other factors. You just need to unshackle yourself.
Dwarkesh Patel
What kind of company would you start if you left Microsoft?
Satya Nadelladraw
The company I'm going to start? Oh, my God. That's the "company man" in me thinking, "I'll never leave Microsoft."
If I were to consider doing something, I think I would choose someone with ...... When I look at the dream of technology, we always say that technology is the biggest and greatest democratizing force.
I feel like we finally have that capability. If you're talking about the number of tokens per dollar per watt that we're able to achieve, I'd like to find some areas where we can apply it, some areas that are severely underserved.
Health care, education ...... public sector would also be an option. If you choose these under-served areas where I, as a citizen of this country or a member of this community, or a citizen anywhere, would my life be better off if all of this amply supplied intelligence could be translated into better healthcare, better education, and better public sector institutions to serve me? That would be an area worth exploring.
1:10:46 - Does Satya Nadella believe in AGI?
Dwarkesh Patel
Listening to your answers to different questions, I'm not so sure you think AGI is real. Will there ever be something that can automate all cognitive labor , like what anyone can do on a computer?
Satya Nadella
This is where I get confused with people talking about the definition of AGI. Cognitive labor is not static. Cognitive labor exists today. If I have an inbox to manage all my agents, is that a new kind of cognitive labor?
Today's cognitive labor may be automated. But what about the newly created cognitive labor? Both need to be considered, and this is a shift ......
That's why I made the distinction, at least in my mind, not to confuse knowledge workers with knowledge work. Today's knowledge work may be automated. Who says my goal in life is to sort my email? Let an AI agent sort my email.
But after sorting through my emails, give me a higher level cognitive labor task like "Hey, here are the three drafts I really want you to review." It's a different level of abstraction.
Dwarkesh Patel
But will AI ever reach the second level?
Satya Nadella
It might, but when it reaches the second level, there will be a third. Why do we think we're worried about all cognitive labor disappearing when we've dealt with the tools that have historically changed the definition of cognitive labor?
Dwarkesh Patel
I'm sure you've heard the examples before, for example, that horses can still be useful in certain areas, and that there are terrains where cars can't travel. But the question is, do you see horses on the streets? Would you employ millions of horses? That's not going to happen.
So the question is, will something similar happen to humans?
Satya Nadella
But is that only in a very narrow dimension? Humans have only been valuing what we understand to be some kind of narrow "cognitive labor" for 200 years.
Let's take chemistry as an example. If quantum computing + AI really helps us do a lot of new materials science research and so on, then having AI to do new materials science research is great. But does that take away all the other things that humans can do?
Why can't we live in a world where we have a powerful cognitive machine and know that our cognitive agency is not deprived?
Dwarkesh Patel
I'm going to ask a question that doesn't pertain to you, but rather a different scenario so you can answer without fear. Assuming a Microsoft board meeting, would you consider adding an AI member to the board? Is it likely to have enough judgment, background knowledge, and overall understanding to be a useful advisor?
Satya Nadella
This is a good example. One of the features we've added is a facilitator agent in Teams. The goal is that, although it's still early days, can this facilitator agent be a good facilitator by utilizing long-term memory, not only the context of the meeting, but also the context of the project I'm working on, the team, etc.?
I even wish there were such lead agents at board meetings, because it's easy to get distracted at board meetings. After all, board members only come in once a quarter and they try to digest what's going on in a company as complex as Microsoft. A lead agent that can help all the human members stay focused and concentrate on the important issues is fantastic.
It's actually equivalent to having, as you said in your previous question, having something that has an infinite memory that can even help us. You know, what was Herbert Simon's theory? We are all bounded rationality. So it would be great if human bounded rationality could be solved by an external cognitive amplifier.
Dwarkesh Patel
Speaking of the field of materials and chemistry, I think you recently said. You want the progress of the next 250 years to be realized in the next 25 years.Now, when I imagine what might happen in the next 250 years. Now, when I imagine what might happen in the next 250 years, I think of space travel, space elevators, immortality and cures for all diseases. What do you think about the next 25 years?
Satya Nadella
I ask this question because I like the idea that the industrial revolution lasted 250 years. We had to go through the whole process of going from a carbon-based system to something different.
That means you have to radically overhaul everything that's happened in chemistry in the last 250 years. That's why I want us to have quantum computers that can help us get new materials that we can then make to help us meet all the challenges that we face here on Earth. And then I'm all for interstellar travel.
Dwarkesh Patel
Great. Satya, thank you so much for taking the time to do this interview.
Satya Nadella
Thank you very much. This interview was great. Thank you.
Dwarkesh Patel
Okay, thank you.
English 
简体中文 
日本語 
Deutsch 
Português do Brasil