Stanford OnlineStanford CS153 Frontier Systems | Building the Frontier Ecosystem
EVERY SPOKEN WORD
50 min read · 10,176 words- 0:10 – 3:11
Why Microsoft bet big on OpenAI (2019): prepared mind, scaling laws, compute concentration
- MAMichael Abbott
Welcome. Uh, today we have Satya Nadella. Thank you so much for coming.
- SNSatya Nadella
Absolutely. It's my pleasure.
- MAMichael Abbott
And, you know, realizing it's, uh, it's finals week, so it's a-
- SNSatya Nadella
Yeah
- MAMichael Abbott
... uh, a different type, different time of the day and whatnot, and but, you know, I appreciate the students that, that y'all, y'all made it here. Um, you know, I was thinking this morning how it's, it's kind of fitting that you're the last person, uh, that we're having to this class. And I say that because if I think back, like, your bet of putting a billion dollars into OpenAI in 2019 feels like that really set the stage for, like, this ex- Cambrian explosion, if you will, around AI. So I'm curious to just kick it off, like, how, like, what was the thought process to take that, to make that bet?
- SNSatya Nadella
Yeah, I mean, I think it is fascinating to r- look back, whatever, now six, seven, seven, eight years in some sense. Um, I think the, the thing that I, I feel at least got me convinced that this was the right thing to go try at that time because, um, is q- quite frankly, what I describe as a prepared mind, right? But Microsoft's obsession, um, has always been in natural language. Um, and, um, and of course, we were mostly focused on, uh, trying to get to natural language with, you know, some machine learning, some NLP. Uh, but fundamentally, if you had asked us even in 2017, 2018, it'd be some combination of some symbolic logic plus, um, uh, machine learning, right? So we were perhaps at that stage not the truest believers that deep learning can even get you NL- uh, NLP breakthroughs. Uh, but that was something we wanted to have happen. So I would take shots. In fact, most people talk about just the OpenAI bet, but you know, we bought bunch of companies and we invested in a whole lot, uh, of others. Uh, because the fundamental thing we were conditioned to do was anyone who had an ambitious angle, uh, on natural language, it would... Irrespective of what sort of lineage they came from, we would always take it, whether it was in organically inside the company or outside. Um, and that is when, when I, you know, when Sam and... In fact, you know, we were one happy family at that time, right? So Dario was there. [laughs]
- MAMichael Abbott
[laughs] That's what I'm saying, though, like it really did kick off so many things. [laughs]
- SNSatya Nadella
So everyone was in the same place, and so to some degree, um, the, the scaling laws paper came out and, you know, their, their ambition on pushing the transformer with more compute, um, and data, uh, was an appealing thing for us to take a shot at. And of course, what's happened, uh, it's pretty stunning that the fact that the capability graph sort of has just stayed, uh, at that, um, scaling law Pareto is just pretty amazing.
- 3:11 – 4:53
Culture, partnerships, and the ecosystem playbook: why “we’ll build + partner” is in Microsoft’s DNA
- MAMichael Abbott
Mm-hmm. It's... I mean, as someone who worked at Microsoft, I guess, 20 years ago, um, you've changed the culture, like, so much, and it... I think it's, it's striking to me that when you took that bet, I mean, was there an uprising within Microsoft saying, "Hey, like, we can do this ourselves," like?
- SNSatya Nadella
Yeah, I mean, I, I think... You see, Microsoft, uh, over the years, I, I always say that at, at the end of the day, the core bet has to be the organic bet and what one does inside, and then there are partnerships. Uh, there's M&A. Um, and I think any company... Like, in some sense, you grew... When you grow up in Microsoft, you learn that you can create a lot of enterprise value, uh, by building, by partnering, right? If I look back, even, you know, if you take what, you know, the PC revolution wouldn't have been possible but for, uh, what people describe as the Gates Grove model, right? Which is Intel, Microsoft coming together, um, to create essentially what was a PC ecosystem. So I, you know... Or, you know, I worked on SQL Server, and so what we did with SAP to build our database business and for them to build the ERP application, right? So we are conditioned, quite frankly, for these type of ecosystem partnerships, um, as well as organically build. And so I, I would say there was not an uprising in some sense, you know? There would have always been like, "Hey, you know, whenever you allocate your scarce resource of whether it's capital or more..." In this case, it was more than capital. The biggest decision, uh, was about compute concentration on a particular effort, right? I mean, that was more, uh, the big, uh, bet, and that's why, um, you know, we made it because this was the group that wanted to sort of go drive it and, you know, we benefited obviously immensely from it.
- 4:53 – 5:31
Build announcements and the 'Frontier Intelligence Ecosystem': why Microsoft launched seven new models
- MAMichael Abbott
Hmm. And now, um, one of the reasons why he's in town is that there was a big developer conference called Build, and yesterday you announced this Frontier intelligence ecosystem, which kind of right in line what you're saying, and you had a bunch of other pretty huge announcements. Um, and I'd love to talk about some of those. So, um, you launched seven new models, and I thought what was really interesting about at least Mustafa's kind of description of how you did those models is that, you know, all the data is very clean. There was, like, a lot of focus on, um, we'll say not breaching any copyright things. I'm just kind of interested to hear from you why seven? What was kind of the thought behind that?
- SNSatya Nadella
Yeah.
- MAMichael Abbott
Obviously, you want your own models, which makes tons of sense.
- 5:31 – 9:04
Clean data lineage and licensing: enabling company-owned 'hill-climbing machines' (RLE, evals, traces)
- SNSatya Nadella
Yeah. I think it's, um, uh... If you step back, in fact, I think you call this class the Frontiers, uh, class, and so I think that one of the, the challenges of this, uh, um- ... conceptualizing how does anyone, any individual entrepreneur or developer company participate at the frontier, right? There are frontier models, but how does one have real agency, um, to add value, derive value, and protect value, right? Because that's the question. Like, if you have a model that basically learns from data, uh, what's the future of the firm even, right? Which is the firm today is about tacit knowledge inside the company that comes about because of its operations and human capital. And, uh, in a world where there is going to be tokens, uh, and humans collaborating together, uh, what's the future of the firm, right? So there are some substantially big questions. And so what our vision for this is simple, which is we believe a frontier ecosystem is one where every company can actually operate at the frontier with their own IP compounding over time, not just the human capital, but even their token capital. So that is the motivation we have, right? Just like, um, so for example, when the models we built, um, I'll come back to the lineage. There's a nice technical report that I would encourage, uh, folks to go read, uh, because I think it's probably one of the most transparent, good detailed document on the entire pipeline, uh, that's been written lately, uh, from a, a model of this size, and I think you'll learn a lot from it. But the purpose of both let's take our thinking and coding models was to be able to do this in such a way that we can license it, uh, along with the weights in... and, and really allow every company to build their own hill climbing machine, right? So we ourselves and I'll, you know, climbed our hill using, as you said, very clean lineage of data, m- making sure we were not, you know, adding a bunch of synth data in, in the mix. Uh, so all that was very much true so that we could truly have a model, uh, that where reasoning emerged. Um, and, uh, and so we now have a fantastic, good, efficient model. But inside of a hill climbing machine that any company sets up, it can go learn using the traces, uh, of that company and those tasks, right? So our goal is every company starts thinking strategically, uh, about what's the RLE environment that they set up? What is the private evals that they have? How do they then welcome any model, uh, into that gym, so to speak, uh, and then allow them to retain the IP, uh, and not leak value? Uh, so to me, that's kind of what I think every company will need to start doing. Because if you're just a consumer of a foundation model, um, then I'm not sure how you can retain, uh, enterprise value, uh, let alone create, right? So, so the only way I see this ecosystem, quite frankly, m- being non-, uh, zero-sum or positive-sum, where lots of participants can all, uh, be at the frontier, is they're able to take frontier models, take open weight models,
- 9:04 – 11:07
Turning Microsoft 365 into a multi-tenant AI improvement platform: bootstrapped RLE and enterprise ownership
- SNSatya Nadella
take a, a model like ours, which is a licensed IP, uh, and then hill climb on their own environment, uh, and then build out their own IP. So that's the core premise. And in, and we unpacked that in a lot of detail and all the tooling around it. How-- For example, one of the coolest things is if you're a Microsoft 365 customer, we can bootstrap even, right? Because after all, what is Microsoft 365? Today, you use it, uh, to run your business, right? People communicate with other people related to a business process. So you can imagine we can bootstrap the RLE. Uh, in fact, we can even generate the evals, uh, for, let's say, an HR onboarding process based on the observation of what you're doing-
- MAMichael Abbott
It's unique to that
- SNSatya Nadella
... that's unique to the company. And first of all, it's their data. It's just that this... Think of we built a multi-tenant SaaS application. We now can turn that into a multi-tenant hill climbing service, where the data and the environment and the models and the traces and the outcomes are owned by the company.
- MAMichael Abbott
Do you think that most companies, though, have the right talent to be able to build those hill climbing machines?
- SNSatya Nadella
Yeah, it's a great one. So that's why I think this is the easy button on it, right? So we are now not saying you need to build. So you have the hill climbing machine, uh, that has been instantiated for you. All you need is a bit of strategic discipline in making sure that these models, the harness, the context, the evals are all artifacts and constructs that you understand, and you manage them as assets. Uh, just like how you have it historically done, where you cared about privacy, you cared about confidentiality, you cared about security. I think in a world where AI comes into your company, these things will become as important architectural and strategic considerations.
- MAMichael Abbott
And one of the other products you announced was Scout, uh, was like around enterprise class. And I'm kind of interested to hear-
- SNSatya Nadella
Yeah. So-
- MAMichael Abbott
... like the, the vision thinking behind it.
- 11:07 – 12:59
Scout and the next Copilot form factors: from chat → cowork → autopilot agents with enterprise identity
- SNSatya Nadella
Yeah. So one of the things that, uh, we're very excited about is like when I look at Copilot and its evolution, right? It started at, uh, at, with chat, um, and chat became very powerful, especially with reasoning models, because you could not, um, not only, you know, get Uh, just essentially use it more like search, but you can now really do it as... You can use it as a thinking assistant essentially, and so that became powerful. Then Cowork was the next form factor, and Cowork is pretty neat as a way to delegate tasks, right? It's a multi-step, uh, reasoning tool calling agentic loop. Uh, and so therefore you're able to do, uh, long-running or, or sh- a, a short task, uh, assignment. It's very much like what we were doing, uh, with GitHub Copilot, let's say even two years ago when the agent loop started coming, right? So we're now doing it for knowledge work. But now with Scout, you essentially are g- have the third form factor, which is autopilot. So now you have the long-running agent, uh, where, um, it continuously is operating, uh, it's monitoring, it's got a heartbeat, it's got, you know, it's dreaming. All the things that you expect from a claw, um, you can now have. Uh, and you can create it. You can have one that with your identity, right? So I can essentially, if I have a Entra ID, I can give Scout my Entra ID as a delegated ID, and it's sort of essentially my digital, uh, twin that's, uh, working on my behalf continuously. Um, but not that, just that, but we can also allow you to mint more, uh, autopilots. Uh, and those things can have their own identity and their own sandboxes, and so, so it's a pretty complete system. So I think of it as an op- enterprise open claw, um, and a UI that fits in nicely with the rest of the Copilot system.
- 12:59 – 14:44
Securing long-running agents: OpenClaw collaboration, sandboxing, and Windows containment (MXC)
- MAMichael Abbott
And it makes sense because you have those identities. You can really, you can address the security question or what-- I mean, obviously, I mean, I don't know how many of you've, you know, set up OpenClaw, but, like, I kind of struggle to YOLO and, and give all my credentials 'cause I'm like, I-
- SNSatya Nadella
Yeah
- MAMichael Abbott
... I just, I don't trust it. It-
- SNSatya Nadella
Yeah, I mean, I think that the most, uh... Yeah, we even announced, in fact, Peter was on stage with us at, uh, Build as well because one of the other things is we're even working with the OpenClaw Foundation to make sure, uh, that it can be run securely. We will have, in fact, on Windows an out-of-the-box, uh, experience, uh, where you can install OpenClaw and have it secured or contained in what this, this new, um, essentially a container called MXC, uh, which is ess-essentially a way to sandbox the environment, right? And, uh, so I think containment is key, right? Because after all, you now have these long-running agents that are able to generate code and execute code. Um, and so therefore it'll become very important to govern the execution. And so we have a container, uh, that then you can set policy and isolation boundaries, right? It can be process level isolation, session level isolation. You can even have a VM boundary if you wanted. Uh, for me, for example, I run, if anything I wanted to ever run even, uh, I'll just run it on Windows which is my cloud instance, right? So a complete cloud instance, uh, that's fully isolated for long-running agents. So I think we are all gonna learn how to work with many agents. Uh, and we are also gonna learn how to isolate the environments for these agents, just like how we r- you know, back in the day, we thought about processes. We're gonna think about the process boundaries, session boundaries, uh, and container boundaries for agents.
- 14:44 – 17:04
Unmetered intelligence on the edge: NVIDIA-powered PCs, dev boxes, and running trillion-parameter models locally
- MAMichael Abbott
Yeah. One of the other things that you also announced was around bringing, uh, we'll say, AI to consumers. And, um, and I'm kind of curious, like, what does that mean? I mean, there was a lot of big announcements on, uh, with NVIDIA.
- SNSatya Nadella
Yeah, no, there are a couple of things on that, right? One is, um, we're very [chuckles] excited about this concept of unmetered intelligence. Uh, so you know, if you think about it, right, every PC, uh, has, you know, historically the install base had a lot of GPUs. If you count the number of PCs with GPUs, it's pretty substantial, uh, you know, the dGPU install base. So one of the things that we are trying to make sure is that in, in a world where, um, you know, these models are there, there is applications that are being written, that tokens are in short supply, we wanna tap into essentially the edge compute, uh, silicon. Um, and, uh, so and in that context, obviously NVIDIA announced a new SoC, which we are very excited about, their RTX. So we have a Surface laptop which is gonna come out, um, in the fall, which is built on it. In fact, all our OEMs will have fantastic, uh, you know, designs for it. Uh, we also announced a dev box. I mean, think about it. It's gonna have a petaflop of AI compute. It's gonna have 20 CPU cores, 128 gigabytes of memory, unified memory for both the, the CPU and the, uh, the AI compute. Um, and, uh, uh, and it's gonna run something like a, a trillion parameter model locally, right? I mean, think about, uh... And by the way, we and, and Jensen even-- And we also worked with Jensen to get, uh, Windows working on a GB300, so we even have a DGX workstation. So I think of it as a, a, a data center desktop, right? I mean, um, and so we're e- I think that there's gonna be real demand for all this, right? Because people will want, especially when you install something like Scout or Claw or what have you, and I want it to just keep working, um, you know, 24 by seven, uh, and I don't want to get bill- billed for it, the best way to do that is to run it on your laptop or on your desktop. Uh, so we are very excited about just even the rebirth of the existing PC form factor with this new unbelievable functionality brought, uh, because, brought forth because
- 17:04 – 19:32
New agent-era hardware form factors: Project Solara badge and desk companion as ambient agent endpoints
- SNSatya Nadella
of both the silicon innovation and the model capabilities that now we can have locally. So that was sort of, uh, a lot of what we talked about. But the other thing, uh, that we also said is- Uh, just as there's new functionality in the old form factors, uh, I think there's a real opportunity to create new form factors for the agent era. Uh, so that's where Project Solara comes in. And, uh, what our goal there is to say, you know, we showed two, uh, reference designs. One is a badge, uh, and the other one was, um, a desk, uh, companion, if you will. Um, but the b-the badge is pretty interesting, right? So you can imagine an agent that has a fingerprinter read- fingerprint reader and, uh, a badge that has a fingerprint reader as well as a camera, uh, and has enough onboard compute, it's a MediaTek, uh, processor, um, to be able to wake up something like Copilot. Uh, and I can literally get notified. I can, you know, in fact, give it-- Like, I can even give it, say, a, a coding task or whatever. Uh, I can dictate to it. It will take the input, uh, and then go execute it, uh, in the cloud, notify me back. Uh, you can imagine in healthcare, a nurse. If I was a nurse, I was moving, uh, station to station, I could use that to badge in the data, uh, right? Versus the phone. Like, right now we are conditioned, either we are entering in the PC or we are, you know, using the phone. And in an agent era where you really have ambient intelligence and ubiquitous computing, uh, you can imagine these form factors now, uh, that are just endpoints for long-running agents, uh, that wake up, notify, um, and help you get both output, input, uh, that's right there in the real world. And so we're very excited about sort of bringing even a platform for it, right? So we will build some, but the goal here is also to have even, by the way, new platform rules, right? So Windows has always been, it's kind of, uh, you know, fascinating that we are the only open platform out there, right? There is no... You, you can go through our app store or not. You can install, uh, anything on Windows, right? It's always had that ethos of being, uh, not something that only Micros- You need, you don't need to call Microsoft to build applications for Windows, right? How about that? So that's the, uh, the openness we want even in this new agent platform, so that we don't have the carryover of these platform rules that were written, um, uh, for the previous era.
- 19:32 – 22:56
Making AI feel like 'light,' not 'electricity': value, jobs, healthcare, and social permission
- MAMichael Abbott
I'm gonna switch gears a little bit. So here we're at Stanford University, um, probably, you know, the center of the world in terms of AI pil-pilled people. And, um, when you get outside of the, you know, the Bay Area, Seattle, um, you know, people are looking at AI and saying, like, "What, what's good for me?" And I think there was a prior speaker that used a metaphor which I found quite powerful, which was, as electricity came about, we didn't sell electricity, we sold light.
- SNSatya Nadella
Mm-hmm.
- MAMichael Abbott
And what do you think is that equivalent for AI? Because right now there's not a lot of good messaging around AI of how it's gonna benefit people.
- SNSatya Nadella
Yeah, I think that, I think that's right in the sense that we have, um, perhaps gone too into the, uh, you know, it-- the bubble that I guess we all live in, um, is more about hyping the tech and the tech progress for its sake. Um, and it, you know, we live in it, and it's great to be impressed by it and push the frontiers of it and what have you. But at the end of the day, the world will evaluate us in what was the value we created for the world, one community at a time. Uh, I mean, that's kind of how it sh- it should always be the case. And so unless I can see the true benefits, uh, of this technology be broad spread, uh, right? Health-- We talked about healthcare. Uh, when we suddenly start seeing AI in healthcare change the cost equation, the care, uh, one can get, uh, not in an abstract, uh, you know, sense, but when it happens to someone in our community, in our family, uh, when... Or even take economic opportunity, right? Talking about this as something where, um, it takes away jobs, uh, it's w- it's clear that any im- technology that's disruptive will have, you know, real displacement. Uh, but at the same time, there is going to be new economic activity where humans will have agency, which will have wages, which in fact, if you think about it, right, if what is current intelligence gets commoditized, humans are the one species that are most adaptive in the sense of creating new value on top of what's the new commodity. Uh, and so the way I look... And, and, and that has to not be abstract, but it has to be real. Um, and it will happen. Uh, but until that transition happens, to your point, as we go from electricity to light, and the light is not seen only by s- the AGI-pilled people in the zip code, uh, but it's seen by the world, uh, as something that they can thrive in. And even my point about that frontier ecosystem, right? That it-- When every company is not sitting there thinking that, "Oh my God, I'm just going to be... You know, if I let any one of these frontier models into my organization, it's just gonna run over my, you know, f- all the IP I've created," that's not a... Why would they welcome that? They, you know, by definition, they should not. And so, uh, I think that's why as entrepreneurs, as students, uh, I think we have to shape... A-and, and as incumbents, uh, we have to shape this to an ecosystem which is positive sum by definition. If we are not, and it's about like a, a few firms that have all the returns and everybody else, it's all, you know, um, uh, in bad shape, that's just... You, you'll, you by def- you will absolutely lose social permission or we will lose social permission.
- MAMichael Abbott
Yeah. Okay, I'm gonna switch over to qu-questions.
- 22:56 – 27:54
Custom silicon and cloud architecture: training vs inference vs agent workloads, Maia/Cobalt, heterogeneous fleets
- SPSpeaker
So I'm generally curious about your custom silicon program and how you've learned from other hyper tailors like Google and, um, Amazon who have kind of made some progress there. Um, it seems like at the hardware side, right, uh, you know, their chips that they offer are actually pretty bifurcated. Like they have like training chips and inference chips versus NVIDIA and AMD have kind of kept a unified chip. On the networking side, we had an invited class talk about like the optical MEMS, uh, system that they built. That was really interesting. And at the software side, it seems like they've built their own kind of versions of CUDA with Neuron and XLA respectively. Um, whereas you guys are kind of building on something, uh, based off of Trident from my understanding. So I'm curious, like given these sort of different design decisions that these guys have made, what have you learned and where are you taking, uh, your own custom?
- SNSatya Nadella
Yeah. I think, uh, I think the, the key thing is to sort of recognize, um, what are the new workloads, right? Whenever you think about any new, um, uh, sy- system, you want to be motivated by what's the new software or what's the new workload. And the good, the good news here is that there are these three dominant new workloads, right? There's the training workload, there's the inference workload, and now we can sort of say there is the long-running agent, uh, that uses inference, um, uh, and regular compute. So if you sort of said that's kind of what, uh, you have, then you can start from a first principles, uh, like looking at... And these are interesting type of workloads, right? They're not like the previous scale-out workloads. These are synchronous data parallel workloads, um, where you need to, to your point, to Amin's I guess point, which is you got to even think about the scale-up part. Uh, some of the tri- tricks that worked for us for scale-out in the past won't work, so therefore you now need to even innovate on the scale-up and the scale-out to really keep things coherent, um, and, uh, you know, uh, and the MFUs on a training run are maximized and so on. So therefore, the way we come at this and say, okay, uh, even just last yesterday we announced there's a n- there's Maia 200 that's, uh, essentially being co-designed with our own, uh, uh, models, plus the OpenAI, uh, models because we have that IP. And right now, in fact, Maia 200 is running, uh, GPT 55, uh, you know, in multiple data centers powering Copilot, right? So the-- And giving us total TCO advantage. So that's a great way to round trip, uh, for an inference workload, what's the advantage of that is. Uh, we not only did that, but we also for, um, uh, uh, uh, built Cobalt, uh, which is our ARM, uh, uh, processor, uh, for compute, and we are benchmarking it to improve both for latency performance, uh, when it comes to, you know, for example, the agentic loop, right? Because the place where you need great cores, uh, are for these agent loops. Uh, so we're using all the GitHub Copilot traces to optimize our ARM processor even, and bringing all this together with even the networking stack. Um, and so our approach would be to no-- And at the same time, we love, uh, to have the GPUs because they're general purpose to your point, right? Which is in fact we are using the GPUs. In fact, we are using the old GPUs in our fleet to accelerate, um, our, uh, data warehouse. Uh, so Fabric, uh, is seeing 7x plus performance gains, uh, because of GPU acceleration. So we think of our fleet as a heterogeneous fleet where we will use software to get the maximum benefit out of it and do smart workload placement. Um, at the same time optimize for the high volume workloads like inference and training and agent loop with our own ground-up system. And there's lots of design points, right? Most people get fixated on the AI accelerator, but the AI accelerator is one, the CPU is one, the network, uh, sma- you know, accelerator, the storage accelerator, the AI WAN is another one. Uh, right, you want to be able to sort of really do multi data center, uh, hops even. So lots of stuff. Uh, it's a great time, by the way, f- to be in computer architecture. Um, you know, I think, you know, when I started in, you know, in the industries when the Patterson book first came out and that was the RISC versus CISC debate, I feel like we are back at, uh, sort of a time like this where, uh, you can really rethink, uh, from the physical design of a data center, uh, to, uh, to... By the way, the electrons, I mean, one of the places where I'm very excited about is the, the efficiency with which we can bring the electrons all the way, uh, to the CPU, uh, so that the tokens are, uh, that much more efficient, right? Without all the losses in between. So I think there's just a tremendous design space for computer architecture.
- 27:54 – 32:46
Quantum at Microsoft: near-term traces for science models and long-term fault-tolerant Majorana roadmap
- SPSpeaker
And I know there-- another announcement you had yesterday was around quantum.
- SNSatya Nadella
Yeah.
- SPSpeaker
Which is kind of adjacent to what we're talking about. So I'm kind of curious, you know, what was the announcement and what was the, the kind of recent advances there?
- SNSatya Nadella
Yeah. So I-- So look, this quantum, you know, uh, we have been on at this for now the last twenty plus years and, um, uh, it's sort of really exciting to see the progress. I'll just say one of the things is, uh, even independent of the quantum program, um, even with what we were able to achieve in the last couple of years, uh, with even the natural atom-based, uh, quantum computers with our stack, we worked with partners on it. We're able to generate now these very good traces, um, and which those traces, like for... Basically, if you think about what's the purpose of a quantum computer, uh, a quantum computer can simulate nature, right? I mean, it's, since the w- the w- the n- nature is quantum. And so therefore if it's, uh... So instead of, you know, relying on DFTs or what have you, you can now have a lot better fidelity, um, of, uh, say, uh, chemistry or, um, or molecular dynamics or what have you. And those traces then can be taken back and you can train a model. In fact, we, uh, are doing that with our material science models, uh, where you can take the traces from even a, whatever, an early stage quantum computer, uh, to improve the data on which you train a model, uh, for something like, uh, material science or chemistry. Um, now our quantum program itself is, as I said, there's a software side to it Which we will put on ion trap machines, which we're putting with partners. We're putting it on a photonics-based machine. Uh, we're also putting it on natural atoms. We have a partnership with, uh, you know, in Denmark called Q.North, where we will even have, uh, a quantum computer powered by atom computing with our stack, uh, within the year and so on. So that's sort of one side of it. The second side is ultimately, in order to build a quantum computer at scale, at utility scale, um, you need fault tolerance. Um, our bet on that has been that, uh, we ha-- there was a, a theoretical physicist, um, uh, who theorized the, uh, essentially, um, a, a state of matter, uh, called, uh, Majorana, uh, in the 1930s. Um, and so one of the things that we felt that that was the state of matter that we needed to make, e-e-essentially fabricate and make real. So we launched our first QPU, which was Majorana 1, a year ago, which essentially proved out the fundamental physics breakthrough, uh, that you can actually have this and then, um, instantiate it. And now we've got Majorana 2, which allows this, uh, to s- you know, be built at industrial scale. Uh, and so there's a lot of detail in terms of, you know, how long these qubits can be stable for, you know. Uh, and by the way, one of the other things is we have perfected the digital control of this quantum computer because that's going to be super important. Um, so overall, we feel that the quantum program at Microsoft's progressing on two dimension. One is in the near term, uh, with even what are the quantum computers that I think are most easy to fabricate and build today with these things like natural atoms. Uh, and then in the long run, we want to build out, uh, what we think is, uh, you know, the, what is needed in order for true, uh, quantum computers to act, uh, like, uh, utility scale qua- computers.
- MAMichael Abbott
On, on that latter, if you had to guess a timeline.
- SNSatya Nadella
You know, you know, I'm the third CEO at Microsoft, uh, to keep going on the quantum journey. Um, I would say that the-- what I'm now a lot more bullish is it may not... It, it's kind of like the AI. It's kind of like the previous discussion. I think of quantum as the new accelerator. Uh, and, and remember, by the way, quantum is not going to replace classical, right? Quantum is not going to be great at storage and memory and so on. It's going to be great at computation. And so you kind of have to marry classical plus quantum in order to do things. And so therefore, I think of this as, uh, maybe a lot more staged even. So if you have a hundred logical qubits with good error correction, we can start using it to generate synth data for, uh, science models. Like, that'll be a pretty important milestone. That may be even more achievable in the short run. Um, so we'll see. But I'm like, you know, we, we have, you know, we, we, I think, made the claim even yesterday that by end of the decade, uh, we believe we will be able to build a quantum computer that starts solving some real, uh, challenges.
- MAMichael Abbott
Real problems.
- SNSatya Nadella
Yeah.
- MAMichael Abbott
Amazing. Next question.
- 32:46 – 41:59
Talent pipelines and culture: Microsoft’s rotating programs, growth mindset, and empathy practices
- SPSpeaker
Thanks so much for coming. Um, I spent twenty years at Microsoft. Um, I joined the year after you joined as a te- a-after you and became a CEO, um, as a Mac employee. So I went through cloud transformation, went through pandemic, went through AI transformation, and I think, um, like Mac program was such a, such an amazing experience for me and it shaped me as a person. Um, how do you think that it contributed, um, and is it still contributing to the culture and, uh, and success of Mac?
- SNSatya Nadella
Yeah, I mean, first of all, we're very thrilled about obviously students coming in and joining and having essentially the Mac program. There are a couple of programs like that at Microsoft we created where people can even rotate through, uh, various functions. Um, you know, at the end of the day, any company, for it to be at the frontier, so to speak, has to be able to get people coming with fresh ideas, fresh energy, and sh- reshaping. I always say to anyone joining Microsoft, of course, you want to come in and learn about how Microsoft works, but we also want Microsoft to learn from you. Um, and more importantly, for you to have the agency to reshape, uh, what is Microsoft's culture. It's not a static thing. Um, it's a, it's an organic thing that gets shaped by the behaviors, the decisions, uh, of people at the company. And so we always would, uh, welcome students coming in, uh, building their career, uh, at Microsoft. Uh, people, you know, one of the things as, as a fifty-year-old company, I mean, Mike's an alum at Microsoft and still engaged with us. Uh, we have people come, had a tour of duty, gone out, come back. And so at this point, um, I think, you know, I think what, what-- the way to think about it is, uh, the uniqueness of the Microsoft is our core DNA has remained. Uh, right. We are a developer tools, platform, knowledge worker tools company. That's kind of what we've done for fifty years. But the interesting thing about is-us is that we have been able to reinterpret that with every new platform, right? In fact, I joined the company back in the '90s when my existential competition was Novell. Uh, and now, you know, it's some foundation lab, right? I'd not even heard of five years ago. Uh, but that is the thing that I think keeps us vibrant, which is our existential challenge or what we need to compete with, uh, is new and fresh versus it's the same old. And I think that that's sort of an attractive part, right? When-- and you come to Microsoft, um, you will be able to sort of Go at that mission of being able to empower people and organizations all over the planet, which means a lot, um, uh, to us. Uh, but to do so, uh, you know, recognizing that we as a company can bring a lot to that mission.
- MAMichael Abbott
Well, I have a follow-up question. So one of the many attributes I've admired about you is you have a growth mindset, and you really look at, like, your leadership team and, and drive it. How, how have you instilled that across the company? 'Cause you clearly have. Like, you just pointed that out, that you've been able to deal with these platform transformations.
- SNSatya Nadella
Yeah, I mean, at some level, the... I think it's not, uh, something you instill per se, Mike. It's sort of you invoke what is innate in all of us, I think. Uh, uh, I think the... I mean, s- and by, uh, you have to do it more out of practice, right? It's, it's not like, uh, you know, mostly what I have to exhibit more than anything else is my, um, ability to confront my fixed mindset, right? Because at the end of the day, all of us, it's easy to talk about growth mindset, but it's very difficult to exercise it individually, right? As somebody said to me, which I have always liked as a sort of a nice quip, is, "Everybody likes, um, uh, to change. Oh, everybody likes change, except they want the other person to change, not themselves," right? Um, and that, I think, is the challenge of growth mindset. So it's not about growth, talking about growth mindset. It's about having the courage to confront one's own fixed mindset. So it can't become corporate dogma, to your point, right? Which is, so one of the keys, the reason why it's worked at Microsoft is we never made it like, oh, you know, some new-
- MAMichael Abbott
Some mandate
- SNSatya Nadella
... yeah, it's not a mandate.
- MAMichael Abbott
But it, it started with you.
- SNSatya Nadella
Yeah, it is not. And, and also, it's like, it does, it's not like trademarked Microsoft, right? I mean, if you exercise growth mindset or you confront a fixed mindset, you'll be a, a better human being first. You'll be a better colleague, a better friend, a better neighbor, a better parent, a better student, everything. So you're not even doing this for Microsoft's sake. You're doing this for yourself. And I think le- giving that oxygen, leaving that at that, as opposed to some new corporate thing, uh, has been very, very helpful. So I'm an advocate of it, uh, not just at Microsoft, anywhere. Uh, and more importantly, it's sort of that practice of, uh... There are two things that I feel as, uh, that were pretty influential for me, which I learned through sort of my wife's, uh, you know, readings, quite frankly. One was this thing, uh, called non-violent communications, which is also another form, uh, of un- having, developing a sense of empathy, understanding where the other person is coming from, not having your amygdala always triggered [laughs] and what have you. Um, so that's sort of one which I think is, uh, it's a great read if you've not read it. Uh, and then of course Carol Dweck's work, um, on, uh, growth mindset. These are two things that, you know, are, I think, relevant for children and students and child psychology. Uh, but I think they apply to corporate, uh, cultures. Uh, because I think one of the fascinating things is what Herbert Simon described as the bounded rationality, right? Um, I think humans are great, uh, but we have this unfortunate, um, um, uh, you know, um, you know, we don't see what's in our interest all the time. Uh, we get hijacked, uh, often, uh, without sort of being able to do, uh, the c- the simple calculus of what's the, to what does it mean to be at the frontier of our own behavior. Um, and I think that these are nice practices that gets us and pushes us. So as, think of it as your training run [laughs] that you need.
- MAMichael Abbott
That's great guidance.
- SPSpeaker
Sir, uh, thank you so much for the answer. And, uh, I have my dissertation started 10, 10, 10 minutes ago. But I really wanted to meet you. Thank you so much.
- SNSatya Nadella
All right.
- MAMichael Abbott
Thank you. And good luck. [laughs] I realize it's finals, realize it's finals week. Um, next question.
- SPSpeaker
Hi, Satya. Thanks so much for coming. I was wondering, like, how do you become such a good public speaker, and what do you-
- SNSatya Nadella
[laughs]
- MAMichael Abbott
[laughs]
- SNSatya Nadella
I don't know, man. I mean, I, uh-
- MAMichael Abbott
[laughs]
- SNSatya Nadella
... I'm glad you think, uh, I'm a good public speaker. Let's leave it at that.
- MAMichael Abbott
[laughs]
- SNSatya Nadella
You know, I, I mean, look, I think in, like anything else, um, uh, it's not that I, I, I think about public speaking as sort of a key thing that I'm trying to develop or what have you. Uh, but, uh, uh, uh, the, the, the lucky part that I find myself is, uh, um, in particular even, I think one of the things is when I became CEO, um, you kind of had to talk about things perhaps that you didn't get the opportunity to talk about previously. Maybe that's a better way to characterize it. But the good news is, it's not as if I was not thinking about those things previously. Uh, and that, you know, I reflected on it, is that why is it that I was thinking about those things previously? Uh, and I think that comes out of just natural interest, uh, right? For example, uh, thinking about technology, but its impact. Uh, what does it mean? Uh, you know, and I have, like, many pet sort of passions. Like, what is any technological progress, uh, like, uh, AI mean to the Global South? Uh, what does it mean to even, uh, what has been a dream... And I grew up as a son of a development economist, so he instilled in me that, hey, this convergence growth is gonna happen, and it's gonna be great, and so on. And so I'm, like, obsessed about it. Uh, and so as long as I think you have these passions, uh, that allow you to think broadly, um, and then for you to be able to talk freely about it, write is the other one. Um, and so, uh, I'm not particularly an expert on public speaking. Uh, but I think the more any of us, uh, can have broad interests that we can articulate... And in today's day and age, the media allows us, um, to be able to have our own, uh, outlets. And so I think that this is a great time to both build that interest and then to be able to have different medium, whether it's speaking, whether it's writing, whether it's podcasts, what have you. Uh, there are a variety of ways I think we can express, reach, debate, uh, which I think is fantastic.
- SPSpeaker
Thanks, sir.
- 41:59 – 57:12
Learning and building in the agent era: student advice, cognitive coverage, future UI, and open vs licensed models
- SPSpeaker
Hi, thank you for coming. I was wondering, if your undergraduate self was sitting in the audience right now, what advice would you give him? And like knowing what you know now, I guess, and what would you tell him like to put his energy into and to like maybe avoid?
- SNSatya Nadella
Yeah, I mean, see, it's a great question. I mean, it's such a, you know, such a privilege in some sense. I wish I could, um, right? Uh, because everything is in front of you, you're risk on 100% of the time. Um, um, maybe that's what it is, right? Which is, um, the, uh, let me just say two interesting things. Last, yeah, last, yesterday on Hacker News, I came across, I forget, one of the CS classes here, it had the guidelines on how to use coding agents, um, which I thought was well done, right? They had the do's and don'ts and, um, and, uh... And so the, the fascinating thing I find right now is the ability to learn, uh, new things has become so much easier, right? Uh, because you have this very accessible, personalized tutor that's deep, uh, that you can go, uh, and work with. And so I would say more so than any assignment anxiety or, I don't know, grade anxiety or what have you, um, you can have, one of the terms one of my colleagues uses, um, real cognitive coverage, right? Uh, like test coverage. Um, you can now have cognitive coverage that really follow you through your curiosity, right? If, if I were back as an undergrad, I would be trying to-- Like, it's kind of like what I do with GitHub sessions today, right? GitHub app, which is what are all the coding agents and what are all they doing, right? So that's what I would do. I would be sitting with what are the ten hundred agents and me at Stanford learning. But I need to have cognitive coverage. It's not I've offloaded to the hundred, right? The, the key is what am I instructing them? And then when they get something done, can I understand what they did, uh, in order to have learned, right? It's kind of like a, like it's hundred classes. Uh, so it's a fascinating-- And that I think is what I think will happen. One of these days, somebody's gonna break a new, uh, pedagogy that goes with, um... Like the, the tools, for example, are evolving, right? Like think about what happened in developer tools. We went from saying, "Hey, we have like hundred CLIs," to now we need a thing that to manage our CLIs com- CLI complexity, which is the new ADE, which is kind of like, for example, the GitHub app is fantastic in that context because it's like the new inbox for managing my sessions, right? What's the moral equivalent of that, that allows a student to navigate through their learning experience, um, and be max curious, uh, but really getting deeper faster, uh, on things that, uh, you're trying to cover? Uh, and I think that that's what I would sort of do and not have anxiety, right? Because, you know, you can always push a button to get an assignment done, right? So that's no longer the case, and the grades may or may not matter, right? Um, and so therefore, there's a lot of relitigation on the things we valued.
- MAMichael Abbott
That's a really good answer. Next question.
- SPSpeaker
Hey, thank you so much for coming. So, like, whenever we're interacting with a computer, pretty much we always interact with a GUI interface. But, uh, agents, uh, they, they just so happen to be good at coding, not so happen to be not that good at, uh, interacting with GUI interfaces. So for example, if I want to design a poster and I want an agent to do it, uh, it is easier for me to get it to generate like HTML, CSS code rather than to use like a GUI design product. Uh, in that case, like what do you think, uh, are like implication of like agent abilities, uh, for like, uh, uh, GUI interfaces versus a CLI?
- SNSatya Nadella
Yeah, I mean, I think you're bringing up a couple of different things. One is, uh, it is essentially code gen is powerful and therefore HTML and, um, web UI, uh, as an artifact creation process, uh, I think is going to really, you know, proliferate, right? So basically we've gone from, you know, we, we in fact always, uh, Bill used to have this thing where what's the difference between, you know, building an app, writing a document or creating, um, a website. You know, uh, at this point there's none. You can just sort of basically do all three by using code. Uh, so that's sort of one side of it. Um, uh, but I think that, you know, the direct manipulation is the challenge, right? But, but I think of in, in the intermediate timeframe, what's gonna happen is you're gonna have an intermediate format, right? So you're going to do the HTML and then you can convert into Excel, PowerPoint, PowerPoint into intermediate format, and then have agents. So I think that that's what you see in Copilot and elsewhere, um, when you think about artifact creation. But the ultimate thing is can you truly, uh, teach, uh, even the agent on the model, uh, the canvas and the direct manipulation of the canvas, uh, which has to be done fundamentally by teaching it, um, the semantics of that canvas. Um, and so it has to be exposed, whether it's through APIs, whether it is through, um, uh, a protocol or what have you. And so therefore I think you will see innovation like that. But it is true that direct manipula- By the way, you, you talked about one of the other things that struck me is one of the nice little features we added to GitHub yesterday was, uh, a thing called Canvas Uh, and the reason was not because, um, we wanted to, uh, the agents need UI, but we need UI, right? Because it's now become too dense to just keep tracking my CLI session or the chat session, uh, because it's, you know, it's kind of-- first of all, it's linear and, uh, it's painful, like you're trying to sort of scroll through it. And so one of the things that we, uh, said is now we can... For example, I can have a Kanban board as a visualization, uh, which both the agent and I are working on, uh, right? So I think that this idea of generated UI, uh, becoming the new way for human-agent interaction, uh, might be one of the coolest things that'll happen across all product lines.
- MAMichael Abbott
Next question.
- SPSpeaker
Oh, hi, Satya. Thank you so much for coming. Um, if you were at our age, like college freshman, and you have like the world at your fingertips, what's a problem that you would, um, encourage us to try to attack?
- SNSatya Nadella
Oh, man, I wish, I wish I was that. But, um, um, I don't know. I mean, I-i-it's sort of always interesting to look back and say, "What would I pick?" Right? Um, I don't know. I mean, I, you know, in, in an interesting, a world like ours right now, uh, they-- I think you have to go and say, what is the thing that you have inherent interest in, uh, and the world n- will value, right? I think always whenever people are making choices, I think they have two things they're trying to intersect, right? They're trying to intersect something that they believe, uh, they have a real passion for, but they're also doing, quite frankly, the calculus on what does the wa- world value, uh, right? They always have some destination in mind. "I want that career. I want that job. I want to start that company," or what have you. And so I would sort of focus on that. Like answering those two questions, uh, is what I think will lead. And in my case, I would probably go step back and look at that, right? You know, it may, may be even outside, quite fr-- like, right, we, you know... One of the things I-- if I go back to the computer industry, I was an electrical engineer, and sort of I then drifted into software. Uh, but if I went back now, I may go back into hardware, um, you know, just because, uh, there is just such an unbelievable time, um, in, uh... Like there are a lot of things that, uh, you know, I would love to go deep on understanding what sort of, uh, the optical side of networking would look like, uh, some of the systems of that. So I think that's kind of how I think things will pan out. People will pick, uh, the thing that they're good at, then they see the trajectory of what they're good at and say, "Wow, I'm gonna bet on myself, uh, to get good at this and start something." Uh, and, um, and, oh, b- or policy, right? Uh, um, and, you know, people talk about safety engineering, but I kind of was thinking, wow, there are so many aspects of, um, what does it mean to have safety around AI, uh, that require people to think through deeply. Um, and so anyway, so there are lots of choices out there.
- MAMichael Abbott
Along those lines, have you guys hired like philosophers to help you with the d- AI kind of guidance?
- SNSatya Nadella
I think Mustafa is a philosopher dropout or something.
- MAMichael Abbott
Okay.
- SNSatya Nadella
Yeah. So he's-
- MAMichael Abbott
That's so good. [laughs]
- SNSatya Nadella
At least, uh, he's... So we have a quasi-
- MAMichael Abbott
[laughs]
- SNSatya Nadella
... wannabe philosopher. Um, um, but yeah, but he thinks ve- you know, clearly, uh, uh, from that sta-- I mean, he's been obviously since, uh, being a founder of DeepMind to now, he's always thought about it, um, you know, uh... And we've always had folks in MSR who have brought a real deep multidisciplinary approach to it, whether it's the economists, the moral philosophers, um, the sociologists. Um, uh, and I think, I think we'll always have that.
- MAMichael Abbott
Mm-hmm. Next question.
- SPSpeaker
Hi, Satya. I was curious, what do you think about space data centers? Because I've had a lot of startup, uh, you know, CEOs that come here and talk and, uh, try to convince me that we should be looking at that. And the research I'm doing is showing that like Elon's probably the only one that's gonna be able to, you know, do it profitably just from the launch costs, you know, per kilo. I was just curious what your-
- SNSatya Nadella
I'm, I'm, I'm not an expert to sort of on, um, any of that in both the supply side of it and the economic side of it. Um, but it's possible, at least what I've read and what I've talked to who are people who are experts in it. Seems like, um, you know, it makes sense. Um, the question really is you now need to not only solve both how do you get there, uh, but how do you build the stack, uh, that operates there, and then solve all the practical issues of RMA and others, uh, right? Because, uh, therefore, I think there's a whole supply chain, uh, right? When you think about the data center, most people, you know, it's a complex project, uh, right? We have built on the shoulders of unbelievable engineering depth, right? Starting from civil engineering, uh, on, and to electrical engineering, to mechanical engineering, to then ultimately have it meet, um, uh, the needs of computing. Um, and so that level of sophistication, uh, for this new payload in space has to get built, and it could get built. Um, and as far as I'm concerned from a Microsoft standpoint, I would love to. I mean, uh, uh, I think we have a few, right, instances where we have already had some, you know, Azure SKUs. We had a program where we even, um, put Azure SKUs in space and what have you. But they were more like Edge. Um, and so to the degree to which, you know, um, if somebody says to me that there's a s- uh, there's, um, you know, gigawatts available or even megawatts available, I'm happy to plug myself in
- SPSpeaker
Yeah. So, um, it seems that Meta is now pulling back from building frontier open models, and I understand that Google and Google are doing some work with open models like Flash and the Phi class and stuff, but they tend to be, like, pretty small, like, you know, for, for, like, running locally. Um, given how you've personally embraced open source, the biggest example being Linux, I'm curious, do you see Microsoft and AI, OpenAI building frontier open models, like, at the parameter count, you know, like-
- SNSatya Nadella
Yeah, I think the thing that, uh, we're focused on... We definitely will always have open weight models and t-to your point, uh, they will be more for what we will ship. In fact, we launched two, um, even yesterday, both, um, an instruct model and a plan model for local agent loop and what have you. So they are derivatives of what we have done with Phi Silica before, uh, called Ion Instruct and Ion Plan, and they'll run on Windows, and they'll be open weight. Um, the, uh, thing that we are focused on with the MAI lineage of models is again, think of them as licensed, but we are going to license them pretty broadly, right? So for example, you can go to Base10, you can go to Fireworks, and you can then fine-tune even using their inference stack, um, and so on. So we are-- But the reason why we are doing that is because we want, quite frankly, every company, whether it's a SaaS company, an AI native or an enterprise company, to have their own, uh, model, uh, that they can then, uh, post-train, they, they can RL, uh, and what have you, right? So therefore, that'll be our goal, um, is to build an ecosystem around, uh, the MAI lineage of models. Um, um, and so-- And, and the reason why we want to ensure, uh, they're still licensed is because at this point, you know, there's going to be real need for inspection, safety, uh, you know, so there's... At the, the accounts at which, uh, uh, we are, and even if you look at the Chinese models, they're also quickly becoming closed source, and so on. So I think there, I think there will be, and, and I know Jensen's working on, uh, some open weight models and we're, we're definitely supporters of it. But, uh, we want to make sure, uh, that we are all leading... The ultimate goal here is to have everyone have real agency in being able to take some model and to be able to then add to it, uh, and then protect it from having it sort of go back.
- SPSpeaker
So you think that MAI is open? Like what-
- SNSatya Nadella
It won't be open. It'll be licensed.
- SPSpeaker
What, what does that-
- SNSatya Nadella
So we license the weights.
- SPSpeaker
Okay. Okay.
- SNSatya Nadella
Hmm.
- SPSpeaker
And, and then how does... If someone's using it on Fireworks or Together or something like that, how does that help y'all at, at Microsoft?
- SNSatya Nadella
You know, it, it will be licensed, and so therefore, we will have an economic model, uh, in all of this. Yeah.
- MAMichael Abbott
Mm-hmm. Well, I think we're out of time, but thank you so much for coming.
Episode duration: 57:17
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode d0Pfu6B7gIM
