# Day 54 — YouTube 自动生成字幕（AUTO_GENERATED）

覆盖训练段 000–15 分钟。已去滚动字幕重叠。ASR 生成可能含识别错误。

==================================================

[Music] [Music] It's great to be here with you Satya. Thank you so much. We did this a couple of years back. That is right and this has become our annual thing and I'm really glad to be back in Bangalore with you and uh it's wonderful. You know, one thing I was thinking, you know, the last two years there's been an exponential rate of change and you know, we're all every day we get up and something new is happening. How do you deal with that and how

do you deal with what do you think is coming now? Yeah, I mean that's a I think think the this entire industry in fact backstage we were talking about this um the last whatever 35 years that I've been in this industry um the thing that has powered us is this Moore's law. Yeah. Right. I mean it's just been the most unbelievable um empirical observation. Right. We call it a law but it's really uh an empirical observation that is held true for a long period of time.

Um and I remember distinctly like we used to have these um um you know he builds to get the top Microsoft folks every year and he would do only one thing. He would just put a chart. He would just show us Moors law. Uh he would say here's what's happening with memory. Uh go fill it with software. That was sort of the instruction right that was basically what we've been doing all these decades. And then comes this AI revolution and it fundamentally accelerates because we were

all bemooning the fact that oh my god Moors law is maybe ending or it's not like the same as before. Uh and then here is one algorithmic breakthrough DNN's here is sort of the graph GPUs and AI accelerators now and you put those things together what was an 18month doubling has become a six-month doubling. Wow. Right. So you can think think about it right which is we were already on a fast pace and now we've done it six month and now you have another thing this

inference time comput's law this well I I think I'll avoid u you know the you know naming anything except I just want to ride the wave um and so that is sort of the change but the question you asked is the interesting one right which is I think a little bit of where what's going to be the phase shift going forward is I don't think we're going be sitting here next year, the year after admiring either quite frankly uh the Moors law or the new Moors

law or even the LLMs or the models. It will be what we're doing with all that abundance because the interesting thing is that's kind of go I go back to that bill's statement, right, which is fill it with software. And so it goes back to what do people in this room and everywhere else do with it, right? If there's going to be an abundant commodity, you don't sit there and pray to the abundant commodity. You use it. It's kind of like when Excel came out, we

didn't say, "Oh my god, here is an Excel spreadsheet. Let's sort of put a Morty of it and play." We just said, "Let's create spreadsheets." That's right. And I think that that's what's going to happen uh with and that's I think how we as humans will understand to manage the change. Sure. And so one of the things that brings for me Nandan is you know you you and I just talked briefly backstage even you take all this contextualize it for me in India like how how

does this apply to the Indian economy and Indian society broadly uh to totally I think India will be the use case capital of AI in the world. Uh I think we have a number of things working for us. One is uh I think we have 15 years of experience in building population scale digital infrastructure which we know how to make it work cheap you know high volume billions of transactions all that stuff. So we know that game we have a political leadership which is very tech-savvy.

I know you met from Prime Minister Modi yesterday and they understand that we have to strike the right balance between AI innovation and safeguards because in some parts of the world they're saying safeguards first without worrying about the innovation. So I think we know the right balance between responsible AI and u and and and and innovation and we have a population which has learned to accept technology. I mean, you know, when you think about it, UPI was launched about pretty unbelievable. Seven years back now 400

million now 400 million users, 16 billion transactions a month. I mean, unbelievable this can happen, right? So, I think AI is at that spot. We have to make it work and we already seeing early uh you know signs for example if you look at Aadhaar authentication, it's all AI based because we had to do livveness detection for biometrics so people don't spoof it. It's all AI based. If you look at our tax systems, back end is all AI, you know, figuring out who's doing fraud and

all that's why tax revenues are going up. So I think India is now ready for that and we are going to see many many applications coming. So I think this is the place for you to see this stuff working at population scale. No, 100%. I mean we that's why we are very very excited about you know our investments here. We see the I mean like the combination you said it's a pretty unique place where you have this virtual cycle between the entrepreneurial energy the government sort

of yojenas the India stack and then the population scale so you add those four things into a virtuous cycle like the UPI or Aadhaar or now what's going to happen in education and in health and what have you uh I don't think there is a place uh where the atcale benefit so in an interesting way You know in the past people talked about this convergence right between developing countries and developed countries in an interesting way there's abundance of uh tokens can probably reduce that convergence oh

I I think in many leaprog I think in many of these areas you're going to see leap frog because the advantage of no legacy is you can leaprog so I think that's going to happen and but it'll have to be at uh you know the have to be efficient I think inference costs have to be super frugal frugal because if you're going to have a billion people doing you know all kinds of queries or agentic stuff it has to work at at that scale so I

think there's a lot but I see it happening but you know you you meet and so do I but we meet a lot of CEOs they're all excited but little you know what to do next what is your advice to global CEOs yeah I think the the fundamental challenge um is that first question you asked which is change management right so if I if you think about um when the PCs first came out, right? In fact, I met the CEO um of General Ali, which is

in Italian. Yeah. Milan. So, he started, you know, an insurance company and he was telling me about how in the early 80s when he started, um he'd have to go ask for permission to send a fax. So, I said, "Oh my, it must have been some compliance issue." He said, "No, it was not a compliance issue. It was like expensive to send a fax." Uh and then PCs came out, right? And then you could literally say, "Okay, let me put an attachment in an email and

send it." And it changed the entire process of how his agent, brokers, and them communicated. Same thing. Take forecasting, right? How did forecasting happen in a multinational company like ours? Uh preC, right? I mean, faxes went around, somebody then did an inter office memo, and then eventually 3 months later, you had a forecast. Whereas then suddenly there was a new workflow which was email attachments and an Excel spreadsheet. So I think that process change right when I think about what's happening even with copilot and agents

inside a copilot as the UI layer it's a new workflow. Um and so the work process has to change. So good old in in the mid '90s we used to talk about business process re-engineering. it's back again where you got to go look at how does any process right uh happen today and you got to re-engineer so that means it's moving the cheese around a little bit yeah and so that change management is the hard thing but then let me give you an example even just

at Microsoft when I look at my marketing efficiency of dollars spent double digit customer service double digit uh internal IT ops double digit right so you take any one of these so I'm putting that right into my budget. So to as a CEO what I do is I literally say okay 10 10 points or 100 basis points of um operating leverage next year and then go take that for the next five years you compound it right so that means if that's going to be what the

efficient frontier of any company is and once that is in the water every CEO has to wake up and say you know because after all thanks to capital markets what they expect of CEOs is miracles every 90 days and So uh so if you don't produce that miracle well you're one of those who actually delivered every 90 but I think you know we we see it because we are the probably one of the world's largest users of co-pilot on GitHub you can see the impact already

but it's not just about technology it's about process improvement it's about training people it's about changing mindsets you know the whole human issue that have to deal with it uh but I think this is something which we'll we'll take off but one thing which people are I think you know I remember in the early days of the last two year there was this huge hype about AI will rule the world and all that but that seems to have died down and I think people are realizing

that responsible AI can be done with innovations what's your take on all that I I think that yeah multiple things there right there's two sets of if I sort of had to characterize what's the big debates today one debate is of course hey are the scaling laws still working right because and I'm a believer that the scaling laws are working it's just that you know with scale even things become harder so pre-training pre-training uh as you scale becomes harder because there's data challenge which you have

to overcome just even cluster size I mean it's a massive distributed I mean this workload which is a synchronous data parallel workload is a unlike anything we have seen before too so even the systems problems are unique so there's pre-training itself then there is post-raining or you could say inference time compute uh which essentially is like there is if you take even pre-training there's a massive sample piece to it. We now have a better way to do that with all this chain of thought in a

monologue and autograding it and what have you. So therefore I think we'll continue to see some good algorithmic and systems innovation. So I feel good about that. Then that same capability that's being used to produce capabilities in the model is what is needed for safety as well. Okay. So when you think about guardrails reasoning when I think about reasoning when you can inter you know when you can do the sampling look at the chain of thought autograde it that's the best way to do alignment in

okay so you think some of these hallucination stuff will come hallucination like for example one of the services I'm most excited about even in Azure is this grounding service right so you can kind of use AI to check AI uh how grounded it is right so I think one of the frontiers right now will be a lot of how good are we at our eval for performance, eval for groundedness, eval for safety. Uh, and I think that the state-of-the-art is moving quite frankly like one of

the other things I locked like if you remember every platform shift has had both the core system but also a core app server that we built whether it is for the web or the uh mobile and now we have a full-blown app server uh that we're building in our case called Foundry you know on everything. How do I do fine-tuning? How do I do evaluation? How do I do safeguards? And that way I think so that application developers don't have to recreate it. And you have

made this famous statement that SAS is gone or dead. No, no. I I that's what I mean that's kind of what the social media says. I said I didn't say that. Um but what I said is you know like all things the the application architecture is changing, right? So it's kind of like saying did applications change with relational databases? Absolutely. I mean if you remember I you know when I started even it was in the beginning of relational database and all yeah and so but the

bottom line is before that everybody built a vertical database inside the app right so basically they said oh here's a b tree let me build it in u and that was sort of the state-of-the-art and then suddenly said oh well we can have the separation between a database and have SQL and then I can have the app logic so right now with agents what's going to happen is a lot of the business logic will get separated so in some sense there's going to be databases which

you can do crowd operations on um and then you will sort of even manage eventual consistency between multiple of these SAS backends but a lot of the business logic will move uh to a new tier uh which then will be a multi- aent tier that needs to be orchestrated so it's going to be not I'm not going to one SAS application to another SAS applications I'm going to go to an agent that will orchestrate across multiple SAS applications so it's going to be a very big

change. That's why I think of this copilot as the UI for AI. Like for example, just to give you a real world feel. I just go to copilot every day to get to my dynamic CRM. Okay. I never go to dynamic CRM because all I go is at sales give me what's happening with Infosys and Microsoft and how are we working doing well and it'll come back with by the way not just data inside the CRM system but in all the communications a little and you

and I may have had I get back. So that ability to aggregate for a user query information across the front office system or a productivity system, a back office system, multiple of them, that's an agentic behavior that I think is going to be fairly disruptive in how users use things. No, I agree with you. I think agents will really now drive enterprise consumption. I think uh because it's very easy to tell a CEO that we can create an agent, he can be a digital worker or

he can amplify human potential. I think it's it's easy also to communicate and I think that's going to be a big thing and I think this year will be a big it's kind of like humans and the swarm of agents. Yeah. Yeah. you know uh that I think is the next frontier as you point rightly pointed out and that's where I think a lot of the productivity uh will come from right which is the operating leverage for any function uh will come where the agency just

like I I say you know and by the way the creation of agents today people mystify it and they will always be there will be the high end of you know what's the uh the specific agent about but just like I can build a spreadsheet I will build thousands hundreds of agents that

# Day 54 — 火山引擎 录音文件识别（volc.seedasr.auc）

豆包大模型 ASR，异步接口。可能含识别错误，仅供听力对照。

==================================================

It's great to be here with you, Satya.

Thank you so much.

We did this a couple of years back.

That is right.

And this has become our annual thing.

And I'm really glad to be back in Bangalore with you.

And it's wonderful.

You know, one thing I was thinking in the last two years, there's been an exponential rate of change and, you know, we're all, every day we get up and something new is happening.

How do you deal with that?

And how do you deal with what do you think is coming now?

Yeah, I mean, that's a, I think the, this entire industry, in fact, backstage we were talking about this, um, the last, whatever, 35 years that I've been in this industry, um, the thing that has powered us is this Moore's Law, right?

I mean, it's just been the most unbelievable, um, empirical observation, right?

We call it a law, but it's really, uh, an empirical observation that has held true for a long period of time.

Um, and I remember distinctly, like we used to have these, um, um, you know, he bills to get the top Microsoft folks every year and he would do only one thing.

He would just put a chart.

He would just show us Moore's Law.

He would say, here's what's happening with memory.

Go fill it with software.

That was sort of the instruction, right?

That was basically what we've been doing all these decades.

And then comes this AI revolution and it fundamentally accelerates because we were all bemoaning the fact that, oh my God, Moore's Law is maybe ending or it's not like the same as before.

And then here is one algorithmic breakthrough, DNS.

Here is sort of the graph.

GPUs and AI accelerators now and you put those things together, what was an 18 month doubling has become a six month doubling.

Wow.

Right.

So you can think, think about it, right, which is we were already on a fast pace and now we've done it six month and now you have another thing, this inference time computer.

I think I'll avoid, you know, the, you know, naming anything except I just want to ride the wave.

And so that is sort of the change.

But the question you asked is the interesting one, right, which is I think a little bit of where what's going to be the face shift going forward is.

I don't think we're going to be sitting here next year, the year after, admiring either, quite frankly, uh, the Moore's Law or the new Moore's Law or even the LLMs or the models, it will be what we're doing with all that abundance because the interesting thing is.

That's going to go.

I go back to that Bill's statement, right, which is fill it with software.

And so it goes back to what do people in this room and everywhere else do with it, right?

If there's going to be an abundant commodity, you don't sit there and pray to the abundant commodity, you use it.

It's kind of like when Excel came out.

We didn't say, oh, my god, here is an Excel spreadsheet, let's sort of put a Murthy of it and play.

We just said let's create spreadsheets.

That's right.

And I think that that's what's going to happen, uh, with and that's, I think, how we as humans will understand to manage the change.

And so one of the things that brings for me Nandan is, you know, you, you and I just talked briefly backstage even you take all this, contextualize it for me in India.

Like how, how does this apply to the Indian economy and Indian society broadly?

Totally.

I think India will be the use case capital of AI in the world.

I think we have a number of things working for us.

One is I think we have 15 years of experience in building population scale digital infrastructure, which we know how to make it work cheap, you know, high volume, billions of transactions, all that stuff.

So we know that game.

We have a political leadership which is very tech savvy.

I know you met from Prime Minister Modi yesterday.

And they understand that we have to strike the right balance between AI innovation and safeguards because in some parts of the world, they're saying safeguards first without worrying about the innovation.

So I think we know the right balance between responsible AI and, and, and, and, and innovation.

And we have a population which has Learned to accept technology.

I mean, you know, when you think about it, UPI was launched about seven years back now, 400 million users, 16 billion transactions a month.

I mean, unbelievable this can happen, right?

So I think AI is at that spot.

We have to make it work.

And we're already seeing early, uh, you know, science, for example, if you look at Adhar authentication, it's all AI based because we had to do liveness detection for biometrics so people don't spoof it.

It's all AI based.

If you look at our tax systems, back end is all AI, you know, figuring out who's doing fraud and all.

That's why our tax revenues are going up.

So I think India is now ready for that and we are going to see many, many applications coming.

So I think this is the place for you to see this stuff working at population scale.

No, hundred percent.

I mean, we, that's why we are very, very, very excited about, you know, our investments here.

We see the, I mean, like the combination you said it's a pretty unique place where you have this virtuous cycle between the entrepreneurial energy, the government sort of yojanas, the India stack and then the population scale.

So you add those four things into a virtuous cycle, like the UPI or Adhaar or now what's going to happen in education and in health and what have you.

I don't think there is a place where the at scale benefit.

So in an interesting way, you know, in the past, people talked about this convergence, right, between developing countries and developed countries in an interesting way, this abundance of, uh, tokens.

Yeah.

Can probably reduce that convergence.

Oh, I, I think in many ways it's a leapfrog.

I think in many of these areas you're going to see leapfrog because the advantage of no legacy is you can leapfrog.

So I think that's going to happen and but it'll have to be at, uh, you know, the have to be efficient.

I think inference costs have to be super frugal because you're going to have a billion people doing, you know, all kinds of queries or agentic stuff.

It has to work at at that scale.

So I think there's a lot, but I see it happening.

Uh, but you know, you, you meet and so do I.

But we meet a lot of CEOs.

They're all excited, but little you know what to do next.

What is your advice to global CEOs?

Yeah, I think the, the fundamental challenge, um, is that first question you asked, which is change management, right?

So if I, for, if you think about it, um, when the PCs first came out, right, in fact, I met, uh, the CEO, um, of Generali, which is in, uh, yeah, yeah, Italian.

Yeah, Milan.

So he started, you know, an insurance company and he was telling me about how in the early 80s when he started, um, he'd have to go ask for permission to send a FAX.

So I said, oh, my way must have been some compliance issue.

He says, no, it was not a compliance issue.

It was like expensive to send a FAX.

Uh, and then PCs came out, right?

And then you could literally say, okay, let me put an attachment in an email and send it.

And it changed the entire process of how his agent brokers and them communicated.

Same thing, take forecasting, right?

How did forecasting happen in a multinational company like ours?

Pre PC, right?

I mean, faxes went around somebody, then you did an interoffice memo and then eventually three months later you had a forecast, whereas then suddenly there was a new workflow which was email attachments and an Excel spreadsheet.

So I think that process change, right, when I think about what's happening even with copilot and agents inside a copilot as the UI layer, it's a new workflow.

Um, and so the work process has to change.

So good old in the mid 90s we used to talk about business process reengineering.

It's back again where you got to go look at how does any process, right, uh, happen today and you got to reengineer.

So that means it's moving the cheese around a little bit.

Yeah.

And so that change management is the hard thing.

But then let me give you an example.

Even just at Microsoft, when I look at my marketing efficiency of dollar spent, double digit, customer service, double digit, internal it ops, double digit, right?

So you take any one of these.

So I'm putting that right into my budget.

So to as a CEO, what I do is I literally say, okay, 10 minute, 10 points or 100 basis points of operating leverage next year and then go take that for the next five years.

You compound it, right?

So that means if that's going to be what the efficient frontier of any company is.

And once that is in the water, every CEO has to wake up and say, you know, because after all, thanks to capital markets, what they expect of CEOs is miracles every 90 days.

And so, ah, so if you don't produce that miracle, well, you're one of those who actually delivered every 90.

But I think, you know, we, we see it because we are the probably one of the world's largest users of copilot.

On Github, you can see the impact already.

But it's not just about technology.

It's about process improvement.

It's about training people.

It's about changing mindsets.

You know, the whole human issue that have to deal with it.

But I think this is something which will, will, will, will take off.

But one thing which people are, I think, you know, I remember the early days of the last two years, there was this huge hype about AI will rule the world and all that, but that seems to have died down.

And I think people are realizing that responsible AI can be done with innovations.

What's your take on all that?

I think that there, yeah, multiple things there, right?

There is two sets of, if I sort of had to characterize what's the big debates today.

One debate is, of course, hey, are the scaling laws still working?

Yeah, right.

Because.

And I'm a believer that the scaling laws are working.

It's just that, you know, with scale, even things become harder.

So pre training as you scale becomes harder because there's data challenge which you have to overcome just even cluster size.

I mean, it's a massive distributed.

I mean, this workload, which is a synchronous data parallel workload, is unlike anything we have seen before too.

So even the systems problems are unique.

So there's pretraining itself, then there is post training, or you could say inference time compute, which essentially is like there is, if you take even pretraining, there's a massive sampling piece to it.

We now have a better way to do that with all this chain of thought in a monologue and auto grading it and what have you.

So therefore, I think we'll continue to see some good algorithmic and systems innovation.

So I feel good about that.

Then that same capability that's being used to produce capabilities in the model is what is needed for safety as well.

OK.

So when you think about guardrails, reasoning, when I think about reasoning, when you can interest, you know, when you can do the sampling, look at the chain of thought, auto grade it.

That's the best way to do alignment in some sense.

So you think some of these hallucination stuff will come?

Hallucination.

Like, for example, one of the services I'm most excited about even in Azure is this grounding service, right?

So you can kind of use AI to check AI, uh, how grounded it is, right?

So I think one of the frontiers right now will be a lot of how good are we at our evals, E valves for performance, E valves for groundedness, e valves for safety.

And I think that the state of the art is moving, quite frankly, like one of the other things I locked, like if you remember, every platform shift has had both the core system but also a core app server that we built.

Whether it is for the web or the mobile, and now we have a full blown app server that we're building.

In our case called Foundry, you know, on everything, how do I do fine tuning, how do I do evaluation, how do I do safeguards?

And that way I think so that application developers don't have to recreate it.

And you all made this famous statement that SAS is gone or dead or something.

No, no, I, I.

That's all.

I mean, that's kind of what the social media says.

I said, I didn't say that.

Um, but what I said is, you know, like all things, the, the application architecture is changing, right?

So it's kind of like saying, did applications change with relational databases?

Absolutely.

I mean, you remember, I mean, when I started even it was in the beginning of relational database.

Yeah.

And so, but the bottom line is before that, everybody built a vertical database inside the app.

That's right.

So basically they said, oh, here's a B tree, let me build it in.

And that was sort of the state of the art.

And then suddenly said, oh, we can have the separation between a database and have SQL and then I can have the app logic.

So right now with agents, what's going to happen is a lot of the business logic will get separated.

So in some sense, there's going to be databases which you can do crud operations on.

Um, and then you will sort of even manage eventual consistency between multiple of these Saas back ends, but a lot of the business logic will move, uh, to a new tier, uh, which then will be a multi agent tier that needs to be orchestrated.

So it's going to be not, I'm not going to one Saas application to another Saas applications, I'm going to go to an agent that will orchestrate across multiple Saas applications.

So it's going to be a, a very big change.

That's why I think of this copilot as the UI for AI.

Like for example, just to give you a real world feel, I just go to copilot every day to get to my dynamic CRM.

I never go to dynamic CRM because all I go is at sales.

Give me what's happening with Infosys and Microsoft and how are we working?

Doing well.

Yeah.

You know, and you'll come back with, by the way, not just data inside the CRM system, but in all the communications.

A little and you and I may have had.

I get back.

So that ability to aggregate for a user query information across the front office system or a productivity system, a back office system, multiple of them, that's an agentic behavior that I think is going to be fairly disruptive in how users use things. No, I agree with you. I think agents will.

Really now drive enterprise consumption.

I think, uh, because it's, it's very easy to tell a CEO that we can create an agent.

He can be a digital worker.

He can amplify human potential.

I think it's it's easy also to communicate and I think that's going to be a big thing.

And I think this year will be a big, it's kind of like humans and the swarm of agents.

Yeah, yeah.

You know.

Uh, that I think is the next frontier as your point rightly point out.

And that's where I think a lot of the productivity will come from, right, which is the operating leverage for any function will come where the agency, just like I say, you know, and by the way.

The creation of agents today, people mystify it and they will always be.

There will be the high end of, you know, what's the the specific agent about.

But just like I can build a spreadsheet, I will build.