1
00:00:00,520 --> 00:00:02,917
Well, Satya, thank you for joining me

2
00:00:03,000 --> 00:00:04,797
and thank you for hosting me at your

3
00:00:04,880 --> 00:00:08,637
beautiful offices here at Microsoft HQ.

4
00:00:08,720 --> 00:00:10,557
I'm super grateful to be here and I'm

5
00:00:10,640 --> 00:00:12,557
looking forward to talking to you about

6
00:00:12,640 --> 00:00:15,997
what the future of AI in the enterprise

7
00:00:16,080 --> 00:00:18,997
actually looks like. Our mission is to

8
00:00:19,080 --> 00:00:21,197
help companies own their intelligence by

9
00:00:21,280 --> 00:00:23,437
being able to train their own models and

10
00:00:23,520 --> 00:00:26,037
improve those the more they use them.

11
00:00:26,120 --> 00:00:28,317
So, when you recently wrote this article

12
00:00:28,400 --> 00:00:30,877
that got a lot of traction on X, I think

13
00:00:30,960 --> 00:00:35,957
it was titled "A Frontier Without
an Ecosystem Is Not Stable", it really

14
00:00:36,040 --> 00:00:40,277
resonated with a lot of the things
that we think about at at AC.

15
00:00:40,360 --> 00:00:43,797
And, you know, I wanted to maybe start
the conversation talking a little bit

16
00:00:43,880 --> 00:00:46,597
about uh you know, your core arguments

17
00:00:46,680 --> 00:00:48,877
in in that piece, which is

18
00:00:48,960 --> 00:00:50,877
you know, the stable equilibrium is

19
00:00:50,960 --> 00:00:53,837
really companies building on top of

20
00:00:53,920 --> 00:00:56,037
frontier capabilities rather than being

21
00:00:56,120 --> 00:01:01,037
attached to one or two frontier model
labs. I was curious if you could maybe

22
00:01:01,120 --> 00:01:05,157
talk a little bit about what that looks
like in practice, um how Microsoft is

23
00:01:05,240 --> 00:01:09,797
approaching this ecosystem play, and what
are the maybe the practical things

24
00:01:09,880 --> 00:01:12,917
that are being built on the
infrastructure side to make it happen.

25
00:01:13,000 --> 00:01:16,837
>> Yeah, no, first of all, yes, it's great
to have you here on the Microsoft campus

26
00:01:16,920 --> 00:01:21,117
and also it's great to be partnered
with you as you're building out

27
00:01:21,200 --> 00:01:23,277
um AC and uh we're building out

28
00:01:23,360 --> 00:01:25,717
essentially this frontier ecosystem.

29
00:01:25,800 --> 00:01:30,517
>> Um to me, the uh the motivation

30
00:01:30,600 --> 00:01:32,477
for both that article and what we're

31
00:01:32,560 --> 00:01:36,237
doing comes stems from

32
00:01:36,320 --> 00:01:38,277
understanding, I would say, two uh

33
00:01:38,360 --> 00:01:40,197
things or answering two questions,

34
00:01:40,280 --> 00:01:43,557
perhaps. One is what's the future of a

35
00:01:43,640 --> 00:01:48,757
firm in the ecosystem or in the economy

36
00:01:48,840 --> 00:01:50,957
uh in the age of AI,

37
00:01:51,040 --> 00:01:54,277
as well as what's the nature of the

38
00:01:54,360 --> 00:01:57,277
ecosystem and platforms

39
00:01:57,360 --> 00:02:01,877
uh with AI, right? So, let's start
with the first one which which is

40
00:02:01,960 --> 00:02:04,957
AI is a learning system, right? So, it's

41
00:02:05,040 --> 00:02:08,237
sort of a not a traditional digital

42
00:02:08,320 --> 00:02:10,877
technology or it's not a technology like

43
00:02:10,960 --> 00:02:14,277
all the other digital tools we built.

44
00:02:14,360 --> 00:02:16,197
And so, therefore, one has to sort of

45
00:02:16,280 --> 00:02:21,477
think about in a world where
a model can learn anything.

46
00:02:21,560 --> 00:02:22,197
>> Yeah.

47
00:02:22,280 --> 00:02:25,837
>> What are the limits
to that? Or in other

48
00:02:25,920 --> 00:02:28,797
words, how should the firm itself

49
00:02:28,880 --> 00:02:31,797
think of itself going forward? And so,

50
00:02:31,880 --> 00:02:36,997
my simple thing is there should be as many
models in the world as firms in the

51
00:02:37,080 --> 00:02:38,437
world, right? Because after all, what is

52
00:02:38,520 --> 00:02:42,517
a firm? A firm is a learning system that

53
00:02:42,600 --> 00:02:45,197
today is mostly about human capital with

54
00:02:45,280 --> 00:02:49,157
digital tools every day compounding the

55
00:02:49,240 --> 00:02:53,437
tacit knowledge that the firm
has and how it produces a

56
00:02:53,520 --> 00:02:55,717
good or a service that's
valuable in the economy.

57
00:02:55,800 --> 00:02:57,917
>> Totally. >> Right. So,
if that is the definition of

58
00:02:58,000 --> 00:02:59,957
a firm, my question is then in the age

59
00:03:00,040 --> 00:03:02,437
of AI, it should compound because of

60
00:03:02,520 --> 00:03:06,877
both its ability to create the token

61
00:03:06,960 --> 00:03:09,877
capital that then is being used in

62
00:03:09,960 --> 00:03:14,317
conjunction with the human capital
and then the compounding loop between the

63
00:03:14,400 --> 00:03:17,677
two, right? So, that's
sort of one part of it.

64
00:03:17,760 --> 00:03:19,957
In order for that to be true,

65
00:03:20,040 --> 00:03:22,877
then you do need to have a platform that

66
00:03:22,960 --> 00:03:27,357
is really enabling
every firm in the economy

67
00:03:27,440 --> 00:03:31,197
to be able to create its own token
capital. It can't be, "Hey, look, I have

68
00:03:31,280 --> 00:03:35,477
two frontier models or three frontier
models or whatever, some finite set that

69
00:03:35,560 --> 00:03:38,037
have learned everything that is

70
00:03:38,120 --> 00:03:41,117
differentiated today in the economy."

71
00:03:41,200 --> 00:03:45,477
Because then it collapses, right?
So, there is no stability

72
00:03:45,560 --> 00:03:49,717
to even the economy and definitely
the political economy won't tolerate it

73
00:03:49,800 --> 00:03:53,917
because you kind of need a world
that is, you know, there may be very

74
00:03:54,000 --> 00:03:55,637
successful companies. This is not about

75
00:03:55,720 --> 00:03:58,717
sort of any one company's success.

76
00:03:58,800 --> 00:04:01,637
It is about an economy being stable

77
00:04:01,720 --> 00:04:03,637
because there is a lot of economic

78
00:04:03,720 --> 00:04:06,917
growth that is more broad-spread. And so

79
00:04:07,000 --> 00:04:10,837
that's sort of the the two part
and that's why I wrote that.

80
00:04:10,920 --> 00:04:14,277
And you know, and then we can unpack that.
I mean, in terms of each of the layers.

81
00:04:14,360 --> 00:04:16,397
>> Totally. Yeah. And and I guess like

82
00:04:16,480 --> 00:04:18,757
maybe drilling in on that point about

83
00:04:18,840 --> 00:04:21,637
the stability of the economy, what do

84
00:04:21,720 --> 00:04:27,037
you think just breaks down when you have
all the value accruing to maybe a small

85
00:04:27,120 --> 00:04:29,917
set of model companies? >> Yeah,
I mean, if you think about it like

86
00:04:30,000 --> 00:04:33,997
definition of these tech platforms
at least that I subscribe to is more value

87
00:04:34,080 --> 00:04:37,197
has to be created about the platform
for the platform to be stable. Because

88
00:04:37,280 --> 00:04:40,517
>> Yeah, it's like the definition of >>
Yeah, it's otherwise it's kind of like,

89
00:04:40,600 --> 00:04:45,357
you know, if you you need
customers in order to be successful.

90
00:04:45,440 --> 00:04:49,317
And so that means customers have to be
successful. And that is a little bit of

91
00:04:49,400 --> 00:04:52,277
the the simple logic on it. >> Yeah.

92
00:04:52,360 --> 00:04:55,757
>> But more than that, right? I kind
of look at this and say today, first of

93
00:04:55,840 --> 00:04:58,757
all, I think we've got some very cool

94
00:04:58,840 --> 00:05:02,357
attributes to these models and we now

95
00:05:02,440 --> 00:05:05,357
know that what fuels that is of course

96
00:05:05,440 --> 00:05:10,917
compute, some great talent,
but also a lot of

97
00:05:11,000 --> 00:05:14,437
data. >> Yes. >> And so I
sort of look at it and say,

98
00:05:14,520 --> 00:05:18,517
hey, we have now created a lot of value
in these models using let's call it

99
00:05:18,600 --> 00:05:21,837
public data and great
talent and lots of compute.

100
00:05:21,920 --> 00:05:24,517
>> Yeah. >> But where
is the vast amount of data

101
00:05:24,600 --> 00:05:27,797
still? It's all in the enterprises or in

102
00:05:27,880 --> 00:05:31,797
the companies. And so where
the unlock at least if you

103
00:05:31,880 --> 00:05:35,117
sort of it's not it's it's not a zero-sum,
right? It's about what is the

104
00:05:35,200 --> 00:05:38,077
positive sum way for, let's say, to

105
00:05:38,160 --> 00:05:40,077
achieve what I think has to be the goal

106
00:05:40,160 --> 00:05:45,837
here collectively. Can the GDP
inflect to 10%, 15% or what have you?

107
00:05:45,920 --> 00:05:48,477
>> Yeah. >> And I think
that will come when every

108
00:05:48,560 --> 00:05:53,597
firm is able to take even a frontier
model or an open weight model or

109
00:05:53,680 --> 00:05:55,077
whatever

110
00:05:55,160 --> 00:06:00,997
and then hill climb using their R and E,
right? Which is the company itself.

111
00:06:01,080 --> 00:06:03,317
>> Yeah. >> And in that
process even build maybe

112
00:06:03,400 --> 00:06:05,437
even their own model or two which is

113
00:06:05,520 --> 00:06:10,397
much more fine-tuned to the their needs
and doesn't have to spend all the tokens

114
00:06:10,480 --> 00:06:13,277
>> Totally. >> on a frontier
model, right? So that I

115
00:06:13,360 --> 00:06:16,557
think is where the direction of travel is.

116
00:06:16,640 --> 00:06:20,157
>> That makes makes total sense and it's
a it's a perfect lead into what I was

117
00:06:20,240 --> 00:06:24,717
going to ask next which is how do
you make this decision about when

118
00:06:24,800 --> 00:06:28,477
to own versus rent? Because I think,
you know, you've said this before

119
00:06:28,560 --> 00:06:32,557
people will use other people's AI, they
will build their own AI, but when do you

120
00:06:32,640 --> 00:06:36,797
decide to do what? And from our
experience at AC, we've like seen kind

121
00:06:36,880 --> 00:06:41,317
of like four, you know, four sort
of different tailwinds that are kind of

122
00:06:41,400 --> 00:06:45,037
driving this all kind of recent actually.
You know, one is just the

123
00:06:45,120 --> 00:06:48,237
token apocalypse, things
are getting really expensive.

124
00:06:48,320 --> 00:06:52,637
Two is, you know, with maybe the most
recent Claude Fable release and the

125
00:06:52,720 --> 00:06:56,877
stipulations around it,
data protections, government stepping in,

126
00:06:56,960 --> 00:07:01,477
maybe the models on the frontier APIs
won't always be there. The third kind of

127
00:07:01,560 --> 00:07:05,677
being like labs are moving more into
the application layer and people are like,

128
00:07:05,760 --> 00:07:09,357
oh, I want to differentiate on multiple
levels than just depend on frontier

129
00:07:09,440 --> 00:07:12,597
models. And the fourth
being open source models are

130
00:07:12,680 --> 00:07:13,877
starting to get really, really good.

131
00:07:13,960 --> 00:07:16,277
>> Yeah. >> I'm curious
and this data point was a

132
00:07:16,360 --> 00:07:19,397
really, really good point. What are the

133
00:07:19,480 --> 00:07:24,077
are there any other major reasons
why you think companies might

134
00:07:24,160 --> 00:07:25,837
choose to own their intelligence?

135
00:07:25,920 --> 00:07:29,037
>> Yeah, I mean, I I think you you you hit
on the most of the criteria, but the way

136
00:07:29,120 --> 00:07:31,397
I I would think about is maybe a firm

137
00:07:31,480 --> 00:07:37,157
should say, okay, here is, I don't know,
an agentic system I want to build for my

138
00:07:37,240 --> 00:07:41,077
supply chain. >> Yeah. >> Right?
And optimizing the supply chain

139
00:07:41,160 --> 00:07:44,397
planning, let's call it.
After all, supply chains for an

140
00:07:44,480 --> 00:07:48,277
industrial company or for anyone is,
you know, or a data center company like us

141
00:07:48,360 --> 00:07:49,957
is pretty critical stuff, right? Which

142
00:07:50,040 --> 00:07:54,117
is it's it's also idiosyncratic based on

143
00:07:54,200 --> 00:07:58,437
what is unique about us and how we
operate and what have you. So, you I

144
00:07:58,520 --> 00:08:02,397
would say that one of the things that
is least talked about, although it's very

145
00:08:02,480 --> 00:08:04,717
critical for any agentic or AI system,

146
00:08:04,800 --> 00:08:07,797
is the outcome or the eval that matters

147
00:08:07,880 --> 00:08:11,117
for any such business process, right? Uh

148
00:08:11,200 --> 00:08:12,997
and so, I would start there, which is

149
00:08:13,080 --> 00:08:19,037
what's the outcome?
Uh and my ability to grade that outcome.

150
00:08:19,120 --> 00:08:24,357
>> [snorts] >> Um and start with that and
then say, "Okay, I welcome all models."

151
00:08:24,440 --> 00:08:27,877
>> Yeah. >> Then,
the important thing is the the the

152
00:08:27,960 --> 00:08:31,117
all part, which is I don't want
to be locked into anyone model.

153
00:08:31,200 --> 00:08:33,277
>> Totally, yeah. >> Right?
I'm happy to have all the

154
00:08:33,360 --> 00:08:38,757
frontier models. I want to use them
to see where they are on the eval.

155
00:08:38,840 --> 00:08:41,717
>> Yeah. >> Then, I want
to be able to use my own

156
00:08:41,800 --> 00:08:43,997
context, my own data, in fact, my own

157
00:08:44,080 --> 00:08:47,037
traces to maybe even take a much more

158
00:08:47,120 --> 00:08:51,797
open weight cost-efficient model
or a fine-tuned model which is not using

159
00:08:51,880 --> 00:08:55,397
general intelligence, but is using
the intelligence I need in order to

160
00:08:55,480 --> 00:08:58,597
basically hit the same eval
that a frontier model had.

161
00:08:58,680 --> 00:09:02,277
>> Yeah.
>> That is the empirical way I think

162
00:09:02,360 --> 00:09:04,717
companies are going to go about getting

163
00:09:04,800 --> 00:09:06,957
value because what that achieves is one,

164
00:09:07,040 --> 00:09:09,957
you're grounded on what the result is.

165
00:09:10,040 --> 00:09:14,877
Second is you are also ensuring that
you're not locked in to any one model.

166
00:09:14,960 --> 00:09:17,437
>> Yeah. >> And that means
you're in control of your

167
00:09:17,520 --> 00:09:21,317
token economics long term. That is how I

168
00:09:21,400 --> 00:09:26,277
think companies can decide, which is,
you know, really be outcome-oriented

169
00:09:26,360 --> 00:09:28,997
>> Yeah. >> uh and continuously
optimizing. There's

170
00:09:29,080 --> 00:09:31,997
no such thing as, "Hey, this model and

171
00:09:32,080 --> 00:09:35,877
this model family is the only thing that

172
00:09:35,960 --> 00:09:38,237
I am locked into." If anything, you're

173
00:09:38,320 --> 00:09:39,677
locked into one thing, which is the

174
00:09:39,760 --> 00:09:43,917
monotonic improvement of your outputs.

175
00:09:44,000 --> 00:09:48,077
>> Yeah, I think that's that's really
really well said and and I actually

176
00:09:48,160 --> 00:09:50,077
hadn't occurred to me before is by

177
00:09:50,160 --> 00:09:53,437
simply having the benchmark, you get all

178
00:09:53,520 --> 00:09:57,677
of this control cuz you can see
how things relate to one another.

179
00:09:57,760 --> 00:09:59,877
Um I think you know, in the article you

180
00:09:59,960 --> 00:10:04,957
were talking about this idea of like
a learning loop outside of your model. And

181
00:10:05,040 --> 00:10:09,117
I think this kind of touches on that is
like once you have these private emails,

182
00:10:09,200 --> 00:10:13,117
which you know, they're very unique
to the company, you can start to make your

183
00:10:13,200 --> 00:10:16,277
systems better. I'm curious
out of the companies that

184
00:10:16,360 --> 00:10:18,717
Microsoft works with, how what are the

185
00:10:18,800 --> 00:10:21,677
sort of initial foundations of learning

186
00:10:21,760 --> 00:10:23,997
loops that you're seeing people start to

187
00:10:24,080 --> 00:10:26,197
build around their AI systems?

188
00:10:26,280 --> 00:10:28,957
>> Yeah, in fact,
one of the biggest things

189
00:10:29,040 --> 00:10:31,397
that we are trying to get done is like

190
00:10:31,480 --> 00:10:33,277
after all we built

191
00:10:33,360 --> 00:10:36,917
uh you know, Office 365, Microsoft 365

192
00:10:37,000 --> 00:10:40,917
as this multi-tenant SaaS service in

193
00:10:41,000 --> 00:10:43,477
which some of the most important data

194
00:10:43,560 --> 00:10:46,797
{slash} context of a given enterprise is

195
00:10:46,880 --> 00:10:50,797
present, right? Which is if you think
about it, all their communications

196
00:10:50,880 --> 00:10:53,517
uh data and information, their projects,

197
00:10:53,600 --> 00:10:58,957
their timelines, their calendars,
their documents, all of that

198
00:10:59,040 --> 00:11:02,517
>> Yeah.
>> is the unstructured database uh where

199
00:11:02,600 --> 00:11:05,837
all that tacit knowledge. In fact,
it's sort of not only that the the

200
00:11:05,920 --> 00:11:09,477
unstructured data is there, but it's
the pointers to all the other of the

201
00:11:09,560 --> 00:11:13,517
business process, right? Because humans
are communicating, creating documents,

202
00:11:13,600 --> 00:11:15,237
all this unstructured data is in the

203
00:11:15,320 --> 00:11:17,717
context of even structured processes,

204
00:11:17,800 --> 00:11:21,797
right? And so, to me, taking that

205
00:11:21,880 --> 00:11:27,317
and really in fact, we describe it
as Work IQ, we want that to be the context

206
00:11:27,400 --> 00:11:31,357
layer along with any model,
right? So, when we

207
00:11:31,440 --> 00:11:33,157
even built Copilot, that's kind of how

208
00:11:33,240 --> 00:11:34,957
we built it, right? Which is Copilot

209
00:11:35,040 --> 00:11:38,717
takes the Work IQ context, mixes up with

210
00:11:38,800 --> 00:11:43,677
any model, and then helps you with the
outcomes you want and the intents you

211
00:11:43,760 --> 00:11:46,717
want. >> Yeah. >> And we are
even exposing that, like so

212
00:11:46,800 --> 00:11:50,117
for example, you know, one of the things
that we did even at Microsoft was to

213
00:11:50,200 --> 00:11:52,757
take the HR onboarding process.

214
00:11:52,840 --> 00:11:55,397
>> Yeah. >> The HR onboarding
process has a system

215
00:11:55,480 --> 00:11:59,677
of record somewhere, but a lot of it
also happens in email, in calendars, and

216
00:11:59,760 --> 00:12:01,957
and what have you. It's much So, if we

217
00:12:02,040 --> 00:12:05,117
can take that context, and then create

218
00:12:05,200 --> 00:12:07,757
an agentic system which streamlines the

219
00:12:07,840 --> 00:12:10,157
H, you know, onboarding process,

220
00:12:10,240 --> 00:12:12,557
and it's token efficient, right? Because

221
00:12:12,640 --> 00:12:17,277
the context is being used
for both input tokens, um

222
00:12:17,360 --> 00:12:21,717
being you know, being input properly,
and then you're not using frontier

223
00:12:21,800 --> 00:12:26,197
models to generate and regenerate output
tokens that don't need to be regenerated.

224
00:12:26,280 --> 00:12:29,677
>> Yeah. >> And that to me
is a great way to in some

225
00:12:29,760 --> 00:12:33,837
sense create a learning loop that
continuously gets better before we even

226
00:12:33,920 --> 00:12:36,597
have come to fine-tuning
a model or what have you.

227
00:12:36,680 --> 00:12:40,157
>> Yeah. >> Then on top of it, you can,
right? You can You can also have an all

228
00:12:40,240 --> 00:12:41,477
>> That's the layer they stack.

229
00:12:41,560 --> 00:12:43,397
>> Exactly, right?
And you can always like,

230
00:12:43,480 --> 00:12:46,077
you know, whether it's you know, I think
that that's sort of what I think people

231
00:12:46,160 --> 00:12:50,117
will go is have evals, have context,

232
00:12:50,200 --> 00:12:53,837
>> Yeah. >> then have
models all hill climb, and as

233
00:12:53,920 --> 00:12:58,517
the models are hill climbed and you say,
"Oh, but I'm I've got all this new

234
00:12:58,600 --> 00:13:02,957
trace data. >> Yeah. >> Let me
take all that, tune a a more a

235
00:13:03,040 --> 00:13:04,957
computer-efficient model >> Yeah.

236
00:13:05,040 --> 00:13:09,637
>> for this particular outcome, and then
keeps doing the cycle continuously.

237
00:13:09,720 --> 00:13:14,157
>> And and I I I think one of the points
you made there is is is actually, you

238
00:13:14,240 --> 00:13:16,037
know, kind of a centerpiece of the the

239
00:13:16,120 --> 00:13:19,597
whole original article is this idea of,

240
00:13:19,680 --> 00:13:24,037
you know, human capital and token capital.
And when I and that they're

241
00:13:24,120 --> 00:13:27,397
very intertwined and interlaced. And
and when I was looking at some of the you

242
00:13:27,480 --> 00:13:29,997
know reactions to your post, one of them

243
00:13:30,080 --> 00:13:31,797
that caught my eye was like this is like

244
00:13:31,880 --> 00:13:34,997
an incredibly positive some take on AI,

245
00:13:35,080 --> 00:13:37,237
which is actually not what the narrative

246
00:13:37,320 --> 00:13:40,317
is usually around around AI. It's that

247
00:13:40,400 --> 00:13:44,717
tokens are going to eat into what
humans do and you know, it's a little bit

248
00:13:44,800 --> 00:13:49,037
fatalistic that like AI is just going to
do everything. But your view is that you

249
00:13:49,120 --> 00:13:51,357
know, humans are actually essential to

250
00:13:51,440 --> 00:13:54,677
making you know, the token economy grow

251
00:13:54,760 --> 00:13:56,797
and these models actually do effective

252
00:13:56,880 --> 00:14:01,757
work. So Evals is is one way, but you
know, when you're thinking about these

253
00:14:01,840 --> 00:14:03,357
these learning loops that are being

254
00:14:03,440 --> 00:14:06,477
built, how are how are how is Microsoft

255
00:14:06,560 --> 00:14:10,877
thinking about designing learning loops
that are very human-centric so that they

256
00:14:10,960 --> 00:14:14,597
learn >> It's a great thing,
right? Because at some level,

257
00:14:14,680 --> 00:14:18,637
you know, I I go back a little bit
to Uh first of all, I think we should be

258
00:14:18,720 --> 00:14:23,997
clear-eyed about the change to the
knowledge economy and the knowledge work.

259
00:14:24,080 --> 00:14:28,077
>> Yeah. >> And so therefore,
even the displacement

260
00:14:28,160 --> 00:14:30,517
with something as disruptive as this.

261
00:14:30,600 --> 00:14:33,837
>> Yeah. >> But if I look
back, um let's say in the

262
00:14:33,920 --> 00:14:36,317
early '80s, long before you were born.

263
00:14:36,400 --> 00:14:40,397
>> [laughter] >> Yeah. >> Um if
somebody had come to me and said,

264
00:14:40,480 --> 00:14:44,597
"Hey, 3 billion people are going to wake
up and they're all going to be typing

265
00:14:44,680 --> 00:14:47,877
every day." And we would have said,
"What are we talking about? Like why

266
00:14:47,960 --> 00:14:50,437
would the world need 3 billion typists?"

267
00:14:50,520 --> 00:14:51,717
>> Yeah. >> But

268
00:14:51,800 --> 00:14:54,597
But we did invent a complete thing

269
00:14:54,680 --> 00:14:57,957
called knowledge work. And it is turns

270
00:14:58,040 --> 00:15:00,117
out that all of us in different forms,

271
00:15:00,200 --> 00:15:04,997
right? Whether you're an accountant
or a supply chain analyst or a coder,

272
00:15:05,080 --> 00:15:08,397
uh we're all doing knowledge work,
right? So we kind of figured out how to

273
00:15:08,480 --> 00:15:12,997
use these computers uh
to accelerate what was work that was

274
00:15:13,080 --> 00:15:15,237
being done in different ways. >> Yeah.

275
00:15:15,320 --> 00:15:17,637
>> I think we're at the We now need that

276
00:15:17,720 --> 00:15:20,117
same new reinvention of knowledge work.

277
00:15:20,200 --> 00:15:24,157
And I I mean, take even
coding. Right? I mean

278
00:15:24,240 --> 00:15:29,757
Let's you know coding for the longest
time even for you and for me it was code

279
00:15:29,840 --> 00:15:33,717
as input. >> Yeah. >> Suddenly
code has become output. Right?

280
00:15:33,800 --> 00:15:35,957
So like you know in some sense we know

281
00:15:36,040 --> 00:15:40,957
but but at the same time I do
think that you know what does a

282
00:15:41,040 --> 00:15:43,957
coder do today? I you know
one of the things that one of

283
00:15:44,040 --> 00:15:47,317
my colleagues you know thought me
and he's in fact written a skill is called

284
00:15:47,400 --> 00:15:49,357
cognitive coverage right? Which is the

285
00:15:49,440 --> 00:15:54,637
human is still on the loop.
Just like test coverage in the past.

286
00:15:54,720 --> 00:15:59,437
>> Yeah. >> Humans are needed to have
cognitive coverage over machine output.

287
00:15:59,520 --> 00:16:03,757
>> Mhm. >> So that you know
what the machine produced.

288
00:16:03,840 --> 00:16:06,557
You're able to interrogate it.
That's a job.

289
00:16:06,640 --> 00:16:11,277
>> Totally. >> That we will in fact value.
That you will then need skills to do.

290
00:16:11,360 --> 00:16:13,517
>> Yeah. >> That you
will go to school for.

291
00:16:13,600 --> 00:16:15,557
>> Yeah. >> And that will
have wage support.

292
00:16:15,640 --> 00:16:18,117
>> Yeah. >> It may be
different than what a software

293
00:16:18,200 --> 00:16:23,477
developer's wages in the past were.
It may be more or less the same who knows.

294
00:16:23,560 --> 00:16:25,797
>> Yeah. >> But there is a job to be done.

295
00:16:25,880 --> 00:16:29,837
>> Totally. >> And so one
of my things is hey let's now

296
00:16:29,920 --> 00:16:32,117
get into this thing about where are all

297
00:16:32,200 --> 00:16:37,597
the human capital interventions
to create more token capital.

298
00:16:37,680 --> 00:16:40,677
>> Yes. >> And I look
around the empirical evidence

299
00:16:40,760 --> 00:16:44,917
is there will be no chance of creating
any token capital without human capital.

300
00:16:45,000 --> 00:16:49,557
>> Yeah. >> And so we are talking
about as if this just came from nowhere.

301
00:16:49,640 --> 00:16:52,557
>> Yeah. >> Even if you
go to you know let's let's

302
00:16:52,640 --> 00:16:54,157
say we're all AGI paired we're going to

303
00:16:54,240 --> 00:16:58,477
hit RSI. Even with RSI the question is

304
00:16:58,560 --> 00:17:02,437
you know unless we're talking about us
as humans losing control. I look at it

305
00:17:02,520 --> 00:17:04,677
and say the only thing valuable in the

306
00:17:04,760 --> 00:17:07,277
age of RSI is the human capital that's

307
00:17:07,360 --> 00:17:10,237
controlling the AI. So I feel like

308
00:17:10,320 --> 00:17:13,437
that's the imagination we should unlock

309
00:17:13,520 --> 00:17:18,517
and starting atomically at the firm
level and saying how do I really design

310
00:17:18,600 --> 00:17:20,437
my organization now where the human

311
00:17:20,520 --> 00:17:24,477
capital, its ambition, its agency

312
00:17:24,560 --> 00:17:28,957
is in fact driving
the accumulation of token capital.

313
00:17:29,040 --> 00:17:31,837
>> Yeah. >> And if we
framed it that way, it's

314
00:17:31,920 --> 00:17:35,197
positive sum for the individual, for the

315
00:17:35,280 --> 00:17:38,797
firm, for the economy, and by the way,

316
00:17:38,880 --> 00:17:42,517
we'll be fine with tokens
doing a lot of the work.

317
00:17:42,600 --> 00:17:45,477
>> Yeah. And you've said
this before, which is

318
00:17:45,560 --> 00:17:47,477
a lot of it is about finding the right

319
00:17:47,560 --> 00:17:49,837
hill to climb. And that'll always be a

320
00:17:49,920 --> 00:17:52,917
very human thing. And when you feedback

321
00:17:53,000 --> 00:17:57,317
to these AI systems and you define what
good and bad looks like, out pops that

322
00:17:57,400 --> 00:18:01,157
private email that you're actually
trying to go and optimize your token

323
00:18:01,240 --> 00:18:03,757
spend towards. But one thing I was

324
00:18:03,840 --> 00:18:05,877
actually curious about was, you know,

325
00:18:05,960 --> 00:18:09,397
there's a big supply and demand mismatch

326
00:18:09,480 --> 00:18:13,197
right now between how people
are using these AI systems.

327
00:18:13,280 --> 00:18:16,277
And one thing I I was just kind
of thinking about this last night, so I

328
00:18:16,360 --> 00:18:20,277
wanted to ask you is, is in a world where,

329
00:18:20,360 --> 00:18:22,477
you know, demand is far outpacing

330
00:18:22,560 --> 00:18:24,797
supply, but there are huge companies

331
00:18:24,880 --> 00:18:29,917
like these model labs buying up a bunch
of capacity for compute, and compute

332
00:18:30,000 --> 00:18:31,677
equals intelligence,

333
00:18:31,760 --> 00:18:34,157
do you worry about a world where there

334
00:18:34,240 --> 00:18:36,997
may be any like unequal distribution of

335
00:18:37,080 --> 00:18:39,797
intelligence across these firms? And

336
00:18:39,880 --> 00:18:42,397
what are the ways that people can start

337
00:18:42,480 --> 00:18:47,597
investing in their future today? Is it
like buying capacity to actually go and

338
00:18:47,680 --> 00:18:50,637
run these models? Is it working
with a hyperscaler like you guys?

339
00:18:50,720 --> 00:18:51,597
>> Yeah, it's a great question. I mean I

340
00:18:51,680 --> 00:18:55,317
mean we do have definitely because of

341
00:18:55,400 --> 00:18:57,877
where the demand has come from and and

342
00:18:57,960 --> 00:19:01,117
then the supply does require time

343
00:19:01,200 --> 00:19:05,597
to get built. So we're definitely in the
throws of a what is a massively supply

344
00:19:05,680 --> 00:19:10,197
constrained market.
But I would say like if I were Uh,

345
00:19:10,280 --> 00:19:12,157
whether it's an AI startup or a large,

346
00:19:12,240 --> 00:19:14,957
sort of bank.

347
00:19:15,040 --> 00:19:20,557
I would start with, "Hey, how do
I create my own hill climbing loop

348
00:19:20,640 --> 00:19:23,877
system?" And that learning loop

349
00:19:23,960 --> 00:19:26,997
in which every day I'm sort of really

350
00:19:27,080 --> 00:19:29,437
creating that private eval benchmark,

351
00:19:29,520 --> 00:19:32,317
right? That is our and my IP.

352
00:19:32,400 --> 00:19:37,637
What's that context layer that is
optimized in fact to use tokens most

353
00:19:37,720 --> 00:19:42,037
efficiently? How do I then
use even uh, the models

354
00:19:42,120 --> 00:19:45,637
that exist today uh, to help me

355
00:19:45,720 --> 00:19:50,797
uh, climb up that benchmark? Uh,
and then maybe take even there and say

356
00:19:50,880 --> 00:19:54,037
what's a way for me to fine-tune a model,
right? So, that's the place to

357
00:19:54,120 --> 00:19:58,237
start. And and that way then you're
participating. And I think that if we

358
00:19:58,320 --> 00:20:00,717
did that I think that some of the supply

359
00:20:00,800 --> 00:20:03,077
challenges will over time get solved. I

360
00:20:03,160 --> 00:20:07,237
think that this concentration of compute

361
00:20:07,320 --> 00:20:10,317
in, you know, is sort of by definition

362
00:20:10,400 --> 00:20:14,477
of physical constraint that over time

363
00:20:14,560 --> 00:20:17,917
will have to solve except for the
network effects. So, that's one of the

364
00:20:18,000 --> 00:20:20,557
reasons why I'm advocating for this

365
00:20:20,640 --> 00:20:23,357
ecosystem piece is, "Hey, the everyone

366
00:20:23,440 --> 00:20:28,197
who's a participant in the economy
has to recognize the network effects of

367
00:20:28,280 --> 00:20:30,037
intelligence are different than even the

368
00:20:30,120 --> 00:20:32,757
classic network effects of digital

369
00:20:32,840 --> 00:20:35,477
technologies." This is not just about

370
00:20:35,560 --> 00:20:37,397
some tech people competing with each

371
00:20:37,480 --> 00:20:41,317
other. It's much more invasive. Yeah.

372
00:20:41,400 --> 00:20:42,757
>> Intelligence is a
different ballgame, right?

373
00:20:42,840 --> 00:20:44,317
>> It's a very different ballgame. And so,

374
00:20:44,400 --> 00:20:47,757
as long as you understands it, then I

375
00:20:47,840 --> 00:20:50,437
think every firm will figure out how to

376
00:20:50,520 --> 00:20:52,997
participate in it versus just be a

377
00:20:53,080 --> 00:20:55,757
bystander renting a model only to be

378
00:20:55,840 --> 00:20:58,477
locked in and never have the even the

379
00:20:58,560 --> 00:21:02,837
freedom even when compute is abundant
to do anything, right? Like that's why I

380
00:21:02,920 --> 00:21:06,637
say like one of the things I wrote
in their article was you can always buy a

381
00:21:06,720 --> 00:21:10,917
tool. You can even outsource a task or
even a job, but you can't outsource your

382
00:21:11,000 --> 00:21:13,277
learning. >> Yes. >> Right?
If you do outsource your

383
00:21:13,360 --> 00:21:15,957
learning, then why exist?

384
00:21:16,040 --> 00:21:18,517
>> Yeah, and and I I
I like the comparisons

385
00:21:18,600 --> 00:21:20,637
you were making between like

386
00:21:20,720 --> 00:21:23,397
technologies in the past and uh you

387
00:21:23,480 --> 00:21:26,277
know, what this new AI paradigm is. Uh

388
00:21:26,360 --> 00:21:30,437
you know, you you you stepped into
the role of of CEO, but you've always kind

389
00:21:30,520 --> 00:21:34,797
of operated with this like founder energy.
People call you a refounder,

390
00:21:34,880 --> 00:21:38,477
right? I'm curious like when it comes
and and and the hallmark of an amazing

391
00:21:38,560 --> 00:21:42,677
leader is sort of leading companies
through these paradigm shifts, much like

392
00:21:42,760 --> 00:21:47,117
you did with the move to cloud. Uh when
you're thinking about this new paradigm

393
00:21:47,200 --> 00:21:51,517
shift to AI, how are you
thinking about practically,

394
00:21:51,600 --> 00:21:53,677
maybe even within Microsoft itself,

395
00:21:53,760 --> 00:21:58,717
changing the way the company operates
in order to really catch the wave and make

396
00:21:58,800 --> 00:22:02,477
sure that um you guys are building
the right things and and sort of

397
00:22:02,560 --> 00:22:05,917
>> Yeah, I think see the magic
of founders are they create

398
00:22:06,000 --> 00:22:09,557
something from nothing, right? I mean,
that's you know, what by definition it's

399
00:22:09,640 --> 00:22:13,397
sort of full risk on and if you're
successful, that means you've created

400
00:22:13,480 --> 00:22:15,477
something that didn't exist at all and

401
00:22:15,560 --> 00:22:20,557
was conceptualized and then executed
on and that takes that energy and and that

402
00:22:20,640 --> 00:22:22,317
completeness of thought and the

403
00:22:22,400 --> 00:22:24,877
relentless pursuit of it. Uh and for

404
00:22:24,960 --> 00:22:27,677
companies that have been successful, uh

405
00:22:27,760 --> 00:22:30,917
right? You sometimes forget what made

406
00:22:31,000 --> 00:22:33,277
you successful initially and that's why

407
00:22:33,360 --> 00:22:36,157
I think uh this refounding as a concept

408
00:22:36,240 --> 00:22:38,437
is empowering, right? As I say to for

409
00:22:38,520 --> 00:22:43,717
mere mortal CEOs like me who come
in later after having, you know, you know,

410
00:22:43,800 --> 00:22:47,037
being in a company that was super
successful, but not being a founder.

411
00:22:47,120 --> 00:22:49,357
>> Yeah. >> But the point
is the opportunity for

412
00:22:49,440 --> 00:22:53,717
anybody, whether you're a CEO or part
of the management team or even an IC in

413
00:22:53,800 --> 00:22:56,277
this company, is to be part of the

414
00:22:56,360 --> 00:23:00,757
learning culture of this place and be
part of that refounding team, right?

415
00:23:00,840 --> 00:23:03,117
That's I think a great empowering thing.

416
00:23:03,200 --> 00:23:06,317
>> Sure. >> Um and the
reason why you have to do

417
00:23:06,400 --> 00:23:09,357
that is, you know, you remember like our

418
00:23:09,440 --> 00:23:13,877
industry in particular always has
been pretty harsh around technology

419
00:23:13,960 --> 00:23:17,957
transitions, right? Either you make it
or you don't, right? You So, you can't

420
00:23:18,040 --> 00:23:20,157
say, "Hey, I'm building

421
00:23:20,240 --> 00:23:23,117
um mainframes in the age of mini or

422
00:23:23,200 --> 00:23:28,277
minis in the age of the PC." Uh it just
moves. And so, when it moves, if you're

423
00:23:28,360 --> 00:23:30,597
not on >> It won't wait.
>> Yeah, it's sort of sort of one-way

424
00:23:30,680 --> 00:23:36,237
doors, right? And so, you have to sort of
move with the technology secular change.

425
00:23:36,320 --> 00:23:40,757
The At some level, I would
say that is the easy part.

426
00:23:40,840 --> 00:23:43,957
>> Yeah. >> This next hard part
is the business model shift.

427
00:23:44,040 --> 00:23:46,557
>> Totally. >> Right?
Because you fall in love with how

428
00:23:46,640 --> 00:23:51,237
you made your margin, you succeeded
in the mainframe or in the mini or in the

429
00:23:51,320 --> 00:23:54,557
PC, and suddenly something new
shows up and you say, "Wow, >> Yeah.

430
00:23:54,640 --> 00:23:58,117
>> this thing doesn't come with the same
margin profile, and I'm so yet I'm

431
00:23:58,200 --> 00:24:00,997
supposed to love it." And so, that's
kind of it's, you know, and the and the

432
00:24:01,080 --> 00:24:05,157
street and others will not give you
permission, right? Everybody in this you

433
00:24:05,240 --> 00:24:09,237
know, in the the street will talk
the long game, but then everyone is short.

434
00:24:09,320 --> 00:24:11,197
>> Yeah. >> Uh but that's
kind of what you have to

435
00:24:11,280 --> 00:24:14,837
earn their permission, figure
out a path. It's always hard.

436
00:24:14,920 --> 00:24:18,197
>> Yeah. >> Uh but so, that you
have to conquer the business model.

437
00:24:18,280 --> 00:24:20,957
And sometimes the change will even come

438
00:24:21,040 --> 00:24:26,237
like this AI change, where even
the way you produce the magic,

439
00:24:26,320 --> 00:24:28,717
>> Yeah. >> which is human
capital using digital

440
00:24:28,800 --> 00:24:32,477
tools to create more software. Like
that's you know, for the first 50 years

441
00:24:32,560 --> 00:24:35,877
of Microsoft, that's kind of what it is,
right? We had to deal with a lot of

442
00:24:35,960 --> 00:24:40,317
changes in technology, right? From the
Altair to the cloud, there was a lot of

443
00:24:40,400 --> 00:24:44,557
change. >> Yeah.
>> But at some level, the production

444
00:24:44,640 --> 00:24:46,837
function was just create more software.

445
00:24:46,920 --> 00:24:49,637
>> Yeah. >> But suddenly
now software has gone from

446
00:24:49,720 --> 00:24:52,837
just being pure input to just output.

447
00:24:52,920 --> 00:24:56,157
>> Yeah. >> And so,
that production function change

448
00:24:56,240 --> 00:25:00,597
is also something now we not only have
to deal with what is going to be changes

449
00:25:00,680 --> 00:25:04,477
in technology. Uh, the good
news here is we were early to

450
00:25:04,560 --> 00:25:07,397
it. So, we are I think we
are good at you know at it.

451
00:25:07,480 --> 00:25:09,597
>> Yeah. >> The business
model is actually pretty

452
00:25:09,680 --> 00:25:14,197
conducive, right? After all the cloud
is a consumptive business model and AI has

453
00:25:14,280 --> 00:25:17,197
got marginal cost. So, it's all
a consumptive business model.

454
00:25:17,280 --> 00:25:19,277
>> Yeah. >> So, the bigger
change for us this time

455
00:25:19,360 --> 00:25:23,077
around is going to be that last change,
which is hey how do we ourselves at

456
00:25:23,160 --> 00:25:26,117
Microsoft and you know completely change

457
00:25:26,200 --> 00:25:30,837
the way we work across every function,
not just software, right? Because you

458
00:25:30,920 --> 00:25:34,677
can't change one function because you
run into the Amdahl's law, right? Which

459
00:25:34,760 --> 00:25:37,037
is you got to have the entire

460
00:25:37,120 --> 00:25:41,877
organization not have impedance mismatch

461
00:25:41,960 --> 00:25:46,237
and work like an AI native company
because an AI native company is born

462
00:25:46,320 --> 00:25:50,077
today, they will think about their human
capital, token capital in the most

463
00:25:50,160 --> 00:25:52,757
effective compounding loop. And we have

464
00:25:52,840 --> 00:25:56,357
to now go from our old to the new new

465
00:25:56,440 --> 00:26:00,717
fast. >> That makes total sense.
And and I think

466
00:26:00,800 --> 00:26:04,517
you know Microsoft enterprises run
on Microsoft. You guys have built this

467
00:26:04,600 --> 00:26:07,597
whole productivity stack on top of it.

468
00:26:07,680 --> 00:26:10,237
And I I I know you've
said in the past that

469
00:26:10,320 --> 00:26:14,917
there's a lot of that that is still very
applicable in the in the age of AI and

470
00:26:15,000 --> 00:26:17,037
people have asked you like you know,

471
00:26:17,120 --> 00:26:19,437
what is the durability of software and

472
00:26:19,520 --> 00:26:24,477
whatnot. Um, I'm curious if if if you
could talk to like how does software in

473
00:26:24,560 --> 00:26:29,077
today's world and people who
build software, durable software

474
00:26:29,160 --> 00:26:32,997
you know, we've seen the markets say oh
software is dead, but really like people

475
00:26:33,080 --> 00:26:35,517
are still using these things. Um,

476
00:26:35,600 --> 00:26:40,117
what do you think happens to software
companies and the products that they

477
00:26:40,200 --> 00:26:44,917
produce? Will they flourish in the age
of AI or eventually people will just

478
00:26:45,000 --> 00:26:47,957
build everything in house?
>> No, I think that

479
00:26:48,040 --> 00:26:53,037
the good way to think about it is perhaps
if you're starting something new.

480
00:26:53,120 --> 00:26:56,277
>> Yeah. >> It's fascinating,
right? Which is

481
00:26:56,360 --> 00:27:00,157
if you believe the firm is going
to have human capital >> Yeah.

482
00:27:00,240 --> 00:27:04,557
>> and token capital. >> Mhm.
>> You then say, "What's the interface

483
00:27:04,640 --> 00:27:07,317
between the human capital
and the token capital?"

484
00:27:07,400 --> 00:27:11,237
>> Yeah. >> And then say,
"Wow, if the next big hit

485
00:27:11,320 --> 00:27:16,197
thing that we produce which
is a fully embodied agent

486
00:27:16,280 --> 00:27:18,997
>> Yeah. >> that needs
to interface, has to show up

487
00:27:19,080 --> 00:27:22,797
in Teams or some other communications
medium." That that means there is some

488
00:27:22,880 --> 00:27:24,797
need for some software. >> Totally.

489
00:27:24,880 --> 00:27:28,517
>> Right? So, that's kind of one. Like oh,
even Excel is fascinating, right? Which

490
00:27:28,600 --> 00:27:29,717
is

491
00:27:29,800 --> 00:27:35,597
Let's say in the model created us most
sophisticated supply chain or financial

492
00:27:35,680 --> 00:27:40,317
model. >> Yeah. >> But I
as a human now need to sort of

493
00:27:40,400 --> 00:27:42,517
make sense of it. Uh

494
00:27:42,600 --> 00:27:46,277
then the best way for that AI to

495
00:27:46,360 --> 00:27:49,277
generate the artifact is to produce a

496
00:27:49,360 --> 00:27:53,637
great Excel model, which by the way
is just not a grid, right? It's got all

497
00:27:53,720 --> 00:27:55,437
sorts of positional >> behind it.

498
00:27:55,520 --> 00:27:57,797
>> got all the positional
information. I as

499
00:27:57,880 --> 00:28:02,477
a human know how to parse the rows and
columns. I can do summations and so on.

500
00:28:02,560 --> 00:28:07,357
And give me an AI agent even right next to
it as a side car that can drive the model.

501
00:28:07,440 --> 00:28:07,797
>> Yeah.

502
00:28:07,880 --> 00:28:10,237
>> Right? So, I feel like software still

503
00:28:10,320 --> 00:28:15,117
has a lot of value. And as long as the

504
00:28:15,200 --> 00:28:18,357
but the pricing adjusts, right?
Which is at some level if you are a very

505
00:28:18,440 --> 00:28:22,397
high-priced piece of software which
was delivering low value >> Yeah.

506
00:28:22,480 --> 00:28:24,877
>> people are going to ask the question,
"Should I just generate it?"

507
00:28:24,960 --> 00:28:27,277
>> Yeah. >> Uh generating
it is one part. After that

508
00:28:27,360 --> 00:28:31,397
you got to keep up with it. You have
to make sure that all the cyber stuff is

509
00:28:31,480 --> 00:28:33,477
being taken care of not one time but

510
00:28:33,560 --> 00:28:36,917
continuously. But that said, I think the

511
00:28:37,000 --> 00:28:42,477
physics have changed and so therefore no
one can just sit on what they produced.

512
00:28:42,560 --> 00:28:45,397
>> Yeah. >> Uh and customers
will question what's

513
00:28:45,480 --> 00:28:47,797
the value of all of this. Uh but at the

514
00:28:47,880 --> 00:28:52,517
end of the day, as long as there
is humans and tokens compounding together,

515
00:28:52,600 --> 00:28:56,477
>> Yeah. >> the need for software is going
to actually explode. I mean, another

516
00:28:56,560 --> 00:28:59,197
example, like we're seeing it.
We're seeing it in GitHub.

517
00:28:59,280 --> 00:29:01,357
>> Yeah. >> Uh painfully so.

518
00:29:01,440 --> 00:29:06,117
Uh and also in M365 where agents as they

519
00:29:06,200 --> 00:29:11,117
explode, what was systems that
were built for human workflow,

520
00:29:11,200 --> 00:29:15,317
>> Exactly. >> now have to completely
like I mean like I mean think >> Yeah.

521
00:29:15,400 --> 00:29:20,277
>> the Git protocol >> Yeah.
>> itself was built for a human workflow.

522
00:29:20,360 --> 00:29:24,037
And so now we have to think about wow,
what does, you know, a work tree mean?

523
00:29:24,120 --> 00:29:26,797
What does agentic workflow mean? Uh with

524
00:29:26,880 --> 00:29:30,477
code. Same thing with Work IQ. Work IQ

525
00:29:30,560 --> 00:29:34,397
cannot be on the same path
as my dial tone when I'm in Outlook.

526
00:29:34,480 --> 00:29:36,717
>> Yeah. >> So we literally
are creating a complete

527
00:29:36,800 --> 00:29:39,077
different middle tier such that the

528
00:29:39,160 --> 00:29:42,317
reasoning loops of AI that are also

529
00:29:42,400 --> 00:29:47,477
interrogating the same data are not
in fact sort of crowding out what is the

530
00:29:47,560 --> 00:29:51,477
human interaction uh with
the data and what have you.

531
00:29:51,560 --> 00:29:53,917
>> I love it.
And And I just maybe the last

532
00:29:54,000 --> 00:29:58,917
question cuz I know I know you have
to run is as Microsoft thinks about, you

533
00:29:59,000 --> 00:30:00,477
know, the next the next couple of years

534
00:30:00,560 --> 00:30:02,997
of this AI shift, there's sort of this

535
00:30:03,080 --> 00:30:06,037
question of what is first-party product?

536
00:30:06,120 --> 00:30:10,317
What is platform? And then what,
you know, as the dynamics and physics are

537
00:30:10,400 --> 00:30:15,237
changing, we should expect companies
to just go and build entirely themselves.

538
00:30:15,320 --> 00:30:19,597
Um do you have some rough sort
of guidelines for the types of products

539
00:30:19,680 --> 00:30:22,557
that Microsoft's going to build, the types
of infrastructure they're going to

540
00:30:22,640 --> 00:30:26,437
support, and what, hey, you know,
companies, you should just go and and

541
00:30:26,520 --> 00:30:29,597
tackle this head on and it's good
for you in the the long run to do that?

542
00:30:29,680 --> 00:30:33,077
>> Yeah, I mean, I think, you know, the
way we've always thought about it and I at

543
00:30:33,160 --> 00:30:34,997
least the the way I've conceptualized

544
00:30:35,080 --> 00:30:37,117
these ecosystems are is like, hey, you

545
00:30:37,200 --> 00:30:40,557
build various layers of the stack.

546
00:30:40,640 --> 00:30:45,037
But each layer of the stack, one our case,
should compose, but it should be

547
00:30:45,120 --> 00:30:46,877
extensible, right? So, what I'm like

548
00:30:46,960 --> 00:30:49,877
even take something like

549
00:30:49,960 --> 00:30:52,637
Co-work, right? Co-work for us is one of

550
00:30:52,720 --> 00:30:57,197
the modalities. We have chat, we have
Co-work, we have autopilot or scout, and

551
00:30:57,280 --> 00:30:59,677
so on, which are all
the modalities of Co-pilot.

552
00:30:59,760 --> 00:31:03,237
>> Yeah. >> But the point
is all of those also got

553
00:31:03,320 --> 00:31:05,237
extensibility, right? So, you can build

554
00:31:05,320 --> 00:31:08,317
an agent that is discoverable inside a

555
00:31:08,400 --> 00:31:12,117
chat or inside Co-work can use it as a

556
00:31:12,200 --> 00:31:15,317
tool and so on. So, that's kind of So,

557
00:31:15,400 --> 00:31:17,957
in other words, we want to build every

558
00:31:18,040 --> 00:31:22,477
layer of the stack as something
that is ecosystem extensible.

559
00:31:22,560 --> 00:31:26,637
>> Yes. >> model diverse >> Yeah.

560
00:31:26,720 --> 00:31:30,597
>> Right? Those are some of the principles
we will have, which is at the end of the

561
00:31:30,680 --> 00:31:35,317
day we want to make sure we are
making each layer competitive

562
00:31:35,400 --> 00:31:40,037
and at the same time making sure that
the customers or third-party developers

563
00:31:40,120 --> 00:31:42,117
have choice. >> Yeah.

564
00:31:42,200 --> 00:31:46,037
>> And not because,
you know, competitively

565
00:31:46,120 --> 00:31:49,117
that is always the best answer for us.

566
00:31:49,200 --> 00:31:52,437
But in the long run, what we've learned

567
00:31:52,520 --> 00:31:58,557
is by interoperating by having
this positive sum construct

568
00:31:58,640 --> 00:32:03,277
>> Yeah. >> we've been able to grow
the market and grow our share.

569
00:32:03,360 --> 00:32:06,037
>> Yeah. People want choice
and they want People want choice and

570
00:32:06,120 --> 00:32:08,597
>> sort of in the long
run it's much better

571
00:32:08,680 --> 00:32:11,717
and more stable going all the way
back to the beginning >> Totally.

572
00:32:11,800 --> 00:32:16,197
>> which is if I said, "Hey, look,
let's not Let's have a frontier >> Yeah.

573
00:32:16,280 --> 00:32:18,717
>> where there is a real ecosystem

574
00:32:18,800 --> 00:32:21,277
participation. Because if you do that,

575
00:32:21,360 --> 00:32:26,717
that positive sum is got lot
more stability for the economy.

576
00:32:26,800 --> 00:32:28,157
>> Well, thank you so much for joining me.

577
00:32:28,240 --> 00:32:31,717
It's I really appreciate your time
and yeah, looking forward to working more

578
00:32:31,800 --> 00:32:33,837
together. >> Likewise.
Thank you so much, Yash.

579
00:32:33,920 --> 00:32:36,360
Thanks.

580
00:32:37,160 --> 00:32:39,180
>> [music]