Satya Nadella： Microsoft's Products Will Soon Access Open AI Tools Like ChatGPT ｜ WSJ

We will incorporate this in our own application. So you can fully expect us to obviously, every product of Microsoft will have some of the same AI capabilities completely transform the product. To jump right into it, let's talk about artificial intelligence. It's a big topic already in everybody's mind here. There's already been quite a bit of speculation given to me that in seven or eight years, this interview will be conducted by some kind of artificial intelligence, which might make it better. That there won't be journalists here. There'll be machines doing all of the work, which also could be an improvement for all I know. You guys obviously are an investor in open AI. You've invested more. There are many reports out there that you are looking to invest more. I think 10 billion has been thrown around as the number out there. You have an inside view of the technology. Can you just put in some broad perspective? We saw the demo, we've all experienced it. We see the dramatic impact it has. How do you assess really holistically where the technology is and isn't, what it does, what it doesn't do, and how big a breakthrough moment we're at? Yeah, first of all, yeah, I wish I had put chat GPT through that security line myself. Let me just give you two anecdotes and then try to generalize it. On New Year's Day this year, I just looked in my newsfeed and I saw this tweet that Andrij Karpathy, who was the guy at Tesla who did the autopilot, tweeted saying 80% of the code he writes today is being written by GitHub Copilot. Think about this. This is one of the most elite AI developers who is saying that they're being that much more productive. It's not that Andrij is not doing great AI work. It's just that he's getting a ton of leverage. And that's perhaps been the breakthrough product of this generation of AI in the last year. And it's a real product. It's not some research project somewhere, what have you. I mean, this is some literally changing the productivity curve for some of the most, for software engineers, which was considered the knowledge worker, or rather the most difficult to automate knowledge work, so to speak. And so that's one. Then the next day I was in India and I saw this demo, which was just completely stunning for me. I saw someone use one of the Indian vernacular languages talking to a bot that Microsoft Research, along with our partners from the Ministry of Electronics and the program called Bashini, which is about trying to basically use AI to get language translation to be democratized in India. So here is essentially someone who is in rural India using one of a local language, talking to a bot on WhatsApp. And this bot, basically this person wanted to have access to some government program. And so he asks about the program. The bot comes back and says, you should go to the portal and do X, Y, and Z. And he then goes on to say that, look, I don't wanna go to the portal, just do this for me. And so this is someone in rural India sort of expressing a pretty complex thought, even if they may not be all literate about how to navigate some complex set of instructions. And then on the back end of it, what the team had done was used GPT to essentially train over all of government of India's documents and made it accessible in every Indian language and made it automatable. So think about this. I'd never seen in my life, at least in the 30 years I've been in tech, ever a demo where an advance that happened in, say, West Coast to the United States shows up in very real terms for someone in rural India within months. I mean, that is diffusion of technology that I've never seen. I don't think it happened in the Industrial Revolution, but maybe this time around in what is sort of the knowledge worker equivalent of the Industrial Revolution, it's going to help everybody. So these two things, so the most elite AI developer on one end, and on the other end, there is a rural farmer in India being able to access government programs through AI. So those two anecdotes speak to me at least what the power of this thing is. And from what you see and where we're at, are we at the early stage of a curve that's rapidly going to change in the next few years, and we're going to see this kind of exponential development? I think that's the key. So what you just said was, why is there so much more excitement this year than last year? Is because of that exponential growth of the power of these models. I'm not claiming, by the way, that this is the last innovation in AI, or this is the last model architecture. There's going to be lots more to come. But if you look at what has happened between GPT-3 to 3.5 to what comes next, these are not linear, this is not linear progress. Okay, but the slightly pessimist case, you're an optimist on this, would be it's exciting, it's rapidly moving, it's a bit terrifying. Knowledge workers, I heard it already here at Davos this week, already anxious in some ways about their future jobs and what it means. When you see technology spread on that scale, I have to ask, one, are we societally ready for that? And are you really confident that there's the optimist case that it enhances all of our productivity rather than replaces many of us? It's a great, it's the right question to ask. I mean, look, I mean, the first thing I step back and say, when I look at all productivity data, I look at all economic growth data, I sort of look at that and say, God, we need something to help us. It's not like we're, as a world, growing at inflation-adjusted three, 4%. If we really have the dream that the eight billion people plus in the world, their living standards monotonically should keep improving year over year, what is that input that's going to cause that? Like what, you know, we need, and we have some tough challenges, right? And when you think about energy transition, let's just take that for an example. Essentially, what we are saying, at least the way I, as a layperson, understand it, is you gotta take 250 years of chemistry and compress it into some 25 years so that we can then make it to the other side. If that is the case, computational chemistry is gonna be super important, and so applications of AI to that is probably the way we're gonna make it. So I look at it and say, we need something that truly changes the productivity curve so that we can have real economic growth. And then, if you have that, there are all kinds of things that we can talk about, which is what does it mean to displacement, which will be real. After all, there was displacement in the Industrial Revolution. We went through some lot of hardship to come up with even an understanding of how to deal with it. This time around, we have that understanding. So therefore, I would say, first, let's not fall into what is the lump of labor fallacy. If you don't fall to that, then what are all the new jobs? What is the training required? And here's one other thing I'd say, Matt, which, if you think about the way frontline workers today are being enabled with technology to be able to do what was previously perhaps considered knowledge work, it's probably the biggest thing that is underreported, underwritten about. Then, we have at Microsoft a tool called Power Platform. So think of it as what, essentially, it's Excel-level skills at best, even lower, where you can take that tool and create applications, create workflows. In fact, you can use GPT to essentially prompt a workflow. So that means somebody who's in the retail frontline, warehouse frontline, healthcare frontline, pick your favorite frontline job, where they're not IT specialists, they're not skilled in computer science, they can, in fact, participate in the digital transformation of their org. If they do that, what happens next is better wages, because they're now considered to be not just doing frontline work, but are also participating in what was previously knowledge work. So to me, that's an example of, in fact, people participating in the labor force with better support. I'm going to ask you, though, what concerns you when you see seeing things that we haven't seen yet that aren't public yet? You see what's coming. Tell us a little bit about what we haven't seen publicly, and then, you know, tell us a little bit about where it's going. Look, I mean, I think the thing that, all the concerns we have around safety, whether it's harmful content or hallucination or any of these, you just need to buy design. Like, I mean, all software engineering has to shift left. You have to sort of think about the model itself being safe. Then you have to, on top of it, have a safety system that is built. And so you have to really take this as a first-class engineering concern and build. And this is not new to just AI. It's true in any other category of software today as well. But what do we need to be doing as a society, and maybe even for the folks in this room, a lot of them are knowledge workers, as individuals to be thinking about and preparing ourselves for what's coming? Because as you just sort of even hinted, we haven't wrapped our heads around social media and what hit us 20 years ago. This is gonna be exponentially faster and probably have a more dramatic impact in the short term on our society. How do we need, what do we need to be thinking about societally? You know, the way at least I come at it is I grew up with sort of the PC revolution starting in the late 80s. And it was just an empowering moment, right? I mean, even if you think about my own story, I mean, the way I was able to sort of have this technology reach me where I was growing up and be able to sort of get into it, and then, you know, and the rest is history. So therefore, I look at it and say, it doesn't matter if you are a knowledge worker or a frontline worker, use these tools to actually then get more leverage in your job. If Andreej as a software developer in AI can do that, why can't all of us have access to it? If that farmer in rural India can use this technology to get their job done, that's the way. I think the best way to prepare for it is to sort of not bet against this technology and this technology helping you in your job, in your business process. And especially at a time like this, Matt, like here we are, right? We have inflation, we have all of this macroeconomic headwinds. Doing more with less is perhaps more at a premium. And so if I find any technology that allows us to be more productive, let's take that, and then let's create that economic surplus and then talk about how to distribute it in all the ways. And there are many, many mechanisms one can talk about. Well, I mean, when you talk about productivity, I'm gonna ask you about productivity in general sense, but some people hear things like, we have to be more productive, and they think, right, with fewer human beings. So you're obviously very optimistic, you obviously see it. What is, can you just talk a little bit about what might be concerning or cautionary about it as it becomes more widely adopted? Yeah, I mean, I think the point is, if there's more automation, and take even what's happening with tech jobs today. Interestingly enough, if you look at it, there was, as tech supply in what is considered the tech industry is now sort of basically getting more distributed, because every auto company, every pharma company, every retail company needs to hire tech people. So to some degree, we should not think of automation in one company to mean that that person is not required elsewhere. And so the question is, what is the overall capacity for the economy and society to absorb and create new jobs? And that, I think, is going to be the real, and then the other thing that we should also have is, wage support for different jobs. I've always felt like, why is there such disparity today in the job, in the labor market, between, let's say, some care worker, and let's call it sort of a software developer. And it's because we think of, that we as a society have decided to give a different, I'd say, premium. Those premiums will adjust as some of these technologies really, truly get diffused. A bit while the diffusion seems key, when I think about the theme this year at Davos of fractionalization, and there's a lot of talk in the hallways here about the new globalization. The implications of the technology obviously go beyond national borders quite quickly and quite dramatically. When you talk about the impact you see in India, it's not just software jobs moving from Silicon Valley to Boise, Idaho, but really, potentially everywhere around the world very quickly here. Yeah, I mean, look, the idea that every country has to have social cohesion based on the jobs and the economic opportunity, I think, is much better understood today than, let's call it, the first phase of globalization. Whatever happens in this phase of globalization is not gonna repeat the mistakes of the previous era, where when there is displacement happening in a society, I think everybody's gonna focus on what do we do about those people who are getting displaced, getting trained, getting new jobs. So I think now we have to, in some sense, to your point, really account for both forces. A globalization force, you know, David Ricardo was not wrong. There is going to be competitive advantage of nations and things will still exist where there is a reason why people will wanna globalize, but at the same time, every politician gets elected because of local politics, and they're gonna care a lot more about jobs locally. And so therefore, in fact, that's the right balance to have. Last question on this. We will talk about a few other things, but it's a bit hypothetical, of course, particularly given the scale of change of the technology, but have you wrapped your head around it as you look at it? If you thought about a company like Microsoft five years from now, 10 years from now, what this widespread adoption of the technology of chat GPT means for how you're organized, how you're structured, how you do work, what the real actual implications are gonna be on the ground. Have you been able to think that through very much? And when you look ahead, what do you see? Absolutely. I mean, this is one of the hardest things that you learn in a tech company. We are, what, 48 years old today, and it's hard because you have to adjust. The means of production, essentially, in our tech business change dramatically, and it requires us to unlearn and learn, which is always very unforgiving, quite frankly. And if you look at what even these large models or foundational models are, it's very different. In fact, I would say one of the things that is perhaps the toughest thing is the previous era AI expert is the one that's gonna have the hardest time shifting to this new paradigm, because everything, all the assumptions, oh my God, here's my data, I'm gonna train on that data, my data is the only thing that matters, all that's now questioned. You now have a large model which can be used as a retrieval engine on your data that's gonna perform better than sort of something that you trained on on your data. I mean, just the intuition, just so that people have what this is, is suppose you train a model on, let's say, all the math formulas. Yeah, it'll get pretty efficient in math. But let's say you train a model on all the math formulas and all the human text there is on the web. It gets better at math. And you say, wow, how is that possible? Of course it's possible, because when we went to school, it's not like we just learned math, we learned math, we learned history, we learned language. And so yeah, so something, so that emergent capability is a pretty different thing than where we were even a couple of years ago as state of the art in terms of AI. Do we need to learn math anymore? Why learn math? Let me tell you, I'm an electrical engineer who never understood the Maxwell's equations. So now I finally get, thanks to ChatGPT, a better understanding of it. So I think one of the things is, we will all enjoy a lot more math because we will have that personalized tutor who will, in fact, be able to intervene at the crucial time when you're making a conceptual mistake and help you learn. So just imagine that, just math. What if there was a fantastic tutor teacher for every student learning math going forward? That is now possible. Okay, I'm gonna break my word because I wanted to ask you one more thing on this then we'll move on, which is short term. I think we have a story that I think is coming today just that in the short term, you guys are making more tools available through Microsoft with the AI. How will we see it affect Microsoft and our engagement with Microsoft in the near term? Just talk a little bit about what your plan is. Yeah, I mean, look, at Microsoft, the one thing that I think is probably worth stating is this didn't just happen. That is chat GPT and GPT family of models. This is something that we've been partnered with open AI deeply now for multiple years. And we built in Azure, our public cloud infrastructure, an AI supercomputer, which by the way, as a systems architecture has been a massive breakthrough because the way these workloads or the way you train large models is very different than anything out there. It's a much more large scale synchronous distributed job. So we had to build the system from storage to compute to network that really allowed us to even build this capability. So I think at this point, the way Microsoft's going to really commercialize all of this is Azure has become the place for anybody and everybody who thinks about AI and large scale training. And we're way ahead on that. And we continue to plan to sort of really step it up there. Second is we want to make these foundational models available as platforms. So that means anybody who wants to build on them in any domain can build on top of that soon. In fact, today we just made the Azure open AI service available generally, and then even chat GPT will be available as an API. And so we're very, very excited about that. And then we will incorporate this in our own application. So you can fully expect us to obviously, every product of Microsoft will have some of the same AI capabilities completely transform the product. Okay, so we'll be seeing it soon. Let me hit on a couple other topics while I've got you. So productivity, you've mentioned a couple of times in relation to AI, but more broadly, you've been talking a lot about productivity in this workplace moment. Microsoft, which we should say has certainly has a good business of remote work and helping facilitate it but you've sort of suggested that from your research, work from home is not necessarily as unproductive as many bosses are saying that it is. But at the same time, you've also wanted some presence of people in the office. What's the balance in your mind now for being in the office, being at home and maximizing productivity, but still having a healthy culture and connectivity? What's right for you, do you think? Yeah, I mean, look, there are three things at least we observe in the data that we are in aggregate that we use to even inform our own set of policies. The first is what you referenced, which is there is this what I would call paranoia around productivity, right? Every boss thinks that somehow they're not being as productive. Every person working thinks that they're burnt out. So there is this dichotomy of at least how people feel. And I think the only way out of that is to have data, right? Not dogma, which is, ultimately, we are all gonna be responsible for the outcomes and if we're managing to the outcomes, then we'll be better off. So that's sort of one side of it. The second thing I'd say is, we also all have to recognize that people come for other people, not for policy. Just because I said that everyone needs to be at work, you know, it's just not gonna matter. But I think we all have to learn some new soft skills as leaders and managers, where convening people, like if I say I have a meeting, or I have an event, and make that event exciting. I think I would say, I say like the new management technique is all of us are event managers. And I think that's, so we'll have to learn many new soft management techniques in order to get people to connect with people, because it's important. So go ahead, finish your question. And then the last thing I'd say is even recruitment, right? Which is one of the things that is fascinating to see, like people wanna feel connected to the company, its culture, and also feel that they're making progress. And so we have to go back, in the last three years, a lot of what we assumed was happening on campus, so to speak, it was implicit. Now you have to make it explicit. Training, the connection with the managers, all of those things are really things where both digital tools and explicit management time would be required. So if I were a senior executive at Microsoft, and you were my direct supervisor, how often would you wanna see my face in person? You know, as often as I wanted to. So, but the thing, that is the key, I think. It's like, for what purpose? In fact, I think we're all increasingly, when you ask somebody, where are you traveling to? People think about, like, what am I traveling for? What is the purpose? If I come into work, what do I wanna do? So I think being a little more deliberate about things that you do, while exercising your flexibility, I think is what's gonna have to happen. And one other point on this for you as a boss, I think one of the complaints you hear, put the productivity thing to the side, is it's harder to know what everybody's doing. It's harder to have that sense, or to measure some of those things. And of course, there are intangibles that you can't quite measure with data, exactly. So do you feel that tension and stress, because you've got a more diffuse organization with people in different locations? The biggest, to your point, I think you're right, in saying that the productivity data, in fact, I see, like, when I'm on a Teams meeting, I actually meet more people than in a physical meeting, because guess what? I can click on their profile. I know who they are. I can go to their LinkedIn profile. I look them up. And more people in chat on Teams are commenting in addition to what somebody's speaking. So in some sense, the multimodal side of it, I miss it sometimes in in-person meetings. That doesn't mean the in-person meeting doesn't have its own advantage, which is I have a side conversation. I walk out with somebody from a meeting. So I think we're all learning how to use both these sets of capabilities and tools to maximize. And I don't think it's one versus the other. So I don't think there's any going back to 2019. But at the same time, it's not like we're gonna be working as if we're in 2020. And I think that that combination is what we're all trying to crack. I'm gonna go out for a couple questions from the crowd in one second. I'm gonna ask you one thing before I do that. I just wanted to get everybody ready, and then I know you're gonna have to run, but I can't not ask you about Activision, the FTC is sued to stop the purchase. You and Brad Smith have expressed optimism about the deal. There was a report from a competitor of ours, Reuters, yesterday that the EU is looking at challenging the deal. Realistically, I know you publicly are expressing optimism about it. What's the timeframe for a make or break point on really pursuing this deal in your mind? And are there conditions that just would make it impossible to pursue? Look, I mean, being, whatever, a number four player, trying to add some content and create more opportunity for more publishers, more gamers to be able to enjoy. I don't think Lita Khan is watching the live stream. I have to warn you. I mean, it's like, if you believe in competition, you should believe in this deal. And I think that, you know, the fact that people who don't even make games are the largest extractors of rents in gaming, shouldn't that tell the regulators where they should be looking? It's unbelievable to me how, you know, short-sighted some of, I would say, the way people look at competition is. And in fact, I hope the competition authorities get focused more on competition and that'll be a good day. Question right here. Hi, Keith Frazee, I run a research institute on the future of hybrid work. It's very frustrating that fewer than 15% of your peers during this amazing remote laboratory that they had for two years actually looked at reinventing how they work and they still debate where they're working, but they haven't used your extraordinary tech stack to actually reinvent the work. What can you do to build a movement that allows all the tools you have to actually get used when reality organizations are merely scratching the surface? I mean, I think it's like, it looks like what you're doing. But quite honestly, I think it's a great point, which is what's the, if you take, go back to something like business process re-engineering in the early 90s, that was very important to have essentially a management doctrine even in order to help people deploy the ERP systems and what have you. So the question is, what is the sophistication of the sort of modern work doctrine, so to speak, that is much more widely spread across all teams? And so that's, I think, what everybody's understanding. And this is where the complexity here is we have refined, as somebody said, the best collaboration tool we've all developed is the workplace for 200 plus years. And so to suddenly say, let's go add to it, it's sort of a tough challenge. And so, but that's the challenge, I think. We have to learn the management techniques, like even how to hold a meeting going forward, where there are two people remote and nine people, say, in the meeting. The expectations of those two people are very different today than they were two years ago. They're not gonna deal with anything second class. And so if you as a leader don't have the skills on how to incorporate them appropriately, you're gonna have a bad meeting. And so I think that that understanding, like, you know, it's not like the thing that you learn when you join, but I think people who are growing up as leaders today and, you know, workers today are going to learn that inherently. And so it's a generational change. That gentleman right there, the white shirt. Thank you, Satya. I'm Philippe Monnier. I represent a Swiss startup called WayRay, specializing in augmented reality. I have a question. What do you think will be the impact of GPT on newspaper, in particular, on the Wall Street Journal? I think they'll be able to write great articles going into the future using GPT. I think we all know that currently no news organization relies on formula writing at all. So I'll take the gentleman back there on the stand. Right, yes, yes, go ahead. Look, I mean, I am neutral to it, right, which is it's a technology that has a set of use cases and we support it today. And even a distributed database is a good thing. It has its uses. I think the entirety of that space, we have not yet found, like in the context, the difference between, I would say, all this Web3 and AI, and quite frankly, even metaverse, is all of these three things are all going to happen, but you need to have the killer apps. What is the use case that gets broad adoption? What's the chat GPT moment on blockchain? When I first sort of got to know the web, it was through the Mosaic browser. I think all of us remember, like I think at least the difficulty I have in some of these other technologies, what's the moment? Like even on the metaverse, that's why I'm very, very keen, even at WEF, the Global Collaboration Village and Mesh, that's the stuff, right? Which is if we can start saying, hey, this is just the next generation of presence, here is how teams will evolve to include more presence, and then we're all using it and it's mainstream, then we'll know that moment has arrived. So I think we have to keep pushing. I believe in all of these technologies and it's up to us on the technology side to make it more real for more people. I think we have to stop it there, Satya. Thank you very much for all of your thoughts today. I think we all feel we're really on the cusp of a lot of changes. So thanks for talking so much this morning. Good to see you.