WEBVTT

1
00:00:00.160 --> 00:00:02.520
<v Speaker 1>Welcome to the deep dive, where the place where we

2
00:00:02.640 --> 00:00:05.919
<v Speaker 1>unpack complex topics, really getting into the sources you provide

3
00:00:05.960 --> 00:00:09.240
<v Speaker 1>to pull out the core insights. Today we're tackling something

4
00:00:09.279 --> 00:00:11.439
<v Speaker 1>that well, it touches pretty much every part of your

5
00:00:11.480 --> 00:00:15.080
<v Speaker 1>digital life, often completely invisibly. It's how deep learning is

6
00:00:15.119 --> 00:00:18.600
<v Speaker 1>actively shaping actually building the intelligent web. You know, the

7
00:00:18.640 --> 00:00:21.320
<v Speaker 1>Internet is always changing, but the biggest shifts recently they've

8
00:00:21.320 --> 00:00:24.640
<v Speaker 1>been driven by artificial intelligence. So our mission today is

9
00:00:24.679 --> 00:00:26.839
<v Speaker 1>to take the source material you gave us, hands on

10
00:00:27.000 --> 00:00:30.000
<v Speaker 1>Python deep learning for the Web, and really illuminate how

11
00:00:30.039 --> 00:00:33.399
<v Speaker 1>deep learning with Python as its engine, is creating these smart,

12
00:00:33.520 --> 00:00:36.479
<v Speaker 1>responsive web apps we kind of expect now think of

13
00:00:36.479 --> 00:00:40.439
<v Speaker 1>it as a shortcut maybe to understanding the magic behind

14
00:00:40.439 --> 00:00:43.679
<v Speaker 1>the modern web. Get ready for some aha moments, I think,

15
00:00:43.719 --> 00:00:45.840
<v Speaker 1>and maybe a few surprises as we explore how the

16
00:00:45.840 --> 00:00:49.200
<v Speaker 1>web gets its smart Okay, so let's unpack this AI.

17
00:00:49.320 --> 00:00:51.679
<v Speaker 1>The idea. It's been around since the fifties, right, Tearing

18
00:00:51.840 --> 00:00:54.799
<v Speaker 1>McCarthy asking if machines could think, But why now? Why

19
00:00:54.880 --> 00:00:56.679
<v Speaker 1>is it suddenly everywhere online?

20
00:00:56.799 --> 00:00:59.479
<v Speaker 2>Yeah, that's a great question. What's really fascinating is it

21
00:00:59.600 --> 00:01:03.280
<v Speaker 2>was one single thing. It was more like a perfect storm.

22
00:01:03.640 --> 00:01:06.680
<v Speaker 2>Several critical factors just converged at the right time. Probably

23
00:01:06.680 --> 00:01:09.760
<v Speaker 2>the biggest one just the sheer amount of data available now,

24
00:01:10.159 --> 00:01:13.599
<v Speaker 2>it's staggering. How very in Google's chief economists he put

25
00:01:13.599 --> 00:01:15.400
<v Speaker 2>it really well. He said something like, between the dawn

26
00:01:15.439 --> 00:01:19.599
<v Speaker 2>of civilization in two thousand and three, we created five exabllites.

27
00:01:19.480 --> 00:01:21.359
<v Speaker 1>Which is huge, right, five exatabytes.

28
00:01:21.359 --> 00:01:24.000
<v Speaker 2>But now we're apparently creating that much every two days,

29
00:01:24.239 --> 00:01:28.319
<v Speaker 2>and the prediction for twenty twenty was fifty three zetabytes.

30
00:01:28.680 --> 00:01:30.680
<v Speaker 2>Is an astronomical amount of information.

31
00:01:30.840 --> 00:01:33.719
<v Speaker 1>Wow, every two days. Where is all that data actually

32
00:01:33.799 --> 00:01:36.680
<v Speaker 1>coming from? And specifically what kind of data is useful

33
00:01:36.680 --> 00:01:37.719
<v Speaker 1>for deep learning on the web?

34
00:01:37.799 --> 00:01:40.400
<v Speaker 2>Is it just you know, clicks, Oh, it's way beyond clicks.

35
00:01:40.439 --> 00:01:44.079
<v Speaker 2>It's fueled by it, well, storage getting cheaper, faster data transmission,

36
00:01:44.120 --> 00:01:47.719
<v Speaker 2>cloud computing, becoming common sensors everywhere with the Internet of things,

37
00:01:48.159 --> 00:01:53.159
<v Speaker 2>and critically it's us right, our constant use of phones, apps, websites,

38
00:01:53.319 --> 00:01:57.760
<v Speaker 2>we generate this incredibly rich stream of interaction logs, text, images, audio.

39
00:01:58.200 --> 00:02:01.439
<v Speaker 2>It's complex, often messy on labeled data.

40
00:02:01.120 --> 00:02:03.840
<v Speaker 1>Unlabeled data, so stuff that AI has to make sense

41
00:02:03.840 --> 00:02:04.400
<v Speaker 1>of on its.

42
00:02:04.239 --> 00:02:07.079
<v Speaker 2>Own exactly data that AI can learn patterns from in

43
00:02:07.159 --> 00:02:10.400
<v Speaker 2>ways that you know, simple statistics just couldn't handle before.

44
00:02:10.840 --> 00:02:14.960
<v Speaker 2>And then alongside this data flood, the algorithms themselves got

45
00:02:15.039 --> 00:02:18.280
<v Speaker 2>much smarter, much more powerful. This directly helped neural networks

46
00:02:18.319 --> 00:02:21.599
<v Speaker 2>become practical. And you can't forget the hardware leap. Think

47
00:02:21.599 --> 00:02:25.520
<v Speaker 2>about memory. Intel's first dynamic RAM in nineteen seventy held

48
00:02:25.520 --> 00:02:29.280
<v Speaker 2>what one kilobyte one kb tiny tiny, Today you can

49
00:02:29.319 --> 00:02:34.000
<v Speaker 2>get one hundred and twenty eight gigabyte modules. That's let

50
00:02:34.080 --> 00:02:36.400
<v Speaker 2>me think about one point two eight times ten to

51
00:02:36.439 --> 00:02:39.199
<v Speaker 2>the eighth more memory. That's the kind of power you

52
00:02:39.240 --> 00:02:41.159
<v Speaker 2>need for deep learnings, heavy calculations.

53
00:02:41.280 --> 00:02:43.919
<v Speaker 1>Okay, so a sea of data, smarter algorithms, and the

54
00:02:44.039 --> 00:02:46.840
<v Speaker 1>raw computing power. That really does sound like the perfect storm.

55
00:02:46.919 --> 00:02:47.439
<v Speaker 2>It really was.

56
00:02:47.520 --> 00:02:47.719
<v Speaker 1>Yeah.

57
00:02:47.719 --> 00:02:51.560
<v Speaker 2>One more piece the democratization of high performance computing. Cloud

58
00:02:51.560 --> 00:02:54.800
<v Speaker 2>platforms like eight of US, Google Cloud, Azure. They made

59
00:02:54.840 --> 00:02:56.759
<v Speaker 2>all this power accessible. Suddenly, you didn't need be a

60
00:02:56.879 --> 00:03:00.960
<v Speaker 2>massive research lab to experiment with AI. Startup individual developers

61
00:03:01.000 --> 00:03:01.680
<v Speaker 2>they could tap.

62
00:03:01.520 --> 00:03:02.879
<v Speaker 1>Into it right accessibility.

63
00:03:03.080 --> 00:03:06.000
<v Speaker 2>Yeah. So the key takeaway is AI's ubiquity on the

64
00:03:06.000 --> 00:03:11.080
<v Speaker 2>web wasn't just one invention. It was this convergence, data, algorithms,

65
00:03:11.159 --> 00:03:14.560
<v Speaker 2>hardware and access. That's what made it scalable and practical

66
00:03:14.599 --> 00:03:15.520
<v Speaker 2>for the everyday web.

67
00:03:15.680 --> 00:03:18.240
<v Speaker 1>Okay, so with that foundation, late, let's look at how

68
00:03:18.280 --> 00:03:21.520
<v Speaker 1>this AI has actually changed the web experiences we have

69
00:03:21.599 --> 00:03:24.759
<v Speaker 1>every day, often, you know, totally behind the scenes. Let's

70
00:03:24.759 --> 00:03:27.599
<v Speaker 1>start with chatbots. Everyone's bumped into those, uh chatbots.

71
00:03:28.000 --> 00:03:30.039
<v Speaker 2>Yeah, it's almost funny thinking back to the early ones

72
00:03:30.159 --> 00:03:34.439
<v Speaker 2>like Eliza back in sixty six, very rule based. You know, sorry,

73
00:03:34.439 --> 00:03:36.800
<v Speaker 2>I did not get that. They hit dead ends really fast.

74
00:03:36.879 --> 00:03:40.120
<v Speaker 1>Oh I remember those so frustrating. Sorry, I did not

75
00:03:40.240 --> 00:03:42.199
<v Speaker 1>get that over and over exactly.

76
00:03:42.360 --> 00:03:46.120
<v Speaker 2>Today's chatbots, powered by neural networks. They can understand context,

77
00:03:46.439 --> 00:03:50.000
<v Speaker 2>even emotions. Sometimes they can pull information from the web

78
00:03:50.080 --> 00:03:53.479
<v Speaker 2>in real time to personalize the conversation. I mean, look

79
00:03:53.520 --> 00:03:56.639
<v Speaker 2>at Facebook Messenger. The source mentioned over one hundred thousand

80
00:03:56.639 --> 00:03:59.159
<v Speaker 2>bots were created there in just the first seventeen months

81
00:03:59.240 --> 00:04:02.000
<v Speaker 2>or so. And WhatsApp bots are booking apployments. Now it's

82
00:04:02.039 --> 00:04:03.080
<v Speaker 2>a huge leap.

83
00:04:03.080 --> 00:04:05.800
<v Speaker 1>It really is. It feels like night and day. Okay,

84
00:04:05.800 --> 00:04:09.360
<v Speaker 1>what about something less obvious like web analytics. How has

85
00:04:09.439 --> 00:04:11.960
<v Speaker 1>AI changed tracking website visitors?

86
00:04:12.560 --> 00:04:15.800
<v Speaker 2>Complete transformation there too? It started super simple, like those

87
00:04:15.840 --> 00:04:19.600
<v Speaker 2>old odometer style page hit counters.

88
00:04:19.279 --> 00:04:20.480
<v Speaker 1>Right, just counting clicks?

89
00:04:20.600 --> 00:04:23.160
<v Speaker 2>Yeah, then maybe tracking where visitors came from. But now

90
00:04:23.800 --> 00:04:26.759
<v Speaker 2>AI tools don't just report what happened. They predict future performance.

91
00:04:27.240 --> 00:04:30.879
<v Speaker 2>They can suggest specific content changes to boost engagement. Get this.

92
00:04:31.040 --> 00:04:33.240
<v Speaker 2>The source even mentions that how your mouse pointer just

93
00:04:33.279 --> 00:04:35.920
<v Speaker 2>sits idle on a page might get reported back to

94
00:04:35.959 --> 00:04:38.920
<v Speaker 2>a Google Analytics dashboard trying to figure out user intent.

95
00:04:39.079 --> 00:04:43.079
<v Speaker 1>WHOA, Okay, my idle mouse movements. That's slightly creepy, but

96
00:04:43.160 --> 00:04:45.360
<v Speaker 1>I see the point. So it's not just reporting, it's

97
00:04:45.399 --> 00:04:48.480
<v Speaker 1>predicting intent, optimizing the whole experience proactively.

98
00:04:48.720 --> 00:04:52.959
<v Speaker 2>Precisely, it's about moving from looking backwards to actively shaping

99
00:04:53.000 --> 00:04:56.519
<v Speaker 2>the future interaction. AI could spot patterns, maybe predict if

100
00:04:56.519 --> 00:04:59.079
<v Speaker 2>someone's about to abandon their shopping cart, or figure out

101
00:04:59.079 --> 00:05:02.000
<v Speaker 2>which blog post will resonate most with this specific user.

102
00:05:02.680 --> 00:05:06.399
<v Speaker 2>That allows for real time personalization, dynamic content, even running

103
00:05:06.439 --> 00:05:10.560
<v Speaker 2>sophisticated ab tests automatically, much more targeted, very powerful.

104
00:05:10.959 --> 00:05:14.480
<v Speaker 1>Okay, Another big one spam filtering. We all rely on it,

105
00:05:14.519 --> 00:05:16.639
<v Speaker 1>but it feels like spammers are always trying to get

106
00:05:16.639 --> 00:05:17.079
<v Speaker 1>around it.

107
00:05:17.120 --> 00:05:19.000
<v Speaker 2>Oh, it's absolutely a constant arms race.

108
00:05:19.120 --> 00:05:19.399
<v Speaker 1>Yeah.

109
00:05:19.480 --> 00:05:23.160
<v Speaker 2>It started with basic IP blacklisting, which spammers quickly learned

110
00:05:23.160 --> 00:05:26.519
<v Speaker 2>to defeat. Then came bejan filtering around the early two thousands.

111
00:05:26.519 --> 00:05:28.519
<v Speaker 2>That was a big step up, so much so that

112
00:05:28.560 --> 00:05:31.120
<v Speaker 2>Bill Gates back in two thousand and four famously said

113
00:05:31.680 --> 00:05:33.800
<v Speaker 2>two years from now, spam will be solved.

114
00:05:33.920 --> 00:05:36.480
<v Speaker 1>Yeah, I remember that prediction didn't quite pan.

115
00:05:36.360 --> 00:05:39.680
<v Speaker 2>Out, not even close, because the spammer's just adapted again.

116
00:05:39.959 --> 00:05:42.639
<v Speaker 2>Now it's large scale neural networks doing the heavy lifting.

117
00:05:42.879 --> 00:05:46.839
<v Speaker 2>They're constantly scanning emails looking for complex, often non obvious

118
00:05:46.879 --> 00:05:50.360
<v Speaker 2>patterns of spammy behavior. They learn and adapt almost as

119
00:05:50.360 --> 00:05:52.319
<v Speaker 2>fast as the spammers invent new treks.

120
00:05:52.639 --> 00:05:54.879
<v Speaker 1>That makes sense. It has to be adaptive, and search

121
00:05:54.920 --> 00:05:58.000
<v Speaker 1>engines the absolute core of the web for most people.

122
00:05:58.120 --> 00:06:01.480
<v Speaker 1>How deeply is AI involved there monumentally?

123
00:06:02.120 --> 00:06:05.439
<v Speaker 2>Think about the very beginning. Tim berners Lee's Worldwide Web

124
00:06:05.519 --> 00:06:09.279
<v Speaker 2>Virtual Library in ninety one was basically a hand curated list.

125
00:06:09.600 --> 00:06:12.480
<v Speaker 2>Then you had pioneers like Jonathan Fletcher with JumpStation in

126
00:06:12.560 --> 00:06:16.360
<v Speaker 2>ninety three doing crawling and indexing more like modern search.

127
00:06:16.800 --> 00:06:20.680
<v Speaker 2>But today it's a whole different beast. Google Search, for example,

128
00:06:20.839 --> 00:06:24.439
<v Speaker 2>uses deep neural networks extensively. They use natural language processing

129
00:06:24.519 --> 00:06:28.319
<v Speaker 2>or NLP to understand the meaning and relevance of content,

130
00:06:28.399 --> 00:06:29.199
<v Speaker 2>not just keywords.

131
00:06:29.319 --> 00:06:32.000
<v Speaker 1>NLP so understanding language itself.

132
00:06:31.720 --> 00:06:36.160
<v Speaker 2>Exactly, understanding intent, context, and they use things called convolutional

133
00:06:36.160 --> 00:06:39.120
<v Speaker 2>neural Networks CNNs, which are brilliant for image analysis, powering

134
00:06:39.160 --> 00:06:42.959
<v Speaker 2>image search. Google even generates direct answers now using its

135
00:06:43.000 --> 00:06:45.959
<v Speaker 2>knowledge graph, not just blue links. It's trying to understand

136
00:06:46.040 --> 00:06:48.600
<v Speaker 2>and answer your query directly and Google Translate.

137
00:06:48.680 --> 00:06:50.959
<v Speaker 1>That one feels like pure sci fi sometimes. The way

138
00:06:51.000 --> 00:06:53.079
<v Speaker 1>it works in real time, it's like the Internet itself

139
00:06:53.120 --> 00:06:54.399
<v Speaker 1>is becoming multilingual.

140
00:06:54.560 --> 00:06:57.639
<v Speaker 2>It's an amazing application, truly. Google switched to their neural

141
00:06:57.680 --> 00:07:02.279
<v Speaker 2>machine translation system back in November twenty sixteen. It now

142
00:07:02.279 --> 00:07:06.399
<v Speaker 2>supports over one hundred languages, often with startling accuracy integrated

143
00:07:06.439 --> 00:07:09.279
<v Speaker 2>right into your browser. The translations are much more natural,

144
00:07:09.360 --> 00:07:12.839
<v Speaker 2>more context awar than the old phrase based systems. It's

145
00:07:12.839 --> 00:07:15.160
<v Speaker 2>really changed how we can communicate online globally.

146
00:07:15.439 --> 00:07:17.639
<v Speaker 1>Okay, so we've seen the impact of AI. Let's shift

147
00:07:17.680 --> 00:07:19.839
<v Speaker 1>gears and get into the nuts and bolts. How are

148
00:07:19.839 --> 00:07:23.639
<v Speaker 1>these intelligent web applications actually built? What's the architecture, what

149
00:07:23.680 --> 00:07:25.920
<v Speaker 1>are the tools? This is where it gets really interesting

150
00:07:25.959 --> 00:07:27.519
<v Speaker 1>for anyone wanting to understand.

151
00:07:27.160 --> 00:07:30.000
<v Speaker 2>The how right, And this brings up a key question,

152
00:07:30.399 --> 00:07:32.920
<v Speaker 2>what exactly is deep learning in this web context? And

153
00:07:32.959 --> 00:07:34.160
<v Speaker 2>how does Python fit in?

154
00:07:34.439 --> 00:07:37.199
<v Speaker 1>So deep learning is a specific type of machine learning.

155
00:07:37.519 --> 00:07:41.480
<v Speaker 1>It's based purely on these artificial neural networks with multiple layers,

156
00:07:41.680 --> 00:07:45.879
<v Speaker 1>hence deep Instead of programmers manually defining features in the data,

157
00:07:46.360 --> 00:07:51.000
<v Speaker 1>deep learning models learn these features automatically in hierarchies. Think

158
00:07:51.040 --> 00:07:54.800
<v Speaker 1>of recognizing an image. Layers might learn edges and contours

159
00:07:54.839 --> 00:07:58.600
<v Speaker 1>than textures, than objects. It learns directly from raw web data,

160
00:07:58.839 --> 00:08:01.560
<v Speaker 1>which is crucial given and how messy and unstructured that

161
00:08:01.680 --> 00:08:02.439
<v Speaker 1>data often is.

162
00:08:02.600 --> 00:08:04.720
<v Speaker 2>Okay, so it's all about these neural networks. Are they

163
00:08:04.759 --> 00:08:06.600
<v Speaker 2>really modeled after our brains?

164
00:08:07.160 --> 00:08:10.199
<v Speaker 1>Conceptually yes, though it's a highly simplified model. Our brain

165
00:08:10.240 --> 00:08:12.959
<v Speaker 1>has something like ten billion neurons, each connected to maybe

166
00:08:12.959 --> 00:08:16.279
<v Speaker 1>ten thousand others. It's incredibly complex, and artificial neuron is

167
00:08:16.319 --> 00:08:18.839
<v Speaker 1>much simpler. It receives inputs. Each input gets a weight

168
00:08:18.920 --> 00:08:22.360
<v Speaker 1>representing its importance, they're summed up, a bias is added,

169
00:08:22.600 --> 00:08:25.319
<v Speaker 1>and then an activation function decides if the neuron fires.

170
00:08:25.519 --> 00:08:28.839
<v Speaker 2>And you mentioned activation functions like sigmoid, not just on

171
00:08:28.879 --> 00:08:31.759
<v Speaker 2>and off, right, instead of a simple step function, we

172
00:08:31.879 --> 00:08:35.120
<v Speaker 2>use things like the sigmoid function. It's smoother, more sensitive

173
00:08:35.120 --> 00:08:37.960
<v Speaker 2>to nuances and nonlinear patterns in the data, which is

174
00:08:38.039 --> 00:08:41.919
<v Speaker 2>vital for complex tasks like understanding language or user behavior,

175
00:08:42.519 --> 00:08:46.039
<v Speaker 2>and the network learns by adjusting all those weights and biases.

176
00:08:46.559 --> 00:08:50.200
<v Speaker 2>It uses clever mathematical techniques like gradiate descent and backpropagation

177
00:08:50.519 --> 00:08:54.519
<v Speaker 2>to iteratively tweak the connections, minimizing the difference between its

178
00:08:54.519 --> 00:08:57.720
<v Speaker 2>predictions and the actual outcomes in the training data.

179
00:08:57.399 --> 00:09:00.320
<v Speaker 1>It sounds like you'd need different kinds of networks for

180
00:09:00.320 --> 00:09:03.679
<v Speaker 1>different jobs, like analyzing an image must be different from

181
00:09:03.720 --> 00:09:05.480
<v Speaker 1>predicting the next word in a sentence.

182
00:09:05.600 --> 00:09:10.600
<v Speaker 2>Absolutely, there are specialized architectures. Convolutional neural networks or CNNs

183
00:09:10.639 --> 00:09:14.000
<v Speaker 2>are fantastic for grid like data, which makes them perfect

184
00:09:14.000 --> 00:09:17.759
<v Speaker 2>for images. That's how Facebook spots faces in photos, or

185
00:09:17.759 --> 00:09:21.720
<v Speaker 2>how product recommendation systems analyze pictures. Then you have recurrent

186
00:09:21.720 --> 00:09:25.360
<v Speaker 2>neural networks or RNNs. These are designed to handle sequences

187
00:09:25.440 --> 00:09:28.159
<v Speaker 2>where order matters. Think predicting the next word in a

188
00:09:28.200 --> 00:09:31.679
<v Speaker 2>sentence like Google completing your search query, or understanding the

189
00:09:31.679 --> 00:09:33.159
<v Speaker 2>flow of dialogue and a chatbot.

190
00:09:33.399 --> 00:09:37.080
<v Speaker 1>Okay, CNNs for images. RNNs for sequences.

191
00:09:36.879 --> 00:09:39.080
<v Speaker 2>Makes sense, and a really important type of RNN is

192
00:09:39.080 --> 00:09:43.519
<v Speaker 2>the long short term memory network or LSTM. They're particularly

193
00:09:43.600 --> 00:09:47.200
<v Speaker 2>good at remembering information over longer sequences. Finding those long

194
00:09:47.279 --> 00:09:51.960
<v Speaker 2>term dependencies crucial for sophisticated language understanding. We're spotting subtle

195
00:09:52.000 --> 00:09:54.440
<v Speaker 2>anomalies in user activity logs over time.

196
00:09:54.759 --> 00:09:57.440
<v Speaker 1>So given all this power and complexity, why Python? What

197
00:09:57.519 --> 00:09:59.960
<v Speaker 1>makes Python the go to language for building these things

198
00:10:00.159 --> 00:10:00.559
<v Speaker 1>on the web.

199
00:10:00.720 --> 00:10:04.519
<v Speaker 2>It's really the ecosystem around Python. Python itself is relatively

200
00:10:04.559 --> 00:10:07.600
<v Speaker 2>easy to learn and use, but its libraries are the key.

201
00:10:08.159 --> 00:10:11.039
<v Speaker 2>The source recommends Python three point six or later, often

202
00:10:11.120 --> 00:10:14.480
<v Speaker 2>using the Antaconta distribution because it packages many useful tools.

203
00:10:15.000 --> 00:10:18.080
<v Speaker 2>You have numb PI, which is fundamental for efficient numerical

204
00:10:18.120 --> 00:10:21.879
<v Speaker 2>operations on a raise basically fast math on large data

205
00:10:21.879 --> 00:10:26.759
<v Speaker 2>sets essential for mL. Then pandas built on NUMBPI gives

206
00:10:26.799 --> 00:10:29.879
<v Speaker 2>you powerful data structures like data frames and tools for cleaning,

207
00:10:29.960 --> 00:10:33.840
<v Speaker 2>manipulating and analyzing data. Absolutely vital for preparing web data

208
00:10:33.840 --> 00:10:36.200
<v Speaker 2>and for building the neural networks themselves. You have libraries

209
00:10:36.240 --> 00:10:39.440
<v Speaker 2>like Keras. Keras is a high level API, makes it

210
00:10:39.519 --> 00:10:42.679
<v Speaker 2>much faster and easier to define and train complex networks,

211
00:10:43.080 --> 00:10:45.759
<v Speaker 2>often using Google's tensorfol library as the underlying engine.

212
00:10:46.039 --> 00:10:49.919
<v Speaker 1>Right, so Numpi, Panda's keras tools for the data in

213
00:10:49.960 --> 00:10:52.879
<v Speaker 1>the model. But how do these models trained in Python

214
00:10:53.320 --> 00:10:55.759
<v Speaker 1>actually connect to a live website? How does a user

215
00:10:55.799 --> 00:10:56.559
<v Speaker 1>interact with them?

216
00:10:56.759 --> 00:10:59.360
<v Speaker 2>That's where web frameworks come in. Python frameworks like Jango

217
00:10:59.360 --> 00:11:01.840
<v Speaker 2>and flask are very popular for building rest APIs.

218
00:11:01.919 --> 00:11:04.720
<v Speaker 1>Rest APIs those are like the messengers between the website

219
00:11:04.759 --> 00:11:06.360
<v Speaker 1>and the AI model exactly.

220
00:11:06.960 --> 00:11:10.000
<v Speaker 2>The website front end sends data like text from a

221
00:11:10.080 --> 00:11:12.759
<v Speaker 2>chatbot input or an image to be analyzed to the

222
00:11:12.799 --> 00:11:15.799
<v Speaker 2>API end point. The Python back end running the deep

223
00:11:15.879 --> 00:11:19.399
<v Speaker 2>learning model processes the data, generates a prediction or result,

224
00:11:19.559 --> 00:11:21.799
<v Speaker 2>and sends it back via the API to the website

225
00:11:21.840 --> 00:11:24.519
<v Speaker 2>to display to the user. But you don't always have

226
00:11:24.559 --> 00:11:28.159
<v Speaker 2>to build everything yourself. The major cloud providers offer pre build,

227
00:11:28.399 --> 00:11:30.279
<v Speaker 2>battle tested deep learning APIs.

228
00:11:30.440 --> 00:11:33.399
<v Speaker 1>AH the cloud APIs that sounds like a massive shortcut.

229
00:11:33.440 --> 00:11:34.679
<v Speaker 1>What kinds of things do they offer?

230
00:11:34.840 --> 00:11:39.200
<v Speaker 2>Huge time savers. Google Cloud GCP has its vision API

231
00:11:39.240 --> 00:11:42.600
<v Speaker 2>for image analysis, a translation API and dialogue flow for

232
00:11:42.639 --> 00:11:48.320
<v Speaker 2>building chatbots. Amazon Web Services AWS has recognition for detecting objects, faces,

233
00:11:48.320 --> 00:11:52.039
<v Speaker 2>even celebrities and images, and the Alexa API for voice applications.

234
00:11:52.159 --> 00:11:54.960
<v Speaker 2>Microsoft Azure offers a FASE API for detection and a

235
00:11:55.000 --> 00:11:59.320
<v Speaker 2>motion recognition, a text analytics API for sentiment analysis, keyphrase extraction.

236
00:11:59.679 --> 00:12:03.039
<v Speaker 2>Plus Microsoft has its own deep learning framework called Cognitive

237
00:12:03.039 --> 00:12:04.559
<v Speaker 2>Toolkit or CNTK, so.

238
00:12:04.639 --> 00:12:08.080
<v Speaker 1>Developers can just plug into these powerful pretrain models. That

239
00:12:08.159 --> 00:12:10.759
<v Speaker 1>really lowers the barrier to entry definitely.

240
00:12:10.799 --> 00:12:15.000
<v Speaker 2>It lets web developers integrate sophisticated AI features without needing

241
00:12:15.039 --> 00:12:16.840
<v Speaker 2>to be deep learning PhDs themselves.

242
00:12:17.000 --> 00:12:20.080
<v Speaker 1>Okay, so building the model is complex but achievable with

243
00:12:20.120 --> 00:12:23.399
<v Speaker 1>these tools, but putting it into production on a live website,

244
00:12:23.639 --> 00:12:28.519
<v Speaker 1>making it robust, scalable, secure, that feels like a whole

245
00:12:28.519 --> 00:12:32.039
<v Speaker 1>different challenge. It's engineering, not just science.

246
00:12:31.799 --> 00:12:35.440
<v Speaker 2>Absolutely connecting it to the bigger picture. The standard machine

247
00:12:35.480 --> 00:12:38.679
<v Speaker 2>learning workflow isn't just build and deploy. It's a cycle.

248
00:12:39.200 --> 00:12:42.120
<v Speaker 2>It starts with getting the data, then meticulous data preparation

249
00:12:42.200 --> 00:12:46.000
<v Speaker 2>that involves things like exploratory data analysis, cleaning, feature engineering.

250
00:12:46.399 --> 00:12:49.000
<v Speaker 2>Then you train the model, but then comes deployment and

251
00:12:49.080 --> 00:12:51.679
<v Speaker 2>crucially continuous monitoring. It's never really.

252
00:12:51.519 --> 00:12:54.919
<v Speaker 1>Finished and the data itself, I imagine real world web data

253
00:12:54.960 --> 00:12:57.720
<v Speaker 1>can be messy and potentially biased. How do you deal

254
00:12:57.759 --> 00:12:58.000
<v Speaker 1>with that.

255
00:12:58.279 --> 00:13:01.639
<v Speaker 2>Bias is a huge, huge chol It can easily creep

256
00:13:01.639 --> 00:13:04.679
<v Speaker 2>in from the data you train on, reflecting existing societal

257
00:13:04.720 --> 00:13:07.480
<v Speaker 2>biases or quirks in how the data was collected. The

258
00:13:07.600 --> 00:13:10.519
<v Speaker 2>source gives a great example the Amazon find Food reviews

259
00:13:10.600 --> 00:13:14.080
<v Speaker 2>data set. They found that positive reviews often had more text,

260
00:13:14.679 --> 00:13:17.039
<v Speaker 2>so a model might learn to associate longer reviews with

261
00:13:17.120 --> 00:13:20.200
<v Speaker 2>positive sentiment, even if a long review is actually a

262
00:13:20.200 --> 00:13:21.039
<v Speaker 2>detailed complaint.

263
00:13:21.080 --> 00:13:23.799
<v Speaker 1>Oh interesting, So it learns the wrong correlation exactly.

264
00:13:23.840 --> 00:13:26.919
<v Speaker 2>It highlights how real world data has these hidden complexities.

265
00:13:26.919 --> 00:13:29.480
<v Speaker 2>In edge cases, you have to be really careful during

266
00:13:29.559 --> 00:13:32.600
<v Speaker 2>data prep and model evaluation to spot and try to

267
00:13:32.639 --> 00:13:33.840
<v Speaker 2>mitigate these biases.

268
00:13:33.960 --> 00:13:37.600
<v Speaker 1>So beyond just spotting bias, how do developers actively reduce it?

269
00:13:38.000 --> 00:13:39.840
<v Speaker 1>Especially with diverse users on the web.

270
00:13:40.200 --> 00:13:44.080
<v Speaker 2>It's tough, but there are techniques things like carefully sampling

271
00:13:44.159 --> 00:13:48.720
<v Speaker 2>data to ensure representation, sometimes mathematically adjusting feature weights to

272
00:13:48.759 --> 00:13:52.799
<v Speaker 2>reduce bias, or even using specific algorithms designed for fairness

273
00:13:52.879 --> 00:13:56.000
<v Speaker 2>during training. After training, you have to rigorously test the

274
00:13:56.000 --> 00:13:59.639
<v Speaker 2>model's performance across different groups, and for web apps, strict

275
00:13:59.679 --> 00:14:04.639
<v Speaker 2>input validation helps prevent weird inputs from skewing things. Transparency

276
00:14:04.679 --> 00:14:07.440
<v Speaker 2>is also key. Sometimes explaining why an AI made a

277
00:14:07.480 --> 00:14:10.000
<v Speaker 2>decision helps build trust and catch issues.

278
00:14:10.279 --> 00:14:13.120
<v Speaker 1>That makes sense. What are some other common mistakes? Maybe

279
00:14:13.120 --> 00:14:16.600
<v Speaker 1>some how not to build an AI back end? Tips?

280
00:14:16.759 --> 00:14:19.919
<v Speaker 2>Oh? Definitely A big one is expecting every AI component

281
00:14:19.960 --> 00:14:23.799
<v Speaker 2>to respond instantly in real time. Deep learning can be

282
00:14:23.919 --> 00:14:26.799
<v Speaker 2>computationally heavy. It's often much better designed to have the

283
00:14:26.799 --> 00:14:30.480
<v Speaker 2>AI processing happen asynchronously, separate from the main website back

284
00:14:30.559 --> 00:14:32.759
<v Speaker 2>end that talks to the user. Let the website be fast,

285
00:14:32.879 --> 00:14:34.360
<v Speaker 2>let the AI think in the background.

286
00:14:34.440 --> 00:14:36.600
<v Speaker 1>Decouple the AI from the user response time.

287
00:14:36.720 --> 00:14:40.320
<v Speaker 2>Got it right? Another mistake assuming the data coming from

288
00:14:40.320 --> 00:14:43.440
<v Speaker 2>the website will be clean and perfect. Never assume that

289
00:14:43.759 --> 00:14:47.200
<v Speaker 2>you must build robust validation and cleaning steps into your pipeline.

290
00:14:47.320 --> 00:14:51.519
<v Speaker 2>And a third one neglecting model Versioning models change, they

291
00:14:51.519 --> 00:14:54.960
<v Speaker 2>get updated. Your API needs to handle different versions smoothly

292
00:14:55.039 --> 00:14:57.279
<v Speaker 2>so you don't break the website every time you improve

293
00:14:57.320 --> 00:14:57.679
<v Speaker 2>the AI.

294
00:14:57.919 --> 00:15:02.559
<v Speaker 1>Okay, important practical points out. Security always a massive concern

295
00:15:02.600 --> 00:15:05.399
<v Speaker 1>on the web. How is deep learning helping there?

296
00:15:05.480 --> 00:15:08.159
<v Speaker 2>It's actually playing a big role. Think about recap tccha.

297
00:15:08.240 --> 00:15:10.360
<v Speaker 2>We all remember those distorted words you could barely rerate?

298
00:15:10.440 --> 00:15:13.799
<v Speaker 1>Oh? Yes, sometimes impossible exactly now.

299
00:15:13.879 --> 00:15:17.159
<v Speaker 2>It's often invisible. AI is working in the background analyzing

300
00:15:17.159 --> 00:15:20.480
<v Speaker 2>your behavior, mouse movements, timing, et cetera to figure out

301
00:15:20.519 --> 00:15:23.279
<v Speaker 2>if you're human or a bot, often without needing any

302
00:15:23.279 --> 00:15:28.600
<v Speaker 2>explicit test. That's AI making security less intrusive. Deep learning

303
00:15:28.639 --> 00:15:32.120
<v Speaker 2>is also great for malicious user detection. Those LSTM networks

304
00:15:32.120 --> 00:15:34.960
<v Speaker 2>we talked about. They're really good at spotting unusual patterns

305
00:15:34.960 --> 00:15:38.600
<v Speaker 2>in user activity over time, things like logins from weird locations,

306
00:15:38.679 --> 00:15:42.200
<v Speaker 2>super fast clicking that suggests the script. LSTMs can learn

307
00:15:42.240 --> 00:15:45.279
<v Speaker 2>normal behavior and flag these anomalies in real time to

308
00:15:45.320 --> 00:15:46.639
<v Speaker 2>block potential attacks.

309
00:15:46.840 --> 00:15:49.799
<v Speaker 1>That's pretty neat using AI to spot the bad actors.

310
00:15:49.919 --> 00:15:52.759
<v Speaker 1>Are there specific security risks developers need to watch out

311
00:15:52.759 --> 00:15:55.279
<v Speaker 1>for when using AI tools, especially in pythons.

312
00:15:55.399 --> 00:15:59.679
<v Speaker 2>Yes, definitely. The biggest danger always is untrusted input data

313
00:15:59.679 --> 00:16:02.120
<v Speaker 2>coming the web cannot be trusted by default. The source

314
00:16:02.200 --> 00:16:05.200
<v Speaker 2>highlights the Python Pickle library. It's used for saving and

315
00:16:05.240 --> 00:16:08.519
<v Speaker 2>loading Python objects, including models sometimes, but if you load

316
00:16:08.519 --> 00:16:10.879
<v Speaker 2>a Pickle file from an untrusted source, it can be

317
00:16:11.000 --> 00:16:12.960
<v Speaker 2>crafted to execute arbitrary code on.

318
00:16:12.960 --> 00:16:16.519
<v Speaker 1>Your server, arboratory code like deleting files exactly.

319
00:16:16.960 --> 00:16:21.080
<v Speaker 2>The example given is oss dot system rmpstree, which could

320
00:16:21.120 --> 00:16:24.919
<v Speaker 2>potentially wipe a user's entire home directory. It underscores why

321
00:16:25.000 --> 00:16:28.720
<v Speaker 2>you must rigorously validate and sanitize any input, especially data

322
00:16:28.799 --> 00:16:32.480
<v Speaker 2>used to load models or configure systems. Never trust external

323
00:16:32.559 --> 00:16:33.679
<v Speaker 2>data blindly, A.

324
00:16:33.720 --> 00:16:36.840
<v Speaker 1>Very stark warning. Okay, so you've built it, secured it, deployed.

325
00:16:36.840 --> 00:16:39.799
<v Speaker 1>It is the job done. Then you mentioned monitoring.

326
00:16:39.399 --> 00:16:42.279
<v Speaker 2>Right, The job is never truly done. Continuous monitoring is

327
00:16:42.360 --> 00:16:46.840
<v Speaker 2>vital because models go stale, the world changes, user behavior changes,

328
00:16:47.240 --> 00:16:51.039
<v Speaker 2>language evolves. Think about an NLP model trained on texts

329
00:16:51.080 --> 00:16:54.240
<v Speaker 2>from say, two thousand and five. It wouldn't understand someone asking,

330
00:16:54.320 --> 00:16:57.360
<v Speaker 2>can you what's app? Me the wikilink for Avengers endgame?

331
00:16:57.759 --> 00:17:00.320
<v Speaker 2>It wouldn't know WhatsApp wikilinks or that move title.

332
00:17:00.399 --> 00:17:02.320
<v Speaker 1>Good point, language drift exactly.

333
00:17:02.440 --> 00:17:05.279
<v Speaker 2>Model drift happens. The model's performance degrades over time if

334
00:17:05.279 --> 00:17:07.519
<v Speaker 2>it's not updated. So you need systems in place to

335
00:17:07.559 --> 00:17:11.079
<v Speaker 2>continuously monitor performance, retrain models on new data, and redeploy

336
00:17:11.119 --> 00:17:13.640
<v Speaker 2>them to keep them relevant and effective. It's a life cycle.

337
00:17:13.880 --> 00:17:16.680
<v Speaker 1>What a journey we've taken. We started with the why

338
00:17:16.799 --> 00:17:21.119
<v Speaker 1>now of AI, that perfect storm of data algorithms, hardware access,

339
00:17:21.519 --> 00:17:25.920
<v Speaker 1>then saw its invisible hand reshaping search, chatbots, analytics. We

340
00:17:26.079 --> 00:17:29.440
<v Speaker 1>dove into the how the neural networks, the CNNs, the RNNs,

341
00:17:29.519 --> 00:17:33.279
<v Speaker 1>the amazing Python ecosystem with numb PI, pandas Karras, and

342
00:17:33.319 --> 00:17:36.519
<v Speaker 1>how rest APIs and cloud services bridge the gap to

343
00:17:36.559 --> 00:17:41.160
<v Speaker 1>the web, and finally the crucial real world stuff deployment,

344
00:17:41.359 --> 00:17:45.160
<v Speaker 1>the dangers of bias, security threats like untrusted input, and

345
00:17:45.200 --> 00:17:46.799
<v Speaker 1>the need for constant monitoring.

346
00:17:46.960 --> 00:17:49.079
<v Speaker 2>Yeah, you really should have a much clearer picture now

347
00:17:49.160 --> 00:17:51.480
<v Speaker 2>of how deep learning gets woven into the fabric of

348
00:17:51.480 --> 00:17:54.559
<v Speaker 2>the web and just how powerful that Python ecosystem is

349
00:17:54.559 --> 00:17:57.240
<v Speaker 2>for making it happen. And maybe this raises a question

350
00:17:57.240 --> 00:17:59.000
<v Speaker 2>for you listening. We're moving into this era people are

351
00:17:59.000 --> 00:18:01.720
<v Speaker 2>calling software two point zero, where intelligence is baked into

352
00:18:01.759 --> 00:18:04.440
<v Speaker 2>applications from the start. So how might you use these ideas,

353
00:18:04.480 --> 00:18:07.000
<v Speaker 2>maybe to build something new or just understand the smart

354
00:18:07.039 --> 00:18:08.440
<v Speaker 2>tools you use every day a bit better.

355
00:18:08.599 --> 00:18:11.640
<v Speaker 1>Absolutely, whether you're thinking about building that next gen chat,

356
00:18:11.680 --> 00:18:14.839
<v Speaker 1>bought a smarter security system, or maybe you just have

357
00:18:14.880 --> 00:18:17.920
<v Speaker 1>a newfound appreciation for the AI humming away behind your

358
00:18:17.920 --> 00:18:20.920
<v Speaker 1>favorite apps. The concepts and tools we explore today are

359
00:18:21.000 --> 00:18:23.920
<v Speaker 1>right at the center of it all. Keep digging, keep learning.

360
00:18:24.119 --> 00:18:25.839
<v Speaker 1>Thanks for joining us on the Deep Dive, and we'll

361
00:18:25.839 --> 00:18:26.559
<v Speaker 1>see you next time.