WEBVTT

1
00:00:00.080 --> 00:00:02.200
<v Speaker 1>You know, if you look at a microprocessor under an

2
00:00:02.200 --> 00:00:09.199
<v Speaker 1>electron microscope, there's this this expectation of just absolute crystalline perfection.

3
00:00:09.359 --> 00:00:12.320
<v Speaker 2>Oh for sure. I mean it's engineering in its purest,

4
00:00:12.359 --> 00:00:13.720
<v Speaker 2>most rigid form, right.

5
00:00:13.880 --> 00:00:16.440
<v Speaker 1>You see these immaculate grids of silicon, and you just

6
00:00:16.559 --> 00:00:19.960
<v Speaker 1>know exactly what the hardware is doing, like millions of

7
00:00:19.960 --> 00:00:23.359
<v Speaker 1>transistors are flipping between one and zero true or false,

8
00:00:23.440 --> 00:00:24.000
<v Speaker 1>on or off.

9
00:00:24.120 --> 00:00:27.679
<v Speaker 2>Yeah, it's an architecture that's fundamentally built on absolute certainty exactly.

10
00:00:28.399 --> 00:00:31.719
<v Speaker 1>But the paradox we're unpacking in today's deep dive is

11
00:00:31.760 --> 00:00:36.479
<v Speaker 1>that this rigid certainty, it actually creates a structural.

12
00:00:35.960 --> 00:00:40.479
<v Speaker 2>Trap, right, because we construct these flawlessly deterministic environments for

13
00:00:40.560 --> 00:00:42.719
<v Speaker 2>our machines, and then.

14
00:00:42.520 --> 00:00:44.359
<v Speaker 1>We ask them to step outside of.

15
00:00:44.240 --> 00:00:46.960
<v Speaker 2>That environment exactly. We ask them to process the real

16
00:00:47.000 --> 00:00:50.200
<v Speaker 2>world like we expect them to understand a casual conversation

17
00:00:50.320 --> 00:00:53.600
<v Speaker 2>or a Parsi poem, or you know, navigate the nuances

18
00:00:53.640 --> 00:00:55.679
<v Speaker 2>of a simple joke, and the moment.

19
00:00:55.439 --> 00:00:59.359
<v Speaker 1>That microscopic perfection collides with the chaotic, messy reality of

20
00:00:59.439 --> 00:01:02.840
<v Speaker 1>human coog ignition, the system just breaks down. Because human

21
00:01:02.880 --> 00:01:08.359
<v Speaker 1>intelligence isn't a spreadsheet, it's murky it's totally subjective.

22
00:01:08.079 --> 00:01:11.159
<v Speaker 2>It really is, and the gap between the flawless binary

23
00:01:11.200 --> 00:01:14.920
<v Speaker 2>machine and the wonderfully messy human mind is basically the

24
00:01:14.959 --> 00:01:17.400
<v Speaker 2>central hurdle of modern computing.

25
00:01:17.000 --> 00:01:19.439
<v Speaker 1>Which brings us to the stack of notes and excerpts

26
00:01:19.480 --> 00:01:22.280
<v Speaker 1>we have for today. We're looking at a deeply technical,

27
00:01:22.519 --> 00:01:27.719
<v Speaker 1>really paradigm shifting academic book called Artificial Intelligence with Uncertainty,

28
00:01:27.840 --> 00:01:28.719
<v Speaker 1>second edition.

29
00:01:28.640 --> 00:01:33.120
<v Speaker 2>Yes, by Researchers Daily, and you do. It's a dense read,

30
00:01:33.280 --> 00:01:35.000
<v Speaker 2>but fascinating.

31
00:01:34.560 --> 00:01:36.959
<v Speaker 1>So dense, But our mission for you today is to

32
00:01:37.000 --> 00:01:40.519
<v Speaker 1>explore why, after literally decades of trying to make computers

33
00:01:40.560 --> 00:01:44.879
<v Speaker 1>perfectly precise, scientists are now intentionally teaching AI how to

34
00:01:44.920 --> 00:01:45.640
<v Speaker 1>be uncertain.

35
00:01:45.760 --> 00:01:48.920
<v Speaker 2>Yeah. They're using this revolutionary framework called the cloud model

36
00:01:49.280 --> 00:01:52.640
<v Speaker 2>to force precise machines to think in the beautiful, uncertain

37
00:01:52.760 --> 00:01:54.359
<v Speaker 2>gray areas of human language.

38
00:01:54.439 --> 00:01:57.359
<v Speaker 1>Okay, let's unpack this because to really glasp why injecting

39
00:01:57.439 --> 00:02:00.400
<v Speaker 1>uncertainty into AI is a necessity and it's not like

40
00:02:00.439 --> 00:02:03.239
<v Speaker 1>a flaw. We have to start with the universe itself right.

41
00:02:03.120 --> 00:02:05.280
<v Speaker 2>Before we even touch algorithms. We have to recognize that

42
00:02:05.319 --> 00:02:08.400
<v Speaker 2>reality refuses to be put into a neat, predictable.

43
00:02:07.920 --> 00:02:11.439
<v Speaker 1>Box, which wasn't always the scientific consensus, right, I mean,

44
00:02:11.479 --> 00:02:14.879
<v Speaker 1>for a long time, the dominant paradigm was totally deterministic. Oh.

45
00:02:14.879 --> 00:02:19.120
<v Speaker 2>Absolutely. Think back to the nineteenth century. You had Newton

46
00:02:19.199 --> 00:02:22.240
<v Speaker 2>in la Place and this classical view of the universe

47
00:02:22.280 --> 00:02:26.680
<v Speaker 2>as this giant, perfect clockwork mechanism, a clockwork universe exactly.

48
00:02:26.919 --> 00:02:29.960
<v Speaker 2>The assumption was that every particle moves according to fixed

49
00:02:30.199 --> 00:02:34.719
<v Speaker 2>early rules, and the uncertainty we experience was just well

50
00:02:34.759 --> 00:02:37.680
<v Speaker 2>a byproduct of our own ignorance, Like if we just

51
00:02:37.719 --> 00:02:41.240
<v Speaker 2>built better measuring tools, the uncertainty would completely vanish, right,

52
00:02:41.319 --> 00:02:43.879
<v Speaker 2>But modern physics dismantled that pretty quickly.

53
00:02:44.000 --> 00:02:46.599
<v Speaker 1>Yeah. The source text pulls this amazing quote from the

54
00:02:46.599 --> 00:02:49.599
<v Speaker 1>philosopher Carl Popper that serves as the thesis for this

55
00:02:49.639 --> 00:02:53.039
<v Speaker 1>whole problem. He said, all clocks are clouds.

56
00:02:53.400 --> 00:02:56.879
<v Speaker 2>I love that quote. It perfectly captures the illusion of determinism,

57
00:02:57.360 --> 00:03:01.000
<v Speaker 2>because even the most incredibly precise clocks we can engineer,

58
00:03:01.120 --> 00:03:05.120
<v Speaker 2>like atomic clocks measuring the decay of isotopes.

59
00:03:04.520 --> 00:03:07.400
<v Speaker 1>They're actually just clouds of probability at their core.

60
00:03:07.639 --> 00:03:12.439
<v Speaker 2>Exactly at the quantum level, it's all clouds. Werner Heisenberg's

61
00:03:12.520 --> 00:03:17.159
<v Speaker 2>uncertainty principle prove that you literally cannot know both the

62
00:03:17.199 --> 00:03:20.240
<v Speaker 2>exact position and momentum of particles simultaneously.

63
00:03:20.439 --> 00:03:23.919
<v Speaker 1>So uncertainty isn't just a lack of knowledge, No, it's

64
00:03:23.960 --> 00:03:27.599
<v Speaker 1>an objective, mathematical attribute of nature itself. And the book

65
00:03:27.680 --> 00:03:30.800
<v Speaker 1>gives this phenomenal example of how this uncertainty scales, which

66
00:03:30.840 --> 00:03:32.800
<v Speaker 1>I think is so cool. Like it points out that

67
00:03:32.840 --> 00:03:36.400
<v Speaker 1>the dimension of a system is entirely relative to your

68
00:03:36.479 --> 00:03:37.560
<v Speaker 1>vantage point of the Earth.

69
00:03:37.599 --> 00:03:38.000
<v Speaker 2>Example.

70
00:03:38.159 --> 00:03:40.639
<v Speaker 1>Yeah, if you look at the Earth from outside the

71
00:03:40.680 --> 00:03:44.680
<v Speaker 1>Milky Way, it's a single certain point, a zero dimensional dot.

72
00:03:45.120 --> 00:03:47.400
<v Speaker 1>But zoom into the Solar system and it becomes a

73
00:03:47.439 --> 00:03:49.240
<v Speaker 1>one dimensional elliptical orbit.

74
00:03:49.400 --> 00:03:51.599
<v Speaker 2>Right. If you're standing on the surface, it operates as

75
00:03:51.639 --> 00:03:53.800
<v Speaker 2>a flat two dimensional plane exactly.

76
00:03:54.240 --> 00:03:57.120
<v Speaker 1>Or look at the coastline paradox, The length of a

77
00:03:57.159 --> 00:03:59.879
<v Speaker 1>coastline actually changes depending on the size of the ruler

78
00:04:00.159 --> 00:04:02.280
<v Speaker 1>use to measure it. The closer you look, the more

79
00:04:02.360 --> 00:04:04.120
<v Speaker 1>chaotic the boundary becomes.

80
00:04:03.800 --> 00:04:07.240
<v Speaker 2>Which is exactly why the scientific community developed entropy is a.

81
00:04:07.199 --> 00:04:09.520
<v Speaker 1>Metric to mathematically quantify that disorder.

82
00:04:09.599 --> 00:04:12.919
<v Speaker 2>Right, yeah, I mean it started in thermodynamics back in

83
00:04:12.960 --> 00:04:16.519
<v Speaker 2>the eighteen fifties with Rudolph Clausis tracking how heat dissipates,

84
00:04:17.000 --> 00:04:19.759
<v Speaker 2>and then o Lidvig Boltzmann mapped it to the random

85
00:04:19.759 --> 00:04:20.800
<v Speaker 2>movement of molecules.

86
00:04:21.199 --> 00:04:24.040
<v Speaker 1>But the big leap for AI happened later.

87
00:04:24.120 --> 00:04:30.079
<v Speaker 2>Right right. In nineteen forty eight, Claude Shannon introduced information entropy.

88
00:04:30.279 --> 00:04:33.519
<v Speaker 2>He proved that you could mathematically measure the uncertainty of

89
00:04:33.519 --> 00:04:36.079
<v Speaker 2>information as it travels through a communication channel.

90
00:04:36.240 --> 00:04:38.160
<v Speaker 1>Okay, but hang on, let me push back on this

91
00:04:38.199 --> 00:04:39.040
<v Speaker 1>premise for a second.

92
00:04:39.079 --> 00:04:40.279
<v Speaker 2>Sure, good, I.

93
00:04:40.279 --> 00:04:43.079
<v Speaker 1>Understand that the physical universe is a cloud of chaos,

94
00:04:43.560 --> 00:04:47.720
<v Speaker 1>but computers are literally artificial environments. They are the clocks

95
00:04:47.720 --> 00:04:49.439
<v Speaker 1>we built to escape the clouds.

96
00:04:50.079 --> 00:04:51.160
<v Speaker 2>I see where you're going.

97
00:04:51.040 --> 00:04:54.680
<v Speaker 1>Right, Like, if our hardware is fundamentally binary, isn't injecting

98
00:04:54.759 --> 00:04:59.160
<v Speaker 1>chaos into a computational system counterproductive? I mean, if I'm

99
00:04:59.160 --> 00:05:03.279
<v Speaker 1>building an AI for medical diagnostics or autonomous driving, forcing

100
00:05:03.279 --> 00:05:06.680
<v Speaker 1>a perfect machine to act imperfectly sounds like a recipe

101
00:05:06.680 --> 00:05:08.000
<v Speaker 1>for catastrophic failure.

102
00:05:08.120 --> 00:05:11.279
<v Speaker 2>It seems counterintuitive. I'll give you that. But what's fascinating

103
00:05:11.319 --> 00:05:14.079
<v Speaker 2>here is that rigid systems are actually more brittle in

104
00:05:14.120 --> 00:05:17.480
<v Speaker 2>the real world. Brittle if you force a machine to

105
00:05:17.720 --> 00:05:22.000
<v Speaker 2>only view the world through strict, deterministic boundaries, it will

106
00:05:22.040 --> 00:05:25.240
<v Speaker 2>fail the second it encounters an edge case that doesn't

107
00:05:25.240 --> 00:05:26.639
<v Speaker 2>fit its exact parameters.

108
00:05:26.759 --> 00:05:29.600
<v Speaker 1>Oh, because the real world isn't binary exactly.

109
00:05:29.959 --> 00:05:33.160
<v Speaker 2>Think about a self driving car. It doesn't encounter a

110
00:05:33.199 --> 00:05:37.360
<v Speaker 2>neat pedestrian or no pedestrian scenario. It encounters like a

111
00:05:37.399 --> 00:05:39.600
<v Speaker 2>plastic bag blowing across the street. That might be a

112
00:05:39.639 --> 00:05:42.519
<v Speaker 2>small animal under lighting conditions, that might be a shadow.

113
00:05:42.759 --> 00:05:44.399
<v Speaker 1>Right right, It's a gray area.

114
00:05:44.800 --> 00:05:48.480
<v Speaker 2>And for an AI to safely navigate that reality, it

115
00:05:48.519 --> 00:05:50.639
<v Speaker 2>can't just throw an error code when things get fuzzy.

116
00:05:50.680 --> 00:05:55.160
<v Speaker 2>It must be mathematically equipped to process and compute that entropy.

117
00:05:55.279 --> 00:05:57.240
<v Speaker 1>That makes a lot of sense. And I mean since

118
00:05:57.240 --> 00:05:59.800
<v Speaker 1>the physical world is uncertain, the way human beings p

119
00:06:00.360 --> 00:06:03.560
<v Speaker 1>and describe it is also totally uncertain, which leads us

120
00:06:03.600 --> 00:06:05.879
<v Speaker 1>directly to the ultimate carrier of human.

121
00:06:05.639 --> 00:06:07.279
<v Speaker 2>Intelligence, natural language.

122
00:06:07.319 --> 00:06:10.240
<v Speaker 1>Exactly natural language. If we want to simulate human thought,

123
00:06:10.240 --> 00:06:12.480
<v Speaker 1>we have to decode our language, and human language is

124
00:06:12.519 --> 00:06:15.399
<v Speaker 1>just full of randomness and fuzziness.

125
00:06:15.759 --> 00:06:19.600
<v Speaker 2>It really is the ultimate expression of cognitive fuzziness, and

126
00:06:19.639 --> 00:06:24.120
<v Speaker 2>it's the essential distinction between a machine processing data and

127
00:06:24.199 --> 00:06:26.959
<v Speaker 2>an entity demonstrating actual common sense.

128
00:06:27.279 --> 00:06:30.639
<v Speaker 1>The text highlights this beautifully by looking at how language

129
00:06:30.680 --> 00:06:35.560
<v Speaker 1>completely resists numerical substitution. There are these uh poetic examples

130
00:06:35.639 --> 00:06:36.199
<v Speaker 1>they use, oh.

131
00:06:36.199 --> 00:06:37.399
<v Speaker 2>With Tang dynasty poems.

132
00:06:37.480 --> 00:06:40.720
<v Speaker 1>Yeah, like Wang Bo writing about a proud eagle flying

133
00:06:40.759 --> 00:06:45.920
<v Speaker 1>with rosy clouds, or Cowzyugen's description of Lindaiu's eyebrows which

134
00:06:46.000 --> 00:06:47.879
<v Speaker 1>seem to knit and yet not to knit.

135
00:06:48.120 --> 00:06:51.240
<v Speaker 2>I mean, you cannot map an RGBX code to rosy.

136
00:06:51.480 --> 00:06:54.000
<v Speaker 1>No, you really can't, and you certainly can't assign a

137
00:06:54.040 --> 00:06:56.800
<v Speaker 1>boolean value to eyebrows that seem to knit and yet

138
00:06:56.839 --> 00:06:59.519
<v Speaker 1>not to knit. Like the concept exists entirely in a

139
00:06:59.519 --> 00:07:00.399
<v Speaker 1>state of conradiction.

140
00:07:00.600 --> 00:07:03.800
<v Speaker 2>But any human reading that sentence understands the exact emotional

141
00:07:03.879 --> 00:07:05.439
<v Speaker 2>weight being conveyed exactly.

142
00:07:05.480 --> 00:07:07.000
<v Speaker 1>And we don't even need to look at poetry to

143
00:07:07.000 --> 00:07:09.199
<v Speaker 1>see this. We see it in the most basic functional

144
00:07:09.199 --> 00:07:12.079
<v Speaker 1>definitions of common sense. Like the text breaks down how

145
00:07:12.079 --> 00:07:14.800
<v Speaker 1>we distinguish between a cup, a plate, and a bowl.

146
00:07:14.560 --> 00:07:18.319
<v Speaker 2>Which is fascinating because there's absolutely no strict mathematical ratio

147
00:07:18.720 --> 00:07:21.120
<v Speaker 2>of width to height that defines a bowl.

148
00:07:21.519 --> 00:07:24.800
<v Speaker 1>Right, A cup has a handle, a plate is flat

149
00:07:24.839 --> 00:07:28.319
<v Speaker 1>and holds rice. A bowl has an edge and holds soup.

150
00:07:28.519 --> 00:07:30.680
<v Speaker 2>But if I hand you an object that is like

151
00:07:31.639 --> 00:07:34.040
<v Speaker 2>slightly too wide for a standard cup, but a little

152
00:07:34.040 --> 00:07:37.600
<v Speaker 2>too deep for a standard plate. Your brain doesn't short circuit.

153
00:07:37.839 --> 00:07:40.439
<v Speaker 1>No, you don't need a ruler to categorize it. Your

154
00:07:40.480 --> 00:07:43.519
<v Speaker 1>common sense just steps in and flexes the definition based

155
00:07:43.560 --> 00:07:45.000
<v Speaker 1>on context.

156
00:07:44.439 --> 00:07:47.920
<v Speaker 2>And in AI. This human common sense is highly relative.

157
00:07:48.439 --> 00:07:53.040
<v Speaker 2>It changes by time, place, and people. But that flexibility

158
00:07:53.319 --> 00:07:54.639
<v Speaker 2>is what allows it to survive.

159
00:07:54.920 --> 00:07:57.399
<v Speaker 1>I want you, the listener, to think about this practically

160
00:07:57.439 --> 00:08:00.480
<v Speaker 1>for a second. Think about how often you lie on

161
00:08:00.519 --> 00:08:02.759
<v Speaker 1>this fuzziness just to get through a work day.

162
00:08:02.920 --> 00:08:03.839
<v Speaker 2>Oh, all the time.

163
00:08:04.000 --> 00:08:06.639
<v Speaker 1>Right. You'll sit in a meeting and say the project

164
00:08:06.680 --> 00:08:08.319
<v Speaker 1>is probably going to take about three weeks.

165
00:08:08.399 --> 00:08:10.560
<v Speaker 2>Usually that sentence is a mathematical nightmare.

166
00:08:10.680 --> 00:08:14.759
<v Speaker 1>It really is probably about and usually are massive variables.

167
00:08:15.160 --> 00:08:18.519
<v Speaker 1>But your colleagues understand the exact level of confidence you

168
00:08:18.560 --> 00:08:21.319
<v Speaker 1>are projecting. You don't get confused by the uncertainty. It

169
00:08:21.360 --> 00:08:24.399
<v Speaker 1>actually gives your brain the flexibility to understand the context.

170
00:08:24.600 --> 00:08:27.519
<v Speaker 2>If you instead said the project will conclude in exactly

171
00:08:27.519 --> 00:08:30.079
<v Speaker 2>five hundred and four hours, twelve minutes and six seconds,

172
00:08:30.360 --> 00:08:31.560
<v Speaker 2>they wouldn't think you were smart.

173
00:08:31.839 --> 00:08:35.600
<v Speaker 1>No, they'd think you fundamentally misunderstood the chaotic nature of

174
00:08:35.639 --> 00:08:39.480
<v Speaker 1>project management. The uncertainty in your language doesn't impede understanding,

175
00:08:39.559 --> 00:08:40.919
<v Speaker 1>it actively enables it.

176
00:08:41.120 --> 00:08:44.039
<v Speaker 2>But the problem is the dominant schools of AI development

177
00:08:44.360 --> 00:08:47.679
<v Speaker 2>have basically spent the last sixty years trying to strip

178
00:08:47.720 --> 00:08:49.399
<v Speaker 2>that vital fuzziness out of the system.

179
00:08:49.480 --> 00:08:52.519
<v Speaker 1>Okay, so let's contextualize that. Because if human thought is

180
00:08:52.559 --> 00:08:55.960
<v Speaker 1>fluid and language is fuzzy, how have computer scientists spent

181
00:08:56.000 --> 00:08:59.679
<v Speaker 1>the last six decades trying to shove this messy reality

182
00:08:59.759 --> 00:09:01.039
<v Speaker 1>into rigid code.

183
00:09:01.840 --> 00:09:02.679
<v Speaker 2>What's been a journey?

184
00:09:02.759 --> 00:09:04.960
<v Speaker 1>Yeah, let's do a rapid fire history of AI. Here.

185
00:09:05.320 --> 00:09:08.559
<v Speaker 1>It basically starts at the nineteen fifty six Dartmouth symposium,

186
00:09:08.639 --> 00:09:12.279
<v Speaker 1>right with guys like McCarthy, Minski, and Shannon. That's where

187
00:09:12.279 --> 00:09:14.799
<v Speaker 1>the term artificial intelligence was born exactly.

188
00:09:15.159 --> 00:09:18.120
<v Speaker 2>And over the years we've seen huge milestones the logic,

189
00:09:18.200 --> 00:09:23.440
<v Speaker 2>theorist proving math theorems, gameplaying AI evolving from Samuel's Checkers

190
00:09:23.440 --> 00:09:27.440
<v Speaker 2>to Deep Blue beating Gary kasprov at chess, and eventually

191
00:09:27.559 --> 00:09:29.600
<v Speaker 2>Google's Alpha Go defeating Fanwei.

192
00:09:29.840 --> 00:09:33.320
<v Speaker 1>But academically, the text breaks down three major schools of

193
00:09:33.360 --> 00:09:34.480
<v Speaker 1>thought that emerged from all this.

194
00:09:34.840 --> 00:09:37.600
<v Speaker 2>Right, First, you have symbolism. This is the idea that

195
00:09:37.639 --> 00:09:40.519
<v Speaker 2>thinking is just computation. It's rule based logic.

196
00:09:40.840 --> 00:09:43.759
<v Speaker 1>If A then B, So like giving the AI a

197
00:09:43.799 --> 00:09:45.039
<v Speaker 1>strict recipe book.

198
00:09:45.200 --> 00:09:49.039
<v Speaker 2>Exactly. This gave us early expert systems like mycion for

199
00:09:49.120 --> 00:09:53.639
<v Speaker 2>medical diagnoses, but they were incredibly brittle. Then you have connectionism.

200
00:09:53.879 --> 00:09:57.360
<v Speaker 1>This is simulating the brain structure right, artificial neural networks

201
00:09:57.360 --> 00:09:59.559
<v Speaker 1>and backpropagation like modern deep learning.

202
00:09:59.720 --> 00:10:03.360
<v Speaker 2>Yes, instead of hard coding rules, you let the system

203
00:10:03.399 --> 00:10:07.200
<v Speaker 2>adjust the weights of its own connections across massive data sets.

204
00:10:07.360 --> 00:10:10.240
<v Speaker 1>So connectionism is like trying to build the AI's muscle

205
00:10:10.279 --> 00:10:12.919
<v Speaker 1>memory by mimicking the brain's wiring.

206
00:10:13.039 --> 00:10:14.799
<v Speaker 2>That's a great way to put it. And the third

207
00:10:14.799 --> 00:10:18.759
<v Speaker 2>school is behaviorism. This focus is on the perception behavior model,

208
00:10:18.799 --> 00:10:21.519
<v Speaker 2>intelligent control, and adaptation.

209
00:10:21.320 --> 00:10:24.480
<v Speaker 1>Like giving the AI reflexes to react to its environment

210
00:10:24.639 --> 00:10:27.840
<v Speaker 1>like that inverted pendulum robot mentioned in the text.

211
00:10:27.919 --> 00:10:32.159
<v Speaker 2>Exactly reflexes, muscle memory, and recipe books. But while these

212
00:10:32.159 --> 00:10:35.720
<v Speaker 2>methods conquered chess and go, they still struggle with the fluid,

213
00:10:36.000 --> 00:10:38.960
<v Speaker 2>non mathematical nature of human language in consciousness.

214
00:10:39.039 --> 00:10:42.480
<v Speaker 1>So to bridge this gap between precise math and fuzzy

215
00:10:42.559 --> 00:10:48.120
<v Speaker 1>human language, the authors introduce this groundbreaking framework, the cloud model.

216
00:10:47.960 --> 00:10:50.720
<v Speaker 2>And this is where things get really paradigm shifting.

217
00:10:51.000 --> 00:10:53.279
<v Speaker 1>To explain it, we have to talk about this classic

218
00:10:53.360 --> 00:10:56.639
<v Speaker 1>experiment from the text by the Chinese scholar Nan Lunzang.

219
00:10:57.080 --> 00:11:00.200
<v Speaker 1>He wanted to look at fuzziness versus randomness, so he

220
00:11:00.240 --> 00:11:02.600
<v Speaker 1>asked one hundred and twenty nine people to define the

221
00:11:02.639 --> 00:11:04.360
<v Speaker 1>age range for the concept of youth.

222
00:11:04.639 --> 00:11:07.960
<v Speaker 2>Such a seemingly simple concept, but the data is wild.

223
00:11:08.200 --> 00:11:11.360
<v Speaker 1>It really is. So Zeg measured the membership degree, like,

224
00:11:11.399 --> 00:11:13.879
<v Speaker 1>what percentage of people agree this age is youth? At

225
00:11:13.919 --> 00:11:16.440
<v Speaker 1>age twenty to twenty four, the membership degree is one

226
00:11:17.000 --> 00:11:18.159
<v Speaker 1>one hundred percent agreement.

227
00:11:18.240 --> 00:11:20.679
<v Speaker 2>Everyone agrees a twenty two year old is a youth, right.

228
00:11:20.720 --> 00:11:23.000
<v Speaker 1>But at age thirty it drops two point five nine

229
00:11:23.080 --> 00:11:25.720
<v Speaker 1>sixty nine, and by age thirty six it's just point

230
00:11:25.799 --> 00:11:27.480
<v Speaker 1>zero zero zero seven eight.

231
00:11:27.559 --> 00:11:29.440
<v Speaker 2>What this points out is that the concept of youth

232
00:11:29.480 --> 00:11:31.360
<v Speaker 2>isn't just fuzzy, it's genuinely random.

233
00:11:31.480 --> 00:11:32.799
<v Speaker 1>Wait random? How so?

234
00:11:33.320 --> 00:11:35.679
<v Speaker 2>Well, if you ask the exact same person on a

235
00:11:35.679 --> 00:11:37.960
<v Speaker 2>Tuesday if a thirty year old is a youth, they

236
00:11:38.039 --> 00:11:40.600
<v Speaker 2>might say yes. But if you ask them on a Friday,

237
00:11:40.879 --> 00:11:44.080
<v Speaker 2>depending on their mood or context, their answer might change

238
00:11:44.120 --> 00:11:44.559
<v Speaker 2>oh wow.

239
00:11:44.639 --> 00:11:48.759
<v Speaker 1>So it completely violates traditional probabilities law of excluded middle.

240
00:11:48.960 --> 00:11:52.840
<v Speaker 2>Exactly, a person can't be simultaneously a youth and not

241
00:11:52.879 --> 00:11:56.279
<v Speaker 2>a youth in classical logic, but in human cognition they

242
00:11:56.399 --> 00:11:57.519
<v Speaker 2>absolutely can be.

243
00:11:57.639 --> 00:12:00.480
<v Speaker 1>You're sure? It gets really interesting. The cloud model takes

244
00:12:00.480 --> 00:12:04.159
<v Speaker 1>this qualitative concept like youth, and translates it into a

245
00:12:04.279 --> 00:12:06.440
<v Speaker 1>quantitative distribution of cloud drops.

246
00:12:06.519 --> 00:12:09.159
<v Speaker 2>Right. It doesn't treat a word as a single data point.

247
00:12:09.480 --> 00:12:12.039
<v Speaker 2>It treats it like a literal cloud made of thousands

248
00:12:12.080 --> 00:12:13.600
<v Speaker 2>of randomized drops of data.

249
00:12:13.639 --> 00:12:15.799
<v Speaker 1>But if a qualitative word is a cloud made of

250
00:12:15.879 --> 00:12:19.440
<v Speaker 1>data drops, how do we mathematically describe the shape of

251
00:12:19.440 --> 00:12:22.360
<v Speaker 1>that cloud to a computer? Like? What's the digital DNA

252
00:12:22.440 --> 00:12:23.120
<v Speaker 1>of a concept?

253
00:12:23.240 --> 00:12:27.960
<v Speaker 2>The researchers broke it down into three digital characteristics, expectation, entropy,

254
00:12:28.039 --> 00:12:31.720
<v Speaker 2>and hyper entropy X and he yes. The first one,

255
00:12:31.919 --> 00:12:36.120
<v Speaker 2>expectation or X, is the most representative point the classic sample,

256
00:12:36.440 --> 00:12:37.679
<v Speaker 2>the dead center of the target.

257
00:12:37.879 --> 00:12:40.919
<v Speaker 1>Okay, let me test this. If the concept is a

258
00:12:41.039 --> 00:12:45.559
<v Speaker 1>normal commute to work. The expectation the X is twenty

259
00:12:45.559 --> 00:12:47.600
<v Speaker 1>five minutes, that's the center point perfect.

260
00:12:47.799 --> 00:12:52.039
<v Speaker 2>The second is entropy or INQ. This is the uncertainty measurement.

261
00:12:52.080 --> 00:12:54.159
<v Speaker 2>It dictates how far the drops disperse.

262
00:12:54.399 --> 00:12:58.159
<v Speaker 1>So the entropy is the standard traffic variance, say between

263
00:12:58.200 --> 00:12:59.120
<v Speaker 1>twenty and thirty minutes.

264
00:12:59.240 --> 00:13:02.320
<v Speaker 2>Exactly, it's the width of the cloud. But the third

265
00:13:02.440 --> 00:13:05.600
<v Speaker 2>characteristic is the trickiest hyper entropy or he.

266
00:13:05.919 --> 00:13:08.600
<v Speaker 1>Yeah, the uncertainty of the entropy itself. What is that?

267
00:13:08.720 --> 00:13:11.759
<v Speaker 1>Is that like a sutty snowstorm messing up the commute?

268
00:13:11.919 --> 00:13:12.440
<v Speaker 2>Not quite?

269
00:13:12.639 --> 00:13:13.080
<v Speaker 1>Yeah.

270
00:13:13.159 --> 00:13:15.360
<v Speaker 2>If we connect this to the bigger picture, hyper entropy

271
00:13:15.480 --> 00:13:19.000
<v Speaker 2>is essentially the measure of consensus. Consensus, Yeah, it measures

272
00:13:19.000 --> 00:13:22.200
<v Speaker 2>the thickness of the cloud's fuzziness. If a concept has

273
00:13:22.559 --> 00:13:25.360
<v Speaker 2>hyper entropy, it means it's a concept people fundamentally disagree

274
00:13:25.360 --> 00:13:25.480
<v Speaker 2>on it.

275
00:13:25.559 --> 00:13:27.559
<v Speaker 1>Oh I see, so a normal commute has low hyper

276
00:13:27.720 --> 00:13:29.799
<v Speaker 1>entropy because we mostly agree on what that means.

277
00:13:29.960 --> 00:13:32.519
<v Speaker 2>The line is thin, right, But think about a concept

278
00:13:32.519 --> 00:13:33.559
<v Speaker 2>like a fair salary.

279
00:13:33.759 --> 00:13:36.320
<v Speaker 1>Oh man, the hyper entropy on that would be massive.

280
00:13:36.399 --> 00:13:39.039
<v Speaker 1>My definition of a fair salary variant is going to

281
00:13:39.039 --> 00:13:41.320
<v Speaker 1>be wildly different from a CEO's definition.

282
00:13:41.759 --> 00:13:46.200
<v Speaker 2>Exactly. The conceptual boundary is incredibly thick and heavily debated.

283
00:13:46.840 --> 00:13:50.879
<v Speaker 2>Low hyper entropy means it's universally accepted common sense.

284
00:13:51.600 --> 00:13:54.519
<v Speaker 1>So once we have this DNA, the expectation entropy and

285
00:13:54.600 --> 00:13:57.960
<v Speaker 1>hyper entropy. We can oftually build algorithms that allow computers

286
00:13:58.000 --> 00:13:59.200
<v Speaker 1>to generate these concepts.

287
00:13:59.279 --> 00:14:03.799
<v Speaker 2>Yes, the researchers call them forward cloud generators, and it

288
00:14:03.840 --> 00:14:07.840
<v Speaker 2>turns out these concepts take on very specific visual shapes.

289
00:14:08.360 --> 00:14:12.720
<v Speaker 1>The text describes this forward Gaussian cloud algorithm, and it's

290
00:14:12.759 --> 00:14:17.159
<v Speaker 1>fundamentally different from a standard Gaussian distribution right because it

291
00:14:17.279 --> 00:14:19.759
<v Speaker 1>uses iterative randomness.

292
00:14:19.120 --> 00:14:22.840
<v Speaker 2>Right, randomness built on top of randomness. It takes the hyperanthropy,

293
00:14:23.159 --> 00:14:25.919
<v Speaker 2>uses it to generate a random variance and uses that

294
00:14:26.039 --> 00:14:28.600
<v Speaker 2>variance to plot a single cloud drop.

295
00:14:28.559 --> 00:14:31.159
<v Speaker 1>That is wild and depending on the inputs, it constructs

296
00:14:31.200 --> 00:14:34.679
<v Speaker 1>different shapes of clouds to model real world concepts, like

297
00:14:34.799 --> 00:14:35.799
<v Speaker 1>the symmetric cloud.

298
00:14:36.080 --> 00:14:39.000
<v Speaker 2>Yes, that's for a concept like medium height. It tapers

299
00:14:39.000 --> 00:14:40.960
<v Speaker 2>off evenly in both directions. You can be too tall

300
00:14:40.960 --> 00:14:42.279
<v Speaker 2>to be medium or too short.

301
00:14:42.440 --> 00:14:43.840
<v Speaker 1>But then there's the half cloud.

302
00:14:43.679 --> 00:14:46.440
<v Speaker 2>Right for asymmetrical concepts like having a fever.

303
00:14:46.639 --> 00:14:50.120
<v Speaker 1>Because normal body temp is thirty seven degrees celsius and

304
00:14:50.159 --> 00:14:52.399
<v Speaker 1>a fever only goes up. You can't have a fever

305
00:14:52.440 --> 00:14:55.080
<v Speaker 1>of thirty five degrees. So the cloud is literally sliced

306
00:14:55.080 --> 00:14:56.159
<v Speaker 1>in half exactly.

307
00:14:56.519 --> 00:14:59.720
<v Speaker 2>And for really complex realities you use a combined cloud

308
00:15:00.120 --> 00:15:02.399
<v Speaker 2>like the concept of a white collar salary.

309
00:15:02.080 --> 00:15:05.519
<v Speaker 1>Right because salaries clumped tightly at the entry level plateau

310
00:15:05.679 --> 00:15:09.240
<v Speaker 1>in middle management and then have this long chaotic tail

311
00:15:09.320 --> 00:15:12.639
<v Speaker 1>for executives. A simple bell curve can't model that, but

312
00:15:12.679 --> 00:15:14.720
<v Speaker 1>a combined cloud maps it perfectly.

313
00:15:14.960 --> 00:15:17.960
<v Speaker 2>And these multi dimensional models are what allow AI to

314
00:15:18.000 --> 00:15:19.799
<v Speaker 2>perform soft reasoning.

315
00:15:19.480 --> 00:15:21.519
<v Speaker 1>Thinking with words instead of just numbers.

316
00:15:21.600 --> 00:15:24.720
<v Speaker 2>Exactly. It opens the door for massive leaps in data mining,

317
00:15:25.000 --> 00:15:27.320
<v Speaker 2>intelligent control, and network science.

318
00:15:27.879 --> 00:15:30.360
<v Speaker 1>Like an autonomous drone trying to land in a storm,

319
00:15:30.759 --> 00:15:34.200
<v Speaker 1>a rigid AI would overcorrect it every wind, gust and crash,

320
00:15:34.399 --> 00:15:37.759
<v Speaker 1>but a cloud model controller understands the fuzziness of the environment.

321
00:15:38.200 --> 00:15:41.840
<v Speaker 1>It blends the overlapping data clouds to execute a smooth landing.

322
00:15:41.960 --> 00:15:44.279
<v Speaker 2>It makes decisions based on the shape of the uncertainty.

323
00:15:44.399 --> 00:15:46.480
<v Speaker 1>That is just mind blowing. So what does this all

324
00:15:46.519 --> 00:15:49.840
<v Speaker 1>mean for you, the listener? Let's recap this journey. We

325
00:15:49.919 --> 00:15:52.240
<v Speaker 1>started by seeing that the universe is a cloud, not

326
00:15:52.320 --> 00:15:53.240
<v Speaker 1>a perfect.

327
00:15:52.840 --> 00:15:57.240
<v Speaker 2>Clock, and our natural language reflects that beautiful chaos exactly.

328
00:15:57.720 --> 00:16:00.480
<v Speaker 1>But for sixty years we tried to shove that fluid

329
00:16:00.600 --> 00:16:04.759
<v Speaker 1>language into rigid binary systems. Now by using the cloud

330
00:16:04.799 --> 00:16:08.480
<v Speaker 1>model mapping, the expectation, the entropy and the hyper entropy

331
00:16:08.519 --> 00:16:11.639
<v Speaker 1>of our concepts. We are finally teaching AI that it's

332
00:16:11.720 --> 00:16:13.759
<v Speaker 1>okay not to be one hundred percent.

333
00:16:13.600 --> 00:16:17.440
<v Speaker 2>Certain, because uncertainty is the very foundation of intelligence. And

334
00:16:17.480 --> 00:16:20.399
<v Speaker 2>I think it leads us with a final, really provocative

335
00:16:20.440 --> 00:16:21.279
<v Speaker 2>thought to mull over.

336
00:16:21.559 --> 00:16:22.279
<v Speaker 1>Oh, what's that.

337
00:16:22.559 --> 00:16:26.039
<v Speaker 2>We've spent decades trying to eliminate uncertainty in machines, to

338
00:16:26.080 --> 00:16:30.159
<v Speaker 2>make them perfect, flawless tools. But if true intelligence requires

339
00:16:30.159 --> 00:16:34.000
<v Speaker 2>the fuzziness of clouds, the subjectivity of hyper entropy, and

340
00:16:34.039 --> 00:16:37.480
<v Speaker 2>the flexibility of common sense, Yeah, at what point does

341
00:16:37.519 --> 00:16:40.120
<v Speaker 2>programming a machine to be uncertain make it less like

342
00:16:40.159 --> 00:16:41.360
<v Speaker 2>a tool and more like us?

343
00:16:41.480 --> 00:16:43.840
<v Speaker 1>Wow? That really is the question, isn't it. If we

344
00:16:43.879 --> 00:16:45.600
<v Speaker 1>teach them how to be unsure, we might just be

345
00:16:45.639 --> 00:16:47.440
<v Speaker 1>teaching them how to be human. Thank you so much

346
00:16:47.440 --> 00:16:49.159
<v Speaker 1>for joining us on this deep dive. I hope the

347
00:16:49.200 --> 00:16:52.200
<v Speaker 1>next time you use words like about or probably you

348
00:16:52.279 --> 00:16:56.759
<v Speaker 1>appreciate the hyperintropic chaos in your own mind. Embrace the

349
00:16:56.840 --> 00:16:59.200
<v Speaker 1>gray areas. Everyone. We'll catch you next time.