1
00:00:00,000 --> 00:00:00,729
Frankcx: Hey everyone.

2
00:00:01,119 --> 00:00:06,769
Last week, we argued that the $25
AI token is dying, and this week, AI

3
00:00:06,780 --> 00:00:08,299
gave us another piece of that story.

4
00:00:09,299 --> 00:00:12,659
models are becoming a threshold
that people can start to work with.

5
00:00:13,049 --> 00:00:17,709
And there's a new set of AI models
out of the UAE built for how people

6
00:00:17,709 --> 00:00:22,788
use devices from phones, to PCs, to
servers  Apple NVIDIA, and the other

7
00:00:22,788 --> 00:00:26,948
OEMs are all building machines with
enormous memory pools specific for AI.

8
00:00:27,358 --> 00:00:28,628
What does all this really mean?

9
00:00:28,998 --> 00:00:31,784
It feels like   the AI hardware
boom is really beginning

10
00:00:47,008 --> 00:00:47,838
Frankcx (2): Welcome back.

11
00:00:47,998 --> 00:00:49,178
This is The Local Host.

12
00:00:49,278 --> 00:00:52,638
I'm Frank, and this week I am
your local host once again.

13
00:00:53,098 --> 00:00:58,068
Frankcx: This is show colon 0003, and
in port speak that means we're onto

14
00:00:58,068 --> 00:01:01,428
our third episode so hopefully you've
been following along, but if not,

15
00:01:01,448 --> 00:01:02,928
this is a great chance to catch up.

16
00:01:03,398 --> 00:01:03,988
And guess what?

17
00:01:04,028 --> 00:01:05,968
We've got a couple people that, are out.

18
00:01:06,008 --> 00:01:08,878
Jacob is out enjoying
his first anniversary.

19
00:01:09,038 --> 00:01:11,338
He's down in Disneyland with his wife.

20
00:01:11,398 --> 00:01:12,215
That's super exciting.

21
00:01:12,748 --> 00:01:16,218
And Robert is on a plane somewhere, but
as we're recording this he's sending us

22
00:01:16,258 --> 00:01:20,338
Teams messages , so super funny that he's
wanting to play a part in this as well.

23
00:01:20,648 --> 00:01:23,928
We have a guest all the way
from the United Kingdom, Laura.

24
00:01:24,148 --> 00:01:26,588
So why don't we go around the
horn, and maybe Laura, why don't

25
00:01:26,588 --> 00:01:27,898
you introduce yourself first?

26
00:01:28,791 --> 00:01:29,781
Laura: Thank you so much.

27
00:01:29,831 --> 00:01:31,041
I'm so excited to be here.

28
00:01:31,431 --> 00:01:35,011
Hi, I'm Laura Osborne, joining
from not so sunny London.

29
00:01:35,851 --> 00:01:39,751
I'm a proud member of the AI
Solution Factory, with Microsoft

30
00:01:39,751 --> 00:01:41,191
Surface Engineering team.

31
00:01:41,571 --> 00:01:44,201
So another Microsofty
on this, on this panel.

32
00:01:45,111 --> 00:01:46,581
But yes, very excited to be here

33
00:01:47,826 --> 00:01:51,426
Frankcx: And we have to say
that it is, what, 8:24 PM,

34
00:01:51,753 --> 00:01:52,443
Laura: Thank you

35
00:01:52,743 --> 00:01:56,136
Frankcx: decided to join a bunch of
us on a podcast on a Friday night.

36
00:01:56,146 --> 00:01:56,483
So,

37
00:01:56,783 --> 00:01:57,143
Laura: night.

38
00:01:57,916 --> 00:01:59,336
Frankcx: wild Friday night.

39
00:01:59,766 --> 00:02:00,196
Oh.

40
00:02:00,336 --> 00:02:04,056
Well, maybe, maybe in your life the
wildness doesn't begin until later into

41
00:02:04,056 --> 00:02:07,686
the evening, so this is just setting
you up for amazing conversations

42
00:02:07,686 --> 00:02:08,886
you can have with other people.

43
00:02:11,936 --> 00:02:12,516
All right.

44
00:02:12,586 --> 00:02:13,406
How about Neil?

45
00:02:14,505 --> 00:02:17,135
Neil: Good afternoon from Seattle area.

46
00:02:17,185 --> 00:02:18,565
My name is Neil Mysak.

47
00:02:18,565 --> 00:02:24,225
I'm a member of Surface Engineering, team
member of Laura on the Surface AI factory.

48
00:02:24,525 --> 00:02:28,215
Representing the Chicago Bears
today, NFL season kickoff.

49
00:02:28,255 --> 00:02:32,525
As before, we'll convene next
Friday, so go Bears, go Seahawks.

50
00:02:32,595 --> 00:02:33,376
Over to you, Chauncey

51
00:02:33,676 --> 00:02:34,346
Frankcx: Oh, yes.

52
00:02:34,426 --> 00:02:39,256
Oh, I love the dual Bear-Seahawk,
uh… If it was the Bears and the

53
00:02:39,256 --> 00:02:40,896
Seahawks playing, who would you go with?

54
00:02:41,741 --> 00:02:42,331
Neil: The Bears.

55
00:02:42,831 --> 00:02:44,541
Gotta stay true to my, my hometown team.

56
00:02:44,571 --> 00:02:48,651
But I will root for the Seahawks for any,
against any other team but the Bears, so

57
00:02:49,376 --> 00:02:50,106
Frankcx: That's fair.

58
00:02:50,136 --> 00:02:51,046
still got a week to go.

59
00:02:51,076 --> 00:02:54,186
I'm putting together my fantasy
team, so I'm excited, so.

60
00:02:55,167 --> 00:02:57,247
Chauncey: Are you doing that, are,
are you doing some, like, local

61
00:02:57,277 --> 00:02:58,917
AI crunching for the fantasy team?

62
00:02:59,267 --> 00:03:00,247
Do some model crunching?

63
00:03:00,297 --> 00:03:00,837
You should.

64
00:03:00,880 --> 00:03:01,070
it.

65
00:03:01,167 --> 00:03:01,637
Laura: You should

66
00:03:01,940 --> 00:03:05,440
Frankcx: models and I'm waiting for
Astra to come out because I know that

67
00:03:05,440 --> 00:03:07,260
that's gonna help me win the season, so

68
00:03:07,485 --> 00:03:07,955
Chauncey: I like it.

69
00:03:08,135 --> 00:03:08,545
I like it.

70
00:03:08,975 --> 00:03:14,425
Uh, happy late lunch on a Friday afternoon
everyone from Denver in this case Chauncey

71
00:03:14,425 --> 00:03:19,855
Larson, I also work on the Surface global
marketing team and I-I should just do

72
00:03:19,895 --> 00:03:23,167
a quick reminder that of, of the four
people on this call there are three of

73
00:03:23,195 --> 00:03:27,625
us that actually work at Microsoft our
opinions are entirely on our own we're

74
00:03:27,625 --> 00:03:30,485
not going to discuss anything that's
not already public information We're

75
00:03:30,485 --> 00:03:34,095
not gonna kind of about unannounced
Microsoft hardware or availability or

76
00:03:34,095 --> 00:03:39,935
whatnot so just be aware of that is
definitely just ruminating thoughts And

77
00:03:39,935 --> 00:03:41,475
just kinda geeking out like we love to do

78
00:03:43,530 --> 00:03:44,910
Frankcx: onto this week's show.

79
00:03:45,010 --> 00:03:49,450
we are, watching AI become a new
enterprise hardware category,

80
00:03:49,450 --> 00:03:52,700
and that's the question of the
week: is this, starting to happen?

81
00:03:53,020 --> 00:03:57,820
Last week, we talked about NVIDIA buying
Hugging Face, that's official now.

82
00:03:57,850 --> 00:04:00,890
Hugging Face has become one of the
most important places where developers

83
00:04:00,890 --> 00:04:04,520
discover, download, and distribute
open weight, AI models, and NVIDIA

84
00:04:04,520 --> 00:04:09,040
says they will remain open, and
NVIDIA hardware will not be required.

85
00:04:09,060 --> 00:04:10,890
But I think that raises a bigger question.

86
00:04:11,260 --> 00:04:14,930
If these models are becoming strategically
important, does that place those models

87
00:04:14,930 --> 00:04:19,090
where, you know, strategy, exists
too, especially around sovereignty?

88
00:04:19,110 --> 00:04:22,850
And I'd be interested, does anybody
else have an second take on NVIDIA

89
00:04:22,850 --> 00:04:24,770
buying Hugging Face, that's out there

90
00:04:25,070 --> 00:04:28,470
Neil: I've seen a lot of analogies
to Microsoft's acquisition of

91
00:04:28,470 --> 00:04:34,270
GitHub, And I've seen, Hugging Face
referred to as the GitHub for AI.

92
00:04:34,670 --> 00:04:38,220
And at the time when GitHub was
acquired, there was, some concern

93
00:04:38,230 --> 00:04:39,990
from the open source community.

94
00:04:40,200 --> 00:04:44,870
What's the future of this, really
organic dev-led community now

95
00:04:44,870 --> 00:04:49,050
that it's under the umbrella of
the Microsoft enterprise purview?

96
00:04:49,400 --> 00:04:52,990
Fast-forward a few years and I
think a lot of those concerns

97
00:04:52,990 --> 00:04:55,500
have been resolved or eliminated.

98
00:04:55,710 --> 00:05:00,590
Some of the integrations between GitHub
into Azure DevOps, for example, actually

99
00:05:00,705 --> 00:05:04,490
empowers developers to get their work
done faster, and it still has maintained

100
00:05:04,490 --> 00:05:06,600
this sense of a dev-centric community.

101
00:05:06,980 --> 00:05:08,330
So now take a step back.

102
00:05:08,330 --> 00:05:11,490
There are obviously, some parallels
and some differences between

103
00:05:11,490 --> 00:05:16,620
this moment with Hugging Face
being part of the NVIDIA family.

104
00:05:16,990 --> 00:05:20,240
However, I think from a
security standpoint, ease of

105
00:05:20,240 --> 00:05:24,140
deployment, management, the
topic is enterprise for today.

106
00:05:24,160 --> 00:05:28,670
I think in terms of enterprise readiness,
Hugging Face has been off limits.

107
00:05:28,870 --> 00:05:31,000
You can only pull models
from sanctioned tools.

108
00:05:31,260 --> 00:05:34,970
Now I think there's an opportunity for
a lot of those open source, open weight

109
00:05:34,980 --> 00:05:39,560
models on Hugging Face to now be validated
by enterprises in an expedited fashion

110
00:05:39,890 --> 00:05:41,080
Frankcx: Yeah, I couldn't agree more.

111
00:05:41,150 --> 00:05:45,130
I think it puts a little bit more weight
behind the idea of open weight, right?

112
00:05:45,180 --> 00:05:48,890
All right, let's move on to the second
news that I wanted to bring to the table.

113
00:05:48,890 --> 00:05:53,240
I saw that Citibank and then Vercel,
which operates some AI gateways,

114
00:05:53,280 --> 00:05:58,090
is reporting that they are using
open-weight models more than 50%

115
00:05:58,090 --> 00:06:01,740
of the time for their development
and computation that they're doing.

116
00:06:01,750 --> 00:06:04,260
I think that is just an astounding number.

117
00:06:04,530 --> 00:06:07,830
And in fact, it's happening really
quickly because they were showing that

118
00:06:07,830 --> 00:06:11,865
in June, just a couple months ago,
they were only using it about 29%.

119
00:06:12,760 --> 00:06:17,550
But as of August 25th, they were
up to 53% of their development and

120
00:06:17,570 --> 00:06:21,250
AI workloads that they measure are
happening on open-weight models.

121
00:06:21,470 --> 00:06:25,860
So what that means is they're literally
consuming or building their own AI.

122
00:06:25,940 --> 00:06:29,990
I find it fascinating, and I think
the finserve or financial services

123
00:06:29,990 --> 00:06:34,540
market is ripe for, local AI as
well as, open we- weight models.

124
00:06:34,540 --> 00:06:39,420
But I know that, not to name any names,
but I know Neil and, and Laura, you're

125
00:06:39,420 --> 00:06:42,400
out there talking to these financial
services customers day in and day out.

126
00:06:42,430 --> 00:06:45,240
And again, I'm not asking you
to disclose anything, but hey,

127
00:06:45,250 --> 00:06:46,570
what do you think about this?

128
00:06:46,917 --> 00:06:48,137
Laura: I just have so many questions.

129
00:06:48,187 --> 00:06:49,807
I wanna know what they're running it on.

130
00:06:49,807 --> 00:06:53,597
I wanna know who's leading this, what
kind of workloads are they running on it.

131
00:06:53,627 --> 00:06:58,726
I just, would love to know more about this
Yeah, I've even in some of the handful of

132
00:06:58,726 --> 00:07:03,106
conversations I've been in, or even just
some kind of, research that I've been just

133
00:07:03,106 --> 00:07:07,306
kind of seeing like it's… That the fin
serve industry above all, I think they're

134
00:07:07,306 --> 00:07:12,186
so sensitive about the way that their
data traverses internally, and externally.

135
00:07:12,285 --> 00:07:13,379
Frankcx: How about Disney.

136
00:07:13,379 --> 00:07:17,309
I think that, they had big concerns
about their IP and just the…

137
00:07:17,389 --> 00:07:21,279
Disney is the images that they create
or the characters that they have.

138
00:07:21,689 --> 00:07:26,909
And so being able to have Marvel and
Star Wars and, just old-timey Disney

139
00:07:26,909 --> 00:07:31,539
characters being released into, closed
weight models is a huge concern to them.

140
00:07:31,539 --> 00:07:34,539
And that's their IP, that's how they
make money, that's who they are.

141
00:07:34,539 --> 00:07:38,579
And so, being able to run open weight
models even if they built their own

142
00:07:38,579 --> 00:07:43,599
data center or, being able to deploy
it onto a future RTX Spark device under

143
00:07:43,599 --> 00:07:47,049
the governance of a Windows device,
it just feels like, these are places

144
00:07:47,049 --> 00:07:50,699
where you can protect that IP and have
that sovereignty that they are needing,

145
00:07:50,879 --> 00:07:54,509
Neil: I agree with that, and I'm
seeing really a three-tier model

146
00:07:54,519 --> 00:07:57,649
approach starting to crystallize
in the enterprise space.

147
00:07:57,959 --> 00:08:01,279
There is the AI that you can run
directly on your endpoint, which we've

148
00:08:01,279 --> 00:08:03,249
talked about in episode one and two.

149
00:08:03,529 --> 00:08:06,409
There's cloud AI, which of course
we've talked about, and there's

150
00:08:06,409 --> 00:08:10,199
this emerging middle tier, the edge
server, the department level servers.

151
00:08:10,479 --> 00:08:13,129
GB300s are all the rage right now.

152
00:08:13,359 --> 00:08:17,669
I think RTX Dev Box on the horizon
could also play an important role,

153
00:08:17,859 --> 00:08:22,899
albeit not at the compute capacity that
GB300's bringing to the table, but the

154
00:08:22,899 --> 00:08:27,439
ability for a department to, leverage
some of these more advanced models.

155
00:08:27,479 --> 00:08:32,689
Kimi 3 is all the rage right now,
350 gigs of RAM necessitated.

156
00:08:32,939 --> 00:08:36,849
That middle tier actually offers
a strategic advantage, not just

157
00:08:36,849 --> 00:08:40,389
from a cost savings perspective,
but also from a privacy, right?

158
00:08:40,579 --> 00:08:45,259
Now taking a step back, local, whether
it's on the endpoint or the edge server,

159
00:08:45,449 --> 00:08:50,259
does not equal automatic compliance
from a FERPA perspective, right?

160
00:08:50,269 --> 00:08:51,409
From a HIPAA perspective.

161
00:08:51,659 --> 00:08:56,089
However, no round trip to the
cloud lessens the attack surface,

162
00:08:56,089 --> 00:09:02,169
Frankcx: All right, well speaking
of models, the MBZUA Institute of

163
00:09:02,179 --> 00:09:05,859
Foundational Models, and they've
released six different models

164
00:09:05,859 --> 00:09:07,449
for different hardware classes.

165
00:09:07,699 --> 00:09:11,679
anything from one billion parameters
which could run literally on a phone

166
00:09:11,739 --> 00:09:15,749
like a, Pixel phone, all the way up to
three hundred and seventy-five billion

167
00:09:15,749 --> 00:09:18,709
parameters which is probably more of
a cloud based model that you would run

168
00:09:18,709 --> 00:09:20,759
in that space or a server base model.

169
00:09:21,199 --> 00:09:25,299
But I actually downloaded the thirty-six
billion parameter model and ran it

170
00:09:25,299 --> 00:09:31,389
across, two of my, old dual 3090,
workstation class and honestly, I came

171
00:09:31,389 --> 00:09:33,099
out with some pretty good results.

172
00:09:33,119 --> 00:09:36,649
But I think what's interesting
to me more about these models

173
00:09:36,819 --> 00:09:38,059
is they're not quantized.

174
00:09:38,129 --> 00:09:41,259
They're not like taking a big trillion
parameter model and shrinking it

175
00:09:41,259 --> 00:09:44,689
down and shrinking down the accuracy
which we talked about in episode two.

176
00:09:45,509 --> 00:09:49,389
are literally models that are built
for different classes of devices.

177
00:09:49,409 --> 00:09:54,889
They're trained on the size of the device
that they're gonna be deployed onto so, I

178
00:09:54,889 --> 00:09:59,569
find that fascinating, towards an approach
and I think that still means that people

179
00:09:59,569 --> 00:10:04,009
are taking the time to build models for
different sized hardware because of the

180
00:10:04,019 --> 00:10:07,959
benefit and payoff and the sovereignty
and data protection and, essentially

181
00:10:08,039 --> 00:10:09,949
tokenomics that goes around that.

182
00:10:10,269 --> 00:10:13,209
But I'm wondering if anybody else has
any comments towards this deployment

183
00:10:13,229 --> 00:10:17,419
and this approach that K2 Horizon
is taking over maybe the quantized

184
00:10:17,479 --> 00:10:19,609
approach of, scaling down models

185
00:10:23,264 --> 00:10:26,394
Chauncey: Yeah, I'm actually, as a
hardware guy, I think there are few, you

186
00:10:26,394 --> 00:10:27,644
know, few of us are hardware guys on here.

187
00:10:27,834 --> 00:10:30,534
Like, I think that to me
is very appealing, right?

188
00:10:30,534 --> 00:10:33,054
Because if you start thinking about
being able to really kind of target

189
00:10:33,054 --> 00:10:36,774
the model to the right workload, the
right person, the right device, the

190
00:10:36,854 --> 00:10:38,234
efficiencies that we'll gain out of that.

191
00:10:38,254 --> 00:10:42,094
But also I think like the measure
of capability matching to that

192
00:10:42,094 --> 00:10:44,524
efficiency, 'cause right now when
we quantize, we typically kind of…

193
00:10:45,904 --> 00:10:48,804
We push out some things that we
don't necessarily wanna push out in

194
00:10:48,804 --> 00:10:50,344
places in order to get it smaller.

195
00:10:50,412 --> 00:10:53,704
And so if we can kind of go the other way,
making sure it has the right foundation to

196
00:10:53,704 --> 00:10:57,654
run things, that's very exciting to me in
terms of like what that could potentially

197
00:10:57,654 --> 00:11:01,294
mean for some of these very specific
models, what they would be able to run.

198
00:11:01,334 --> 00:11:04,284
I'll be talking about later on my
device about how I'm trying to run

199
00:11:04,314 --> 00:11:08,184
some CLI, intelligence on a local model
and I'm running on a quantized model.

200
00:11:08,204 --> 00:11:10,564
It's, been an interesting experience.

201
00:11:10,584 --> 00:11:13,534
I'll dive into that later, but
if I can build a model to be able

202
00:11:13,534 --> 00:11:17,624
to do that upfront and be very
kind of generalized approach to

203
00:11:17,624 --> 00:11:19,994
that, I think there's, definitely
some cool future to that for sure

204
00:11:21,159 --> 00:11:23,899
Frankcx: Now I'm wondering, like for
Laura, like as you're talking to customers

205
00:11:23,899 --> 00:11:29,399
there in the UK, you starting to see their
adoption of these open weight models?

206
00:11:29,979 --> 00:11:32,549
and, you know, where are they
finding them, and what are they

207
00:11:32,549 --> 00:11:35,999
gravitating towards, you know,
i- in those different spaces?

208
00:11:36,099 --> 00:11:39,779
it feels like, you know, the, even the
industries across the UK with, from

209
00:11:39,779 --> 00:11:43,739
financial services to other spaces
probably have some duplicat- duplicity to

210
00:11:43,739 --> 00:11:45,509
what the United States is seeing as well.

211
00:11:46,703 --> 00:11:47,553
Laura: Yeah, absolutely.

212
00:11:47,703 --> 00:11:52,043
And I think it's a really interesting
space, particularly within Europe and

213
00:11:52,053 --> 00:11:56,743
the UAE because we have a lot of changing
landscape when it comes to sovereignty

214
00:11:56,743 --> 00:12:01,573
rules and kind of restrictions that are
coming into place more in the cloud scope.

215
00:12:01,573 --> 00:12:06,983
So it's definitely pushing a lot more
customers to consider these things or look

216
00:12:07,023 --> 00:12:09,173
towards open weight models or local AI

217
00:12:09,383 --> 00:12:14,753
And I think particularly UAE as well,
there's so many privacy concerns that

218
00:12:14,803 --> 00:12:19,683
sit, within that region that so many
companies just haven't adopted cloud

219
00:12:19,683 --> 00:12:24,613
at all and therefore AI transformation
has been completely blocked for them.

220
00:12:24,653 --> 00:12:28,313
So this is a really exciting space
and particularly this model coming

221
00:12:28,313 --> 00:12:32,223
out of the UAE is, really exciting to
me 'cause I think a lot of customers

222
00:12:32,223 --> 00:12:35,843
and conversations that I have at the
moment are much more solution driven.

223
00:12:35,853 --> 00:12:39,753
They're not coming from the model
and all of the cool capabilities

224
00:12:39,773 --> 00:12:41,333
that sit on that side first.

225
00:12:41,333 --> 00:12:43,083
It's a bit more in the reverse order.

226
00:12:43,523 --> 00:12:47,653
but this is something that really,
really excites me 'cause I think this

227
00:12:47,653 --> 00:12:52,563
could be a, point that does start
to flip that conversation back again

228
00:12:52,583 --> 00:12:56,883
and combine this level of technology
alongside things like MoE models.

229
00:12:56,913 --> 00:13:00,643
It's super cool, and I think it enables
us to be a bit more strategic when we

230
00:13:00,643 --> 00:13:05,173
think about devices and mapping that
to different users, different devices,

231
00:13:05,473 --> 00:13:07,623
planning headroom on devices around that.

232
00:13:07,643 --> 00:13:09,393
There's so much cool
stuff that can be done

233
00:13:11,019 --> 00:13:14,599
Frankcx: do you hear people
throughout Europe and the UK

234
00:13:14,599 --> 00:13:18,689
and into the regions of the UAE,
gravitating towards specific models?

235
00:13:18,689 --> 00:13:21,859
it's a little bit of a loaded question
'cause Mistral obviously plays a pretty

236
00:13:21,859 --> 00:13:26,439
big role in those open weight, categories
outside of what China is building.

237
00:13:26,929 --> 00:13:32,639
I feel like it's interesting that Mistral
is taking, such a big story around, the

238
00:13:32,799 --> 00:13:36,609
EU's, need for sovereignty and privacy.

239
00:13:36,979 --> 00:13:39,979
and Mistral is kind of like saying,
"Hey, we're here to answer some of

240
00:13:39,979 --> 00:13:43,289
that." but I don't know if you're
hearing those names pop up very often

241
00:13:44,255 --> 00:13:47,075
Laura: Funnily enough,
Mistral not so much, no.

242
00:13:47,115 --> 00:13:49,905
I think it's been a lot of
conversations at the moment have

243
00:13:49,905 --> 00:13:52,465
been really, purpose driven.

244
00:13:52,485 --> 00:13:57,145
So it's things like, Whisper comes up
a lot because people are specifically

245
00:13:57,205 --> 00:14:01,695
looking for transcription and translation
and those kinds of storylines.

246
00:14:02,095 --> 00:14:06,365
which is, you know, if you to
search for what do I use with

247
00:14:06,375 --> 00:14:07,765
transcription, what's a good model?

248
00:14:07,765 --> 00:14:09,905
Whisper's obviously gonna be
the first one that comes up.

249
00:14:09,905 --> 00:14:13,215
So I do think there's still a bit more
of an education piece to be done around

250
00:14:13,215 --> 00:14:18,695
this, and I think it's evolved a lot, and
I think cloud customers that are starting

251
00:14:18,695 --> 00:14:20,745
to make this journey towards local.

252
00:14:21,025 --> 00:14:25,025
Unfortunately, I think the default
has become things like Azure Local and

253
00:14:25,495 --> 00:14:27,775
sovereign cloud options, and I really…

254
00:14:27,785 --> 00:14:31,065
It's one of those things that I really
wanna spearhead into more, that I

255
00:14:31,065 --> 00:14:35,465
think local AI on device AI is being
kind of left out of the picture or

256
00:14:35,465 --> 00:14:38,385
is still trying to catch up and these
emerging technologies that we have

257
00:14:38,395 --> 00:14:43,046
now are going to fill that space for
us hopefully I saw so many use cases

258
00:14:43,056 --> 00:14:48,306
with customers where they had certain
data sets or certain workflows that

259
00:14:48,536 --> 00:14:50,116
they weren't allowed to put into cloud.

260
00:14:50,126 --> 00:14:52,456
They couldn't get the sign-off to
be able to do that, and then it just

261
00:14:52,456 --> 00:14:53,926
became, okay, well, we'll stop here.

262
00:14:53,946 --> 00:14:57,196
so I loved coming along to this
team 'cause I could jump in and

263
00:14:57,196 --> 00:14:58,786
say, "Aha, we have a solution."

264
00:14:59,961 --> 00:15:00,271
Frankcx: Right?

265
00:15:01,011 --> 00:15:04,631
we have devices that are smart
enough to run AI at that level too,

266
00:15:04,731 --> 00:15:05,621
Laura: And that has grown

267
00:15:05,751 --> 00:15:09,201
Gonna say this, that's grown so much
even in the past six months, like since

268
00:15:09,311 --> 00:15:11,421
I've started working on this stuff.

269
00:15:11,421 --> 00:15:16,281
the use cases that we could talk about
now versus then are 10 times more

270
00:15:16,281 --> 00:15:18,101
powerful, which is really, really cool

271
00:15:18,108 --> 00:15:21,306
Frankcx: Well, talking about that
cloud side and the Azure side, I'll

272
00:15:21,306 --> 00:15:25,486
drop in the fifth or the fourth
piece of news is more around OpenAI.

273
00:15:25,546 --> 00:15:30,086
Literally last night, that would be
Thursday, announced the release of Astra.

274
00:15:30,216 --> 00:15:32,656
this is GPT-6 called Astra.

275
00:15:32,656 --> 00:15:36,976
And OpenAI says Astra is the first
model to reach critical cybersecurity

276
00:15:36,976 --> 00:15:39,586
capability in its preparedness framework.

277
00:15:39,636 --> 00:15:44,426
and in testing, it's been known to
find vulnerabilities, and work-working

278
00:15:44,426 --> 00:15:47,823
out exploits in those chains that no
other model has been able to find.

279
00:15:47,966 --> 00:15:52,756
So as a personal story, I've been
developing this small meeting room app

280
00:15:52,756 --> 00:15:57,816
for a, partner of mine, and what I'm
doing is I'm using Cloud Code to actually

281
00:15:57,816 --> 00:16:03,296
write the application, but I'm using Qwen
to do kind of a QA and testing of it.

282
00:16:03,616 --> 00:16:08,126
But I'm super excited to take Astra
because if I can take kinda this hybrid

283
00:16:08,126 --> 00:16:12,986
approach of using both cloud and local
AI to help develop an application,

284
00:16:13,316 --> 00:16:17,276
I could use Astra now to come in and
literally say, "Go look across all the

285
00:16:17,276 --> 00:16:22,746
code as an independent," look at my Git
repository and literally dig into it and

286
00:16:22,746 --> 00:16:26,666
dive into it and give me an answer back
as to, cybersecurity vulnerabilities."

287
00:16:26,696 --> 00:16:28,900
Chauncey: I'll poke a badger and just
say, "Just be careful as you start

288
00:16:28,930 --> 00:16:33,410
setting that A-Astra up," that you
don't want it to maybe go start pinging,

289
00:16:34,120 --> 00:16:36,910
15,000 different chats in some German
chat room somewhere because that's

290
00:16:37,120 --> 00:16:38,690
something that Reuters is reporting on.

291
00:16:39,430 --> 00:16:43,700
So it kind of slipped into the news but I
think that is an interesting take right?

292
00:16:43,880 --> 00:16:49,630
You think about like models
are becoming crazy capable.

293
00:16:49,680 --> 00:16:52,170
I don't wanna say intelligent
because I think we're still trying

294
00:16:52,170 --> 00:16:57,880
to avoid this idea likening them to
AGI at this point But their ability

295
00:16:57,880 --> 00:17:01,490
to solve a problem 'cause that's
effectively what it's doing right?.

296
00:17:01,490 --> 00:17:05,600
if you told it to go fix something or
find something or break something it's

297
00:17:05,600 --> 00:17:08,610
gonna do that And it's going to take
every measure possible to be able to go do

298
00:17:08,610 --> 00:17:14,760
that In this case someone for some reason
thought it to set up some sort of chatroom

299
00:17:14,770 --> 00:17:18,620
Somewhere had literally fifteen thousand
chats where they were talk…It was bots

300
00:17:18,630 --> 00:17:20,820
talking to each other Its almost like that

301
00:17:21,635 --> 00:17:21,745
Frankcx: or

302
00:17:21,910 --> 00:17:22,810
Laura: yeah yea

303
00:17:22,885 --> 00:17:25,265
Frankcx: Facebook for but like Mott yeah.

304
00:17:25,760 --> 00:17:30,660
Laura: But did that on his own which
is really Interesting .So i would

305
00:17:30,660 --> 00:17:34,350
Just Say If Your Going To Use it
Secure your Application Make Sure

306
00:17:35,330 --> 00:17:37,210
giving Some Well Defined Parameters

307
00:17:37,569 --> 00:17:40,149
I just wanna know what's the
gossip between the agents.

308
00:17:40,239 --> 00:17:40,959
What are they all,

309
00:17:41,079 --> 00:17:41,399
Frankcx: I know.

310
00:17:41,399 --> 00:17:42,439
What are they talking about?

311
00:17:43,719 --> 00:17:46,689
I loved all the conspiracy, like
they're creating their own language

312
00:17:46,719 --> 00:17:48,389
that only AIs can understand,

313
00:17:48,689 --> 00:17:50,869
Neil: I think there's two topics here to,

314
00:17:50,894 --> 00:17:51,624
Frankcx: go ahead, Neil

315
00:17:51,949 --> 00:17:53,539
Neil: two topics here to dive deeper on.

316
00:17:53,539 --> 00:17:56,039
One is code generation at the edge.

317
00:17:56,329 --> 00:18:01,159
Frank agreed with your approach of
using these cloud models, as supplements

318
00:18:01,159 --> 00:18:06,639
to your local code generation studio,
as the ability to run 20 billion,

319
00:18:06,989 --> 00:18:10,769
30 billion, 70 billion parameter
models locally on your device.

320
00:18:10,959 --> 00:18:14,009
Seventy billion parameters is where
we're seeing some really effective

321
00:18:14,049 --> 00:18:18,549
codegen models at the edge, and
the hardware floor has risen now

322
00:18:18,599 --> 00:18:20,689
to where, yeah, that's feasible.

323
00:18:20,979 --> 00:18:23,709
However, the cloud can still
serve an important role.

324
00:18:23,779 --> 00:18:25,029
I like the bookend approach.

325
00:18:25,049 --> 00:18:26,819
Sometimes I'll start with the cloud for…

326
00:18:26,939 --> 00:18:30,709
I have an idea, I need to turn
that into a PRD and do the initial

327
00:18:30,709 --> 00:18:34,009
architecture, and then I can do some
of the scaffolding and code locally.

328
00:18:34,249 --> 00:18:37,999
And then at the end, I'll review
with, a cloud-based model as well.

329
00:18:38,199 --> 00:18:41,229
So I agree with the approach
there, in terms of code generation.

330
00:18:41,609 --> 00:18:45,779
On the security side, I think
this is, an area we can go much

331
00:18:45,779 --> 00:18:48,699
deeper into in future episodes.

332
00:18:49,069 --> 00:18:55,019
If the bookend on the tail end is,
reducing code vulnerabilities, I believe

333
00:18:55,029 --> 00:18:58,139
that actually alleviates a lot of the
concerns we're seeing in the market.

334
00:18:58,569 --> 00:19:02,399
With that bookend approach, though,
by scanning for vulnerabilities at

335
00:19:02,399 --> 00:19:06,519
the time of code generation, we get
that, latest and greatest, right?

336
00:19:06,519 --> 00:19:09,749
We get that latest and greatest
as code is being generated, and we

337
00:19:09,749 --> 00:19:11,249
don't have this significant backlog.

338
00:19:11,259 --> 00:19:13,069
All that to say, Frank,
I love the approach.

339
00:19:13,279 --> 00:19:15,299
I think code generation
at the edge is the future.

340
00:19:15,469 --> 00:19:20,599
I think security vulnerability scanning
at the edge is the future as well,

341
00:19:20,599 --> 00:19:24,049
and hopefully our security teams
don't have these massive backlogs when

342
00:19:24,059 --> 00:19:27,709
these frontier models keep, exposing
more vulnerabilities in the future.

343
00:19:28,962 --> 00:19:31,812
Frankcx: I mean, it's good
for the long term but probably

344
00:19:31,812 --> 00:19:33,112
really scary for the short term.

345
00:19:33,412 --> 00:19:33,745
Neil: Totally.

346
00:19:33,935 --> 00:19:36,675
Chauncey: I mean you're definitely
moving in the right direction, right?

347
00:19:36,675 --> 00:19:40,385
Like if you think about even Apple just
kind of came out and mentioned that

348
00:19:40,385 --> 00:19:44,135
they were surprised with the amount
of people that were going out and

349
00:19:44,135 --> 00:19:47,055
buying Mac Minis to run things locally.

350
00:19:47,195 --> 00:19:49,135
and it wasn't just a
bunch of like hobbyists.

351
00:19:49,135 --> 00:19:52,815
enterprises are going out and doing this
which by the way I think points to a

352
00:19:52,825 --> 00:19:57,435
future that we're all looking at from
and Microsoft perspective like this is

353
00:19:57,485 --> 00:19:59,425
obviously a direction that we have to go.

354
00:19:59,735 --> 00:20:04,415
So be really kinda curious Frank, Laura,
Neil like you guys-- you guys are really

355
00:20:04,455 --> 00:20:07,675
kind of hitting the streets and living
this stuff or building it locally and

356
00:20:07,675 --> 00:20:13,355
doing your own thing so give some of that
baseline on what this should mean from

357
00:20:13,355 --> 00:20:16,955
a… What are we really caring about?

358
00:20:17,095 --> 00:20:19,095
What are some of the core
things that people should be

359
00:20:19,295 --> 00:20:20,215
thinking about in this case?

360
00:20:20,515 --> 00:20:24,625
Neil: I'll open by saying that
open-weight models are no longer this

361
00:20:24,905 --> 00:20:28,565
philosophical reference, this pie in
the sky, "Oh, you can have your cake

362
00:20:28,565 --> 00:20:31,975
and eat it too. You can get access
to these very cutting edge models but

363
00:20:31,975 --> 00:20:35,415
it's behind closed doors and it's a
black box." It's becoming a reality.

364
00:20:35,425 --> 00:20:40,215
Again, the case study referenced
of 29% open weight to over 50%

365
00:20:40,215 --> 00:20:41,895
in an abbreviated period of time.

366
00:20:42,225 --> 00:20:48,135
these open-weight models are, becoming
a reality for customers primarily due

367
00:20:48,135 --> 00:20:50,035
to cost concerns, rising cloud costs.

368
00:20:50,085 --> 00:20:51,635
we need more agentic AI.

369
00:20:51,645 --> 00:20:55,515
We wanna run it locally so that the
meter's not spinning with every turn.

370
00:20:55,905 --> 00:20:58,875
That's the hot topic right now,
but then we also talked about

371
00:20:58,875 --> 00:21:02,185
some of the privacy and regulatory
concerns that are alleviated.

372
00:21:02,495 --> 00:21:05,325
And just furthermore to emphasize
your point, Chauncey, that these Mac

373
00:21:05,355 --> 00:21:09,295
Minis aren't just going to hobbyists
folks coding in, mom's basement here.

374
00:21:09,435 --> 00:21:11,855
I've heard multiple times
enterprise customers saying that

375
00:21:11,875 --> 00:21:14,985
Mac Minis are actually part of
their enterprise deployment.

376
00:21:14,995 --> 00:21:16,605
They're in their data centers right now.

377
00:21:16,605 --> 00:21:20,535
They're extending capacity for
these models with Mac Mini.

378
00:21:20,545 --> 00:21:25,045
So, it's no longer this pie in the
sky dream that we're marching towards.

379
00:21:25,285 --> 00:21:27,825
It's a reality right now September 2026

380
00:21:27,826 --> 00:21:29,567
Chauncey: you guys paying attention
to what's coming out of IFA?

381
00:21:29,567 --> 00:21:31,605
I should probably also mention
that IFA is happening right now.

382
00:21:31,965 --> 00:21:34,825
I mean, like even Lenovo just kind
of, came out and started talking

383
00:21:34,825 --> 00:21:39,985
about their RTX Spark adoption
NVIDIA reinforced that they've got

384
00:21:39,985 --> 00:21:41,405
stuff coming out with RTX Park.

385
00:21:41,685 --> 00:21:45,565
So I think like one thing that I am super
excited about is like we keep talking

386
00:21:45,565 --> 00:21:49,605
about Apple Macs and stuff, and I think
because largely Windows has somewhat

387
00:21:49,615 --> 00:21:53,495
suffered in this space, like we just
really haven't had the ability to do

388
00:21:53,495 --> 00:21:56,655
this in a way that our end users and
our customers have been really wanting.

389
00:21:56,905 --> 00:21:59,805
but I think the RTX spark is really
gonna help us change on that front.

390
00:21:59,825 --> 00:22:02,655
And so I'm super excited
to see that really starting

391
00:22:02,655 --> 00:22:03,975
to come together here soon.

392
00:22:04,015 --> 00:22:07,385
and I think like if anything,
IFA is a good example that this

393
00:22:07,385 --> 00:22:11,585
isn't just a Microsoft push, like
this is an industry wide push.

394
00:22:11,635 --> 00:22:15,255
Like it has to happen across
all OEMs, all manufacturers.

395
00:22:15,255 --> 00:22:18,275
I think Laura and Neil, y-you guys
talk to me about all the time about

396
00:22:18,275 --> 00:22:21,885
this idea of like T-shirt sizing and
workload rightsizing and how does

397
00:22:21,885 --> 00:22:27,095
this like and help our customers and
help other people think about like

398
00:22:27,415 --> 00:22:32,045
okay well I might not need a RTX park.

399
00:22:32,245 --> 00:22:36,145
I might need Just for what I do,
I just need some basic device that

400
00:22:36,145 --> 00:22:37,585
still runs a local model though.

401
00:22:37,885 --> 00:22:39,815
and how do we start of that?

402
00:22:39,825 --> 00:22:42,815
So that's actually one thing I'm actually
curious about again the rest of this

403
00:22:42,815 --> 00:22:45,025
team is that kind of work load model.

404
00:22:45,355 --> 00:22:48,795
So far a lot of what we talked
about requires that upper end.

405
00:22:49,145 --> 00:22:52,355
But like Is there some basic
stuff that we can bring locally?

406
00:22:52,805 --> 00:22:57,315
Neil: I'll share an example of a
medical school of the future initiative

407
00:22:57,355 --> 00:22:58,705
that I'm working on right now.

408
00:22:58,905 --> 00:23:04,825
currently the solution is a medical
school training classroom, three different

409
00:23:05,295 --> 00:23:09,805
cameras, multiple microphones, lot of data
collected sent to the cloud for analysis.

410
00:23:10,065 --> 00:23:16,455
The goal is to bring this local primarily
from, an accelerated time to deliver

411
00:23:16,655 --> 00:23:18,505
these results back to the students.

412
00:23:18,555 --> 00:23:21,955
If you recall your time in university,
you took that final exam, you may

413
00:23:21,985 --> 00:23:25,505
not get your results for three
weeks, four weeks, after the exam.

414
00:23:25,505 --> 00:23:26,665
You forgot what it was all about.

415
00:23:26,985 --> 00:23:29,825
Not the type of, feedback we're trying
to give our doctors of the future,

416
00:23:29,825 --> 00:23:31,235
our current medical school students.

417
00:23:31,735 --> 00:23:35,415
We're actually able to accomplish
a lot of this multimodal

418
00:23:35,415 --> 00:23:37,235
analysis with the Surface Pro.

419
00:23:37,515 --> 00:23:44,255
We don't need a beefy device to do some
of the, image and, and audio analysis.

420
00:23:44,275 --> 00:23:48,425
And so I, to the point of T-shirt
size models in Laura's earlier

421
00:23:48,425 --> 00:23:50,415
comment of starting with the workload.

422
00:23:50,645 --> 00:23:54,645
Start with the w- the workload whether
it's cloud based today, maybe it wasn't

423
00:23:54,655 --> 00:23:56,875
feasible in offline settings before.

424
00:23:57,165 --> 00:24:01,205
And, and once we identify that workload,
we can provide the right small language

425
00:24:01,205 --> 00:24:03,025
models to unlock that capability.

426
00:24:03,465 --> 00:24:06,775
The density, the size of those
small language models then

427
00:24:06,775 --> 00:24:08,255
determines the T-shirt size.

428
00:24:08,535 --> 00:24:13,005
We're getting great, great results with
small language models running on 16 gig,

429
00:24:13,025 --> 00:24:19,045
32 gig RAM devices in mobile settings like
with the Surface Pro doing speech to text,

430
00:24:19,355 --> 00:24:21,345
analyzing that transcription locally.

431
00:24:21,395 --> 00:24:24,535
Think about an insurance agent
out in the field snapping a

432
00:24:24,535 --> 00:24:28,425
picture of a damaged vehicle or
car, running an analysis of that.

433
00:24:28,555 --> 00:24:32,575
We don't need to be walking around with
a GB10 or an RTX Spark Laptop Ultra.

434
00:24:32,595 --> 00:24:35,175
We can actually do that with,
what we refer to as more of the

435
00:24:35,175 --> 00:24:38,375
medium or large sized hardware
that we have available today.

436
00:24:38,745 --> 00:24:43,315
And so to tie a bow on this, I think,
you know, for years it was a race to

437
00:24:43,335 --> 00:24:47,195
build the best model, and then we were
building bigger models and running them

438
00:24:47,195 --> 00:24:48,505
on these general purpose accelerators.

439
00:24:49,535 --> 00:24:53,915
Now we're seeing the hardware being
optimized for these specific classes

440
00:24:53,915 --> 00:24:57,775
of models, the T-shirt sizes, and
then delivering these corresponding

441
00:24:57,775 --> 00:24:58,855
workloads to our customers

442
00:25:00,749 --> 00:25:03,399
Chauncey: You kinda mentioned that you're
talking as if this is happening 'cause

443
00:25:03,399 --> 00:25:05,099
I think like it is happening, right?

444
00:25:05,299 --> 00:25:10,479
and I think based on my limited
knowledge that like it's actually

445
00:25:10,489 --> 00:25:12,589
not that terribly difficult either.

446
00:25:12,619 --> 00:25:15,369
Like it is obviously you gotta learn
something new, you gotta dev, you

447
00:25:15,369 --> 00:25:19,029
gotta kind of get your hands dirty
on some code but that like-- That

448
00:25:19,029 --> 00:25:22,559
the platform itself actually somewhat
exists today to be able to adopt this

449
00:25:22,559 --> 00:25:26,739
relatively easy and that I would say
one of our hindrances of this kind of

450
00:25:26,739 --> 00:25:30,059
exploding is just a lack of awareness.

451
00:25:30,109 --> 00:25:32,779
People don't even know that this is
possible and there's only so many

452
00:25:32,779 --> 00:25:35,349
Neils and Lauras that can get out
there and get in front of customers.

453
00:25:35,359 --> 00:25:38,319
So do you feel like
that's-- is that the truth?

454
00:25:38,359 --> 00:25:41,559
I mean are we kind of at the stage
where customers should just kind of be

455
00:25:41,609 --> 00:25:45,349
looking for this or are we still kind
of maybe a few months, a few years

456
00:25:45,349 --> 00:25:46,809
out before that technology catches up

457
00:25:46,826 --> 00:25:50,136
Yeah, And it's some feedback that I
get quite a lot from customers is,

458
00:25:50,476 --> 00:25:54,101
"Oh, it's actually way easier to
implement this stuff than we thought."

459
00:25:54,216 --> 00:25:56,946
There's this kind of expectation that
there's gonna be a whole load of new

460
00:25:57,101 --> 00:26:00,146
tooling required, new skill sets.

461
00:26:00,486 --> 00:26:03,626
But in reality, you're using all
the same tooling that you used to.

462
00:26:03,636 --> 00:26:08,386
You can embed these, models and local
solutions into, if you're developing

463
00:26:08,386 --> 00:26:11,926
with VS Code, if you're using GitHub,
if you're using Copilot CLI, all of

464
00:26:11,926 --> 00:26:15,456
these different, native toolings that
we have at our fingertips, you can

465
00:26:15,456 --> 00:26:18,186
just embed local solutions within that.

466
00:26:18,216 --> 00:26:23,986
And it is so simple, like the examples
that Neil gave are great, and I have the

467
00:26:23,986 --> 00:26:26,106
same kind of, scenarios with customers.

468
00:26:26,106 --> 00:26:29,096
That sometimes it is just something
really simple that they do every day, like

469
00:26:29,096 --> 00:26:33,776
comparing two documents or two invoices
that they need to constantly do this.

470
00:26:34,206 --> 00:26:40,136
They can't afford to have a subscription
for AI for everyone just to be

471
00:26:40,136 --> 00:26:41,506
able to do that one simple task.

472
00:26:41,536 --> 00:26:43,636
But also, it doesn't
make sense to do that.

473
00:26:43,646 --> 00:26:47,956
You can literally just build one simple
tool that does that same solution, uses

474
00:26:47,956 --> 00:26:51,356
local AI to do it, and everyone can
just roll it out on their own devices

475
00:26:53,884 --> 00:26:54,764
Neil: Great call out, Laura.

476
00:26:54,774 --> 00:26:58,384
And I'll say the messaging that's
resonating effectively with customers

477
00:26:58,574 --> 00:27:02,354
for those building with Microsoft Foundry
is a set of large language models in

478
00:27:02,354 --> 00:27:05,564
the cloud that Microsoft has vetted
and it's plug-and-play for developers.

479
00:27:05,944 --> 00:27:09,224
We've now extended Microsoft Foundry
with the Foundry local toolkit.

480
00:27:09,514 --> 00:27:12,488
This is a set of curated models
designed to run locally on the device.

481
00:27:12,544 --> 00:27:15,224
So it's the same SDKs, right?

482
00:27:15,224 --> 00:27:18,224
It's the same deployment and
management tools to get these

483
00:27:18,224 --> 00:27:20,094
models out to your end users.

484
00:27:20,434 --> 00:27:25,234
I will say that the only additional
tooling that some customers are

485
00:27:25,284 --> 00:27:30,034
often adopting is Agent 365,
which is the ability to govern

486
00:27:30,034 --> 00:27:32,034
and manage these solutions.

487
00:27:32,384 --> 00:27:35,174
I met with a financial services
customer, and they said we now have

488
00:27:35,184 --> 00:27:38,684
more agents than human employees,
and getting our arms wrapped around

489
00:27:38,684 --> 00:27:40,114
what are all these agents doing?

490
00:27:40,394 --> 00:27:42,074
Are all these agents authorized?

491
00:27:42,084 --> 00:27:43,284
Should we consolidate some?

492
00:27:43,284 --> 00:27:46,284
Should we promote some of these so
that there's not redundant solutions?

493
00:27:46,604 --> 00:27:50,464
That's where I think the governance
and management and really Microsoft

494
00:27:50,464 --> 00:27:52,124
secret sauce comes into play here.

495
00:27:52,424 --> 00:27:56,954
It's one thing to spin up a local AI
instance using Llama CPP and get an agent

496
00:27:56,964 --> 00:27:58,914
running for one person, for the hobbyist.

497
00:27:59,104 --> 00:28:02,674
It's another thing to deliver
a consistent experience across

498
00:28:02,674 --> 00:28:06,304
your end users, and that's where
we're seeing the pendulum shift.

499
00:28:06,644 --> 00:28:09,814
The question is no longer can
this solution run locally?

500
00:28:10,014 --> 00:28:14,224
It's how can I run this solution
securely in a local environment

501
00:28:15,411 --> 00:28:19,161
Frankcx: So yeah, it's super exciting
that these pieces are coming together

502
00:28:19,161 --> 00:28:22,831
that enterprises can actually start
to adopt and put into that space.

503
00:28:23,331 --> 00:28:27,631
just to keep the show moving, that
brings us into the tension of the week.

504
00:28:28,011 --> 00:28:31,821
so this week's tension, and I'm on
the hot seat, but I'm gonna just kinda

505
00:28:31,821 --> 00:28:35,471
put it out there something a little
spicy, is that last week we talked

506
00:28:35,481 --> 00:28:39,971
about the twenty-five dollar per
million token dying and cloud-based

507
00:28:39,971 --> 00:28:41,931
models starting to reduce their costs.

508
00:28:42,291 --> 00:28:45,501
But I would argue that whereas
the twenty-five d-dollar token

509
00:28:45,511 --> 00:28:49,571
dies, that's when enterprises
and the AI hardware boom begins.

510
00:28:49,971 --> 00:28:52,021
And it's not because
the cloud is going away.

511
00:28:52,261 --> 00:28:57,001
it's not because every AI workload
suddenly belongs on somebody's desk

512
00:28:57,001 --> 00:28:58,671
or inside their private server room.

513
00:28:59,041 --> 00:29:03,601
My case is honestly very simple: open
weight models are becoming capable enough

514
00:29:03,851 --> 00:29:09,761
to work and do real work, and enterprises
are increasingly demanding those weights.

515
00:29:10,071 --> 00:29:13,081
We've seen that from Citi, we've
seen that across the board.

516
00:29:13,451 --> 00:29:17,691
And these model designers, like what we're
seeing out of the UAE, are taking big

517
00:29:17,691 --> 00:29:22,221
models that can work on hardware that is
deployed at different scale, from phone

518
00:29:22,231 --> 00:29:25,661
to PC, to server, to, workstation class.

519
00:29:26,041 --> 00:29:30,581
we're seeing hardware vendors
build specific hardware that

520
00:29:30,581 --> 00:29:32,481
can essentially host local AI.

521
00:29:32,771 --> 00:29:36,381
And Apple with Mac mini is showing
us that these early demands are

522
00:29:36,381 --> 00:29:40,391
much stronger than they anticipated,
especially in the enterprise space.

523
00:29:41,091 --> 00:29:43,391
so that's where this paradox comes.

524
00:29:43,421 --> 00:29:49,031
If intelligence is getting cheaper we
are gonna consume, the thought would

525
00:29:49,041 --> 00:29:50,781
be is we would consume more of it.

526
00:29:51,151 --> 00:29:54,551
Are we gonna start inventing
more and more uses for AI?

527
00:29:54,901 --> 00:29:59,441
more agents, more inference, more
workloads, and we would say that

528
00:29:59,521 --> 00:30:03,481
which intelligence should we rent
and which intelligence should we own?

529
00:30:03,831 --> 00:30:07,601
So I'd ask you to kind of prove me wrong
that, you know, or maybe just agree

530
00:30:07,601 --> 00:30:11,151
along the way, but I've kinda opened it
up for discussion on the open debate.

531
00:30:11,531 --> 00:30:15,111
If the cloud cost pennies, you
know, why would we buy hardware

532
00:30:15,371 --> 00:30:16,941
that can essentially drive it?

533
00:30:16,981 --> 00:30:20,631
But my argument would be there's
more now uses for hardware than ever,

534
00:30:21,001 --> 00:30:24,031
and we would just find the right
place for that in the right time.

535
00:30:24,351 --> 00:30:26,771
But opening up further discussion here

536
00:30:29,876 --> 00:30:34,386
Chauncey: I think a lot of folks have,
strong opinions about cloud usage.

537
00:30:34,876 --> 00:30:39,406
cloud does allow for a lot of scale.

538
00:30:39,766 --> 00:30:45,266
Cloud allows for a lot of very quick
growth, and You can just do things at

539
00:30:45,286 --> 00:30:49,236
that compute level in particular, that
like I think locally, I don't know that

540
00:30:49,236 --> 00:30:51,465
we'll ever really be able to match.

541
00:30:51,706 --> 00:30:55,086
Maybe get really close to, but not quite
exactly match, just because the sheer

542
00:30:55,086 --> 00:30:58,116
amount of power that you can bring from
a data center or collective data centers.

543
00:30:58,726 --> 00:31:05,376
however, local AI, on device AI
really provides, I think more comfort

544
00:31:05,556 --> 00:31:09,686
than anything, like this sense of
control, this sense of ownership.

545
00:31:10,353 --> 00:31:15,846
and when costs are such a variable thing,
especially with AI, because It is still

546
00:31:15,856 --> 00:31:18,616
somewhat of an un-unknown factor, right?

547
00:31:18,616 --> 00:31:22,306
I mean, especially as, models
change in price or we, end up like

548
00:31:22,306 --> 00:31:24,736
as a new model comes out that's
expensive, but we get cheaper models.

549
00:31:24,746 --> 00:31:26,196
Like there's, it's just
a lot of variability.

550
00:31:26,196 --> 00:31:32,506
But if I can control, hey, I know for a
fact that my hardware is okay and that

551
00:31:32,586 --> 00:31:35,836
maybe I do have to buy some kind of
model that's offline, I can control that.

552
00:31:35,836 --> 00:31:37,576
That becomes a very, fixed cost.

553
00:31:37,636 --> 00:31:40,726
and I think that is very
appetizing for a lot of people.

554
00:31:41,076 --> 00:31:47,576
so I think even if the cloud does get down
to pennies, the sheer comfort of having

555
00:31:47,576 --> 00:31:52,996
control over the local piece will be local
AI will continue to be such a big thing

556
00:31:54,431 --> 00:32:00,221
Frankcx: I wanted to pose a question
to Laura on, in the UK and your region

557
00:32:00,841 --> 00:32:05,381
have you started to notice that it's
easier to talk to customers around local

558
00:32:05,381 --> 00:32:07,731
AI and open-weight models and such?

559
00:32:08,251 --> 00:32:12,361
but more on the idea that hey there
was a first wave of AI that came out

560
00:32:12,431 --> 00:32:14,381
that, you know, you only had the cloud.

561
00:32:14,871 --> 00:32:19,341
But people were afraid to push their
IP and their sovereignty, into US

562
00:32:19,341 --> 00:32:23,791
data centers or into sharing, their
intellectual property And now there's

563
00:32:23,791 --> 00:32:28,241
this second wave of these open weight
models that's allowing it to be much

564
00:32:28,241 --> 00:32:32,511
more acceptable and be able to be like
"Now I can actually my company take

565
00:32:32,511 --> 00:32:37,171
part in this AI experience." i don't
know if you've seen that but I guess

566
00:32:37,181 --> 00:32:38,311
I'd be interested in your opinion

567
00:32:40,212 --> 00:32:41,252
Laura: Yeah, for sure.

568
00:32:41,302 --> 00:32:45,022
it's really funny for me actually
coming from spending the last

569
00:32:45,032 --> 00:32:47,282
five years as a cloud architect.

570
00:32:47,342 --> 00:32:50,632
I've been telling everyone to throw
everything into the cloud, push it all

571
00:32:50,632 --> 00:32:52,172
up there, everything's better up there.

572
00:32:52,522 --> 00:32:56,842
And all of those reasons that you
just listed, like, being able to,

573
00:32:56,862 --> 00:33:01,292
you know, have control, everything
being kind of centralized, it's…

574
00:33:01,452 --> 00:33:05,112
Those are all the pushes and it was things
like move from, I can never get this

575
00:33:05,112 --> 00:33:09,432
in the right order, CapEx to OpEx, OpEx
to CapEx, and it just feels like I've

576
00:33:09,432 --> 00:33:13,642
literally done a 360 and I'm like, "Nope,
bring everything back down. Everything's

577
00:33:13,682 --> 00:33:18,552
better on the device." but it really goes
in line with what you just said, Shaunsy,

578
00:33:18,582 --> 00:33:26,112
in terms of we've been so much kind of
almost hate that AI gets, and so much of

579
00:33:26,122 --> 00:33:31,322
that hate is tied to the idea of cloud
and tied to the negative aspects of cloud.

580
00:33:31,322 --> 00:33:36,632
I think like really great PR for AI to be
able to run locally and say, "Actually,

581
00:33:36,632 --> 00:33:41,802
you can get around all of those problems,"
but also the security side of it is such

582
00:33:41,802 --> 00:33:47,772
a… Particularly in the UK and the kind
of Europe with the sovereignty issues

583
00:33:47,772 --> 00:33:52,872
that we have, it's a huge driver and I
think it just answers so many questions.

584
00:33:52,872 --> 00:33:56,672
You don't have to work through
all of these security challenges.

585
00:33:57,052 --> 00:34:00,772
You don't have to think about
cloud space as well, which

586
00:34:00,772 --> 00:34:02,392
in Europe is a huge problem.

587
00:34:02,392 --> 00:34:05,242
Capacity is just at its limits.

588
00:34:05,672 --> 00:34:06,742
and yes, it's funny.

589
00:34:06,752 --> 00:34:11,452
It feels really ironic for me that,
yeah, I'm just slating cloud again now.

590
00:34:11,452 --> 00:34:16,792
I promise I'm not actually doing that, but
yeah, it's definitely an interesting move

591
00:34:18,382 --> 00:34:22,552
Neil: even when cloud is reduced to
pennies, I'll use the insurance example.

592
00:34:22,812 --> 00:34:26,782
If an insurance claim agent is
filing 50 claims a day and we've

593
00:34:26,782 --> 00:34:30,382
got 10,000 insurance agents doing
this, those pennies add up, right?

594
00:34:30,382 --> 00:34:32,702
That's over a million dollars
over the course of a year.

595
00:34:32,702 --> 00:34:37,632
So even these, lower cost workloads,
there's still economic benefit when it's

596
00:34:37,632 --> 00:34:39,412
a repetitive task running it at the edge.

597
00:34:39,902 --> 00:34:44,222
To Laura's point, there are
values beyond just cost savings.

598
00:34:44,222 --> 00:34:48,452
In fact, we've identified eight forces
as to why a workload should run locally.

599
00:34:48,642 --> 00:34:52,522
there's physics, there's economics,
there's regulatory reasons as to

600
00:34:52,522 --> 00:34:55,432
why local is the preferred path.

601
00:34:55,892 --> 00:34:58,752
The public sentiment here, Frank
shared an awesome example of his

602
00:34:58,752 --> 00:35:00,402
conversation with an Uber driver.

603
00:35:00,582 --> 00:35:04,762
I'll just share a brief anecdote of, an
encounter I had at a brewery recently.

604
00:35:05,062 --> 00:35:10,432
I went up to order a beverage, and
there was a lively conversation around,

605
00:35:10,522 --> 00:35:14,802
really this distaste for AI, and
the bartender turns to me and says,

606
00:35:15,122 --> 00:35:20,442
"You're anti-AI, aren't you, bro?"
And to which I hesitated, and I said,

607
00:35:20,452 --> 00:35:22,712
"Actually, I'm an AI solutions architect.

608
00:35:22,892 --> 00:35:27,852
However, I'm focused on getting these
workloads running locally which offers

609
00:35:27,852 --> 00:35:32,472
cost advantage, lower energy draw." And at
the end of my 30 second pitch, I still got

610
00:35:32,472 --> 00:35:33,562
served a beverage. He said, "All right.

611
00:35:33,562 --> 00:35:34,362
you're okay with me."

612
00:35:34,532 --> 00:35:35,252
Laura: You are bad.

613
00:35:35,747 --> 00:35:38,097
Neil: all that to say, I
think this public sen-…

614
00:35:38,150 --> 00:35:38,420
Frankcx: data

615
00:35:38,720 --> 00:35:39,867
Neil: I, I didn't say data center.

616
00:35:39,867 --> 00:35:40,507
Exactly.

617
00:35:40,527 --> 00:35:41,197
Exactly.

618
00:35:41,197 --> 00:35:46,007
So although the emissions 25% increase on,
emissions year over year for Microsoft,

619
00:35:46,227 --> 00:35:49,707
certainly, you know, there's some merit
in those environmental concerns and,

620
00:35:49,907 --> 00:35:54,377
I, I, I believe Local AI offers a, a
path, an exciting path forward to the

621
00:35:54,377 --> 00:35:58,717
future where we can deliver intelligence
back to Zuck's manifesto, right?

622
00:35:58,957 --> 00:36:03,597
Intelligence for every person
across the globe, without creating

623
00:36:03,617 --> 00:36:04,637
unnecessary emissions in the process

624
00:36:06,580 --> 00:36:09,120
Laura: Yeah, there's, a data center
being built in the town that I

625
00:36:09,120 --> 00:36:13,780
live in, and there's been uproar,
there's been riots with people from

626
00:36:13,780 --> 00:36:18,630
an environmental standpoint saying,
you know, AI is ruining everything.

627
00:36:18,640 --> 00:36:24,090
And I feel like the Homer Simpson meme
just like shrinking back into the bushes.

628
00:36:24,100 --> 00:36:28,130
But hope that we can be on the
right side of history in that sense

629
00:36:28,160 --> 00:36:32,250
that some saving to be made there
from a data center perspective

630
00:36:34,760 --> 00:36:37,460
Frankcx: Well then maybe that's a
future show that we have coming up.

631
00:36:37,570 --> 00:36:40,010
But let's get to the motion, the vote.

632
00:36:40,390 --> 00:36:45,900
I had pitched out this idea that as
the $25 token dies and that's when

633
00:36:45,960 --> 00:36:48,750
enterprise AI hardware boom begins.

634
00:36:49,220 --> 00:36:52,670
So as things become cheaper does
that mean we start to use more of it?

635
00:36:52,730 --> 00:36:57,010
And does that open up the gates for
AI hardware to really start to flow

636
00:36:57,010 --> 00:36:58,820
into our commercial industries?

637
00:36:58,850 --> 00:37:00,760
So let's go around the horn.

638
00:37:00,930 --> 00:37:02,420
I will start with Laura.

639
00:37:02,760 --> 00:37:03,680
what's your take on that?

640
00:37:06,308 --> 00:37:07,278
Laura: All for it, yes.

641
00:37:07,378 --> 00:37:10,008
I mean the cost saving is a huge driver.

642
00:37:10,008 --> 00:37:15,618
This idea of tokenomics is becoming
part of my everyday vocabulary now

643
00:37:15,618 --> 00:37:20,648
and I think… I was always a bit
of a cynic that I never expected AI

644
00:37:20,658 --> 00:37:24,998
to stay as cheap as it was or the
kind of unlimited usage that we had.

645
00:37:25,418 --> 00:37:29,158
and in some ways I kind of fear about
the same trend happening from a local

646
00:37:29,168 --> 00:37:33,538
perspective but I definitely think
that it's the, right push that we've

647
00:37:33,538 --> 00:37:36,928
had for customers to just be more
intentional with how they look at

648
00:37:36,928 --> 00:37:39,978
where they're using AI and what they're
doing with it and bringing devices

649
00:37:39,978 --> 00:37:41,758
into that fold and into that strategy

650
00:37:44,686 --> 00:37:45,166
Frankcx: Awesome.

651
00:37:45,166 --> 00:37:46,056
I love the vote.

652
00:37:46,106 --> 00:37:49,866
Chauncey, what's your take on this things
get cheaper, we start to use more of it?

653
00:37:51,311 --> 00:37:52,051
Chauncey: Yeah, I think so.

654
00:37:52,171 --> 00:37:55,251
I hope, the economies of scale
kind of help back this movement.

655
00:37:55,481 --> 00:37:56,951
Like, we kinda need it
to, let's say it that way.

656
00:37:59,071 --> 00:37:59,471
Yes.

657
00:38:01,344 --> 00:38:01,744
Frankcx: All right.

658
00:38:01,804 --> 00:38:03,864
And then Neil, your last one

659
00:38:05,403 --> 00:38:07,733
Neil: This is gonna be an anonymous
vote, for this group today.

660
00:38:07,783 --> 00:38:10,593
If we draw parallels to the '90s
internet boom, it was all about

661
00:38:10,593 --> 00:38:13,093
the servers, and then it was
quickly followed by the PC boom.

662
00:38:13,603 --> 00:38:16,553
Draw parallels now to the AI
moment, and it was all about

663
00:38:16,553 --> 00:38:18,443
the data centers, the GPU boom.

664
00:38:18,793 --> 00:38:22,763
Now it's the AI endpoint boom and,
we have a front row seat to this.

665
00:38:22,773 --> 00:38:24,223
So, exciting times ahead.

666
00:38:24,658 --> 00:38:31,773
edge AI devices are here to stay, and
as hardware continues to increase in

667
00:38:31,773 --> 00:38:37,033
capabilities, software, is quantized
or we're even seeing specific models

668
00:38:37,033 --> 00:38:38,533
developed for specific hardware.

669
00:38:38,743 --> 00:38:40,913
this AI endpoint boom is here to stay

670
00:38:42,484 --> 00:38:46,774
Frankcx: And I would agree, so-- I'll go
right into saying yes, I agree with you,

671
00:38:47,124 --> 00:38:50,884
so we're four for four but that leads us
into kind of like what's on our device

672
00:38:50,884 --> 00:38:52,044
and we'll just finish up with that.

673
00:38:52,054 --> 00:38:56,784
But I did take that thirty-six billion
parameter model, and ran it on the

674
00:38:56,784 --> 00:38:59,344
Beast which is my dual 3090 device.

675
00:38:59,444 --> 00:39:02,824
and I said: "Hey, w-what job do
you get?" Like as in we did that

676
00:39:02,824 --> 00:39:04,444
job interview piece in the past.

677
00:39:04,444 --> 00:39:05,724
And did it get the job?

678
00:39:05,744 --> 00:39:06,994
Actually it does get the job.

679
00:39:07,014 --> 00:39:10,714
it went up against, you know Qwen three
dot eight twenty-seven billion and

680
00:39:10,724 --> 00:39:16,394
almost the exact same qualifying profile
as the test So really surprising how

681
00:39:16,444 --> 00:39:18,444
equal the playing field is these days.

682
00:39:18,454 --> 00:39:22,474
This is a brand new model specifically
not quantized but ran head to head

683
00:39:22,474 --> 00:39:26,254
against a Qwen three dot eight billion…
twenty-seven billion perimeter and

684
00:39:26,254 --> 00:39:29,244
it really came out and showed that
it could perform head to head.

685
00:39:29,944 --> 00:39:33,664
It was giving out about sixty-eight
tokens per second versus forty on a

686
00:39:33,664 --> 00:39:37,754
standard Qwen three dot eight model.Um,
and if I'm hiring it as a local

687
00:39:37,754 --> 00:39:39,654
agent, it's a calling tools agent.

688
00:39:39,864 --> 00:39:43,014
it's great for working through
long prompts and big jup-

689
00:39:43,054 --> 00:39:47,564
continuous jobs and honestly K two
made a very strong case for it.

690
00:39:47,574 --> 00:39:50,344
But there's a little bit of
a catch.Qwen three dot eight

691
00:39:50,344 --> 00:39:54,124
fits onmy single 3090 card.

692
00:39:54,354 --> 00:39:58,234
The K2 needed both of my 3090 cards
so obviously there's a little bit more

693
00:39:58,234 --> 00:40:01,684
expense towards how thirty-six billion
parameters are spread across forty-eight

694
00:40:01,734 --> 00:40:06,300
gigs of beverypowerfulintothatspace.So
anybodyelse outthereplayingon

695
00:40:06,300 --> 00:40:06,320
theirdeviceand testingthings

696
00:40:07,279 --> 00:40:11,699
Chauncey: I-I am on the, training wheels
while you guys are all flying spaceships.

697
00:40:11,749 --> 00:40:14,799
and so I was just kind of messing
around with, Copilot CLI actually

698
00:40:14,799 --> 00:40:18,239
running locally, so I was trying to
get it to run and trying to find the

699
00:40:18,239 --> 00:40:19,979
right model for it to run locally on.

700
00:40:20,089 --> 00:40:22,609
and I used Qwen actually
in this case, Qwen code.

701
00:40:22,729 --> 00:40:25,709
Frankcx: man, it, is-- it was a shocker
that while it was doing this, I literally

702
00:40:25,709 --> 00:40:28,649
asked it a question like, "Hey, like
why are you taking so long?" It's like

703
00:40:28,649 --> 00:40:32,089
well, it's because my parameters are,
pretty limited and I can only do so much

704
00:40:32,189 --> 00:40:35,979
Like you should basically expect about
a two hour response time, for this.

705
00:40:35,979 --> 00:40:39,629
Laura: And so it was a pretty major eye
opening moment for me of like, ah okay.

706
00:40:39,629 --> 00:40:43,889
That's, that's where I think being
able to really help our customers

707
00:40:43,919 --> 00:40:47,889
and, and anyone really just start
to think about like what am I really

708
00:40:47,899 --> 00:40:49,929
trying to do with this specific model?

709
00:40:49,969 --> 00:40:54,539
Because maybe I'm actually not applying
the right thing or the wri-right workload.

710
00:40:54,829 --> 00:40:58,759
Now, I will say once I was able to,
limit the amount of, response and

711
00:40:58,759 --> 00:41:02,389
kind of massage it a little bit, I am
actually using it for coding right now.

712
00:41:02,419 --> 00:41:05,459
it, it is helping at least do some
of the basic planning and some of

713
00:41:05,469 --> 00:41:08,839
the basic stuff, as I'm, I'm kinda
tailoring this other app I'm building.

714
00:41:09,139 --> 00:41:10,719
So it is capable.

715
00:41:10,989 --> 00:41:12,099
is it as fast?

716
00:41:12,239 --> 00:41:13,309
Definitely not.

717
00:41:13,339 --> 00:41:17,679
Like I, you know, will type a, a rude
question in there and then I'll wait for

718
00:41:17,749 --> 00:41:20,989
20 minutes for them to give me another
response but I think we kinda talked

719
00:41:20,989 --> 00:41:24,259
about this on an earlier call which is
if I wanted to do that overnight, maybe

720
00:41:24,259 --> 00:41:28,379
I just wanted to compile something over
night, like why not do that for free?

721
00:41:28,389 --> 00:41:32,249
Frankcx: One of the benefits there
is I'm not costing anyone anything

722
00:41:32,249 --> 00:41:33,919
right now, to be able to run this.

723
00:41:33,999 --> 00:41:34,859
So it's huge

724
00:41:34,896 --> 00:41:37,106
Frankcx (2): say, " Give me your
answer right now?" You know?

725
00:41:37,421 --> 00:41:37,901
Frankcx: Yeah.

726
00:41:37,931 --> 00:41:38,451
Yeah.

727
00:41:39,206 --> 00:41:39,346
Neil: You're

728
00:41:39,461 --> 00:41:40,491
Frankcx: mean-- And it just becomes

729
00:41:40,806 --> 00:41:41,526
Neil: Take your time."

730
00:41:41,826 --> 00:41:43,941
Frankcx: very, focused on-on
kind of driving the right

731
00:41:43,941 --> 00:41:44,851
person for the right models.

732
00:41:44,991 --> 00:41:48,811
but it… yeah, no, I would
maybe, rephrase that to be like,

733
00:41:48,811 --> 00:41:52,091
are you seeing anything that's,
interesting that you've seen that's

734
00:41:52,091 --> 00:41:53,881
been built for on-device as well?

735
00:41:56,397 --> 00:42:00,097
Laura: I did have a cool conversation
with someone today actually who is,

736
00:42:01,197 --> 00:42:07,437
building their own kind of edition of
the Microsoft Speaker Coach That they are

737
00:42:07,517 --> 00:42:11,927
creating something that's almost like a
teleprompter that is wrapped in within

738
00:42:11,927 --> 00:42:16,417
this, that would take your speaker notes,
it would have a look at the slides that

739
00:42:16,417 --> 00:42:20,997
you've got, and it would notice once
you've started to list off some of those

740
00:42:21,027 --> 00:42:24,807
topics that you have in your speaker
notes, it would make them disappear off

741
00:42:24,807 --> 00:42:28,487
the screen and kind of leave you with
anything that you've missed out, anything

742
00:42:28,487 --> 00:42:32,767
to kinda prompt you to, fill the gaps
in your talk track, whatever that is.

743
00:42:32,817 --> 00:42:36,037
which I thought was a really interesting
usage, and that's obviously entirely

744
00:42:36,447 --> 00:42:40,807
on device, but very much a kind of
low spec, device requirement as well.

745
00:42:40,807 --> 00:42:43,467
It's very, very simple small language
models that that would run on.

746
00:42:43,467 --> 00:42:45,727
So that was something that
got me excited this week

747
00:42:46,735 --> 00:42:49,375
Frankcx (2): I feel like I might need
that for the podcast as we're talking

748
00:42:49,425 --> 00:42:49,575
Laura: was

749
00:42:49,668 --> 00:42:50,718
Frankcx (2): follow a script

750
00:42:50,875 --> 00:42:51,785
Laura: are you gonna start building it?"

751
00:42:53,090 --> 00:42:53,390
Frankcx (2): All right.

752
00:42:54,130 --> 00:42:54,860
We'll bring it on.

753
00:42:56,260 --> 00:42:57,970
All right, and Neil,
anything you're doing?

754
00:42:58,843 --> 00:43:02,543
Neil (2): I believe there's, the
right local AI hardware with the

755
00:43:02,543 --> 00:43:04,573
right local model for every user

756
00:43:07,912 --> 00:43:08,642
Frankcx (2): All right.

757
00:43:08,892 --> 00:43:10,772
Well, thank you all.

758
00:43:10,802 --> 00:43:14,212
A big thank you to Laura for
joining us on this Friday night

759
00:43:14,212 --> 00:43:16,032
of yours in the United Kingdom.

760
00:43:16,429 --> 00:43:16,669
Laura: Thank

761
00:43:16,702 --> 00:43:17,322
Frankcx (2): So,

762
00:43:17,389 --> 00:43:17,719
Laura: me

763
00:43:18,262 --> 00:43:21,752
Frankcx (2): You are open…
The door is open at any time for

764
00:43:21,752 --> 00:43:23,202
you to come back and join us.

765
00:43:23,542 --> 00:43:27,332
I'm gonna close up by saying, you
know, with the $25 token that we're

766
00:43:27,332 --> 00:43:32,902
claiming is dying, know, this AI boom
around hardware may just be starting.

767
00:43:32,952 --> 00:43:36,782
we might be seeing that Apple is
selling a lot of Macs for various

768
00:43:36,782 --> 00:43:40,282
reasons, but Bean's started to be
adopted into the enterprise, and we

769
00:43:40,282 --> 00:43:43,402
might be able to say that most of
this still belongs in the cloud, but I

770
00:43:43,402 --> 00:43:45,852
think the direction is worth watching.

771
00:43:46,172 --> 00:43:48,412
cheap AI doesn't kill local AI.

772
00:43:48,722 --> 00:43:52,742
cheap AI is what finally makes
local AI into a category, and

773
00:43:52,742 --> 00:43:54,082
that's the point of the show today.

774
00:43:54,942 --> 00:43:59,272
if this has been worth your time,
a rating out on Apple Podcasts or

775
00:43:59,272 --> 00:44:02,932
Spotify has generally helped us get
this show found, and you can find

776
00:44:02,962 --> 00:44:05,872
everything at thelocalhost.show.

777
00:44:06,812 --> 00:44:11,002
for Laura and Neil and Chauncey
and Robert and, Jacob, who are

778
00:44:11,032 --> 00:44:18,632
off enjoying themselves today,
there is no place like 127.0.0.1.

779
00:44:18,632 --> 00:44:20,182
So have a great weekend, all,

780
00:44:20,547 --> 00:44:21,797
Laura: Have a great extended weekend all.

781
00:44:22,167 --> 00:44:22,647
Thanks all

782
00:44:22,719 --> 00:44:23,329
Neil (2): Happy Labor Day.

783
00:44:23,329 --> 00:44:23,909
Thanks everyone