Transcription
Satya Nadella, thank you so much for doing this. I want to be very, very respectful of your time. The first question is, the internet is abuzz with conversation of Satya Nela said that SAS is dead. Now, I know you know that's not exactly what you said, and the internet has lost nuance over time. But I'd love for you to tell me how that shift happens from SAS all the way to agents.
Yeah, I mean, to me, whenever there's been a real platform shift, the core application architectures have changed, right? I mean, if you look back, let's go all the way back to when the relational database was born, right? That was the first time we really said, "Oh, wow, I can separate out my data tier from my application." Right? Before that, we were building these essentially ISAM databases right into the app. And then we said, "No, no, no, let's have relational algebra, I have SQL, I have data, and then I can build my business logic on top of it." Then, of course, other platforms came like the web and what have you, and then we said, "Oh, what's an N-tier way to write an application? How should I restructure my business logic?"
I think that something of that scale, or if not more, is going to happen again with the application logic, right? And this time around, the thing about agents is they are not going to be bound to essentially anyone's SAS application and its data, right? So I'll have an agentic sort of view where the task, the intent, and I will go operate, uh, and orchestrate, uh, all the logic across multiple SAS applications, right? I'll go call a bunch of APIs through through tools. Use so I'll call a bunch of APIs. I will, in fact, mo, I'll post-train my model to know about multiple SAS applications in the agent tier. And so that's what's going to happen. So I think what'll happen is these CRUD, I mean, SAS applications are a CRUD database with a lot of business logic. So the CRUD database will then get orchestrated outside of the business logic tier of just the SAS application, is what I mean, is going to happen.
Like, right, right now, in my own use case, I go to Copilot, I say, "At sales," which is actually touching Dynamics CRM, brings back whatever the account information, then it brings back information from Office 365. I put it into pages, I share it with people. The entire workflow, right? I never, I mean, everybody talks about their CRM database, but anybody, nobody uses it because, you know, when was the last time I logged into CRM? Never. Except now I'm every day querying my CRM database because it's so much easier because I, it's one agent away, and it's working with all the other agents. So that is what's going to be the change. And when you're hiring people, I think in the future, you'll now hire people plus, you know, their workflows.
That's exactly right. And in fact, that's a way to think about it. It's a swarm of agents, uh, I mean, at a practical level, I kind of look at it and say, it's not like when you hire a data analyst, you don't, you hire them and their spreadsheets. Yeah, that's kind of what it is, right? So it's like agents are going to, I think two years from now, we're going to say, "Yeah, agents, yeah, I build them like all day, like, like I build docs and spreadsheets." And I think that's going to, come with a basket of them. There you go. I have a basket of my agents. And, and, you know, like, get, like I already see that, right? Like, you know, every SharePoint, like, then, like I have a leadership meeting, uh, and a leadership team in which there's like all these documents, the best grounding data is all in there. So I just have a simple agent now, which is a SharePoint agent that I'm always addressing. And, and it's fantastic not to just have to go to a separate entity and, you know, query it, but to have it right there.
A follow-up question here is, how does India stay competitive as this changes? Because you look at when the LLMs first came out, first became popular, uh, we always said that if India can build foundational models and sort of do what the West has done, that'll be an advantage for us. But we're now seeing many of those get commoditized, and we're starting to see that the real moat is the ability to keep coming up with breakthroughs. If you had to pick something that India can do now that is still defensible in the next few years, what would that be?
Yeah, I mean, like, it's not just a, uh, India a common, because it, it's a common for all of us, right? I mean, you know, at the end of the day, um, I think you really want to take what maybe commodity and then do higher value. And then if that higher value thing becomes commodity, you have to be ready to commoditize that and move on to the next, because that's the thing in tech. There is no franchise value, right? So whenever you think, quote unquote, these moats and so on are really overstated, sometimes because you can't really fall in love with your own moat because it'll get attacked. That said, I think India has tremendous opportunity, right? If I think about, uh, the total developer community here, the entrepreneurial energy here, the application space, right? When I think about all this quick commerce work that's happening, that's pretty unique. So whoever builds, you know, AI applied to that here, SAS companies out of here, nearly taking the business model disruption, even right? I mean, I think the next generation SAS company that says, "You know what? I'm going to embrace these agents." Uh, in fact, I'll expose them as first-class agents, right, into Copilot. In fact, change even my business model around it. That's a massive opportunity. In fact, it's a massive attack vector on any existing SAS company, uh, that may have a massive moat, quote unquote, right? So I think that things like that. And by the way, even on the LLM space, right? You know, there's the design space is not narrow. When I talk to customers, that's why we are building out Foundry such that we can even be distribution for people who are building purpose-built models for industries. You're giving them customers.
Yeah, so yeah, we're going to bring them customers. But also, you can build a great LLM or a great foundation model, uh, for different sciences, for different industries, for different roles that are optimized for Cogs, for latency. So I think the design space is large enough, uh, that it may not be just this one model that rules them all, but there will be in every layer of the tech stack opportunities for innovation.
That's fantastic. I want to ask you a more personal question. Let's say if you were 25 years old, you were sitting out of India, let's say you're an engineer by training, and you're seeing all of this happen, you're seeing all these new models come out, you're also seeing the ambiguity that you and everyone else are facing. How would you tackle that? How would you upskill if you were young again?
That's a great one. It's, you know, in fact, the way I think, one has to deal with both the pace and the scale of innovation is to be very good at sampling with agility, is what I describe it like. So that means, for example, the guidance I give ourselves as a company is, keep on the frontier and be ready for the next drop. So experiment. Yeah, not, not experiment like it's kind of like you want to work in multiple gears. So you're all the time sort of looking at what's coming and saying, "What is the impossible thing that I can make possible with what's coming?" While I am then making what I built yesterday optimal, optimized for Cogs, for latency, for deployment. So that two gears, you have to simultaneously work on it. So it's, I mean, experimentation is obviously part of it. But that, you, you can't say, "I'm going to build this and then I'll go to next." You're simultaneously working with some frontier thing where you're making something that is impossible today more possible because of what's coming, while you're optimizing that. I think that's the way, at least as a software developer, you have to sort of work in this age of AI. More so like, even in Moore's Law, like, you know, we used to do that. But this is sort of, you know, when every six months, every three months, when the performance is doubling, that's a, like, even for tech, that's not something we're used to.
Yeah, I mean, I can give you an example of this. I just tried Trellis, which is your text-to-3D model, and I was blown away. I was like, "This is so good." And I was able to try it on my local computer. And I said, "Imagine this in two years." It's sort of like the GPT 3.5 right now, and then in two years, it's going to get good. Are there any other things that you've seen in research level or the demo level where you've seen it and you think it's going to play out, but in five years?
I think the thing that I'm most excited about, or, um, is on what will happen when, you know, this particular model architecture or other model architecture breakthroughs will have on science, right? I mean, if you sort of start saying, "Wow, if, like, even start with chemistry, right? We now have good models for doing novel, um, new materials, right?" I mean, in material science, like when I think about our data center and we say, "Hey, we want to build a more sustainable future," whether it's steel or what, it's all about material science. So having models that impact chemistry, biology, is probably the hardest. That's where I think we're very, very excited about even some of the new models that we built, uh, for doing, you know, molecular dynamics, right? Which is one of the, you know, like, one thing is to, you know, be able to model a protein structure, but to be able to then model the entire dynamic nature, nature of a molecule. That I think is going to be breakthrough in something like drug discovery. So science probably will be the biggest, especially the combination of progress in AI and AI for science and maybe quantum. That to me might be the next big breakthrough, uh, where science itself is computed.
Interesting. I have one last question, right? Which is, you spoke about how these models are getting better every three months, every four months. But I've noticed a problem with legacy businesses. I meet people who say, "I tried this model a year ago, it was me, it had hallucination," and then they never give it another shot. And this seems to be a problem, right? Which is a bunch of people use a specific type of model many years ago or many months ago and then don't try it again. What's the best method? What, what's your advice to them?
Yeah, and that, that, that would be a big mistake, right? So which is, in some sense, you really want to, as I said, that's why you want to be able to find some place where you can deploy something, um, and then it's path dependent, right? All strategy is path dependent. If you don't get started, you will never get shots on goal. Yeah. And like one of my, your friends once says to me, "You can never get fit by watching others go to the gym." So you got to be in the gym every day, lifting weights, uh, if you want to have a shot. So I think the best way to do it is sample the best. You know, figure out how to take the most high ambitious scenario and then start deploying at scale what makes sense for you from an economics and scenario perspective. If you care about hallucination, there are a thousand ways to get rid of hallucinations, like you grounding now. Yeah, that's right. Grounding is one. But quite frankly, if really you don't want to use an LLM, go use just regular machine learning, right? If that is like, if you're so worried about sort of the 99.9 percentile, don't use some of these things where there are errors in, because that's why I think evals are important and really being driven by what's your around for humans or for models, and then using that as the real criteria. But then keep at it.
Interesting. Thank you so much. This was an amazing conversation. I learned a lot. I'm sure people watching learned a lot as well. Thank you, Satya.
Absolutely. Such a pleasure. Thank you.
[Music]