WEBVTT

1
00:00:00.476 --> 00:00:07.160
I think local AI and open models are going to create a ridiculous number of business opportunities over the next 24 months.

2
00:00:07.200 --> 00:00:10.102
And I don't think most people actually have the map yet.

3
00:00:10.802 --> 00:00:12.063
They've used ChatGPT.

4
00:00:12.103 --> 00:00:13.004
They've used Claude.

5
00:00:13.024 --> 00:00:23.550
But when they hear local AI, Hugging Face, Olama, LM Studio, AI Edge, it sounds like it's for this developer world and that normal founders are just not supposed to touch it.

6
00:00:23.990 --> 00:00:28.053
And I think that's a mistake because the opportunity here is actually pretty endless.

7
00:00:28.613 --> 00:00:46.329
By the end of today's episode, you're going to understand what local AI is, when it matters, how to run open models at work, where Hugging Face fits in here, which Gemma model I'd start with, how I'd run a model locally with LM Studio or Olama, and how this turns into real business ideas.

8
00:00:46.649 --> 00:00:56.635
And I'll give you three startup ideas I'd actually consider building using local AI, including who the customer is, what the first version does, why local matters, and how I'd sell it.

9
00:00:57.015 --> 00:01:03.839
Basically, this is gonna be a masterclass around local AI, how to run models, how to build apps, how to make money from it.

10
00:01:03.859 --> 00:01:07.301
And I'm gonna explain it for the average person who isn't technical.

11
00:01:07.901 --> 00:01:14.427
Quick shout out to Google for sponsoring today's episode and for caring about local AI and open models for entrepreneurs.

12
00:01:14.968 --> 00:01:25.277
Today's episode, I'm going to use Gemma and Google AI Edge as the main examples, but the goal is to give you a full map so you can actually understand the space and build with it and use whatever model suits you.

13
00:01:25.717 --> 00:01:27.019
Okay, let's dive in.

14
00:01:35.921 --> 00:01:40.729
So put simply, local AI means the model runs on hardware you control.

15
00:01:41.150 --> 00:01:42.612
The hardware could be your MacBook.

16
00:01:42.652 --> 00:01:45.757
It could be your Windows laptop, an Android phone, an iPhone.

17
00:01:46.098 --> 00:01:47.941
It could be a browser, a Raspberry Pi.

18
00:01:48.925 --> 00:01:51.226
It could be in a workstation in your office.

19
00:01:51.266 --> 00:01:53.947
I just got a DGX Spark, which is like a high-end one.

20
00:01:54.408 --> 00:01:59.470
But the important part to note is a phone now could run local AI.

21
00:02:00.011 --> 00:02:06.154
Cloud AI means the model runs somewhere else and you access it through a website or an API.

22
00:02:06.194 --> 00:02:07.974
And that's the basic difference.

23
00:02:08.555 --> 00:02:12.297
The business question is, where should the intelligence live?

24
00:02:13.217 --> 00:02:23.070
If I'm doing deep research and strategy and hard reasoning or something where I want the strongest possible model, I'm probably going to be using a frontier cloud model.

25
00:02:23.711 --> 00:02:30.800
If the work involves private files like sensitive customer data, offline usage, fieldwork, low latency,

26
00:02:32.023 --> 00:02:40.030
audio input, or an internal workflow that runs again and again and again, local AI starts to make a lot of sense.

27
00:02:40.751 --> 00:02:45.395
A smaller model in the right place can actually be very valuable.

28
00:02:45.995 --> 00:02:48.538
That is the idea I want you to keep in your head.

29
00:02:48.678 --> 00:02:55.384
The first question most people ask is, is this model smarter than the biggest model in the cloud?

30
00:02:55.864 --> 00:03:01.005
The actual more useful question to ask actually is, is this model good enough for the job?

31
00:03:01.365 --> 00:03:04.286
And does running it locally make the product better?

32
00:03:04.786 --> 00:03:08.947
Once you ask it that way, you start seeing these business opportunities, which we'll go into.

33
00:03:09.607 --> 00:03:13.568
So there's four pieces to the local AI landscape.

34
00:03:14.088 --> 00:03:16.229
The model, which is the brain file.

35
00:03:16.569 --> 00:03:19.990
That could be something like Gemma, Llama, or Mistral.

36
00:03:20.590 --> 00:03:23.371
The warehouse, which is where you find the model.

37
00:03:23.431 --> 00:03:24.691
You might've heard of Hugging Face.

38
00:03:25.231 --> 00:03:28.373
I think they're trying to get acquired right now at $13 billion.

39
00:03:29.573 --> 00:03:30.394
That's what they do.

40
00:03:30.974 --> 00:03:36.477
The software, which is what runs the model, that's something like LM Studio or Ollama.

41
00:03:36.877 --> 00:03:40.799
And then the workflow, which is the product you're building around all of it.

42
00:03:41.240 --> 00:03:43.161
And those are the real four pieces.

43
00:03:43.741 --> 00:03:45.102
The model is the brain file.

44
00:03:45.562 --> 00:03:46.963
Gemma is a model family.

45
00:03:47.343 --> 00:03:48.764
Lama is a model family.

46
00:03:48.784 --> 00:03:49.244
Kodama.

47
00:03:49.364 --> 00:03:51.666
Quinn or Mistral, you might have heard of Phi.

48
00:03:51.946 --> 00:03:53.347
These are model families too.

49
00:03:54.048 --> 00:03:58.531
Some of these are actually better at reasoning and some of them are better at coding.

50
00:03:58.571 --> 00:03:59.632
Some of them are smaller.

51
00:03:59.652 --> 00:04:00.633
Some of them are faster.

52
00:04:00.713 --> 00:04:02.114
Some of them are better for images.

53
00:04:02.514 --> 00:04:05.056
Some are easier to run on your own machine.

54
00:04:05.716 --> 00:04:09.139
And then you need somewhere to find these models.

55
00:04:09.159 --> 00:04:10.580
That's what Hugging Face is.

56
00:04:10.600 --> 00:04:11.961
That's the first place I would go.

57
00:04:12.021 --> 00:04:14.043
They're the biggest at it.

58
00:04:15.052 --> 00:04:18.695
The easiest way to explain Hugging Face is that it's a model warehouse.

59
00:04:19.195 --> 00:04:32.325
You go there and you can find model cards, licenses, file formats, examples, benchmarks, community versions, and sometimes versions that have already been compressed so they're way easier to run locally.

60
00:04:33.025 --> 00:04:40.691
If you're new to local AI, one of the best exercises is actually just to open Hugging Face and read a model card really slowly.

61
00:04:40.711 --> 00:04:41.592
You're going to learn a lot.

62
00:04:42.372 --> 00:04:49.175
At first, though, you can ignore half the scary-looking details and just look for a few basic things, in my opinion.

63
00:04:50.036 --> 00:04:51.316
What is the Model 4?

64
00:04:51.957 --> 00:04:52.997
How big is it?

65
00:04:53.417 --> 00:04:54.978
What license does it use?

66
00:04:55.358 --> 00:04:57.239
What hardware are people running it on?

67
00:04:57.619 --> 00:05:01.861
Does it support text, images, audio, tool use, or embeddings?

68
00:05:02.321 --> 00:05:04.302
Are there quantized files available?

69
00:05:04.823 --> 00:05:09.285
Once you can answer those questions, the space gets a lot more intimidating.

70
00:05:09.305 --> 00:05:09.685
Okay.

71
00:05:10.065 --> 00:05:14.647
Because I know when I first looked at these cards initially, I was like overwhelmed.

72
00:05:14.707 --> 00:05:17.308
So just those are the key questions to ask.

73
00:05:18.008 --> 00:05:20.309
Then you need software that runs the model.

74
00:05:20.709 --> 00:05:24.371
For most people, I would just start with LM Studio or Olama.

75
00:05:24.951 --> 00:05:27.252
LM Studio feels like a normal desktop app.

76
00:05:27.392 --> 00:05:32.094
You download it, you search for the model, you click download, and then you can just chat with it.

77
00:05:33.034 --> 00:05:38.757
My opinion is it's probably one of the most friendly first time user experiences if you're non-technical.

78
00:05:39.719 --> 00:05:44.421
Ollama is a little more builder-oriented or developer-oriented.

79
00:05:44.942 --> 00:05:56.387
You install it, you run a command like olama run gemma 4 colon e4b, and now you have a model running locally with an API your apps could talk to.

80
00:05:56.988 --> 00:06:04.792
Then underneath those tools, you're going to start hearing about things like llama.cpp and mlx.

81
00:06:05.112 --> 00:06:06.813
And I'll explain what those two things are.

82
00:06:07.834 --> 00:06:12.077
Llama.cpp powers a lot of the local model inference.

83
00:06:12.778 --> 00:06:15.580
MLX matters if you're on Apple Silicon.

84
00:06:16.101 --> 00:06:25.088
And if you're thinking about shipping real on-device apps in the Google ecosystem, that's where Google AI Edge and Lite RT LM come in.

85
00:06:25.708 --> 00:06:36.319
Basically, Google AI Edge is the broader on-device AI development world, and Lite RT LM is the runtime layer for the language models.

86
00:06:36.700 --> 00:06:47.911
This is what you study when you want to move from, I ran a model on my laptop to, I want this model inside an iOS app or an Android app or a web app, desktop app, whatever it is.

87
00:06:49.000 --> 00:06:59.444
We got to talk about some key vocabulary, just about the most important things you need to know about these words that come up time and time again in local AI.

88
00:06:59.464 --> 00:07:02.145
I'm just going to give you simple, clear definitions of what they are.

89
00:07:02.285 --> 00:07:08.707
By the end of this part, you'll know just the core basics of local AI vocab.

90
00:07:09.147 --> 00:07:13.550
So I'm sure you've heard this one before, parameters like 2 billion, 4 billion.

91
00:07:13.990 --> 00:07:16.412
These are what's called the internal weights of the model.

92
00:07:16.872 --> 00:07:23.716
And more parameters just usually means more capacity for harder tasks, but it does require more memory.

93
00:07:24.196 --> 00:07:27.418
So parameters are the internal weights of the model.

94
00:07:27.799 --> 00:07:33.042
The beginner shortcut is that more parameters usually means more capacity.

95
00:07:33.402 --> 00:07:36.224
And more capacity can help with the harder tasks.

96
00:07:36.244 --> 00:07:36.304
So.

97
00:07:36.704 --> 00:07:40.426
So the trade-off is usually memory, speed, and hardware.

98
00:07:40.846 --> 00:07:49.511
So a 2 billion or 4 billion model is the kind of thing you might use for edge devices, phones, fast workflows, and smaller tasks.

99
00:07:50.091 --> 00:08:03.238
A 12 billion parameter model is more of a middle ground, and a 26 or 31 billion model is getting to the stronger workstation territory, depending on your hardware and how the model is built.

100
00:08:05.240 --> 00:08:10.924
I recommend not going out there and spending $5,000, $10,000, $20,000 on a workstation just yet.

101
00:08:11.805 --> 00:08:20.591
By the end of this episode, you're going to understand how to just set up some of these things on your phone or on a spare laptop that you have from 2021.

102
00:08:21.151 --> 00:08:22.272
Then there are tokens.

103
00:08:22.332 --> 00:08:26.835
So tokens are the chunks of text that the model reads and writes.

104
00:08:27.415 --> 00:08:31.818
Locally, you care about speed and memory rather than the per token bill.

105
00:08:32.399 --> 00:08:33.920
Then there is the context window.

106
00:08:34.380 --> 00:08:39.205
The context window is basically how much information the model can work with at once.

107
00:08:39.646 --> 00:08:40.847
Then there is quantization.

108
00:08:41.287 --> 00:08:43.269
The word sounds more technical than it needs to.

109
00:08:43.329 --> 00:08:44.951
Honestly, I can barely pronounce it.

110
00:08:45.532 --> 00:08:48.935
Quantization is the compression for models.

111
00:08:49.336 --> 00:08:52.779
It allows giant models to fit on normal laptops.

112
00:08:53.139 --> 00:08:55.462
For example, you might have heard of Q4, Q8 formats.

113
00:08:55.502 --> 00:08:56.243
That's quantization.

114
00:08:58.083 --> 00:09:05.047
If the full model is the giant version, the quantized model is the version that can actually fit on a normal laptop.

115
00:09:05.407 --> 00:09:09.869
So you might lose a little quality, but suddenly this thing magically runs.

116
00:09:10.369 --> 00:09:13.291
You will see things like Q4 or Q8.

117
00:09:13.871 --> 00:09:17.593
And as a beginner rule, Q4 is just usually easier to run.

118
00:09:17.973 --> 00:09:21.515
And Q8 keeps more quality, but it needs more memory.

119
00:09:21.915 --> 00:09:25.277
If you're just getting started, Q4 is just a reasonable place to begin.

120
00:09:25.797 --> 00:09:26.898
So I would start there.

121
00:09:27.258 --> 00:09:29.421
Then there is GGUF.

122
00:09:29.902 --> 00:09:36.511
It's a common file format for local models that make inference easier on normal machines like you and I have.

123
00:09:36.832 --> 00:09:42.299
And in the Google AI Edge world, you'll see something called the Lite RTLM.

124
00:09:42.920 --> 00:09:49.485
This is the model format and runtime path you care about when building on-device apps with LightRTM.

125
00:09:50.005 --> 00:09:51.447
So the simple map is this.

126
00:09:51.907 --> 00:09:55.189
Hugging Face helps you find and understand models.

127
00:09:55.810 --> 00:10:00.514
Gemma is Google's open model family and Google's a trusted brand.

128
00:10:01.454 --> 00:10:04.577
I run my business on top of Google, so it just makes sense.

129
00:10:05.297 --> 00:10:09.060
LM Studio helps you try models locally without much friction.

130
00:10:09.660 --> 00:10:15.985
Olama helps you run models locally in a way that a bit more technical people can plug into apps.

131
00:10:16.665 --> 00:10:20.648
GGUF is a common local model format.

132
00:10:21.109 --> 00:10:28.174
And Google AI Edge and Lite RTLM are the path toward shipping on-device AI products.

133
00:10:28.454 --> 00:10:29.295
That's what you need to know.

134
00:10:29.855 --> 00:10:34.996
So let's talk about Google's open model family because I feel like there's a lot here.

135
00:10:35.156 --> 00:10:36.356
It's a bit overwhelming.

136
00:10:36.916 --> 00:10:43.518
And I'm just going to break it down so you understand what you need to know about the whole Google AI open model family.

137
00:10:44.258 --> 00:10:47.578
So Gemma is Google's family of open models.

138
00:10:47.938 --> 00:10:53.079
And Gemma 4 is built for the efficient, local, and on-device use.

139
00:10:53.099 --> 00:10:57.020
So they have Google 4 E2B, which is the smaller edge model that's built for the efficient, local, and on-device use.

140
00:10:57.480 --> 00:10:58.721
for phone workflows.

141
00:10:59.101 --> 00:11:08.724
You have a bigger E4B, Gemma 4E4B, which is pretty much the most practical starting point for most local tests.

142
00:11:09.145 --> 00:11:15.147
Then you have Gemma 412B, which is a middle ground with more capability for laptops.

143
00:11:15.287 --> 00:11:23.110
And then you have Gemma 426B slash 31B, which is the stronger local workstation territory.

144
00:11:23.710 --> 00:11:26.031
That is the main model picker.

145
00:11:26.671 --> 00:11:32.412
Then, and a lot of people don't know this, there's specialized Gemma models that are just really useful to know.

146
00:11:32.872 --> 00:11:36.213
So you have things like embedding Gemma, which is just for search.

147
00:11:36.913 --> 00:11:42.455
So specifically, it helps you turn text into embeddings, which lets you search by meaning.

148
00:11:42.895 --> 00:11:51.117
If you want to search your own docs or customer notes or support tickets, sales calls or knowledge-based localies, embeddings matter a lot.

149
00:11:51.497 --> 00:11:56.619
Then they have something called Function Gemma, and that's a tool use in structured function calling.

150
00:11:57.339 --> 00:12:02.681
That means the model can help software take actions in a way more structured way.

151
00:12:03.301 --> 00:12:09.783
It is part of the path from the model gave me an answer to the model help the product do the next step.

152
00:12:10.143 --> 00:12:15.084
Then you have a few more like Pali Gemma, which is more vision focused.

153
00:12:15.104 --> 00:12:18.224
You have Shield Gemma, which is more safety focused.

154
00:12:18.484 --> 00:12:23.105
Then you have Gemma Scope, which is more understanding how models work under the hood.

155
00:12:23.585 --> 00:12:26.006
You can leave most of the family alone on day one.

156
00:12:26.386 --> 00:12:31.387
The practical path, like on day one, if you're a beginner, start with Gemma 4E4B.

157
00:12:31.827 --> 00:12:32.987
Understand the workflow.

158
00:12:33.007 --> 00:12:37.608
Then you can move up or down or sideways, actually, depending on what you are building.

159
00:12:38.585 --> 00:12:44.589
So the way I understand the whole Google AI ecosystem is you have Gemma as the open model family.

160
00:12:44.889 --> 00:12:49.552
You have Google AI Edge, which is the on-device AI development ecosystem.

161
00:12:49.993 --> 00:12:57.117
You have Lite RTLM, which is the runtime for running languages models across all the devices.

162
00:12:57.498 --> 00:13:04.642
You have AI Edge Gallery, which lets you try on-device models and see the experience just more directly.

163
00:13:05.043 --> 00:13:06.504
And if you need huge scale...

164
00:13:07.444 --> 00:13:18.411
things like strong managed infrastructure or frontier level cloud reasoning, you still have Gemini and Google Cloud that you can use or another frontier LLM that you can use.

165
00:13:19.231 --> 00:13:26.196
The reality is a lot of big products and serious products are going to use a hybrid setup.

166
00:13:26.236 --> 00:13:31.279
They're going to use cloud for certain things and you're going to use local for other things.

167
00:13:31.798 --> 00:13:35.579
As an example, imagine a local AI tool for a professional service firm.

168
00:13:36.059 --> 00:13:41.461
So the local model is going to read the sensitive drafts, checking for the issues.

169
00:13:41.521 --> 00:13:47.763
It's going to strip or summarize all the private details and prepare a clean version of the problem.

170
00:13:48.303 --> 00:13:52.625
Then when the customer wants deeper reasoning, a cloud model will...

171
00:13:53.385 --> 00:13:55.126
can help with the sanitized version.

172
00:13:55.687 --> 00:14:02.291
That to me feels like a more natural architecture than just putting everything into the cloud, which a lot of people don't want.

173
00:14:02.632 --> 00:14:08.035
You basically have local handling, the private files as a first pass.

174
00:14:08.536 --> 00:14:12.038
And then cloud handles the heavy thinking when you need it.

175
00:14:12.458 --> 00:14:16.301
A human can improve the work before anything important goes out.

176
00:14:16.801 --> 00:14:20.024
This is how I'm starting to think about building a lot of these products.

177
00:14:20.983 --> 00:14:29.790
Beyond Google Gemma, I'll give you a quick primer on the other families or other open model families you'll hear about and some of the pros and cons.

178
00:14:30.462 --> 00:14:41.809
Lama is a Meta's model family, and it's probably the default open model reference point for a lot of developers because it's a pretty big ecosystem.

179
00:14:42.230 --> 00:14:45.732
The upside is the community, the tooling, the examples, support.

180
00:14:46.172 --> 00:14:53.177
The downside is you still need to read the license and the model card, especially if you're building a serious commercial product.

181
00:14:54.527 --> 00:15:03.774
QN is Alibaba's model family and has become very strong, especially around coding, multilingual work, long contacts, and agentic tasks.

182
00:15:04.374 --> 00:15:05.675
The China thing is real.

183
00:15:05.696 --> 00:15:22.008
A lot of people use QN because it performs really well, but if you're in an enterprise, a government, healthcare, finance, or sensitive data environment, you need to separate running open weights locally from sending data to a hosted service and

184
00:15:22.488 --> 00:15:26.172
You need to check what your company is comfortable with or what you're comfortable with.

185
00:15:27.033 --> 00:15:37.963
Deep Seek is similar in the sense that it's made a lot of people realize how strong Chinese-based open models really could be.

186
00:15:39.019 --> 00:15:41.840
especially for reasoning and coding.

187
00:15:42.340 --> 00:15:44.601
The upside is performance and cost.

188
00:15:44.661 --> 00:15:45.341
It's pretty cheap.

189
00:15:45.781 --> 00:15:53.063
The trade-off is that some buyers will have procurement, security, or geopolitical concerns.

190
00:15:54.163 --> 00:16:00.645
So I'd be thoughtful about where I'd use it, how I'd deploy it, even if you want to use it and go down that path.

191
00:16:01.467 --> 00:16:05.308
There's also GLM, or people know it as Z.AI.

192
00:16:06.588 --> 00:16:08.688
It's another one you'll see pop up a lot.

193
00:16:08.708 --> 00:16:10.148
I actually did an episode on it.

194
00:16:10.869 --> 00:16:16.610
Especially if you spend time on Hugging Face and Olama or just local model Twitter, you're going to see it a lot.

195
00:16:17.530 --> 00:16:21.830
The thing to know is that some of these models can be weirdly good for specific jobs.

196
00:16:22.431 --> 00:16:26.151
So I wouldn't ignore them just because they're not the obvious brand name.

197
00:16:26.611 --> 00:16:30.132
You can test them, you can read the model card, you can check the license.

198
00:16:30.732 --> 00:16:31.773
and just play with them.

199
00:16:32.514 --> 00:16:37.661
Um, but what you might deploy in the sense of for your business or for what you're doing might be very different.

200
00:16:38.142 --> 00:16:41.406
There's also a mistral, which is the European model family.

201
00:16:41.426 --> 00:16:43.168
I think they're based in France.

202
00:16:44.050 --> 00:16:44.210
Um,

203
00:16:45.112 --> 00:16:52.278
If you care about efficient models, they do a lot of releasing, a lot of practical developer use cases, they're pretty good.

204
00:16:53.859 --> 00:17:01.705
It's a strong model with a pretty builder-friendly posture, but the downside is the lineup is a little confusing.

205
00:17:02.226 --> 00:17:07.030
Some models are open, some are commercial, so some question marks there.

206
00:17:07.910 --> 00:17:11.034
Microsoft also has their open model family.

207
00:17:11.054 --> 00:17:11.754
It's called Fi.

208
00:17:11.774 --> 00:17:17.460
I think it's interesting if you care about smaller, faster, lower latency models.

209
00:17:18.782 --> 00:17:24.227
But for a lot of use cases, I haven't seen it work very well.

210
00:17:25.448 --> 00:17:28.470
Honestly, there are new models showing up all the time.

211
00:17:28.510 --> 00:17:29.791
It feels like every other day.

212
00:17:30.871 --> 00:17:32.392
That's why hugging face matters.

213
00:17:33.113 --> 00:17:35.234
You're not going there just to find the big models.

214
00:17:35.534 --> 00:17:50.963
You're going to find these weird specialist models, these community fine tunes, quantized versions of stuff, these forks, and model cards that tell you whether something's actually usable for the workflow.

215
00:17:51.403 --> 00:18:13.472
So you don't need to memorize all of this, but the takeaway basically is that there's these ecosystems and your job as a founder or just someone who's playing with these models is to pick a model family that fits your workflow, that you connect with that company, you like how they do things, and then go from there.

216
00:18:13.752 --> 00:18:17.314
You can play with a lot, learn a lot, and then pick a family.

217
00:18:17.974 --> 00:18:20.156
So how do we make this whole thing real?

218
00:18:20.196 --> 00:18:23.478
If you actually want to run Gemma, here's how I would do it.

219
00:18:23.578 --> 00:18:25.480
I would start with LM Studio.

220
00:18:25.500 --> 00:18:27.641
I would download LM Studio.

221
00:18:27.701 --> 00:18:28.702
It's free to download.

222
00:18:29.102 --> 00:18:30.383
You open the app.

223
00:18:30.803 --> 00:18:32.425
You search for Gemma 4.

224
00:18:32.945 --> 00:18:35.707
If your machine is solid, try E4B.

225
00:18:36.247 --> 00:18:40.450
But if your machine is a bit slower or older, I would look for E2B.

226
00:18:41.051 --> 00:18:43.313
And then I would look for the quantized version.

227
00:18:44.333 --> 00:18:49.978
If you're using the GGUF path because you want the model to run just a lot more comfortably.

228
00:18:50.819 --> 00:18:55.083
Once it downloads, open a chat and just ask it something really simple.

229
00:18:56.188 --> 00:19:00.832
I would use a business prompt because I want you to feel the value immediately.

230
00:19:00.892 --> 00:19:02.593
It's sort of an aha moment.

231
00:19:03.273 --> 00:19:11.419
Maybe it's something like, read these customer notes and turn them into a one-page memo about what customers are struggling with.

232
00:19:11.920 --> 00:19:15.843
What has changed and what the business should fix this week.

233
00:19:16.663 --> 00:19:23.688
And then just paste some customer notes or just fake customer notes just if you want to see the value.

234
00:19:23.708 --> 00:19:23.809
Okay.

235
00:19:24.648 --> 00:19:26.649
The point of this exercise is just basic.

236
00:19:27.750 --> 00:19:33.352
The model is now running on your machine and you're using AI without sending that prompt to a cloud model.

237
00:19:33.452 --> 00:19:42.137
I believe everyone should try that and feel what that is because I do think that it's just going to be a lot more common and it's going to unlock your brain in a completely new way.

238
00:19:42.900 --> 00:19:48.104
After that, go to LM Studio's developer section and start the local server.

239
00:19:49.085 --> 00:19:51.067
Because that just gets a lot more interesting.

240
00:19:51.107 --> 00:19:55.230
Because other apps can talk to the model on your laptop.

241
00:19:55.690 --> 00:19:58.132
Your computer becomes this little AI server.

242
00:19:58.673 --> 00:20:08.661
So you can have a script or a prototype or just an internal tool that can call the model through local host and you get an answer back.

243
00:20:09.941 --> 00:20:15.264
I think that's when you start to see how products are going to get built in the modern age.

244
00:20:15.644 --> 00:20:17.405
The second path is Olama.

245
00:20:17.865 --> 00:20:23.068
So install Olama and run olama.pol.jema4.

246
00:20:23.248 --> 00:20:25.789
Then run olama.run.jema4.e4b.

247
00:20:25.809 --> 00:20:26.329
Now...

248
00:20:32.773 --> 00:20:35.675
You have Gemma running locally from a command line.

249
00:20:36.255 --> 00:20:39.877
Olama also gives you a local API port.

250
00:20:40.037 --> 00:20:43.479
I think it's on 11434.

251
00:20:44.159 --> 00:20:47.761
It is useful because you can connect your own app or script to it.

252
00:20:48.541 --> 00:20:58.907
If you want to test a larger model later, you can try the 12 billion, 26 billion, or 31 billion versions, assuming your hardware can handle it.

253
00:21:00.023 --> 00:21:06.646
And you can ask an LLM if your hardware can handle it, or you can do it yourself and just suffer through the slowness and the pain of it.

254
00:21:08.206 --> 00:21:12.888
The third path is Google AI Edge and LiDAR TLM.

255
00:21:13.409 --> 00:21:18.771
I would only use this path if I wanted to build an actual app and a model inside of it.

256
00:21:19.211 --> 00:21:24.033
For example, maybe I'm building a mobile app and the model is running on the phone.

257
00:21:24.693 --> 00:21:28.075
Or it could be like a browser app where the model runs locally.

258
00:21:29.315 --> 00:21:34.558
Or maybe it's a desktop app with a private workflow or something on an edge device.

259
00:21:35.418 --> 00:21:39.441
Light RT LM is designed for that world.

260
00:21:40.761 --> 00:21:45.444
Android, iOS, web, desktop and edge environments.

261
00:21:45.564 --> 00:21:51.767
That is the path from local AI as a demo to local AI as a product.

262
00:21:52.696 --> 00:21:55.419
So here's the hardware cheat sheet that I would use.

263
00:21:55.779 --> 00:22:00.344
If you have 8 gigabytes of RAM, start small and keep the first test simple.

264
00:22:00.825 --> 00:22:08.974
But if you have something like 16 gigabytes of RAM, you can do some useful experiments with models like E4B and smaller quantized models.

265
00:22:09.594 --> 00:22:11.416
If you have 32 gigabytes of RAM,

266
00:22:13.118 --> 00:22:16.622
you have way more room to do larger local workflows.

267
00:22:17.062 --> 00:22:25.651
If you have a strong GPU or a workstation like a DGX Spark, the bigger models become just much more realistic.

268
00:22:25.871 --> 00:22:30.395
And for phones, I would think a lot less about model size and more about the job.

269
00:22:30.496 --> 00:22:32.758
So can the model understand a photo?

270
00:22:33.078 --> 00:22:34.500
Can it summarize audio?

271
00:22:34.560 --> 00:22:36.221
Can it classify something quickly?

272
00:22:37.155 --> 00:22:39.156
Can it help a worker in the field?

273
00:22:39.436 --> 00:22:41.237
Can it run without a strong connection?

274
00:22:41.717 --> 00:22:47.720
Can it do something useful inside the app before the user even thinks to ask?

275
00:22:48.380 --> 00:22:51.782
Now, let's build the first workflow in our heads.

276
00:22:52.450 --> 00:22:56.693
So I would make a folder on your desktop called customer notes.

277
00:22:56.953 --> 00:23:02.117
And inside that folder, I'd put 10 support tickets for a specific business.

278
00:23:02.737 --> 00:23:08.882
Let's say it's a home health agency or a med spa or a water damage restoration company.

279
00:23:09.407 --> 00:23:13.630
The notes might say something like, I tried to reschedule but couldn't find the link.

280
00:23:13.650 --> 00:23:16.712
Or the technician didn't explain what happens next.

281
00:23:16.812 --> 00:23:19.534
Or, hey, no one actually confirmed my appointment.

282
00:23:19.694 --> 00:23:21.316
Or, I was charged twice here.

283
00:23:21.756 --> 00:23:32.364
Then I would run a local model like Gemma and ask it to create a file called whatcustomersaretellingus.md, the markdown file.

284
00:23:33.024 --> 00:23:46.039
The output should include the repeated complaints, the exact customer language, the likely root cause, the part of the business that seems broken, and the one thing the operator should test this week, the high-priority stuff.

285
00:23:46.419 --> 00:23:50.985
This is a good first local AI workflow because it's useful and it's simple.

286
00:23:51.465 --> 00:23:52.667
What do you have here, right?

287
00:23:52.687 --> 00:23:54.829
You have this private, messy data.

288
00:23:55.170 --> 00:23:56.711
The model runs next to it.

289
00:23:56.751 --> 00:24:00.556
And the output is a memo someone could actually use.

290
00:24:01.097 --> 00:24:07.684
And then once you actually go and, you know, you're going to go and do this and get the output, you're going to, like the unlock I was talking about.

291
00:24:08.625 --> 00:24:31.845
before like it's going to unlock something in your brain you're going to see this pattern everywhere a folder of customer calls become a market research memo a folder support tickets become a product roadmap signal a folder of pdfs become like a risk checklist a folder of drafts become a pre-send reviewer this is why i always start with workflows before i'm fine-tuning anything

292
00:24:32.585 --> 00:24:38.589
People hear open model and immediately want to train their own model.

293
00:24:38.689 --> 00:24:39.530
And I get it.

294
00:24:39.670 --> 00:24:40.210
I get why.

295
00:24:40.250 --> 00:24:41.291
I was actually the same way.

296
00:24:42.051 --> 00:24:43.272
It sounds really cool.

297
00:24:43.592 --> 00:24:45.954
But I feel like that's like an advanced move.

298
00:24:46.214 --> 00:24:51.157
The practical move, the beginner move, where you should start is just to find a repeated workflow.

299
00:24:51.237 --> 00:24:57.502
First, you pick one folder, one model, one output, and you run it like 10 times.

300
00:24:58.220 --> 00:24:59.862
You see where it gets confused.

301
00:24:59.882 --> 00:25:03.066
You see where you can improve the prompt and add examples.

302
00:25:03.106 --> 00:25:06.430
You add a checklist and then you create a small eval.

303
00:25:07.111 --> 00:25:07.972
You know, what's an eval?

304
00:25:07.992 --> 00:25:13.819
An eval is just a small, it's just a test that tells you whether the model did the job well enough.

305
00:25:14.460 --> 00:25:19.024
For this workflow, for example, the eval could be really simple.

306
00:25:19.044 --> 00:25:31.256
It could be like, take the same 10 customer nodes and run them through Gemma locally, and then run them through a strong cloud model, a frontier model, and then just compare the outputs.

307
00:25:31.856 --> 00:25:36.060
And then you ask, did Gemma catch the same complaints?

308
00:25:36.120 --> 00:25:38.562
Did Gemma pull the right quotes and...

309
00:25:39.283 --> 00:25:40.724
Did it follow the format?

310
00:25:40.744 --> 00:25:41.845
Did it miss something?

311
00:25:42.366 --> 00:25:53.394
The comparison actually teaches you where local is already useful and where you still want that stronger cloud model and how you should think about the hybrid model I was talking about.

312
00:25:53.815 --> 00:25:57.718
That's really how I think about local versus cloud decisions.

313
00:25:58.318 --> 00:26:07.366
Use local for private, repetitive, fast, offline, device native, and high volume workflow, stuff that you want to run all the time.

314
00:26:07.926 --> 00:26:16.673
You use cloud for deep reasoning, giant context, broad research, and cases where the strongest model changes the quality of the answer.

315
00:26:17.113 --> 00:26:22.297
So you use both when the product has sensitive data and hard reasoning.

316
00:26:22.758 --> 00:26:25.580
A lot of valuable products will work that way.

317
00:26:25.760 --> 00:26:31.044
You know, local first pass, you do the cloud escalation, human approval for anything important.

318
00:26:31.064 --> 00:26:33.466
I think that's the way work's going to get done.

319
00:26:34.002 --> 00:26:39.125
So I want to give you three startup ideas where local AI actually matters.

320
00:26:39.205 --> 00:26:48.611
And these are the kind of businesses I would look for, niche, useful, cash flowing businesses that you don't need to raise venture for, and tied to a painful workflow.

321
00:26:49.091 --> 00:26:51.332
The filter is pretty straightforward.

322
00:26:51.372 --> 00:26:57.296
So I look for a customer with sensitive data, repeated review work, bad software usually,

323
00:26:58.176 --> 00:27:05.660
expensive mistakes, mistakes that will cost them a lot, and a workflow that happens close to the device.

324
00:27:06.240 --> 00:27:09.642
That combination is the interesting zone for me.

325
00:27:10.082 --> 00:27:11.443
So let's go through the three ideas.

326
00:27:12.303 --> 00:27:21.808
I want you to steal these ideas, and at the very least, it'll get your creative juices flowing with how you can use local AI to run model, build apps, and make money.

327
00:27:22.410 --> 00:27:28.353
Idea number one is a local QA reviewer for home health agencies.

328
00:27:28.573 --> 00:27:39.038
So home health agencies have nurses and caregivers and they go into people's homes and they write, you know, visit notes and updating care plans and dealing with billing and compliance.

329
00:27:39.609 --> 00:27:40.670
The paperwork is a pain.

330
00:27:40.710 --> 00:27:45.612
It takes a lot of time if you've ever witnessed it in person, but it matters so, so much.

331
00:27:45.712 --> 00:27:52.215
Like a missing detail can create a billing delay and a vague note can create extra admin work.

332
00:27:52.495 --> 00:27:58.958
And a mismatch between the visit and the care plan can create a ton of risk and we don't want that.

333
00:27:59.338 --> 00:28:03.882
So the first version is a local desktop app for the agency.

334
00:28:04.383 --> 00:28:09.407
The agency drops in visit notes and care plans and dictated transcripts.

335
00:28:09.767 --> 00:28:14.752
And then the model is going to review them before the submission and should look for flags.

336
00:28:15.252 --> 00:28:19.555
So it's going to flag things like this note mentions dizziness, but vitals are missing.

337
00:28:19.835 --> 00:28:25.638
Or the caregiver described a medication change, but the follow-up instructions is pretty unclear.

338
00:28:26.019 --> 00:28:29.381
Or the note may not support the billed service level.

339
00:28:29.921 --> 00:28:36.925
The buyer mostly cares about fewer documentation problems before the billing or the audit or a supervisor review.

340
00:28:37.385 --> 00:28:39.707
So if you solve that, you have their attention.

341
00:28:39.727 --> 00:28:40.167
Okay.

342
00:28:40.327 --> 00:28:42.469
Now, I don't want to just give you the idea.

343
00:28:42.510 --> 00:28:44.572
I mean, how would you actually grow this?

344
00:28:44.672 --> 00:28:46.774
If I was starting this business, how would I grow this business?

345
00:28:47.155 --> 00:28:48.816
I would actually start it as a service.

346
00:28:49.277 --> 00:28:55.143
So I would find five small home health agencies and I would offer to review a batch of notes.

347
00:28:55.484 --> 00:28:59.768
I would do the review with AI, helping behind the scenes with the local AI.

348
00:29:00.649 --> 00:29:05.110
And I would inspect everything manually with human beings, myself first.

349
00:29:05.630 --> 00:29:08.591
I would write down the 20 issues that keep showing up.

350
00:29:09.031 --> 00:29:10.691
And those issues become the checklist.

351
00:29:10.991 --> 00:29:13.072
And then the checklist eventually becomes the product.

352
00:29:13.112 --> 00:29:20.493
So you have this wedge, it's pretty simple, where you're catching documentation problems before they cost the agency time or money.

353
00:29:20.833 --> 00:29:21.933
And then you build from there.

354
00:29:22.514 --> 00:29:24.554
I love this business and totally would start it.

355
00:29:24.963 --> 00:29:31.809
The second startup idea is an offline field report co-pilot for restoration contractors.

356
00:29:31.889 --> 00:29:37.414
So think water damage or fire damage or mold remediation, things like that.

357
00:29:37.814 --> 00:29:46.582
Those teams are out there field taking photos, recording notes, documenting damage and creating reports for homeowners and insurance adjusters.

358
00:29:46.962 --> 00:29:49.545
I unfortunately had this, so I know a little bit about it.

359
00:29:50.146 --> 00:29:52.208
The job is actually pretty visual.

360
00:29:53.209 --> 00:29:54.690
It's also physical, right?

361
00:29:55.071 --> 00:29:56.652
It happens away from a desk.

362
00:29:56.993 --> 00:30:05.542
And the report matters because the report becomes the handoff between the technician, the customer, the office, and the insurance process.

363
00:30:05.562 --> 00:30:07.304
So how would we build a product here?

364
00:30:07.824 --> 00:30:09.305
The first version is a mobile app.

365
00:30:09.485 --> 00:30:17.411
So a technician walks through the property, takes photos, record voice notes, and the app drafts the report before they leave the site.

366
00:30:17.871 --> 00:30:22.114
So it can flag missing pieces while the technician is still there walking around.

367
00:30:22.494 --> 00:30:24.816
You mentioned the basement, but there are no basement photos.

368
00:30:25.296 --> 00:30:30.439
You took a photo of ceiling damage, but there are no moisture meeting reading, things like that.

369
00:30:30.720 --> 00:30:33.882
Or maybe like the affected room is like missing.

370
00:30:34.322 --> 00:30:38.066
Could be the homeowner explanation is way too technical.

371
00:30:38.406 --> 00:30:41.089
Here's a clearer version they can understand.

372
00:30:41.449 --> 00:30:43.311
And the last part of that is underrated.

373
00:30:43.552 --> 00:30:48.457
In a stressful home damage situation, clear communication is part of the product, right?

374
00:30:49.558 --> 00:30:51.339
So if you had that, that would be key.

375
00:30:52.000 --> 00:30:53.282
How would I grow this business?

376
00:30:53.322 --> 00:30:55.004
Well, I would pick one niche first.

377
00:30:55.024 --> 00:30:56.205
I wouldn't go after everything.

378
00:30:56.265 --> 00:30:59.829
So say I'm going after water damage restoration.

379
00:31:00.229 --> 00:31:01.811
I would talk to owner operators.

380
00:31:01.891 --> 00:31:11.502
I'd look at their current report templates, study the software they use, which is some old stack, and I'd build around the checklist that's already in their head.

381
00:31:12.243 --> 00:31:14.945
The demo is actually the easy part.

382
00:31:14.985 --> 00:31:20.529
Send me three old jobs and I'll show you how fast your techs could create reports.

383
00:31:21.169 --> 00:31:23.611
If that works, then the product could expand from there.

384
00:31:23.651 --> 00:31:24.692
That's just the wedge, right?

385
00:31:25.472 --> 00:31:31.656
It can go into QA and estimates and insurance packets, customer updates, and training new technicians.

386
00:31:31.716 --> 00:31:35.519
But I would start with the field report because it's specific and obviously super annoying.

387
00:31:36.300 --> 00:31:38.041
And I just think that...

388
00:31:40.942 --> 00:31:44.572
When you look at some of these old softwares that these people are using,

389
00:31:45.745 --> 00:31:53.530
I recently had some water damage at my apartment and I was seeing some of the software and it's antiquated.

390
00:31:53.771 --> 00:31:56.032
It's stuff from the early 2000s.

391
00:31:56.072 --> 00:32:03.738
So I think that there's this opportunity to create local AI native software and wedge now.

392
00:32:03.758 --> 00:32:10.382
And that's why I said in the beginning, I think there's a 24-month window and opportunity to do some of these products.

393
00:32:10.542 --> 00:32:12.244
Let's go into startup idea number three.

394
00:32:13.044 --> 00:32:19.945
So startup idea number three is a local pre-send reviewer for professional services.

395
00:32:20.426 --> 00:32:26.847
So every professional service firm, or 99.9% of them, has a version of this workflow.

396
00:32:27.327 --> 00:32:34.949
Someone writes a client email, a proposal, a memo, a contract summary, an investment note, an HR note, and then asks someone,

397
00:32:36.689 --> 00:32:39.790
else to check it out before it goes out, like a review.

398
00:32:39.950 --> 00:32:41.671
And it happens constantly.

399
00:32:42.071 --> 00:32:49.213
Law firms, accounting firms, wealth advisors, recruiting firms, even consultants have a version of this.

400
00:32:49.633 --> 00:32:57.816
So the first version is you build a local desktop app that reviews outbound drafts before they leave the company.

401
00:32:58.276 --> 00:33:05.498
So for a wealth advisor, it could be flagging language that sounds like a guaranteed return, which is a definite no-no.

402
00:33:06.018 --> 00:33:09.828
For a law firm, it'll flag a sentence that sounds too definitive.

403
00:33:11.082 --> 00:33:16.323
For HR, it's going to flag sensitive employee information that should stay out of the threat.

404
00:33:16.743 --> 00:33:21.164
For an agency, it flags a promise that the scope does not support.

405
00:33:21.504 --> 00:33:25.985
And for an accountant, it flags a number that doesn't match the attached file.

406
00:33:26.005 --> 00:33:27.745
You'd be surprised how often that happens.

407
00:33:28.065 --> 00:33:33.326
The product is basically a second set of eyes for sensitive work.

408
00:33:33.366 --> 00:33:36.707
It's basically schmuck insurance is the way I think about it.

409
00:33:36.787 --> 00:33:40.168
And maybe that would be the name, schmuckinsurance.com.

410
00:33:41.068 --> 00:33:41.388
take it.

411
00:33:41.949 --> 00:33:43.291
How would I grow the business?

412
00:33:43.732 --> 00:33:46.976
I would start with one vertical and one document type.

413
00:33:47.517 --> 00:33:52.243
For example, I would do email review for independent wealth advisors.

414
00:33:52.463 --> 00:33:57.229
Not everyone, probably not the big banks to start, independent wealth advisors.

415
00:33:57.750 --> 00:34:02.272
I would interview 10 advisors and ask them which emails make them nervous.

416
00:34:02.492 --> 00:34:04.913
I would collect anonymized examples.

417
00:34:05.273 --> 00:34:07.775
I would turn their real concerns into a review checklist.

418
00:34:08.135 --> 00:34:12.537
And I would build a local tool that checks drafts against that checklist.

419
00:34:13.778 --> 00:34:20.606
Obviously, this is so sellable because the buyer understands this behavior and they already asked someone to check the jaft.

420
00:34:21.026 --> 00:34:28.435
So you're just basically giving them a faster first pass that lives closer to their client data and internal rules.

421
00:34:28.855 --> 00:34:31.499
I love this idea and hope a few of you take it.

422
00:34:32.062 --> 00:34:40.225
By the way, if you're not building one of these ideas tomorrow, I still think you should learn local AI because it does change how you work with your own files.

423
00:34:40.945 --> 00:34:46.968
So I think just like from a personal productivity perspective, it's still super, super helpful.

424
00:34:47.748 --> 00:35:01.673
So if you're working at a company, say, and you just want to be more productive, so you have more time to scroll TikTok or watch movies or hang with your family, make a folder called Local AI Lab.

425
00:35:02.093 --> 00:35:06.155
And then put 10 files that matter to your work in that folder.

426
00:35:06.775 --> 00:35:12.417
It could be anything from sales calls or meeting transcripts, old tweets, ideas that you have.

427
00:35:13.018 --> 00:35:18.760
Then run Gemma, whatever model you choose, to make it produce one useful artifact.

428
00:35:19.340 --> 00:35:26.643
And then ask it to create a weekly business pulse or ask it to find what's changed in customer conversations or meeting notes.

429
00:35:27.683 --> 00:35:31.345
Ask it to group feature requests by the actual pain behind it.

430
00:35:31.945 --> 00:35:36.069
Ask it to review drafts and tell you what your audience keeps responding to.

431
00:35:36.630 --> 00:35:44.657
The key basically is to produce a file, a memo, a checklist, a brief, a report, or a review that you can reuse.

432
00:35:45.358 --> 00:35:50.743
A chat answer is nice, but a useful artifact changes that workflow.

433
00:35:51.103 --> 00:35:53.585
This is the first rep I would recommend.

434
00:35:53.605 --> 00:35:53.726
Okay.

435
00:35:54.426 --> 00:36:01.048
A model reads the folder, the model writes the file, you inspect it, you improve the workflow, and then you run it again.

436
00:36:01.108 --> 00:36:05.289
If you do that a few times, your brain really starts to connect the dots.

437
00:36:05.949 --> 00:36:09.710
You start noticing where private data is trapped in folders.

438
00:36:09.790 --> 00:36:13.651
You notice which reviews happen over and over again.

439
00:36:24.433 --> 00:36:25.173
week after week.

440
00:36:25.574 --> 00:36:33.037
Hopefully this episode got your creative juices flowing because once you see the pattern, you start spotting local AI businesses everywhere.

441
00:36:33.357 --> 00:36:40.201
You can learn enough of the map to spot where these models belong without turning yourself into a local engineer overnight.

442
00:36:41.137 --> 00:36:47.507
I believe some local AI belongs in the cloud and some AI belongs in the device.

443
00:36:47.988 --> 00:36:52.314
And a lot of the best products over the next couple of years are going to combine them both.

444
00:36:52.334 --> 00:36:53.596
So if I was starting today,

445
00:36:54.341 --> 00:36:56.623
What I would do is I would run Gemma locally.

446
00:36:56.803 --> 00:36:59.124
I'd read model cards on Hugging Face.

447
00:36:59.305 --> 00:37:02.447
I'd learn the difference between LM Studio and Olama.

448
00:37:02.887 --> 00:37:10.553
I'd play with Google AI Edge and then look for one boring workflow where local AI actually makes the product better.

449
00:37:11.253 --> 00:37:20.558
Those categories are things like private data, offline work, camera or audio context, or low latency, or if there's a high repeated API cost.

450
00:37:20.878 --> 00:37:30.362
If there's a buyer who feels better when the model is just close to them, a workflow where a small agent team could recheck, summarize, and prepare work every day.

451
00:37:30.402 --> 00:37:31.743
That's like the hunting ground.

452
00:37:32.083 --> 00:37:42.434
Local AI is just way easier to understand once you stop treating it like a model benchmark conversation and start treating it like a product conversation.

453
00:37:42.814 --> 00:37:45.076
You have to ask yourself, where's the work happening?

454
00:37:45.156 --> 00:37:45.957
Where is the data?

455
00:37:45.997 --> 00:37:46.958
Where is the device?

456
00:37:47.118 --> 00:37:48.339
Where is the trust issue?

457
00:37:48.720 --> 00:37:50.602
Where is an annoying review loop?

458
00:37:50.922 --> 00:37:55.266
And then you answer those questions and you just start seeing the ideas.

459
00:37:55.286 --> 00:37:55.326
So

460
00:37:55.947 --> 00:38:09.250
Overall, I hope you understand a little about the core things you need to understand about local AI, some of the models, some of the apps you need to download, some of the workflows that you can build, and some of the business opportunities that exist.

461
00:38:09.810 --> 00:38:14.011
I just don't see that many non-technical people playing with local AI.

462
00:38:15.051 --> 00:38:19.332
And over the last two months or so, I've gotten deeper and deeper into it.

463
00:38:19.392 --> 00:38:19.872
And it's just...

464
00:38:20.687 --> 00:38:24.213
Like I said, it's been connecting the dots and I'm grateful for it.

465
00:38:25.755 --> 00:38:28.179
I hope you have a creative day.

466
00:38:29.020 --> 00:38:34.689
I read every single comment on YouTube and respond to most.

467
00:38:34.749 --> 00:38:36.092
So I'll see you in there.

468
00:38:36.773 --> 00:38:42.021
Share this with a friend who you think could benefit from understanding local AI in a clear way.

469
00:38:42.662 --> 00:38:43.864
And I'll see you next time.

470
00:38:44.645 --> 00:38:45.406
Happy building.