WEBVTT

1
00:00:00.120 --> 00:00:04.240
<v Speaker 1>Being atoma got bugger on the heartanir, the hullah, the

2
00:00:04.400 --> 00:00:09.759
<v Speaker 1>haman a heckol ah ers where auraslance Ladani, the tain

3
00:00:09.919 --> 00:00:13.359
<v Speaker 1>knudov shoo that would kill nagatarhu la hula humperodes and

4
00:00:13.519 --> 00:00:19.800
<v Speaker 1>udera lanse is phaederlatgun's throw tasha tapa siranashka agus nyav

5
00:00:20.120 --> 00:00:22.480
<v Speaker 1>and erfad is fugulm more.

6
00:00:23.920 --> 00:00:24.079
<v Speaker 2>Hi.

7
00:00:24.199 --> 00:00:28.280
<v Speaker 1>A punk ie is called locrilala a on tudoros auraslanse

8
00:00:28.359 --> 00:00:30.839
<v Speaker 1>if we real this in the heron, Hey, Mark.

9
00:00:30.679 --> 00:00:31.440
<v Speaker 3>How are you?

10
00:00:31.879 --> 00:00:34.479
<v Speaker 2>I'm not great a sore throat.

11
00:00:34.960 --> 00:00:36.520
<v Speaker 4>I think I'm coming down with a cold.

12
00:00:36.920 --> 00:00:40.200
<v Speaker 3>Know you should try a Vogel sre throat spray. It's

13
00:00:40.200 --> 00:00:43.000
<v Speaker 3>a natural remedy with achinasia and sage and can treat

14
00:00:43.079 --> 00:00:45.520
<v Speaker 3>symptoms of coals and flus, including sore throats.

15
00:00:45.719 --> 00:00:47.679
<v Speaker 2>Perfect. I'd give that a triso.

16
00:00:47.880 --> 00:00:51.039
<v Speaker 3>We leave sort throat symptoms with a vogel A quinapoor

17
00:00:51.200 --> 00:00:54.799
<v Speaker 3>sore throat spray made from echinasia and sage herbs and

18
00:00:54.920 --> 00:00:58.359
<v Speaker 3>designed to reach irritated parts of the throat, available from

19
00:00:58.399 --> 00:01:00.719
<v Speaker 3>health stores and pharmacies nationwide.

20
00:01:01.000 --> 00:01:04.359
<v Speaker 2>Always read the leafless Oh hey, it's the tag that

21
00:01:04.439 --> 00:01:06.519
<v Speaker 2>you wish you'd cut out of your shirt, Ali Ward,

22
00:01:06.760 --> 00:01:10.719
<v Speaker 2>and for every one of us that has seen AI

23
00:01:10.840 --> 00:01:13.280
<v Speaker 2>become more and more present in our lives and wondered,

24
00:01:13.519 --> 00:01:16.400
<v Speaker 2>is anyone driving this bus? I have here for you

25
00:01:16.439 --> 00:01:18.519
<v Speaker 2>a chat with an expert who tells us exactly who

26
00:01:18.560 --> 00:01:21.000
<v Speaker 2>is driving the bus and where it could be headed.

27
00:01:21.280 --> 00:01:24.640
<v Speaker 2>Is AI evil? Does AI even care about us? Is

28
00:01:24.680 --> 00:01:27.159
<v Speaker 2>it going to kill us? Should we feel bad for it?

29
00:01:27.719 --> 00:01:30.000
<v Speaker 2>Don't ask me, I'm not theologist. We're going to get

30
00:01:30.040 --> 00:01:32.359
<v Speaker 2>to it now. This expert is a Senior Fellow in

31
00:01:32.439 --> 00:01:35.560
<v Speaker 2>Trustworthy AI and an assistant professor at the School of

32
00:01:35.640 --> 00:01:39.560
<v Speaker 2>Computer Science and Statistics at Trinity College in Dublin, Ireland.

33
00:01:39.840 --> 00:01:44.599
<v Speaker 2>And they're cognitive scientists. They research ethics in artificial intelligence,

34
00:01:44.920 --> 00:01:49.239
<v Speaker 2>and they've published papers with such titles as The Forgotten

35
00:01:49.319 --> 00:01:54.599
<v Speaker 2>Margins of AI, Ethics, Toward Decolonizing Computational Sciences, the Unseen

36
00:01:54.680 --> 00:01:58.799
<v Speaker 2>Blackfaces of AI algorithms, and the Values Encoded in Machine

37
00:01:58.879 --> 00:02:01.760
<v Speaker 2>Learning research. So they're on it, and I got to

38
00:02:02.000 --> 00:02:04.799
<v Speaker 2>sit down and visit and chat in person when I

39
00:02:04.840 --> 00:02:08.599
<v Speaker 2>was in Ireland just last month. Also just a pleasing

40
00:02:08.919 --> 00:02:12.360
<v Speaker 2>aesthetic side note. They were born in Ethiopia but lives

41
00:02:12.360 --> 00:02:17.000
<v Speaker 2>in Ireland, and this expert has the most melodic cadence,

42
00:02:17.439 --> 00:02:19.879
<v Speaker 2>just like pyork I was mesmerized. We're going to get

43
00:02:20.000 --> 00:02:21.800
<v Speaker 2>to all that in a sec but first, if you

44
00:02:21.840 --> 00:02:25.479
<v Speaker 2>ever need shorter, kid friendly episodes with no adult language,

45
00:02:25.520 --> 00:02:29.000
<v Speaker 2>we have Smologies their episodes and their own feed wherever

46
00:02:29.039 --> 00:02:32.360
<v Speaker 2>you get podcasts. Also linked in the show notes. That's Smologies.

47
00:02:32.560 --> 00:02:35.719
<v Speaker 2>Also thank you to patrons for supporting ologies and sending

48
00:02:35.759 --> 00:02:37.840
<v Speaker 2>in questions ahead of time. You can join for as

49
00:02:37.879 --> 00:02:40.280
<v Speaker 2>little as a dollar a month at patreon dot com

50
00:02:40.280 --> 00:02:43.080
<v Speaker 2>slash Ologies. Thanks also to everyone who's ever left a

51
00:02:43.120 --> 00:02:45.520
<v Speaker 2>review for the show. They helped so much, and I

52
00:02:45.599 --> 00:02:50.080
<v Speaker 2>read all of them weirdly, including this recent one by Remily,

53
00:02:50.159 --> 00:02:53.560
<v Speaker 2>who wrote, this podcast will expand your horizons and help

54
00:02:53.599 --> 00:02:57.240
<v Speaker 2>you indulge in your hyperfrixations, both known and yet to

55
00:02:57.240 --> 00:03:00.759
<v Speaker 2>be discovered. Remily also says sorry has taken me over

56
00:03:00.840 --> 00:03:04.120
<v Speaker 2>five years to finally write a review. Emily anytime, it's

57
00:03:04.159 --> 00:03:06.199
<v Speaker 2>a good time. Thanks for that. Okay, let's get right

58
00:03:06.240 --> 00:03:11.000
<v Speaker 2>into artificial intelligence ethicology. It's the ethics of machine cognition.

59
00:03:11.199 --> 00:03:14.199
<v Speaker 2>Is a cognition, we'll talk about it. What does chat

60
00:03:14.280 --> 00:03:17.639
<v Speaker 2>CHPT stand for? Why is Siri a lady? Can you

61
00:03:17.680 --> 00:03:19.520
<v Speaker 2>ask a robot for a cheeseburger?

62
00:03:19.599 --> 00:03:20.039
<v Speaker 4>Yet?

63
00:03:20.080 --> 00:03:23.719
<v Speaker 2>What happens when you're rude to a chatbot? Also? How

64
00:03:23.719 --> 00:03:26.639
<v Speaker 2>do artists prevent getting ripped off? How much energy does

65
00:03:26.680 --> 00:03:30.199
<v Speaker 2>AI take up when we all lose our jobs? Booby traps,

66
00:03:30.360 --> 00:03:35.080
<v Speaker 2>doorbell marks commonly used fallacies. How the creators of AI

67
00:03:35.360 --> 00:03:38.159
<v Speaker 2>feel about AI? What is hype and what is horror?

68
00:03:38.319 --> 00:03:40.879
<v Speaker 2>What are the benefits of AI? What happens if you

69
00:03:41.039 --> 00:03:44.159
<v Speaker 2>assign a chat bot your homework? And whether or not

70
00:03:44.319 --> 00:03:47.639
<v Speaker 2>AI is the root of all evil or a pocket

71
00:03:47.680 --> 00:03:54.800
<v Speaker 2>pal with embodied Cognitive scientist professor, scholar and artificial intelligence ethicologist,

72
00:03:55.120 --> 00:04:34.360
<v Speaker 2>doctor Ababa Burhane.

73
00:04:15.080 --> 00:04:18.279
<v Speaker 4>Some questions that would take hours to one robin, and

74
00:04:18.319 --> 00:04:21.240
<v Speaker 4>they expect you to answer everything like one minute, two

75
00:04:21.240 --> 00:04:21.879
<v Speaker 4>minute marks.

76
00:04:22.839 --> 00:04:25.920
<v Speaker 2>They're rushing it again, So no worries.

77
00:04:26.079 --> 00:04:28.199
<v Speaker 4>I am above Abrahani, she.

78
00:04:28.240 --> 00:04:32.680
<v Speaker 2>Her great and AI. You have been an expert in

79
00:04:32.720 --> 00:04:36.319
<v Speaker 2>this field for a while, but I haven't known about

80
00:04:36.360 --> 00:04:39.279
<v Speaker 2>AI for that long. How long have you been studying it?

81
00:04:39.800 --> 00:04:43.519
<v Speaker 4>Technically speaking? I am a cognitive scientist, so I finished

82
00:04:43.600 --> 00:04:48.600
<v Speaker 4>my PhD about three years ago in cognitive science. So

83
00:04:49.160 --> 00:04:51.800
<v Speaker 4>halfway through my PhD, then around the end of my

84
00:04:51.879 --> 00:04:55.480
<v Speaker 4>second year, I left the cognitive science department and joined

85
00:04:55.519 --> 00:04:59.800
<v Speaker 4>a lab where people do a lot of tasting if

86
00:04:59.839 --> 00:05:02.959
<v Speaker 4>I eating and tasting of you know, chatboats in various

87
00:05:03.000 --> 00:05:04.199
<v Speaker 4>AI models.

88
00:05:04.560 --> 00:05:07.439
<v Speaker 2>And what is exactly cognitive science.

89
00:05:08.199 --> 00:05:13.759
<v Speaker 4>Cognitive science is very broad, so traditional cognitive science tends

90
00:05:13.839 --> 00:05:18.560
<v Speaker 4>to be about you know, understanding cognition, understanding you know,

91
00:05:18.680 --> 00:05:23.360
<v Speaker 4>human behavior, understanding human interaction, and so on. In cognitive

92
00:05:23.399 --> 00:05:28.800
<v Speaker 4>science often is not taught at graduate level. It's either

93
00:05:28.800 --> 00:05:32.079
<v Speaker 4>at a master's livery or a PhD level because cognitive

94
00:05:32.079 --> 00:05:35.079
<v Speaker 4>science is really interdisciplinary.

95
00:05:34.319 --> 00:05:37.560
<v Speaker 2>And doctor Brownie says that cognitive science is actually a

96
00:05:37.600 --> 00:05:41.800
<v Speaker 2>mishmash of disciplines or what sometimes called the cognitive hexagon,

97
00:05:42.120 --> 00:05:50.519
<v Speaker 2>besides representing philosophy, psychology, linguistics, neuroscience, artificial intelligence, and anthropology

98
00:05:50.600 --> 00:05:54.399
<v Speaker 2>according to some institutions, and anthropology can also mean social sciences.

99
00:05:54.800 --> 00:05:58.480
<v Speaker 2>Different institutions will phrase it their own way, but broadly speaking,

100
00:05:58.879 --> 00:06:02.000
<v Speaker 2>cognitive science is a lot of stuff.

101
00:06:02.319 --> 00:06:06.079
<v Speaker 4>So the idea is you will take you know, important

102
00:06:06.240 --> 00:06:11.800
<v Speaker 4>or helpful aspects from these various disciplines, and cognitive science

103
00:06:11.839 --> 00:06:17.399
<v Speaker 4>then allows you to synthesize, to combine these various theories

104
00:06:17.560 --> 00:06:21.600
<v Speaker 4>even computational models to understand human cognition. That's the idea.

105
00:06:21.680 --> 00:06:24.959
<v Speaker 4>So from philosophy, for example, you will learn how to

106
00:06:25.120 --> 00:06:29.399
<v Speaker 4>question assumptions. You will go down into the various questions

107
00:06:29.439 --> 00:06:34.120
<v Speaker 4>around what's cognition, what's intelligence? You know, what's human emotions.

108
00:06:34.120 --> 00:06:39.360
<v Speaker 4>So philosophy really lends you the analytic tools. Same from neuroscience.

109
00:06:39.480 --> 00:06:42.360
<v Speaker 4>The idea is you get to learn how the brain

110
00:06:42.439 --> 00:06:45.759
<v Speaker 4>works and use it in a way to synthesize from

111
00:06:45.959 --> 00:06:49.040
<v Speaker 4>all these different disciplines. So that's traditional cognitive science.

112
00:06:49.279 --> 00:06:52.480
<v Speaker 2>You can call cognitive science cogsie if you want to,

113
00:06:52.879 --> 00:06:56.439
<v Speaker 2>And she says that she's in a really niche specialty

114
00:06:56.519 --> 00:07:00.800
<v Speaker 2>within COGSI, and it's called embodied cognitive science, which isn't

115
00:07:00.879 --> 00:07:03.079
<v Speaker 2>just about understanding the human brain.

116
00:07:03.639 --> 00:07:07.000
<v Speaker 4>Whereas embodied cognitive science is moving away from this idea

117
00:07:07.120 --> 00:07:11.879
<v Speaker 4>of treating cognition in isolation. Your cognition doesn't end at

118
00:07:11.920 --> 00:07:14.920
<v Speaker 4>your brain, and your sense of self doesn't end at

119
00:07:14.920 --> 00:07:19.480
<v Speaker 4>the skin, but rather it's extended into the tools you use.

120
00:07:19.720 --> 00:07:22.839
<v Speaker 4>Anything you do, you do it as an embodied self,

121
00:07:22.879 --> 00:07:27.040
<v Speaker 4>as an embodied person. So your body, your social circle,

122
00:07:27.680 --> 00:07:33.319
<v Speaker 4>your history, your culture, even your gender and sexuality, all

123
00:07:33.399 --> 00:07:39.160
<v Speaker 4>these are important factors that play into you know, your understanding,

124
00:07:39.160 --> 00:07:41.680
<v Speaker 4>your cognition, your intelligence, your emotion and so on.

125
00:07:41.759 --> 00:07:44.360
<v Speaker 2>And when you're studying cognitive science, how do you not

126
00:07:45.680 --> 00:07:48.399
<v Speaker 2>use your brain to think about your brain all the time?

127
00:07:49.759 --> 00:07:53.000
<v Speaker 2>How do you get out of your brain you doing?

128
00:07:53.240 --> 00:07:59.319
<v Speaker 4>Yeah, yeah, yeah, I mean you have to, you have

129
00:07:59.399 --> 00:08:03.680
<v Speaker 4>to use your brain. So you're familiar, you know with

130
00:08:04.160 --> 00:08:08.160
<v Speaker 4>the cards famous quote collegito or gossam, so I think

131
00:08:08.199 --> 00:08:11.800
<v Speaker 4>therefore I am. The idea is you can't know who

132
00:08:11.879 --> 00:08:16.240
<v Speaker 4>you are. You can know you are thinking big, you

133
00:08:16.319 --> 00:08:20.639
<v Speaker 4>can know that you are someone. So that is like

134
00:08:20.800 --> 00:08:25.040
<v Speaker 4>using your brain to understand your brain, so to speak. Whereas,

135
00:08:25.120 --> 00:08:29.920
<v Speaker 4>again the emphasis with embodied cognitive science is that you know,

136
00:08:30.040 --> 00:08:36.480
<v Speaker 4>all that is really very individualistic. Even to confirm our existence,

137
00:08:37.120 --> 00:08:41.960
<v Speaker 4>it's through conversations with others. So that's why embodied cognitive

138
00:08:41.960 --> 00:08:46.279
<v Speaker 4>science emphasizes others and communities and your body as really

139
00:08:46.360 --> 00:08:49.080
<v Speaker 4>important factors in understanding cognition.

140
00:08:49.399 --> 00:08:53.120
<v Speaker 2>So embodied cognitive science kind of goes downstairs a bit,

141
00:08:53.200 --> 00:08:56.279
<v Speaker 2>and it considers how a person's body experiences the world

142
00:08:56.279 --> 00:09:00.679
<v Speaker 2>and how the environment shapes thinking and perception. And cognitive

143
00:09:00.759 --> 00:09:03.759
<v Speaker 2>scientists have this thought experiment that tickled me. And it

144
00:09:03.919 --> 00:09:08.679
<v Speaker 2>envisions a little person in your brain interpreting inputs. But

145
00:09:08.799 --> 00:09:13.080
<v Speaker 2>then who is in the brain inside that little person's brain?

146
00:09:13.159 --> 00:09:16.720
<v Speaker 2>Like does it have its own brain inside the little

147
00:09:16.720 --> 00:09:19.240
<v Speaker 2>person's It's like Russia nesting dolls, and there are just

148
00:09:19.399 --> 00:09:23.679
<v Speaker 2>infinity little humans inside humans brains to interpret what the

149
00:09:23.720 --> 00:09:26.759
<v Speaker 2>brain inside the brain is braining. And this is called

150
00:09:27.279 --> 00:09:33.840
<v Speaker 2>endearingly the homunculous argument. And yes, embodied cognitive science is like,

151
00:09:34.159 --> 00:09:37.360
<v Speaker 2>it's more than a tiny person in your skull. And

152
00:09:37.440 --> 00:09:40.399
<v Speaker 2>how can we truly understand artificial intelligence if we don't

153
00:09:40.399 --> 00:09:46.120
<v Speaker 2>first grasp intelligence intelligently. And when we're thinking about who

154
00:09:46.159 --> 00:09:49.679
<v Speaker 2>we interface with with AI? Way back in the day,

155
00:09:49.879 --> 00:09:52.679
<v Speaker 2>there used to be a search engine called ask Jeeves

156
00:09:52.840 --> 00:09:54.759
<v Speaker 2>and it was like a butler who would find you

157
00:09:55.120 --> 00:10:00.519
<v Speaker 2>the answers to things. And now we have Siri and Alexa.

158
00:10:00.720 --> 00:10:04.919
<v Speaker 2>Why do you think with AI they've gendered and they've

159
00:10:05.080 --> 00:10:11.679
<v Speaker 2>personified these voices that are actually a huge network of

160
00:10:12.200 --> 00:10:16.799
<v Speaker 2>artificial intelligence? Is that to help our brains understand It's

161
00:10:16.840 --> 00:10:17.919
<v Speaker 2>a little bit of like.

162
00:10:18.039 --> 00:10:21.799
<v Speaker 4>A marketing strategy, but it's also a little bit of

163
00:10:22.000 --> 00:10:26.279
<v Speaker 4>appealing to human nature. We tend to kind of gender

164
00:10:26.919 --> 00:10:31.480
<v Speaker 4>and personify objects. So if you are interacting with chat

165
00:10:31.559 --> 00:10:35.879
<v Speaker 4>ChiPT for example, we tend to just naturally, you know,

166
00:10:36.360 --> 00:10:39.440
<v Speaker 4>treat it as another person, an as a being, an

167
00:10:39.480 --> 00:10:42.639
<v Speaker 4>ather entity. On the one hand, as you said, these

168
00:10:42.720 --> 00:10:45.080
<v Speaker 4>chat boats, you know, there is no intention, there is

169
00:10:45.120 --> 00:10:48.919
<v Speaker 4>no understanding, there is no soul in these machines. There

170
00:10:48.919 --> 00:10:53.879
<v Speaker 4>are just pure machines. But also the developers and vendors

171
00:10:53.919 --> 00:10:58.759
<v Speaker 4>of these systems, they tend to market them as kind

172
00:10:58.759 --> 00:11:03.879
<v Speaker 4>of personified entities because it's much more appealing to think

173
00:11:03.919 --> 00:11:06.480
<v Speaker 4>that you are interacting withan asar synthy and think.

174
00:11:06.720 --> 00:11:09.639
<v Speaker 2>I always wonder if they made some of them female voices,

175
00:11:09.679 --> 00:11:15.840
<v Speaker 2>because we're more accepting, we're less threatened by females. We're

176
00:11:16.720 --> 00:11:19.919
<v Speaker 2>socialized to have a mommy figure come and help us

177
00:11:19.960 --> 00:11:24.080
<v Speaker 2>with something. It's not as threatening that it'll turn on us.

178
00:11:24.120 --> 00:11:26.919
<v Speaker 2>And also it's like, oh, Lydia, don't get it for you,

179
00:11:26.919 --> 00:11:27.200
<v Speaker 2>you know.

180
00:11:28.000 --> 00:11:31.320
<v Speaker 4>Yeah, yeah, I mean, we naturally tend to think women

181
00:11:31.480 --> 00:11:35.679
<v Speaker 4>are much more like nurturing and they have the role

182
00:11:35.799 --> 00:11:40.000
<v Speaker 4>of helping you. So it is kind of related to

183
00:11:40.879 --> 00:11:45.440
<v Speaker 4>the social norms that really dictate society. So it's really

184
00:11:45.480 --> 00:11:49.320
<v Speaker 4>also to some extent, leaning on that stereotype that if

185
00:11:49.320 --> 00:11:52.679
<v Speaker 4>it's a woman, you know, it's approachable, it's there to

186
00:11:52.759 --> 00:11:53.159
<v Speaker 4>help you.

187
00:11:53.799 --> 00:11:56.399
<v Speaker 2>And yes, I am far from the first person to

188
00:11:56.519 --> 00:11:59.679
<v Speaker 2>notice that digital assistance tend to be ladies. And some

189
00:11:59.799 --> 00:12:04.039
<v Speaker 2>histrians and media scholars think that it starts early by

190
00:12:04.080 --> 00:12:07.679
<v Speaker 2>hearing a female voice while you're cooking in the womb,

191
00:12:07.919 --> 00:12:10.679
<v Speaker 2>and that's why people love a lady voice, or that

192
00:12:10.759 --> 00:12:14.039
<v Speaker 2>the first telephone operator in the late eighteen hundreds happened

193
00:12:14.039 --> 00:12:16.360
<v Speaker 2>to be a woman with a great voice and then

194
00:12:16.399 --> 00:12:21.000
<v Speaker 2>it just stuck. But as these AI avatars start to

195
00:12:21.039 --> 00:12:26.240
<v Speaker 2>have faces and personalities and pastimes and instagrams, why do

196
00:12:26.320 --> 00:12:30.919
<v Speaker 2>we see mostly younger, hotter avatars. Well, there was a

197
00:12:30.919 --> 00:12:34.039
<v Speaker 2>twenty twenty four article published by Reuter's Institute and it

198
00:12:34.120 --> 00:12:37.200
<v Speaker 2>reached out to this multimedia professor April Newton, who noted

199
00:12:37.240 --> 00:12:41.080
<v Speaker 2>that a gentle, well modulated woman's voice is usually the

200
00:12:41.159 --> 00:12:44.720
<v Speaker 2>default for AI assistants and avatars because, quote, we order

201
00:12:44.759 --> 00:12:47.080
<v Speaker 2>those devices to do things for us, and we are

202
00:12:47.320 --> 00:12:51.159
<v Speaker 2>very comfortable ordering women to do stuff for us. And also,

203
00:12:51.279 --> 00:12:52.759
<v Speaker 2>just a side note for me, did you know that

204
00:12:52.799 --> 00:12:56.480
<v Speaker 2>the word robot it comes from a Slavic root for

205
00:12:56.480 --> 00:13:03.639
<v Speaker 2>forced labor or slaves. Creepy. Historically humans have loved capable service,

206
00:13:03.720 --> 00:13:06.080
<v Speaker 2>but not too capable or else you're just begging for

207
00:13:06.120 --> 00:13:10.120
<v Speaker 2>an uprising. So the future, it's also the past. Am

208
00:13:10.159 --> 00:13:12.960
<v Speaker 2>I rooting for the robots? Now? I don't know. When

209
00:13:13.039 --> 00:13:17.919
<v Speaker 2>it comes to a sery or virtual assistance? How is

210
00:13:17.960 --> 00:13:21.240
<v Speaker 2>that different than the AI that's been ramped up in

211
00:13:21.279 --> 00:13:25.120
<v Speaker 2>the last couple of years. Is Siri and Alexa? Are

212
00:13:25.159 --> 00:13:28.759
<v Speaker 2>those ais? Or are those just search engines? Is Google

213
00:13:28.840 --> 00:13:31.679
<v Speaker 2>an AI? We call something's AI and then other things

214
00:13:31.720 --> 00:13:33.480
<v Speaker 2>just computing? What's the difference?

215
00:13:34.279 --> 00:13:38.799
<v Speaker 4>Yeah, so what's really difference between say traditional AI whizard

216
00:13:39.240 --> 00:13:42.360
<v Speaker 4>it's you know, implemented as a chat boat or as

217
00:13:42.519 --> 00:13:46.120
<v Speaker 4>a predictive system, you know, generative AI that has really

218
00:13:46.200 --> 00:13:49.759
<v Speaker 4>exploded over the past two years, think of chat gipt,

219
00:13:50.200 --> 00:13:54.720
<v Speaker 4>you know, Gemini in cloud and so on. The fundamental

220
00:13:54.799 --> 00:13:58.480
<v Speaker 4>difference is that do I go into the technical details

221
00:13:58.600 --> 00:13:59.960
<v Speaker 4>or remain clear of the.

222
00:14:00.960 --> 00:14:02.360
<v Speaker 2>You can give us some technical details.

223
00:14:02.480 --> 00:14:05.720
<v Speaker 4>I'll try to keep it light. There is no way

224
00:14:05.759 --> 00:14:10.840
<v Speaker 4>of explaining the difference without getting into reinforcement learning. A

225
00:14:10.919 --> 00:14:15.120
<v Speaker 4>typical classification system is an algorithm that you give it

226
00:14:15.360 --> 00:14:19.720
<v Speaker 4>massive amounts of data. Think of like a face identification system,

227
00:14:20.120 --> 00:14:22.679
<v Speaker 4>and these days data sets have to be really big.

228
00:14:23.000 --> 00:14:26.360
<v Speaker 4>You might even have you know, trillions of tokens of

229
00:14:26.480 --> 00:14:30.360
<v Speaker 4>images of faces, and you train your algorism like this

230
00:14:30.519 --> 00:14:32.279
<v Speaker 4>is a face, this is a face, this is not

231
00:14:32.320 --> 00:14:34.080
<v Speaker 4>a face, this is a face, this is not a face.

232
00:14:34.240 --> 00:14:37.600
<v Speaker 4>Called machine learning. If you have succeeded in your training,

233
00:14:37.679 --> 00:14:40.440
<v Speaker 4>then your AI should be able to tell you it's

234
00:14:40.480 --> 00:14:42.200
<v Speaker 4>a face or it's not a face. So this is

235
00:14:42.200 --> 00:14:46.559
<v Speaker 4>like a typical you know, classification system. What we consider

236
00:14:46.679 --> 00:14:51.200
<v Speaker 4>AI is a very broad term that encompasses so many

237
00:14:51.200 --> 00:14:52.480
<v Speaker 4>different subcategories.

238
00:14:52.879 --> 00:14:55.519
<v Speaker 2>Just side out. How alone am I in not knowing

239
00:14:55.720 --> 00:15:01.240
<v Speaker 2>that Apple Virtual assistance theory stands for something each interpretation

240
00:15:01.799 --> 00:15:05.720
<v Speaker 2>and recognition interface. Did you know SIRI was an acronym?

241
00:15:06.159 --> 00:15:10.399
<v Speaker 2>I didn't. Also, Apple is reportedly freaking out behind the

242
00:15:10.399 --> 00:15:13.399
<v Speaker 2>scenes about Chat giput for the last three years being

243
00:15:13.639 --> 00:15:18.480
<v Speaker 2>generative and having Chat bought capabilities and longer conversations and Siri.

244
00:15:18.679 --> 00:15:21.399
<v Speaker 2>So apparently a lot of their efforts in self driving

245
00:15:21.440 --> 00:15:25.000
<v Speaker 2>cars got shuttled over to their AI division, and a

246
00:15:25.000 --> 00:15:26.879
<v Speaker 2>lot of people at Apple are like, I can't even

247
00:15:26.960 --> 00:15:28.320
<v Speaker 2>speak publicly about this.

248
00:15:28.639 --> 00:15:32.639
<v Speaker 4>The other big subcategory under the broad m brilav AA

249
00:15:32.720 --> 00:15:36.200
<v Speaker 4>is NLP or natural language processing. So this is the

250
00:15:36.279 --> 00:15:40.080
<v Speaker 4>area that deals with human language. So you have audio

251
00:15:40.200 --> 00:15:44.039
<v Speaker 4>data that you feed into the AI system, and the

252
00:15:44.159 --> 00:15:46.879
<v Speaker 4>idea there is that you know, you are building an

253
00:15:46.919 --> 00:15:52.240
<v Speaker 4>AI system that even make predictions about human language, and

254
00:15:52.600 --> 00:15:56.759
<v Speaker 4>ptools would learn, for example, predictive texts, so what the

255
00:15:56.799 --> 00:16:01.639
<v Speaker 4>algorithm is doing is kind of predict what the next

256
00:16:01.679 --> 00:16:04.639
<v Speaker 4>world is likely to be. So it's just predicting the

257
00:16:04.720 --> 00:16:08.159
<v Speaker 4>nix stoken the next stoken. So that's kind of traditional

258
00:16:08.360 --> 00:16:10.519
<v Speaker 4>classification or predictive systems.

259
00:16:10.720 --> 00:16:14.679
<v Speaker 2>And that was machine learning or NLP, which is natural

260
00:16:14.759 --> 00:16:18.960
<v Speaker 2>language processing, and those deal with visual data turned into

261
00:16:19.000 --> 00:16:23.480
<v Speaker 2>tokens and they predict language like when your phone knows

262
00:16:23.519 --> 00:16:25.879
<v Speaker 2>you better than you know yourself, and it's heartwarming and

263
00:16:25.919 --> 00:16:29.559
<v Speaker 2>it's scary. It's like love now chat GPT, for example,

264
00:16:29.559 --> 00:16:32.440
<v Speaker 2>I didn't know this, but the GPT stands for generative

265
00:16:32.679 --> 00:16:36.759
<v Speaker 2>pre Trained Transformer, and a transformer is this type of

266
00:16:36.879 --> 00:16:41.639
<v Speaker 2>deep learning system and it converts info into tokens and

267
00:16:41.960 --> 00:16:46.200
<v Speaker 2>can handle more complex processing from language to vision to

268
00:16:46.320 --> 00:16:49.759
<v Speaker 2>games and audio generation. So it's definitely a step up

269
00:16:49.759 --> 00:16:51.559
<v Speaker 2>from just simple prediction.

270
00:16:52.000 --> 00:16:55.360
<v Speaker 4>Whereas over the past year with generative AI, these are

271
00:16:56.399 --> 00:17:01.840
<v Speaker 4>AI systems that do more than classification, more than prediction,

272
00:17:02.360 --> 00:17:06.359
<v Speaker 4>more than aggregation, that are called generative A systems. They produce,

273
00:17:06.480 --> 00:17:10.279
<v Speaker 4>you know, something new. So image generators, for example, you

274
00:17:10.359 --> 00:17:13.480
<v Speaker 4>can put a TICS description as a prompt in the

275
00:17:13.559 --> 00:17:16.960
<v Speaker 4>AI system will produce an image that resembles based on

276
00:17:17.000 --> 00:17:21.920
<v Speaker 4>your description. The same with language systems, so chuch BT

277
00:17:22.000 --> 00:17:25.240
<v Speaker 4>for example, you put in any prompt and it's able

278
00:17:25.279 --> 00:17:30.160
<v Speaker 4>to produce new answers. So this is where this is

279
00:17:30.160 --> 00:17:31.799
<v Speaker 4>what's new with generative systems.

280
00:17:31.880 --> 00:17:33.720
<v Speaker 2>And the systems also need to learn which of the

281
00:17:33.759 --> 00:17:37.240
<v Speaker 2>words in a language model are the important ones, which

282
00:17:37.319 --> 00:17:40.480
<v Speaker 2>is part of the self attention mechanism, and then they

283
00:17:40.559 --> 00:17:43.839
<v Speaker 2>generate based on statistics like what is the most probable

284
00:17:43.880 --> 00:17:47.039
<v Speaker 2>way based on the data setts learned from from completing

285
00:17:47.119 --> 00:17:48.839
<v Speaker 2>a certain sentence or prompt.

286
00:17:49.160 --> 00:17:54.240
<v Speaker 4>So the training data really has a really significant impact

287
00:17:54.880 --> 00:17:59.440
<v Speaker 4>in what outputs, whether it's image or text. What outputs

288
00:17:59.480 --> 00:18:00.880
<v Speaker 4>these those can't generate.

289
00:18:01.400 --> 00:18:04.599
<v Speaker 2>Let's get to the juicy parts. So with any tokens,

290
00:18:05.000 --> 00:18:07.599
<v Speaker 2>would that data be in tokens? Like you were saying,

291
00:18:08.160 --> 00:18:12.440
<v Speaker 2>facial recognition might have a trillion tokens? Does AI kind

292
00:18:12.480 --> 00:18:17.759
<v Speaker 2>of scrub what we know of impressionist art and science

293
00:18:17.799 --> 00:18:20.319
<v Speaker 2>fiction and anime does it scrub it and grab a

294
00:18:20.359 --> 00:18:22.920
<v Speaker 2>bunch of tokens, so then it can have those as

295
00:18:22.960 --> 00:18:23.920
<v Speaker 2>reference points.

296
00:18:24.200 --> 00:18:27.720
<v Speaker 4>So the training data sets is one of the most

297
00:18:27.720 --> 00:18:33.559
<v Speaker 4>contested issues in AI because data is constantly harvested. So

298
00:18:34.240 --> 00:18:39.160
<v Speaker 4>like your search history feeds into some kind of AI

299
00:18:39.240 --> 00:18:40.960
<v Speaker 4>one way or anazon I don't know if you are

300
00:18:41.119 --> 00:18:43.559
<v Speaker 4>like me, if you signed up for some kind of

301
00:18:43.599 --> 00:18:46.839
<v Speaker 4>bonus point, you go to a grocery shop and you

302
00:18:46.960 --> 00:18:50.240
<v Speaker 4>tap that, so that is kind of like in the background.

303
00:18:50.240 --> 00:18:52.480
<v Speaker 4>That's kind of conllicting your behavioral data.

304
00:18:52.640 --> 00:18:57.079
<v Speaker 2>So that data may breeze through infrastructure like Google and

305
00:18:57.119 --> 00:18:59.640
<v Speaker 2>then just be on its merry way to a third

306
00:18:59.680 --> 00:19:03.839
<v Speaker 2>party aggregator or a broker or an AI company itself,

307
00:19:03.880 --> 00:19:06.279
<v Speaker 2>and they're just yam yam yam gobble of the data.

308
00:19:06.359 --> 00:19:10.160
<v Speaker 4>So this kind of practice puts a lot of data

309
00:19:10.359 --> 00:19:15.160
<v Speaker 4>gathering for the purpose of training AI systems either illegal

310
00:19:15.319 --> 00:19:19.079
<v Speaker 4>or borderline illegal. Yes, because you know, first of all,

311
00:19:19.119 --> 00:19:22.039
<v Speaker 4>there is no consent. People are not even aware that

312
00:19:22.200 --> 00:19:24.400
<v Speaker 4>you know their data is being used to train AI.

313
00:19:24.920 --> 00:19:27.880
<v Speaker 2>So that's just number one on a general level. Let's

314
00:19:27.880 --> 00:19:28.400
<v Speaker 2>get to art.

315
00:19:28.720 --> 00:19:32.400
<v Speaker 4>But you will also have noticed over the past year

316
00:19:32.599 --> 00:19:39.960
<v Speaker 4>or two the creative community, writers and artists are realizing

317
00:19:40.119 --> 00:19:45.240
<v Speaker 4>that their work, their novels, their writing, their their arts

318
00:19:45.720 --> 00:19:49.960
<v Speaker 4>are being used to train AI systems, again without any compensation,

319
00:19:50.759 --> 00:19:53.720
<v Speaker 4>or with very little compensation after the fact if they

320
00:19:53.759 --> 00:19:56.559
<v Speaker 4>go and contest you know, the use of their data.

321
00:19:56.720 --> 00:20:02.160
<v Speaker 4>And because large companies that are kind of developing AI,

322
00:20:02.640 --> 00:20:08.160
<v Speaker 4>you know, think of Google, Deepmined, Meta, open AI, anthropic,

323
00:20:08.960 --> 00:20:13.759
<v Speaker 4>they are businesses. They operate under the business model. They

324
00:20:13.920 --> 00:20:18.240
<v Speaker 4>are you know, commercial entities. Their objective is to maximize profits.

325
00:20:18.759 --> 00:20:22.319
<v Speaker 4>My work is auditing, for example, large scale training data sets.

326
00:20:22.680 --> 00:20:26.559
<v Speaker 4>People don't have access to what kind of data set

327
00:20:26.680 --> 00:20:31.400
<v Speaker 4>this company is hold, So we don't have any mechanism

328
00:20:31.920 --> 00:20:34.960
<v Speaker 4>to kind of take a look to scrutinize what's what

329
00:20:35.000 --> 00:20:37.160
<v Speaker 4>you know, what's in the data is it, where does

330
00:20:37.200 --> 00:20:39.000
<v Speaker 4>it come from? How large is it? These are the

331
00:20:39.079 --> 00:20:43.759
<v Speaker 4>kind of questions that simply there is just no mechanism

332
00:20:44.000 --> 00:20:48.240
<v Speaker 4>for tach corporations to be transparent, to open up, for

333
00:20:48.319 --> 00:20:50.880
<v Speaker 4>us to have an objective understanding.

334
00:20:51.160 --> 00:20:53.279
<v Speaker 2>So big companies are like, no, you can't see it.

335
00:20:53.519 --> 00:20:57.799
<v Speaker 2>But auditors like doctor Bhaney and colleagues can observe open

336
00:20:57.880 --> 00:21:02.440
<v Speaker 2>source publicly available similar data sets as proxies or substitutes

337
00:21:02.480 --> 00:21:04.759
<v Speaker 2>to kind of figure out what might be going on

338
00:21:05.400 --> 00:21:09.839
<v Speaker 2>organizationally behind the locked Willi Wanka factory gates, the other

339
00:21:09.920 --> 00:21:10.960
<v Speaker 2>big tech companies.

340
00:21:11.200 --> 00:21:14.920
<v Speaker 4>So we do get an idea, but through proxy data sets.

341
00:21:15.079 --> 00:21:17.839
<v Speaker 4>So to answer your question, I mean, if you are

342
00:21:17.880 --> 00:21:20.519
<v Speaker 4>an artist, a writer, if you are produced novels and

343
00:21:20.559 --> 00:21:23.799
<v Speaker 4>so on, it's very very likely that your work is

344
00:21:23.839 --> 00:21:26.880
<v Speaker 4>being used to train AI. But there's very little legal

345
00:21:27.000 --> 00:21:29.279
<v Speaker 4>mechanism to actually have a clear idea.

346
00:21:30.119 --> 00:21:33.720
<v Speaker 2>Is it like they're stealing a bunch of ingredients and

347
00:21:33.759 --> 00:21:36.160
<v Speaker 2>then making something but they're like, you can't see our recipe,

348
00:21:36.200 --> 00:21:38.359
<v Speaker 2>and you're like, you stole the ingredients and they're like, yeah,

349
00:21:38.400 --> 00:21:43.200
<v Speaker 2>their resume, we made something with it. That is that

350
00:21:43.319 --> 00:21:44.160
<v Speaker 2>sort of great.

351
00:21:44.240 --> 00:21:49.480
<v Speaker 4>Yeah, So the fact that they are protected by proprietary

352
00:21:49.599 --> 00:21:55.160
<v Speaker 4>rights as a commercial entity means that there is virtually

353
00:21:55.240 --> 00:21:58.960
<v Speaker 4>no mechanism to, you know, to force these companies to

354
00:21:58.960 --> 00:22:01.680
<v Speaker 4>open up their data sets. Of course, you can encourage them,

355
00:22:01.920 --> 00:22:04.880
<v Speaker 4>you can kind of you know, appeal to their good

356
00:22:04.920 --> 00:22:06.000
<v Speaker 4>sites and so on.

357
00:22:06.359 --> 00:22:06.440
<v Speaker 2>But.

358
00:22:08.559 --> 00:22:11.720
<v Speaker 4>Yeah, there is no there is no legal mechanism. There

359
00:22:11.759 --> 00:22:16.000
<v Speaker 4>is no law or regulation that says you have to

360
00:22:16.039 --> 00:22:19.359
<v Speaker 4>open access or you have to you know, share your.

361
00:22:19.319 --> 00:22:21.079
<v Speaker 2>Data set, and why not you ask?

362
00:22:21.559 --> 00:22:24.680
<v Speaker 4>Of course, if that is to happen, then you know,

363
00:22:24.720 --> 00:22:28.039
<v Speaker 4>all these AI companies would go out of business. As

364
00:22:28.079 --> 00:22:31.279
<v Speaker 4>it is, a lot of them are under a lot

365
00:22:31.359 --> 00:22:37.240
<v Speaker 4>of lawsuits. Oh yeah, you know, from from Meta to OPENAEI.

366
00:22:37.920 --> 00:22:41.160
<v Speaker 4>Openae itself is under a lot of lawsuits, including from

367
00:22:41.160 --> 00:22:44.440
<v Speaker 4>the New York Times. So the minute they open up

368
00:22:44.440 --> 00:22:48.440
<v Speaker 4>their data, it becomes clear that a lot of people's suspicions,

369
00:22:48.519 --> 00:22:52.440
<v Speaker 4>especially you know, the creative and artists communities. The minute

370
00:22:52.519 --> 00:22:54.920
<v Speaker 4>the data sets are open, why they are going to court?

371
00:22:56.000 --> 00:22:59.079
<v Speaker 2>Is there any recourse that artists writers have like, is

372
00:22:59.119 --> 00:23:02.960
<v Speaker 2>there anything they can do other than trying to open

373
00:23:03.000 --> 00:23:04.480
<v Speaker 2>a really big expensive lawsuit.

374
00:23:04.920 --> 00:23:10.799
<v Speaker 4>Yeah, so writers in creatives are organizing for class action suits.

375
00:23:10.920 --> 00:23:13.319
<v Speaker 4>I know there are a bunch of class action lawsuits.

376
00:23:13.559 --> 00:23:15.279
<v Speaker 4>Was in the US and in the UK.

377
00:23:15.920 --> 00:23:18.839
<v Speaker 2>Now last August there was this landmark case and it

378
00:23:18.880 --> 00:23:22.559
<v Speaker 2>decided in favor of artists. It found that generative AI

379
00:23:22.640 --> 00:23:26.000
<v Speaker 2>systems like mid Journey and Devian Arts, dream Up and

380
00:23:26.119 --> 00:23:30.759
<v Speaker 2>Stability AI's Stable Diffusion, that those were violating copyright law

381
00:23:30.839 --> 00:23:34.880
<v Speaker 2>by using data sets A billions of artistic examples scraped

382
00:23:34.920 --> 00:23:38.720
<v Speaker 2>from the web, and getty Images sued Stable Diffusion for

383
00:23:38.799 --> 00:23:42.880
<v Speaker 2>copyright infringements, and the evidence was like almost funny if

384
00:23:42.920 --> 00:23:46.160
<v Speaker 2>it were such a bummer. But apparently the Generative AI

385
00:23:46.720 --> 00:23:50.240
<v Speaker 2>so relied on getty images that it started adding a

386
00:23:50.279 --> 00:23:53.599
<v Speaker 2>blurry gray box to some of its AI output, which

387
00:23:53.640 --> 00:23:58.559
<v Speaker 2>was learned from the iconic getty images watermarks. Embarrassing now.

388
00:23:58.599 --> 00:24:01.079
<v Speaker 2>In March of this year, a judge ordered that the

389
00:24:01.279 --> 00:24:04.519
<v Speaker 2>lawsuit between the New York Times and open ai can proceed,

390
00:24:04.880 --> 00:24:08.799
<v Speaker 2>despite open ai begging it not to, and the newspaper

391
00:24:08.960 --> 00:24:11.599
<v Speaker 2>New York Times, along with a few other journalism outlets,

392
00:24:11.759 --> 00:24:15.480
<v Speaker 2>alleges that open ai scraped a lot of their work

393
00:24:15.599 --> 00:24:19.319
<v Speaker 2>to train chat GPT. So there are lawsuits, but there

394
00:24:19.319 --> 00:24:21.880
<v Speaker 2>are also just huge glaring holes.

395
00:24:22.160 --> 00:24:26.319
<v Speaker 4>But also in the UK, for example, they are considering

396
00:24:27.000 --> 00:24:33.559
<v Speaker 4>a regulation that leaves massive loophole for intellectual copyright issues

397
00:24:33.640 --> 00:24:39.079
<v Speaker 4>that leaves artists and writers with absolutely no protection at all.

398
00:24:39.440 --> 00:24:42.759
<v Speaker 4>So people are really organizing on the ground and they

399
00:24:42.759 --> 00:24:46.039
<v Speaker 4>have massive amounts of petitions signed and so on. But

400
00:24:46.079 --> 00:24:50.400
<v Speaker 4>also there are technical, i don't want to say solutions,

401
00:24:50.440 --> 00:24:55.880
<v Speaker 4>technical kind of remedies. So you can use data poisoning tools.

402
00:24:55.920 --> 00:24:59.519
<v Speaker 4>For example, you have oh yeah, so you have like

403
00:25:00.160 --> 00:25:03.559
<v Speaker 4>booby trap like one of them is called to night shed.

404
00:25:04.200 --> 00:25:09.759
<v Speaker 4>So for images, for example, you would insert various adversarial

405
00:25:09.880 --> 00:25:15.680
<v Speaker 4>attacks that we make the data unusable for machines because

406
00:25:15.720 --> 00:25:18.599
<v Speaker 4>it's maybe a tiny pixel that's been altered so that's

407
00:25:18.640 --> 00:25:21.160
<v Speaker 4>not visible to the human eye, but it kind of

408
00:25:21.240 --> 00:25:24.759
<v Speaker 4>misses with the automatage system or how machines kind of

409
00:25:24.960 --> 00:25:28.240
<v Speaker 4>use this data as it. So there are various tools

410
00:25:28.359 --> 00:25:28.680
<v Speaker 4>like that.

411
00:25:29.119 --> 00:25:31.559
<v Speaker 2>Another type of AI booby trap is called a tarpit,

412
00:25:32.079 --> 00:25:36.640
<v Speaker 2>and it sends AI crawlers in this infinite loop where

413
00:25:36.680 --> 00:25:39.640
<v Speaker 2>they just get stuck and they can't escape. And I

414
00:25:39.759 --> 00:25:41.720
<v Speaker 2>just love to think of like the AI system on

415
00:25:41.759 --> 00:25:45.119
<v Speaker 2>the toilet, just scrolling, just not being able to exit

416
00:25:45.279 --> 00:25:46.680
<v Speaker 2>and get back to work.

417
00:25:47.079 --> 00:25:50.160
<v Speaker 4>Even if you find tachn care solutions, now, the big

418
00:25:50.200 --> 00:25:53.960
<v Speaker 4>companies are likely to come back with another solution that

419
00:25:54.319 --> 00:25:59.000
<v Speaker 4>makes your own solution defunct. So yeah, you really have

420
00:25:59.119 --> 00:26:02.240
<v Speaker 4>to be adapting with that constantinly, so, I think a

421
00:26:02.319 --> 00:26:05.319
<v Speaker 4>vibrent solution has to come from the regulation from the

422
00:26:05.440 --> 00:26:06.920
<v Speaker 4>legal space.

423
00:26:08.000 --> 00:26:09.960
<v Speaker 2>And how do you feel about is that?

424
00:26:10.039 --> 00:26:12.480
<v Speaker 4>Sam s Oh my gosh, thank you.

425
00:26:12.839 --> 00:26:17.799
<v Speaker 2>I almost said Sam Edelman, which is a shoemaker, how

426
00:26:17.880 --> 00:26:21.319
<v Speaker 2>do you feel about kind of some recent changes, Like

427
00:26:21.519 --> 00:26:25.000
<v Speaker 2>last May. I believe he went before Congress in the

428
00:26:25.079 --> 00:26:29.519
<v Speaker 2>US saying like, hey, we better watch out, and now

429
00:26:30.079 --> 00:26:32.400
<v Speaker 2>I think he was like at the inauguration. So we

430
00:26:32.480 --> 00:26:36.079
<v Speaker 2>chatted about this in twenty twenty three's Neurotechnology episode, because

431
00:26:36.160 --> 00:26:38.920
<v Speaker 2>just a few months prior, in May twenty twenty three,

432
00:26:39.279 --> 00:26:41.960
<v Speaker 2>Sam Altman was in the news a lot as a

433
00:26:42.000 --> 00:26:44.599
<v Speaker 2>cautionary voice, and as we said in that episode, if

434
00:26:44.599 --> 00:26:47.359
<v Speaker 2>you're wondering why this was a big deal. Sam Altman

435
00:26:47.599 --> 00:26:50.680
<v Speaker 2>is the head of open Ai, which invented chat GPT,

436
00:26:50.920 --> 00:26:53.160
<v Speaker 2>and in spring of twenty twenty three, he spoke at

437
00:26:53.160 --> 00:26:57.160
<v Speaker 2>the Senate Judiciary Committee Subcommittee on Privacy, Technology and the Law.

438
00:26:57.200 --> 00:27:01.920
<v Speaker 2>Hearing called Oversight of AI Rules or Artificial Intelligence. He

439
00:27:02.000 --> 00:27:04.759
<v Speaker 2>also signed a statement about trying to mitigate the risk

440
00:27:04.880 --> 00:27:08.440
<v Speaker 2>of extinction, and he told the committee that quote, AI

441
00:27:08.640 --> 00:27:12.799
<v Speaker 2>could cause significant harm to the world.

442
00:27:13.000 --> 00:27:16.359
<v Speaker 5>My worst fears are that we cause significant we the field,

443
00:27:16.400 --> 00:27:19.640
<v Speaker 5>the technology, the industry caused significant harm to the world.

444
00:27:21.039 --> 00:27:22.519
<v Speaker 5>I think that could happen in a lot of different ways.

445
00:27:22.680 --> 00:27:25.279
<v Speaker 5>I think if this technology goes wrong, it can go

446
00:27:25.400 --> 00:27:28.880
<v Speaker 5>quite wrong, and we want to be vocal about that.

447
00:27:28.920 --> 00:27:31.319
<v Speaker 5>We want to work with the government to prevent that

448
00:27:31.359 --> 00:27:31.920
<v Speaker 5>from happening.

449
00:27:32.200 --> 00:27:36.039
<v Speaker 2>And ultimately Altman urged the committee to help establish a

450
00:27:36.079 --> 00:27:39.039
<v Speaker 2>new framework for this new technology. And though in twenty

451
00:27:39.079 --> 00:27:43.720
<v Speaker 2>sixteen Altman declared that Donald Trump was terrible, he recently

452
00:27:44.039 --> 00:27:46.920
<v Speaker 2>backpedaled on that and Altman said that he's changed his

453
00:27:47.000 --> 00:27:51.079
<v Speaker 2>mind and donated one million dollars to Trump's reelection campaign

454
00:27:51.240 --> 00:27:55.599
<v Speaker 2>in twenty twenty four. So Altman's thoughts on AI regulations

455
00:27:56.160 --> 00:27:59.559
<v Speaker 2>likely have pivoted in the last few years since that hearing.

456
00:27:59.799 --> 00:28:03.240
<v Speaker 2>It doesn't seem like regulations are gonna happen very fast.

457
00:28:04.000 --> 00:28:07.319
<v Speaker 4>Yeah, So this idea of oh, we gotta watch out

458
00:28:07.480 --> 00:28:12.759
<v Speaker 4>because our AIS might become sentient and might be out

459
00:28:12.839 --> 00:28:18.359
<v Speaker 4>of control, might cause existential risk and lead to human extinction.

460
00:28:19.119 --> 00:28:25.279
<v Speaker 4>So unfortunately, this is a very popular and commonly dissimilated

461
00:28:25.759 --> 00:28:29.839
<v Speaker 4>worry emerging out of AI, but there is very little

462
00:28:30.079 --> 00:28:33.920
<v Speaker 4>scientific evidence for it. People have done a Toto analysis

463
00:28:34.079 --> 00:28:41.359
<v Speaker 4>how such a possibility is just zero percent likely there,

464
00:28:41.359 --> 00:28:41.960
<v Speaker 4>it's just.

465
00:28:42.000 --> 00:28:46.039
<v Speaker 2>Really yeah, but human beings can be so terrible and

466
00:28:46.079 --> 00:28:47.640
<v Speaker 2>they're learning from us.

467
00:28:47.920 --> 00:28:52.559
<v Speaker 4>Yeah, so there is no intention there is no wish

468
00:28:52.799 --> 00:28:57.039
<v Speaker 4>or there is no desire to act on something, to

469
00:28:57.240 --> 00:28:58.799
<v Speaker 4>do something. But at the end of the day, it

470
00:28:58.920 --> 00:29:03.960
<v Speaker 4>really is a massive, complex algorithm. Of course, and that

471
00:29:04.160 --> 00:29:07.759
<v Speaker 4>is to some extent, I'm predictable, But that doesn't mean

472
00:29:08.039 --> 00:29:11.640
<v Speaker 4>you know AI systems as they develop further, you know,

473
00:29:11.720 --> 00:29:16.240
<v Speaker 4>all of a sudden develop intentionality or wishes or interests

474
00:29:16.319 --> 00:29:19.480
<v Speaker 4>or needs. I mean, you and I are you know,

475
00:29:19.680 --> 00:29:23.440
<v Speaker 4>as human beings, we do something and we get satisfaction

476
00:29:23.559 --> 00:29:26.359
<v Speaker 4>out of it. I have a motivation for doing the

477
00:29:26.440 --> 00:29:30.480
<v Speaker 4>research I do, and if that doesn't happen, I feel disappointed.

478
00:29:30.559 --> 00:29:34.119
<v Speaker 4>I can feel sad. I also feel accountable when I

479
00:29:34.240 --> 00:29:37.720
<v Speaker 4>put out, you know, a research paper. I know if

480
00:29:37.759 --> 00:29:40.519
<v Speaker 4>there are errors in it, I know I'm the one

481
00:29:40.599 --> 00:29:44.559
<v Speaker 4>responsible for it. So there is none of that. So

482
00:29:44.680 --> 00:29:48.440
<v Speaker 4>when an AI system gives you an output, it's not

483
00:29:48.599 --> 00:29:52.160
<v Speaker 4>because it might be worried about if it's an incorrect answer,

484
00:29:52.640 --> 00:29:55.839
<v Speaker 4>or it's because it wants to please you. It's just

485
00:29:55.920 --> 00:30:00.079
<v Speaker 4>a chatbot that is designed to provide, you know, give

486
00:30:00.240 --> 00:30:05.400
<v Speaker 4>answers again based on prompt So this idea of a

487
00:30:05.680 --> 00:30:11.960
<v Speaker 4>systems causing existential risk leans on this huge leap of

488
00:30:12.119 --> 00:30:15.599
<v Speaker 4>faith that requires you to believe that there is intention,

489
00:30:15.920 --> 00:30:19.680
<v Speaker 4>there is emotion, there is motivation, all these human characters,

490
00:30:19.720 --> 00:30:22.119
<v Speaker 4>all these things that makes us human, but it's just

491
00:30:22.160 --> 00:30:25.079
<v Speaker 4>not there. It can't emerge out of nowhere. That's part

492
00:30:25.119 --> 00:30:29.440
<v Speaker 4>of what makes us different and unique as individuals, as

493
00:30:29.680 --> 00:30:33.799
<v Speaker 4>biological organisms. These are things that are hardware on us.

494
00:30:34.279 --> 00:30:36.240
<v Speaker 4>These are also things that makes us human.

495
00:30:36.519 --> 00:30:37.640
<v Speaker 2>This is why I fret.

496
00:30:37.880 --> 00:30:41.240
<v Speaker 4>Still, of course, we have to worry about powerful people

497
00:30:41.880 --> 00:30:45.519
<v Speaker 4>using AI to do terrible things. And what worries me

498
00:30:45.920 --> 00:30:50.359
<v Speaker 4>is over the past year and especially since the rise

499
00:30:50.400 --> 00:30:54.279
<v Speaker 4>of Trump and since the Trump administration came to power,

500
00:30:54.960 --> 00:30:58.960
<v Speaker 4>you have a lot of large corporations really abandoning their

501
00:30:59.039 --> 00:31:05.599
<v Speaker 4>voluntary place to protect basic fundamental rights. So Meta has worked,

502
00:31:05.680 --> 00:31:10.200
<v Speaker 4>for example, walked back their commitment for DEI, their commitment

503
00:31:10.400 --> 00:31:14.799
<v Speaker 4>to fact check and monitor their social media platforms.

504
00:31:14.279 --> 00:31:16.960
<v Speaker 2>Against hate speech two as well, which is.

505
00:31:17.319 --> 00:31:21.960
<v Speaker 4>What exactly what really is worrying also now is you

506
00:31:22.160 --> 00:31:26.559
<v Speaker 4>have superpowers and powerful governments like the US government, the

507
00:31:26.680 --> 00:31:31.119
<v Speaker 4>UK government, even the European Union itself using AI, moving

508
00:31:31.119 --> 00:31:37.039
<v Speaker 4>into AI for surveillance for military purposes, for warfares, and

509
00:31:37.160 --> 00:31:43.240
<v Speaker 4>a lot of AI companies starting from you know, Open Ai, Meta, Amazon, Google,

510
00:31:43.720 --> 00:31:48.240
<v Speaker 4>they had voluntary principles not to use AI for military purposes,

511
00:31:48.319 --> 00:31:51.000
<v Speaker 4>but over the past year all of them have abandoned that.

512
00:31:51.440 --> 00:31:54.440
<v Speaker 4>So even here across in the EU you have a

513
00:31:54.559 --> 00:31:59.440
<v Speaker 4>French AI company called Mistral announcing that they are open

514
00:31:59.480 --> 00:32:03.599
<v Speaker 4>to working with European governments to provide military AI. So

515
00:32:03.680 --> 00:32:06.359
<v Speaker 4>of course we have to worry about governments using AI

516
00:32:06.559 --> 00:32:10.079
<v Speaker 4>under the guise of national security, which really means you know,

517
00:32:10.160 --> 00:32:15.519
<v Speaker 4>monitoring and surveillance and squishing dissent and really this is

518
00:32:15.559 --> 00:32:19.200
<v Speaker 4>against you know, fundamental rights for freedom of expression, freedom

519
00:32:19.240 --> 00:32:21.720
<v Speaker 4>of movement, and so on. So we have to worry

520
00:32:21.759 --> 00:32:26.880
<v Speaker 4>about AI. But AI in the hands of powerful governments

521
00:32:26.920 --> 00:32:30.440
<v Speaker 4>and people in position of power rather than the AI itself,

522
00:32:30.480 --> 00:32:33.920
<v Speaker 4>because the AI can't do anything by itself.

523
00:32:34.440 --> 00:32:36.680
<v Speaker 2>Is it sort of like the guns don't kill people,

524
00:32:36.720 --> 00:32:40.200
<v Speaker 2>people kill people kind of a situation, Yeah, exactly. How

525
00:32:40.279 --> 00:32:42.880
<v Speaker 2>is that working out for us in the US? Though? Well?

526
00:32:43.279 --> 00:32:46.240
<v Speaker 2>Death by firearm is the leading cause of mortality for

527
00:32:46.400 --> 00:32:48.839
<v Speaker 2>teens and children in the US, according to the Pew

528
00:32:49.000 --> 00:32:53.279
<v Speaker 2>Research Center, and over half of our nearly fifty thousand

529
00:32:53.519 --> 00:32:56.839
<v Speaker 2>gun deaths a year in the US are suicides. That's

530
00:32:56.880 --> 00:33:00.599
<v Speaker 2>not going well, and that's because the NRA slogan guns

531
00:33:00.599 --> 00:33:03.519
<v Speaker 2>don't kill people, people do is what is known as

532
00:33:03.720 --> 00:33:08.119
<v Speaker 2>bumper sticker logic or a misdirection, also called a false

533
00:33:08.279 --> 00:33:13.119
<v Speaker 2>dichotomy or plainly speaking, a fallacy according to philosophers. So

534
00:33:13.319 --> 00:33:18.359
<v Speaker 2>giving AI a sense of technological neutrality is a bit misguided.

535
00:33:18.640 --> 00:33:22.599
<v Speaker 2>The regulations being walked back is terrifying, especially trying to

536
00:33:22.640 --> 00:33:26.359
<v Speaker 2>put trust in a government to stop things when a

537
00:33:26.400 --> 00:33:30.000
<v Speaker 2>lot of our people in power, like don't know how

538
00:33:30.000 --> 00:33:32.200
<v Speaker 2>to use their own printers, you know. So some of

539
00:33:32.200 --> 00:33:35.000
<v Speaker 2>the questions, and like the congressional hearings are like, how

540
00:33:35.039 --> 00:33:38.799
<v Speaker 2>does this even work to track my movement?

541
00:33:39.759 --> 00:33:43.839
<v Speaker 3>Does Google through this phone know that I have moved

542
00:33:43.880 --> 00:33:46.119
<v Speaker 3>here and moved over to the left.

543
00:33:45.960 --> 00:33:49.880
<v Speaker 2>Which is terrifying. So maybe they don't have morals and

544
00:33:49.960 --> 00:33:55.880
<v Speaker 2>guilts and things like that and ambitions. But I was

545
00:33:55.920 --> 00:33:59.160
<v Speaker 2>looking at some research showing that AI is being trained

546
00:33:59.160 --> 00:34:02.599
<v Speaker 2>to become more and more sexist, more and more xenophobic,

547
00:34:02.720 --> 00:34:05.839
<v Speaker 2>more and more racist us more and more hate speech.

548
00:34:06.359 --> 00:34:11.559
<v Speaker 2>And is it learning from the worst of humanity? Is

549
00:34:11.559 --> 00:34:16.159
<v Speaker 2>it amplifying it? Is that just exposing how much hate

550
00:34:16.320 --> 00:34:17.159
<v Speaker 2>is in the world.

551
00:34:17.320 --> 00:34:24.880
<v Speaker 4>Yeah, so let's maybe walk back twenty years because that's

552
00:34:25.119 --> 00:34:29.159
<v Speaker 4>when you know, real progress in AI started to emerge.

553
00:34:29.440 --> 00:34:34.000
<v Speaker 4>I mean, we've had a lot of the core principles

554
00:34:34.039 --> 00:34:40.000
<v Speaker 4>for AI, you know, since in the nineteen fifty sixty, seventies, eighties,

555
00:34:40.320 --> 00:34:45.440
<v Speaker 4>Like some of the foundational papers about reinforcement learning deep

556
00:34:45.559 --> 00:34:49.519
<v Speaker 4>learning were written in the nineteen eighties. So jeff Hinton's

557
00:34:49.519 --> 00:34:53.960
<v Speaker 4>famous paper on I think what was convulsional learning was written,

558
00:34:54.119 --> 00:34:56.320
<v Speaker 4>you know, late nineteen eighties.

559
00:34:56.440 --> 00:34:58.159
<v Speaker 2>We don't have to go too deep into it, but

560
00:34:58.199 --> 00:35:00.159
<v Speaker 2>I do want to tell you that Jeffrey Hinton is

561
00:35:00.159 --> 00:35:04.519
<v Speaker 2>apparently considered the godfather of AI and a leading figure

562
00:35:04.599 --> 00:35:07.480
<v Speaker 2>in the deep learning world, and in twenty twenty four

563
00:35:07.920 --> 00:35:10.480
<v Speaker 2>he won the Nobel Prize for his work. He's also

564
00:35:10.480 --> 00:35:13.440
<v Speaker 2>worked for Google Brain, and then he quit Google because

565
00:35:13.480 --> 00:35:17.920
<v Speaker 2>he wanted to quote freely speak about the risks of AI.

566
00:35:18.480 --> 00:35:21.119
<v Speaker 2>Quit Google so he could talk about it now. In

567
00:35:21.159 --> 00:35:23.880
<v Speaker 2>twenty twenty three, during a CBS Saturday Morning news segment,

568
00:35:24.159 --> 00:35:29.760
<v Speaker 2>he warned about deliberate misuse by malicious actors, unemployment, and

569
00:35:29.920 --> 00:35:33.920
<v Speaker 2>existential risk involving AI. He is very much in favor

570
00:35:34.000 --> 00:35:36.760
<v Speaker 2>of research on the risks of what could become a

571
00:35:36.800 --> 00:35:39.519
<v Speaker 2>monster that he helped create. He's like, you know, we

572
00:35:39.559 --> 00:35:43.800
<v Speaker 2>need some safety guidelines and regulations, buddies, and that it's

573
00:35:43.920 --> 00:35:47.000
<v Speaker 2>not really happening, but yes, he is among you who

574
00:35:47.079 --> 00:35:50.239
<v Speaker 2>over the last many decades drove these innovations.

575
00:35:50.719 --> 00:35:56.360
<v Speaker 4>But what really made the AI revolution possible is the

576
00:35:56.360 --> 00:35:59.400
<v Speaker 4>world Wide Web. With the emergence of the world Wide Web,

577
00:36:00.159 --> 00:36:03.239
<v Speaker 4>became possible to kind of script to kind of gather

578
00:36:03.880 --> 00:36:07.320
<v Speaker 4>harvest massive amounts of data from the world Wide Web,

579
00:36:07.360 --> 00:36:12.039
<v Speaker 4>you know, through chat forums or domains like Wikipedia. They

580
00:36:12.079 --> 00:36:15.280
<v Speaker 4>are really a core element of training AI, at least

581
00:36:15.280 --> 00:36:19.000
<v Speaker 4>for text data. So that means that a lot of

582
00:36:19.039 --> 00:36:22.880
<v Speaker 4>our training material for AI comes from the world Wide Web,

583
00:36:22.960 --> 00:36:26.880
<v Speaker 4>whether it's our digital traces, whether you know, it's the

584
00:36:27.000 --> 00:36:30.519
<v Speaker 4>pictures we put on social media, pictures of your kids,

585
00:36:30.599 --> 00:36:34.280
<v Speaker 4>your dogs, yourselves and so on, or the kind of infrastructure.

586
00:36:34.320 --> 00:36:38.360
<v Speaker 4>Digital infrastructure like Google is everywhere has dominated whether you

587
00:36:38.519 --> 00:36:42.440
<v Speaker 4>want to email or you know, prepare a presentation or

588
00:36:42.480 --> 00:36:45.880
<v Speaker 4>write a document. Google has provided the infrastructure that means

589
00:36:45.880 --> 00:36:50.000
<v Speaker 4>they have the infrastructure to constantly harvest training data. This

590
00:36:50.159 --> 00:36:53.880
<v Speaker 4>means that a lot of the data that we are

591
00:36:54.000 --> 00:36:59.960
<v Speaker 4>using for training reflects, you know, humanities beauty, but also our.

592
00:37:00.760 --> 00:37:04.159
<v Speaker 6>In the ugliness of humanity, and just last week tech

593
00:37:04.199 --> 00:37:08.400
<v Speaker 6>report released by Google admitted that its Gemini two point

594
00:37:08.440 --> 00:37:12.199
<v Speaker 6>five flash model is more likely than its predecessor model

595
00:37:12.239 --> 00:37:16.440
<v Speaker 6>two point zero to generate results outside of its safety guidelines,

596
00:37:16.559 --> 00:37:19.280
<v Speaker 6>and images are even worse than text at that and

597
00:37:19.320 --> 00:37:21.280
<v Speaker 6>I mentioned this in a twenty twenty three episode we

598
00:37:21.280 --> 00:37:24.159
<v Speaker 6>did with doctor Nita Farhani about neurotechnology.

599
00:37:24.400 --> 00:37:26.920
<v Speaker 2>But around June teenth of that year, I saw this

600
00:37:27.039 --> 00:37:31.480
<v Speaker 2>viral tweet about chatgept not acknowledging that the Texas and

601
00:37:31.519 --> 00:37:35.920
<v Speaker 2>Oklahoma border the Panhandle was in fact influenced by Texas

602
00:37:36.239 --> 00:37:39.280
<v Speaker 2>desiring to stay a slave state, which is fact that

603
00:37:39.400 --> 00:37:42.559
<v Speaker 2>chat GPT would not acknowledge. So doctor Barhani notes that

604
00:37:42.639 --> 00:37:47.639
<v Speaker 2>when an AI is built on racist, sexist, xenophobic, et

605
00:37:47.719 --> 00:37:51.679
<v Speaker 2>cetera data set, the results, like history itself, are not

606
00:37:51.960 --> 00:37:54.480
<v Speaker 2>kind to minoritized identities.

607
00:37:54.719 --> 00:38:00.320
<v Speaker 4>She says, it reflects, you know, societal norms, its reflections,

608
00:38:00.360 --> 00:38:04.320
<v Speaker 4>you know, historical injustices and so on, unless you really

609
00:38:04.400 --> 00:38:09.199
<v Speaker 4>delve into the data set and ensure that you do

610
00:38:09.280 --> 00:38:13.199
<v Speaker 4>a thorough job of cleaning the data set. We've audited

611
00:38:13.880 --> 00:38:18.480
<v Speaker 4>numerous data sets, and you find content that shouldn't be there.

612
00:38:18.599 --> 00:38:22.440
<v Speaker 4>You find, you know, images of genocide, images of you know,

613
00:38:22.679 --> 00:38:25.840
<v Speaker 4>child rip. One of the early data sets we audited

614
00:38:25.920 --> 00:38:29.199
<v Speaker 4>back in twenty nineteen was a data set called eighty

615
00:38:29.280 --> 00:38:34.199
<v Speaker 4>million Tiny Images. It was held by Mighty, and we

616
00:38:34.280 --> 00:38:39.639
<v Speaker 4>found several thousands of images, really problematic images, Images of

617
00:38:39.719 --> 00:38:43.480
<v Speaker 4>black people labeled with the N word, images of women

618
00:38:43.679 --> 00:38:48.440
<v Speaker 4>labeled with the B word, the C word, and words

619
00:38:48.480 --> 00:38:50.960
<v Speaker 4>I can't really say on air.

620
00:38:51.320 --> 00:38:55.039
<v Speaker 2>So while the upside of AI is detecting cancer from

621
00:38:55.159 --> 00:39:00.400
<v Speaker 2>scans earlier or predicting tornado patterns, there's also so much

622
00:39:00.480 --> 00:39:04.559
<v Speaker 2>concern now. Doctor Martin Luther King Junior observed and proclaimed

623
00:39:04.639 --> 00:39:08.400
<v Speaker 2>that the arc of the moral universe is long, but

624
00:39:08.480 --> 00:39:12.159
<v Speaker 2>it bends toward justice. I think we might consider that

625
00:39:12.519 --> 00:39:15.679
<v Speaker 2>the arc of the Internet is short, and it bends

626
00:39:15.719 --> 00:39:17.400
<v Speaker 2>towards smart and hate.

627
00:39:17.639 --> 00:39:21.639
<v Speaker 4>So you can't assume any datail you collect from the

628
00:39:21.679 --> 00:39:25.639
<v Speaker 4>web is really horrible. And in one of the reason audits,

629
00:39:25.679 --> 00:39:31.239
<v Speaker 4>actually we found an overwhelming amount of women concepts really

630
00:39:31.280 --> 00:39:35.280
<v Speaker 4>represented by images that come from the pornographic space. So

631
00:39:35.760 --> 00:39:38.679
<v Speaker 4>massive amounts of the web is also really you know,

632
00:39:38.800 --> 00:39:43.280
<v Speaker 4>pornographic and really you know, problematic content, So you have

633
00:39:43.360 --> 00:39:46.159
<v Speaker 4>to do a lot of filtering. So as a result,

634
00:39:46.840 --> 00:39:50.480
<v Speaker 4>this is why you know DEI initiatives. This is why

635
00:39:50.960 --> 00:39:55.039
<v Speaker 4>obligations to audit your data a set to ensure that

636
00:39:55.320 --> 00:39:57.800
<v Speaker 4>you know toxic contents have been removed and so on.

637
00:39:58.079 --> 00:39:59.880
<v Speaker 4>This is why it's so critical.

638
00:40:00.440 --> 00:40:04.039
<v Speaker 2>So an AI is only as ethical as its data sets.

639
00:40:04.079 --> 00:40:07.599
<v Speaker 2>And the Internet is a weird dark place where people

640
00:40:07.599 --> 00:40:09.960
<v Speaker 2>say things they would never say in person, So the

641
00:40:10.079 --> 00:40:11.440
<v Speaker 2>data sets are feeding that.

642
00:40:11.840 --> 00:40:13.960
<v Speaker 4>But as we are seeing now a lot of these

643
00:40:14.000 --> 00:40:17.880
<v Speaker 4>companies are abandoning their plages and we're really walking backwards.

644
00:40:18.280 --> 00:40:21.480
<v Speaker 4>But for any given AI system, whether it's a predictive

645
00:40:21.519 --> 00:40:26.159
<v Speaker 4>system or classification or generating, you can't assume that deeply

646
00:40:26.320 --> 00:40:32.159
<v Speaker 4>heled societal injustices and norms will be reflected in how

647
00:40:32.400 --> 00:40:35.320
<v Speaker 4>that AI performs in the kind of output the AI

648
00:40:35.360 --> 00:40:37.760
<v Speaker 4>gives you. So that's the default. So we have to

649
00:40:37.760 --> 00:40:41.559
<v Speaker 4>work backwards to ensure we're removing those biases.

650
00:40:42.360 --> 00:40:45.079
<v Speaker 2>Let's say that some of the comments online, some of

651
00:40:45.079 --> 00:40:50.519
<v Speaker 2>the head online is AI generated comments, which I sometimes

652
00:40:50.559 --> 00:40:53.159
<v Speaker 2>I'll look at now X and I'll say, who are

653
00:40:53.199 --> 00:40:53.920
<v Speaker 2>all these people?

654
00:40:54.000 --> 00:40:54.119
<v Speaker 4>Like?

655
00:40:54.840 --> 00:40:58.599
<v Speaker 2>Why are comments getting meaner and meaner? With Facebook, with

656
00:40:58.639 --> 00:41:02.039
<v Speaker 2>a lack of fact checking, more and more sort of

657
00:41:02.079 --> 00:41:05.760
<v Speaker 2>hateful speech. Does that mean that the next tokens and

658
00:41:05.840 --> 00:41:08.199
<v Speaker 2>data sets pick up on that and say, oh, this

659
00:41:08.360 --> 00:41:10.480
<v Speaker 2>is how people think. And then the next one. So

660
00:41:10.559 --> 00:41:13.639
<v Speaker 2>does it get amplified like mercury toxicity and like a

661
00:41:13.760 --> 00:41:14.360
<v Speaker 2>tuna fish.

662
00:41:14.559 --> 00:41:18.960
<v Speaker 4>That's one way I'm putting you. Yeah, okay, yes, yes,

663
00:41:19.079 --> 00:41:23.480
<v Speaker 4>you are encoding those biases and you are exaggerating them.

664
00:41:23.559 --> 00:41:23.880
<v Speaker 2>Yeah.

665
00:41:24.039 --> 00:41:27.639
<v Speaker 4>The technical drawback is that, so we train given AI

666
00:41:28.320 --> 00:41:31.360
<v Speaker 4>for a next world prediction, for example, it's based on

667
00:41:32.239 --> 00:41:35.440
<v Speaker 4>you know, these massive amounts of data that kind of

668
00:41:35.519 --> 00:41:39.000
<v Speaker 4>tells you how people taket how people use language for

669
00:41:39.199 --> 00:41:42.320
<v Speaker 4>English for example, how people you know, construct a core

670
00:41:42.440 --> 00:41:46.320
<v Speaker 4>and sentence. That data, that training data comes from actual

671
00:41:46.760 --> 00:41:51.639
<v Speaker 4>people activities, people interactions. That is your baseline, so to speak,

672
00:41:51.880 --> 00:41:55.199
<v Speaker 4>in when you are modeling how you know language operates.

673
00:41:55.800 --> 00:41:59.119
<v Speaker 4>But now, as you said, as the world wide wave

674
00:41:59.400 --> 00:42:05.239
<v Speaker 4>is feel more and more with you know, synthetic text

675
00:42:05.400 --> 00:42:09.079
<v Speaker 4>or synthas data that comes from generative AI system itself,

676
00:42:09.239 --> 00:42:12.079
<v Speaker 4>then your AI system has no frame of reference. It

677
00:42:12.280 --> 00:42:16.280
<v Speaker 4>tends to forget, so the quality of the output starts

678
00:42:16.320 --> 00:42:17.199
<v Speaker 4>to deterior it.

679
00:42:19.480 --> 00:42:20.519
<v Speaker 2>That's so scary.

680
00:42:20.800 --> 00:42:22.480
<v Speaker 4>So this is called model collapse.

681
00:42:22.880 --> 00:42:24.920
<v Speaker 2>Okay, does this keep you aver night?

682
00:42:26.000 --> 00:42:26.239
<v Speaker 4>Does that.

683
00:42:26.480 --> 00:42:28.159
<v Speaker 2>I mean, I know that it's like, don't be afraid,

684
00:42:28.199 --> 00:42:32.119
<v Speaker 2>don't be afraid, but it's also like, this is very

685
00:42:32.159 --> 00:42:35.159
<v Speaker 2>new territory for humanity, right, Yeah.

686
00:42:34.880 --> 00:42:37.199
<v Speaker 4>But at the end of the day, I mean, people

687
00:42:37.199 --> 00:42:40.960
<v Speaker 4>should be in control. If an AI system starts producing

688
00:42:41.000 --> 00:42:44.639
<v Speaker 4>outputs that is rubbish, that is irrelevant. I don't think

689
00:42:44.639 --> 00:42:47.719
<v Speaker 4>it should scare us. It should make people like, Okay,

690
00:42:47.719 --> 00:42:50.119
<v Speaker 4>that's not helpful to me anymore, so I'm not gonna

691
00:42:50.159 --> 00:42:51.400
<v Speaker 4>use it.

692
00:42:51.480 --> 00:42:55.320
<v Speaker 2>Maybe the more it unravels and crashes out, the less

693
00:42:55.320 --> 00:42:58.440
<v Speaker 2>people will rely on it. But of course that hinges

694
00:42:58.559 --> 00:43:01.599
<v Speaker 2>on being able to tell that it's spitting nonsense at you.

695
00:43:01.719 --> 00:43:03.719
<v Speaker 2>And in this day and age, the world is so

696
00:43:04.360 --> 00:43:09.199
<v Speaker 2>profoundly absurd that truly anything is believable. And doctor Burhannig

697
00:43:09.239 --> 00:43:12.119
<v Speaker 2>says that public education is key and just getting the

698
00:43:12.239 --> 00:43:14.760
<v Speaker 2>word out that a lot of what we think about

699
00:43:14.800 --> 00:43:19.599
<v Speaker 2>AIS capabilities are just big corporations humping out hype and

700
00:43:19.840 --> 00:43:22.480
<v Speaker 2>pr But the auditors on the inside, like her in

701
00:43:22.519 --> 00:43:25.320
<v Speaker 2>her lab, know that boy, Harry hot damn, it is

702
00:43:25.320 --> 00:43:27.960
<v Speaker 2>a bunch of horse packy flim flam, and not to

703
00:43:28.000 --> 00:43:28.960
<v Speaker 2>believe the hype.

704
00:43:29.199 --> 00:43:35.199
<v Speaker 4>The actual performance is nowhere near what the developers claim,

705
00:43:35.599 --> 00:43:39.199
<v Speaker 4>So these are the facts that we really have to communicate.

706
00:43:39.320 --> 00:43:41.760
<v Speaker 4>A lot of the AI systems that we are interacting

707
00:43:41.800 --> 00:43:45.559
<v Speaker 4>with are actually subpar in terms of performance, in terms

708
00:43:45.559 --> 00:43:47.239
<v Speaker 4>of what they are supposed to do, in terms of

709
00:43:47.280 --> 00:43:50.800
<v Speaker 4>what people expect them to do. Because these big corporations

710
00:43:50.840 --> 00:43:55.440
<v Speaker 4>have really mastered public communication in PR, A lot of

711
00:43:55.559 --> 00:43:59.320
<v Speaker 4>like the failures or the drawbacks of AI systems are

712
00:43:59.440 --> 00:44:02.320
<v Speaker 4>new to people when you actually communicate it. But this

713
00:44:02.400 --> 00:44:05.440
<v Speaker 4>should really be like common knowledge, and if people want

714
00:44:05.480 --> 00:44:08.840
<v Speaker 4>to use AI, they should know both the strengths and

715
00:44:08.920 --> 00:44:11.480
<v Speaker 4>what they can do with it, but also where the

716
00:44:11.559 --> 00:44:14.119
<v Speaker 4>limitations are or what it can't do for them.

717
00:44:14.480 --> 00:44:18.719
<v Speaker 2>Does abstention work? Does not going on matter and giving

718
00:44:18.719 --> 00:44:23.960
<v Speaker 2>them more fadder? Does not using chat GBT? Does any

719
00:44:24.000 --> 00:44:26.840
<v Speaker 2>sort of like boycott work?

720
00:44:28.360 --> 00:44:32.960
<v Speaker 4>Yes? And no. On the one hand, so a lot

721
00:44:33.039 --> 00:44:39.079
<v Speaker 4>of these AI systems have really cleverly been integrated into

722
00:44:39.159 --> 00:44:40.360
<v Speaker 4>the social infrastructure.

723
00:44:40.519 --> 00:44:40.719
<v Speaker 2>Yeah.

724
00:44:40.800 --> 00:44:42.920
<v Speaker 4>So for example, I'm not on Facebook. I haven't been

725
00:44:42.960 --> 00:44:46.639
<v Speaker 4>on Facebook for over ten years, but the apartment complex

726
00:44:46.719 --> 00:44:50.800
<v Speaker 4>I live in can only be communicated via Facebook groups.

727
00:44:50.960 --> 00:44:54.639
<v Speaker 4>I still refuse to create a Facebook account, But situations

728
00:44:54.679 --> 00:44:59.280
<v Speaker 4>like this really gives you very little option to abstain

729
00:44:59.480 --> 00:45:02.840
<v Speaker 4>to not use these platforms, and you can't avoid Google,

730
00:45:02.880 --> 00:45:07.440
<v Speaker 4>for example, Google Search and like you know, gimming and

731
00:45:07.679 --> 00:45:11.480
<v Speaker 4>like Google Docs. It makes it really difficult if you

732
00:45:11.519 --> 00:45:14.599
<v Speaker 4>want to apply for a job. Almost all companies now

733
00:45:14.719 --> 00:45:18.280
<v Speaker 4>use some kind of AI to filter your CV before

734
00:45:18.280 --> 00:45:22.920
<v Speaker 4>it reaches human So in some senses you don't even

735
00:45:23.000 --> 00:45:26.599
<v Speaker 4>have the option to opt out. If you are, you know,

736
00:45:26.719 --> 00:45:30.599
<v Speaker 4>someone looking for a job, you can't say, oh, I

737
00:45:30.599 --> 00:45:33.119
<v Speaker 4>don't want you to use an AI system to sift

738
00:45:33.119 --> 00:45:33.760
<v Speaker 4>through my CV.

739
00:45:34.000 --> 00:45:35.840
<v Speaker 2>It's just like it's gonna happen.

740
00:45:35.920 --> 00:45:36.880
<v Speaker 4>Yeah, it's gonna happen.

741
00:45:37.360 --> 00:45:40.800
<v Speaker 2>Doctor Burhanne says that it's pretty unavoidable. And I have

742
00:45:40.960 --> 00:45:44.880
<v Speaker 2>asked tech lawyers and even they don't read Apple's terms

743
00:45:44.880 --> 00:45:47.800
<v Speaker 2>and conditions. They're like, I just checked the box.

744
00:45:48.000 --> 00:45:51.079
<v Speaker 4>So instead of using WhatsApp, which is owned by Meta,

745
00:45:51.320 --> 00:45:54.559
<v Speaker 4>which you know really gathers all your texts, all your information,

746
00:45:54.960 --> 00:45:58.079
<v Speaker 4>we can't move to other messaging apps like Signal. So

747
00:45:58.159 --> 00:46:01.559
<v Speaker 4>Signal has you know, an to end encryption. There is

748
00:46:01.599 --> 00:46:05.000
<v Speaker 4>no backdoor. Nobody can access it, not even governments. This

749
00:46:05.159 --> 00:46:08.280
<v Speaker 4>is one of the things. Meridith Witiker, the CEO of Signal,

750
00:46:08.320 --> 00:46:12.440
<v Speaker 4>has been really strong and standing up to large governments

751
00:46:12.559 --> 00:46:15.679
<v Speaker 4>is that nobody should have a back end access that

752
00:46:16.039 --> 00:46:18.559
<v Speaker 4>gives them the opportunity to gather data.

753
00:46:18.599 --> 00:46:21.519
<v Speaker 2>And yes, Signal is run by a nonprofit foundation, a

754
00:46:21.719 --> 00:46:26.280
<v Speaker 2>Signal Foundation dot org and Meredith Whittaker is Signal's president

755
00:46:26.320 --> 00:46:28.880
<v Speaker 2>and she had worked at Google for ten years and

756
00:46:28.920 --> 00:46:31.639
<v Speaker 2>she was raising concerns about their AI and she was

757
00:46:31.679 --> 00:46:35.039
<v Speaker 2>also a core organizer of the twenty eighteen Google walkout

758
00:46:35.159 --> 00:46:38.199
<v Speaker 2>in protest of sexual harassment there and pay inequities. She

759
00:46:38.280 --> 00:46:41.639
<v Speaker 2>also advises government agencies on AI safety and privacy laws.

760
00:46:41.760 --> 00:46:46.239
<v Speaker 2>So signal, good, yay signal And many recently laid off

761
00:46:46.320 --> 00:46:49.079
<v Speaker 2>government staff that I know of will only communicate for

762
00:46:49.079 --> 00:46:51.559
<v Speaker 2>you a signal, which is kind of telling in terms

763
00:46:51.599 --> 00:46:54.760
<v Speaker 2>of their own safety concerns. But yes, you signal, so

764
00:46:54.840 --> 00:46:55.280
<v Speaker 2>we can.

765
00:46:55.480 --> 00:46:58.639
<v Speaker 4>Do some things. We can use less and less of

766
00:46:58.679 --> 00:47:02.920
<v Speaker 4>these large corporations infrastructure, and we can use you know,

767
00:47:03.039 --> 00:47:06.599
<v Speaker 4>more open source tools. But also sometimes you know, just

768
00:47:06.639 --> 00:47:10.039
<v Speaker 4>out of your control. But every little helps in every awareness,

769
00:47:10.079 --> 00:47:13.159
<v Speaker 4>you know, it kind of culminates and it will eventually

770
00:47:13.400 --> 00:47:15.800
<v Speaker 4>lead to, you know, this massive switch.

771
00:47:15.880 --> 00:47:20.400
<v Speaker 2>I hope at least that's encouraging. I hope. Yeah, And

772
00:47:20.480 --> 00:47:23.519
<v Speaker 2>can I ask you some questions from listeners. Is that okay,

773
00:47:23.599 --> 00:47:26.159
<v Speaker 2>okay for sure? But before we do, let's give away

774
00:47:26.199 --> 00:47:28.679
<v Speaker 2>some money. And this week doctor Behani selected the cause

775
00:47:28.960 --> 00:47:33.280
<v Speaker 2>the Municipality of Gaza and un r WA, which directly

776
00:47:33.280 --> 00:47:37.280
<v Speaker 2>supports Palestine refugees and displaced families in Gaza. They say

777
00:47:37.320 --> 00:47:40.519
<v Speaker 2>every donation, no matter the amount, helps them reach families

778
00:47:40.559 --> 00:47:44.760
<v Speaker 2>with life saving food, assistance, shelter, healthcare and more. And

779
00:47:44.800 --> 00:47:47.679
<v Speaker 2>for more info you can see donate dot un r

780
00:47:47.840 --> 00:47:50.159
<v Speaker 2>WA dot org, which is linked in the show notes.

781
00:47:50.360 --> 00:47:54.280
<v Speaker 2>And for more on the ongoing humanitarian crisis in Gaza,

782
00:47:54.320 --> 00:47:58.880
<v Speaker 2>please see our Genocideology episode with global expert in Crimes

783
00:47:58.880 --> 00:48:02.000
<v Speaker 2>of a Trosty doctor Dirk Moses, which we will also

784
00:48:02.199 --> 00:48:05.800
<v Speaker 2>link in the show notes. So a quick break. Now, okay,

785
00:48:05.920 --> 00:48:08.960
<v Speaker 2>we are back. Let's run through some questions from your

786
00:48:09.000 --> 00:48:13.320
<v Speaker 2>real squishy brains made of human beings out there. There's

787
00:48:13.400 --> 00:48:19.239
<v Speaker 2>some great ones job replacement Carlyn de Azevo, Alia Meyer's

788
00:48:19.280 --> 00:48:23.320
<v Speaker 2>Red Tongue, Jennifer Grogan, Ian, Jenna Congen, Rosa, Rebecca Rome,

789
00:48:23.880 --> 00:48:27.320
<v Speaker 2>Other Maya, Sam Nelson, Howard Newness. All these people wanted

790
00:48:27.360 --> 00:48:31.239
<v Speaker 2>to know, in Ian's words, will all jobs be obsolete soon?

791
00:48:31.639 --> 00:48:33.920
<v Speaker 2>Do the people working on AI give any thought to

792
00:48:33.960 --> 00:48:37.880
<v Speaker 2>compensating people for the lost income? Jenna Congden said, when

793
00:48:37.880 --> 00:48:40.320
<v Speaker 2>will AI get so good that human writers are basically

794
00:48:40.320 --> 00:48:42.519
<v Speaker 2>crowded out of a job? This goes for visual art

795
00:48:42.519 --> 00:48:45.440
<v Speaker 2>as well. In a capitalist economy, when you got a

796
00:48:45.519 --> 00:48:49.159
<v Speaker 2>hustle to make money as it is, what is going

797
00:48:49.199 --> 00:48:52.840
<v Speaker 2>to happen job wise? Do you think or do AI

798
00:48:52.920 --> 00:48:54.519
<v Speaker 2>experts such as yourself.

799
00:48:55.719 --> 00:49:00.280
<v Speaker 4>So some of the worry about your displacement is is

800
00:49:00.519 --> 00:49:05.440
<v Speaker 4>genuine and grounded on, you know, real worry. You hear

801
00:49:06.239 --> 00:49:11.480
<v Speaker 4>even the so called godfathers saying things like you shouldn't

802
00:49:11.480 --> 00:49:15.559
<v Speaker 4>bozer learning code, or like the job of software engineering,

803
00:49:15.599 --> 00:49:18.039
<v Speaker 4>for example, will become obsolete and so on.

804
00:49:18.199 --> 00:49:21.320
<v Speaker 2>So whether you're a software developer or a writer or

805
00:49:21.400 --> 00:49:23.360
<v Speaker 2>an artist, Doctor Behannie says.

806
00:49:23.199 --> 00:49:28.159
<v Speaker 4>I don't think AI will fully automate AI will fully

807
00:49:28.440 --> 00:49:33.599
<v Speaker 4>replace human task force, because at the end of the day,

808
00:49:35.039 --> 00:49:39.880
<v Speaker 4>what even the most advanced AI systems do is really

809
00:49:39.960 --> 00:49:44.880
<v Speaker 4>kind of aggregate information and kind of outputs a very

810
00:49:45.199 --> 00:49:47.960
<v Speaker 4>mediocre whether it's image or text.

811
00:49:48.400 --> 00:49:50.079
<v Speaker 2>Some of them are so good, though, some of the

812
00:49:50.199 --> 00:49:52.559
<v Speaker 2>artists so good, and you're like, like.

813
00:49:52.639 --> 00:49:56.599
<v Speaker 4>The arts, Yeah, it's not just the pure the row outputs.

814
00:49:56.679 --> 00:50:01.239
<v Speaker 4>People have tweaked probably like a thousand times. People have

815
00:50:01.280 --> 00:50:06.039
<v Speaker 4>twik dates, people have spent hours perfecting the right prompt

816
00:50:06.119 --> 00:50:10.280
<v Speaker 4>and so on. So there is always people in the loop.

817
00:50:10.360 --> 00:50:13.599
<v Speaker 4>There is always whether it's data preparing, you know, data

818
00:50:13.639 --> 00:50:18.079
<v Speaker 4>and notation, data curation, to building the AI system itself,

819
00:50:18.519 --> 00:50:22.920
<v Speaker 4>to them kind of ensuring the output is something appealing.

820
00:50:23.119 --> 00:50:25.199
<v Speaker 4>You really you need people through and through.

821
00:50:25.440 --> 00:50:28.960
<v Speaker 2>So for me, as a former newspaper journalist and I

822
00:50:29.039 --> 00:50:33.920
<v Speaker 2>was also a newspaper illustrator, I I'm not as optimistic.

823
00:50:34.039 --> 00:50:37.840
<v Speaker 2>So so many writers are copywriters who are making content

824
00:50:37.880 --> 00:50:40.920
<v Speaker 2>and articles for websites to raise their profile. And now

825
00:50:41.000 --> 00:50:43.159
<v Speaker 2>I'm hearing from those people that articles are just written

826
00:50:43.159 --> 00:50:45.679
<v Speaker 2>by AI and they are full of shit. And just

827
00:50:45.800 --> 00:50:49.800
<v Speaker 2>doing this acide is making me depressed and my chest hurts.

828
00:50:50.159 --> 00:50:52.320
<v Speaker 2>But doctor Burhani is an expert, so I'm gonna try

829
00:50:52.360 --> 00:50:55.159
<v Speaker 2>to find some bright spots. And before she had mentioned

830
00:50:55.199 --> 00:50:57.760
<v Speaker 2>that lawsuit with open AI and the New York Times,

831
00:50:57.880 --> 00:50:59.400
<v Speaker 2>and I was looking for it and I found a

832
00:50:59.440 --> 00:51:03.400
<v Speaker 2>recent art this was published literally yesterday, which had the

833
00:51:03.400 --> 00:51:07.519
<v Speaker 2>headline AI is getting more powerful, but its hallucinations are

834
00:51:07.559 --> 00:51:10.960
<v Speaker 2>getting worse. A new wave of reasoning systems from companies

835
00:51:11.000 --> 00:51:14.960
<v Speaker 2>like open AI is producing incorrect information more often even

836
00:51:14.960 --> 00:51:18.000
<v Speaker 2>the companies don't know why that's the headline. And this

837
00:51:18.039 --> 00:51:20.599
<v Speaker 2>New York Times article explained that AI systems do not

838
00:51:21.079 --> 00:51:24.920
<v Speaker 2>and cannot decide what is true and what is false,

839
00:51:25.239 --> 00:51:29.239
<v Speaker 2>and sometimes they just make stuff up, a phenomenon that

840
00:51:29.519 --> 00:51:33.679
<v Speaker 2>some AI researchers call hallucinations. And in one test, the

841
00:51:33.760 --> 00:51:37.480
<v Speaker 2>article says, hallucination rates of newer AI systems were as

842
00:51:37.559 --> 00:51:41.199
<v Speaker 2>high as seventy nine percent. And I also want to

843
00:51:41.239 --> 00:51:44.480
<v Speaker 2>note that my spellcheck tried to get me to change

844
00:51:44.559 --> 00:51:47.840
<v Speaker 2>that it's in the headline to one with an apostrophe,

845
00:51:47.960 --> 00:51:51.559
<v Speaker 2>which is incorrect. So computers, what's going on? But yes,

846
00:51:51.679 --> 00:51:54.639
<v Speaker 2>doctor Bahani says that a lot of journalism has been

847
00:51:54.679 --> 00:51:57.679
<v Speaker 2>replaced by AI, even though we all know that the

848
00:51:57.800 --> 00:51:59.480
<v Speaker 2>generative system is unreliable.

849
00:52:00.280 --> 00:52:02.760
<v Speaker 4>Loosen it. A lot of the time. It gives you

850
00:52:03.559 --> 00:52:08.280
<v Speaker 4>information that sounds coherent, that seems fuctual, but it's just

851
00:52:08.360 --> 00:52:12.280
<v Speaker 4>absolutely made up. It just there is. It even sometimes

852
00:52:12.360 --> 00:52:15.920
<v Speaker 4>gives you citations and so on of things that don't exist.

853
00:52:16.440 --> 00:52:21.440
<v Speaker 4>So we always need people to babysit AI, so to speak.

854
00:52:21.559 --> 00:52:24.599
<v Speaker 4>So a writer might be, you know, your hours might

855
00:52:24.639 --> 00:52:28.920
<v Speaker 4>be reduced and you might be getting paid less, and

856
00:52:29.000 --> 00:52:31.559
<v Speaker 4>your company might be bringing AI to kind of do

857
00:52:31.679 --> 00:52:34.480
<v Speaker 4>the bulk of the job. But still you can't put

858
00:52:34.519 --> 00:52:37.960
<v Speaker 4>out the raw outputs because most of the time it's

859
00:52:38.000 --> 00:52:41.599
<v Speaker 4>not even you know, legible. So the role of writers

860
00:52:41.639 --> 00:52:45.519
<v Speaker 4>and artists and journals and so on becomes more of

861
00:52:45.679 --> 00:52:49.360
<v Speaker 4>kind of a babysitter for AI, verifying the information that's

862
00:52:49.400 --> 00:52:52.719
<v Speaker 4>been put out, kind of ensuring it makes sense and

863
00:52:52.760 --> 00:52:53.079
<v Speaker 4>so on.

864
00:52:53.519 --> 00:52:56.079
<v Speaker 2>Right, Kenny, the babysitter is dead to.

865
00:52:56.000 --> 00:52:59.079
<v Speaker 4>Some extent, The answer is yes, and no. Humans will

866
00:52:59.119 --> 00:53:02.400
<v Speaker 4>always remain at the heart of AI. The minut human

867
00:53:02.519 --> 00:53:07.360
<v Speaker 4>involvement seizes the minute AI stops operating, because AI is

868
00:53:07.440 --> 00:53:09.400
<v Speaker 4>human thru and through. As I said, you know, from

869
00:53:09.440 --> 00:53:12.920
<v Speaker 4>the data as gathered from humans, and so much work

870
00:53:13.000 --> 00:53:17.440
<v Speaker 4>goes into data preparation, data annotation, cleaning up the data,

871
00:53:17.559 --> 00:53:21.039
<v Speaker 4>detoxifying the data, and unfortunately a lot of these tasks

872
00:53:21.119 --> 00:53:24.599
<v Speaker 4>allocated to you know, the developing world. So you have

873
00:53:24.679 --> 00:53:28.760
<v Speaker 4>a lot of data workers in Kenya and Nigeria and Ethiopia,

874
00:53:28.840 --> 00:53:32.519
<v Speaker 4>in India for example, that really do the dirty work

875
00:53:32.559 --> 00:53:35.679
<v Speaker 4>of AI. There are even a bunch of stories where

876
00:53:35.960 --> 00:53:40.239
<v Speaker 4>you have Amazon checkout for example AI checkout or self checkout,

877
00:53:40.400 --> 00:53:43.840
<v Speaker 4>where Amazon was introducing this AI where you can just

878
00:53:43.920 --> 00:53:47.800
<v Speaker 4>collect groceries and your items and just walk out, and

879
00:53:47.840 --> 00:53:50.960
<v Speaker 4>the AI is supposed to kind of identify what you

880
00:53:51.039 --> 00:53:54.199
<v Speaker 4>have picked up and charge you from your credit cards

881
00:53:54.239 --> 00:53:57.119
<v Speaker 4>for whatever you have used. But then it turns out

882
00:53:57.280 --> 00:54:01.639
<v Speaker 4>that it was actually data workers in India that's where

883
00:54:02.519 --> 00:54:04.679
<v Speaker 4>scanning every item you are picking up.

884
00:54:04.800 --> 00:54:08.280
<v Speaker 2>So oh man, what word?

885
00:54:08.679 --> 00:54:12.880
<v Speaker 4>Yeah? Yeah, and I mean MacDonald also recently partnered with

886
00:54:13.079 --> 00:54:16.159
<v Speaker 4>IBM or one of those companies to have like an

887
00:54:16.199 --> 00:54:20.599
<v Speaker 4>AI drive through where AI systems take order and they

888
00:54:20.599 --> 00:54:23.440
<v Speaker 4>have to close it within the next few weeks because

889
00:54:23.559 --> 00:54:27.800
<v Speaker 4>people were getting orders of like, you know, bacon on

890
00:54:27.840 --> 00:54:29.960
<v Speaker 4>top of ice cream and things.

891
00:54:29.719 --> 00:54:32.159
<v Speaker 2>Like that, you added bacon to my ice cream? I

892
00:54:32.199 --> 00:54:37.159
<v Speaker 2>don't want I bake it. I raised the national minimum

893
00:54:37.159 --> 00:54:39.400
<v Speaker 2>wage for the first time since two thousand and nine

894
00:54:39.440 --> 00:54:41.800
<v Speaker 2>when you can just spend billions of dollars tweaking unpaid

895
00:54:41.840 --> 00:54:45.519
<v Speaker 2>machines like welcome to the future. Maybe so, I.

896
00:54:45.440 --> 00:54:47.400
<v Speaker 4>Guess the point I'm trying to make is like you

897
00:54:47.519 --> 00:54:51.760
<v Speaker 4>always need humans for AI to to function and operate

898
00:54:52.039 --> 00:54:54.320
<v Speaker 4>as it's supposed to, because at the end of the

899
00:54:54.440 --> 00:54:58.159
<v Speaker 4>stem these days, these are like really mare machines that

900
00:54:58.679 --> 00:55:01.920
<v Speaker 4>don't have you know, intense understanding, motivation and so on,

901
00:55:02.079 --> 00:55:03.719
<v Speaker 4>Like we human star So.

902
00:55:03.719 --> 00:55:07.119
<v Speaker 2>Maybe our jobs will look different, but there will be jobs. Yes,

903
00:55:07.559 --> 00:55:11.119
<v Speaker 2>I know a lot of people myself included, wanted to

904
00:55:11.199 --> 00:55:15.079
<v Speaker 2>know the environmental impact Lily a bunch of folks and

905
00:55:15.119 --> 00:55:17.400
<v Speaker 2>first time question askers elean Or Bundy and Meghan m

906
00:55:17.440 --> 00:55:19.320
<v Speaker 2>And we also did a recent episode of Earth Day

907
00:55:19.320 --> 00:55:22.599
<v Speaker 2>with this climate activist and humanitarian rightsler Adam Met, who

908
00:55:22.639 --> 00:55:26.719
<v Speaker 2>said that AI could be solving some environmental concerns, which

909
00:55:26.760 --> 00:55:29.880
<v Speaker 2>is optimistic. But what does an AI experts take on that?

910
00:55:30.480 --> 00:55:33.519
<v Speaker 2>Megan Walker asked environmentally, how bad is AI when compared

911
00:55:33.559 --> 00:55:36.039
<v Speaker 2>to the current computing we do? Yeah, what's going on?

912
00:55:36.320 --> 00:55:37.239
<v Speaker 4>Yeah? Yeah yeah?

913
00:55:37.360 --> 00:55:39.000
<v Speaker 2>How much energy does it use?

914
00:55:39.760 --> 00:55:40.039
<v Speaker 3>Yeah?

915
00:55:40.280 --> 00:55:45.360
<v Speaker 4>So again, like we have very little information about training data.

916
00:55:45.880 --> 00:55:50.440
<v Speaker 4>The kind of energy consumption used by AI systems is

917
00:55:50.800 --> 00:55:53.239
<v Speaker 4>very OPQ. There is very little transparency.

918
00:55:53.440 --> 00:55:55.559
<v Speaker 2>Okay, but what is the damage generally?

919
00:55:55.960 --> 00:56:00.679
<v Speaker 4>So, generals of AI really conceives massive lout of power

920
00:56:00.800 --> 00:56:04.039
<v Speaker 4>compared to traditional AI. For example, if you are using

921
00:56:04.159 --> 00:56:08.239
<v Speaker 4>Google to put a prompt say you know, how many

922
00:56:08.280 --> 00:56:11.360
<v Speaker 4>glasses of water should I drink per day, and if

923
00:56:11.400 --> 00:56:13.719
<v Speaker 4>you do the exact same prompt and you ask the

924
00:56:13.800 --> 00:56:17.519
<v Speaker 4>generative systems such as chat ChiPT, people estimate you use

925
00:56:17.599 --> 00:56:21.880
<v Speaker 4>about ten times more energy to process that query into

926
00:56:21.920 --> 00:56:23.000
<v Speaker 4>generated answers.

927
00:56:23.159 --> 00:56:25.360
<v Speaker 2>I wanted to go straight to the source, so I

928
00:56:25.519 --> 00:56:29.559
<v Speaker 2>used Google AI and chatchipt for the first time, asking

929
00:56:29.599 --> 00:56:33.280
<v Speaker 2>them both how much energy does chatchipt use as opposed

930
00:56:33.320 --> 00:56:36.800
<v Speaker 2>to Google now. Google AI said in what I hope

931
00:56:36.920 --> 00:56:40.840
<v Speaker 2>is a snotty tone that chat gpt consumes significantly more

932
00:56:40.960 --> 00:56:44.000
<v Speaker 2>energy per query, five to ten times more electricity than

933
00:56:44.000 --> 00:56:47.239
<v Speaker 2>a standard Google search. Now. It cited a twenty twenty

934
00:56:47.239 --> 00:56:50.480
<v Speaker 2>four Goldman SAX report titled AI has poised to drive

935
00:56:50.559 --> 00:56:53.360
<v Speaker 2>one hundred and sixty percent increase in data center power demand.

936
00:56:53.800 --> 00:56:56.320
<v Speaker 2>Then I asked chat schiapt and it said that its

937
00:56:56.400 --> 00:56:59.960
<v Speaker 2>version four can use up to thirty times more energy

938
00:57:00.440 --> 00:57:03.960
<v Speaker 2>than a basic Google search, and it also noted, I

939
00:57:04.039 --> 00:57:07.119
<v Speaker 2>like to think defensively that Google has had decades to

940
00:57:07.199 --> 00:57:11.039
<v Speaker 2>optimize for a lower footprint now. Doctor Behanne says that

941
00:57:11.119 --> 00:57:15.519
<v Speaker 2>the energy consumption of generative AI systems has become indeed

942
00:57:15.920 --> 00:57:18.960
<v Speaker 2>a big issue, and that in countries like Ireland, the

943
00:57:19.039 --> 00:57:21.199
<v Speaker 2>data centers are power hogs.

944
00:57:22.079 --> 00:57:28.840
<v Speaker 4>The computer resources required to running AI systems equals or

945
00:57:29.039 --> 00:57:33.360
<v Speaker 4>is more than the total amount of energy that's required

946
00:57:33.559 --> 00:57:37.000
<v Speaker 4>to run Irish households. But in places like takes Us

947
00:57:37.639 --> 00:57:42.320
<v Speaker 4>sometimes that energy consumption is taken away from households to

948
00:57:42.480 --> 00:57:46.679
<v Speaker 4>run data centers. Wow, so it kind of results in

949
00:57:46.800 --> 00:57:51.480
<v Speaker 4>reduced energy for households. And this is before we even

950
00:57:51.599 --> 00:57:56.400
<v Speaker 4>get into the massive tons of water that you need

951
00:57:56.440 --> 00:57:57.920
<v Speaker 4>to cool down data centers.

952
00:57:58.599 --> 00:58:01.320
<v Speaker 2>Oh yeah, didn't even think about that.

953
00:58:01.400 --> 00:58:04.280
<v Speaker 4>Yeah, and the water also has to be pure because

954
00:58:04.280 --> 00:58:07.719
<v Speaker 4>you can't use say, you know, ocean water or seawater

955
00:58:07.840 --> 00:58:10.760
<v Speaker 4>because of the sea salt that might damage the servers

956
00:58:10.840 --> 00:58:14.320
<v Speaker 4>and so on. So again there is competition. It tends

957
00:58:14.360 --> 00:58:17.679
<v Speaker 4>to be when you use water that is used for households.

958
00:58:17.800 --> 00:58:21.519
<v Speaker 4>As a result, you know, people tend to pay for

959
00:58:21.639 --> 00:58:24.719
<v Speaker 4>the consequences of that. So yeah, water consumption is another

960
00:58:25.599 --> 00:58:26.920
<v Speaker 4>massive area as well.

961
00:58:27.239 --> 00:58:29.679
<v Speaker 2>And do you think that more companies will look toward

962
00:58:30.119 --> 00:58:34.360
<v Speaker 2>some sort of nuclear power for their supercomputers or is

963
00:58:34.400 --> 00:58:36.039
<v Speaker 2>that still too highly regulated?

964
00:58:37.199 --> 00:58:40.840
<v Speaker 4>I think companies like Google are actually talking about using

965
00:58:41.639 --> 00:58:45.000
<v Speaker 4>nuclear power, but yes, that option is being considered.

966
00:58:45.679 --> 00:58:51.199
<v Speaker 2>Yeah, how about healthcare? Several people Benjamin Burnish weared analyst

967
00:58:51.239 --> 00:58:54.559
<v Speaker 2>to young Nikki g asked, how can AI be ethically

968
00:58:54.599 --> 00:58:59.079
<v Speaker 2>apply to healthcare like data analytics, treatment options, medical imaging interpretations.

969
00:58:59.119 --> 00:59:02.559
<v Speaker 2>Second opinions, Is there some hope there for it?

970
00:59:03.639 --> 00:59:05.880
<v Speaker 4>Yeah, I think there is some hope. There is some

971
00:59:06.039 --> 00:59:08.400
<v Speaker 4>hope for sure. I think there is some hope in

972
00:59:08.880 --> 00:59:15.119
<v Speaker 4>you know, numerous domains for AI to be useful. Okay, However,

973
00:59:15.360 --> 00:59:18.679
<v Speaker 4>that just remains a theory. It's it's possible in theory.

974
00:59:19.480 --> 00:59:22.159
<v Speaker 4>But the problem there are a bunch of problems. One

975
00:59:22.199 --> 00:59:27.480
<v Speaker 4>of them is that generative systems are fundamentally unreliable. So,

976
00:59:27.519 --> 00:59:29.840
<v Speaker 4>for example, there is a new audit that came out

977
00:59:29.920 --> 00:59:34.519
<v Speaker 4>I think towards the end of January where they looked

978
00:59:34.559 --> 00:59:39.639
<v Speaker 4>at this new AI tool where the system kind of

979
00:59:39.920 --> 00:59:45.159
<v Speaker 4>records your conversations between say a healthcare provider and a patient,

980
00:59:45.679 --> 00:59:50.000
<v Speaker 4>and it summarizes the conversation, and it kind of it's

981
00:59:50.039 --> 00:59:52.960
<v Speaker 4>supposed to reduce a lot of the work for nurses

982
00:59:53.039 --> 00:59:56.239
<v Speaker 4>and so on. And what they found was that in

983
00:59:56.280 --> 01:00:01.639
<v Speaker 4>some cases, eight out of the teen summaries where hallucinations.

984
01:00:02.519 --> 01:00:02.639
<v Speaker 2>No.

985
01:00:05.039 --> 01:00:08.119
<v Speaker 4>So generative systems tend to be unreliable. And the other

986
01:00:08.239 --> 01:00:12.880
<v Speaker 4>thing is because a lot of these tools that are

987
01:00:12.880 --> 01:00:16.039
<v Speaker 4>supposed to be used in health care tend to be

988
01:00:16.199 --> 01:00:21.079
<v Speaker 4>built by businesses with the objective of maximizing profits. They

989
01:00:21.159 --> 01:00:24.840
<v Speaker 4>tend to have a different kind of objective than say,

990
01:00:25.199 --> 01:00:28.920
<v Speaker 4>you know, what's good for the patient. So another famous

991
01:00:28.960 --> 01:00:32.360
<v Speaker 4>case is United healths that is in court at the

992
01:00:32.400 --> 01:00:35.400
<v Speaker 4>moment where they were using a suit of about fifty

993
01:00:35.440 --> 01:00:41.320
<v Speaker 4>algorithms to look at mental health services, and what they

994
01:00:41.360 --> 01:00:44.719
<v Speaker 4>found was that they were using you know, costs as

995
01:00:44.719 --> 01:00:47.440
<v Speaker 4>a proxy rather than the need of the patient as

996
01:00:47.480 --> 01:00:50.079
<v Speaker 4>a proxy, and they were cutting a lot of services,

997
01:00:50.159 --> 01:00:53.599
<v Speaker 4>a lot of like, you know, therapy services and meditations

998
01:00:53.639 --> 01:00:58.199
<v Speaker 4>and other necessary services. Again because they are you know,

999
01:00:58.239 --> 01:01:01.920
<v Speaker 4>they're looking at the wrong motivation, their own proxy. They're

1000
01:01:01.960 --> 01:01:05.159
<v Speaker 4>looking to save the company money rather than to ground

1001
01:01:05.199 --> 01:01:09.280
<v Speaker 4>what they do in the needs of the patient. So

1002
01:01:09.519 --> 01:01:14.000
<v Speaker 4>if we correct, for example, hallucinations and biases in AI systems,

1003
01:01:14.480 --> 01:01:18.159
<v Speaker 4>and if we kind of it's impossible to strip down

1004
01:01:18.199 --> 01:01:21.880
<v Speaker 4>all you know, capitalist motivations. But if you know, capitalist

1005
01:01:21.880 --> 01:01:26.239
<v Speaker 4>motivations come second to you know, the needs of the patient,

1006
01:01:26.960 --> 01:01:30.639
<v Speaker 4>then it's possible to kind of develop a systems in

1007
01:01:30.840 --> 01:01:36.719
<v Speaker 4>various areas of fALS care that prioritize patients that prioritize

1008
01:01:36.760 --> 01:01:40.239
<v Speaker 4>people as opposed to just you know, inserting technology for

1009
01:01:40.320 --> 01:01:44.920
<v Speaker 4>the sake of having technology, and also for using technology

1010
01:01:44.920 --> 01:01:48.599
<v Speaker 4>to maximize profit, as opposed to ensuring patient safety.

1011
01:01:48.360 --> 01:01:51.760
<v Speaker 2>Which is once again good luck. I mean, our healthcare

1012
01:01:51.920 --> 01:01:55.840
<v Speaker 2>is above and beyond frustrating. But a lot of people

1013
01:01:55.840 --> 01:01:59.639
<v Speaker 2>wanted to know. Sam Wise, Emily heard, Amalia Magda Casawka

1014
01:02:00.400 --> 01:02:03.440
<v Speaker 2>wanted to know. Kier Henrickson asked for some question, asker,

1015
01:02:04.039 --> 01:02:06.719
<v Speaker 2>how do we feel about AI and chatbots and using

1016
01:02:06.719 --> 01:02:10.559
<v Speaker 2>them in high schools schoolwork? Same ways AI used in

1017
01:02:10.559 --> 01:02:14.079
<v Speaker 2>schools thoughts is there a way to flag it? Or

1018
01:02:14.079 --> 01:02:17.519
<v Speaker 2>are we doing education and injustice?

1019
01:02:19.639 --> 01:02:20.079
<v Speaker 1>Yeah?

1020
01:02:20.119 --> 01:02:24.599
<v Speaker 4>So, on the one hand, I know some people that

1021
01:02:24.840 --> 01:02:30.800
<v Speaker 4>find using AI chatboats really helpful. You give it a prompt,

1022
01:02:30.840 --> 01:02:33.639
<v Speaker 4>it gives you just a bunch of answers. Of course,

1023
01:02:33.760 --> 01:02:38.920
<v Speaker 4>these are people that know how to craft the perfect prompts,

1024
01:02:39.280 --> 01:02:42.880
<v Speaker 4>that know where AI can be useful and where it

1025
01:02:43.000 --> 01:02:47.119
<v Speaker 4>might fail you. So with all that in mind, it

1026
01:02:47.159 --> 01:02:49.360
<v Speaker 4>can't be useful, but you need to be an expert.

1027
01:02:50.480 --> 01:02:54.880
<v Speaker 4>Having said that, for young kids, studies are starting to

1028
01:02:54.960 --> 01:02:59.400
<v Speaker 4>emerge that for example, they did a control study I

1029
01:02:59.440 --> 01:03:03.920
<v Speaker 4>think we is over three thousand students where some of

1030
01:03:03.960 --> 01:03:08.360
<v Speaker 4>the students were given chatbots to help them with I

1031
01:03:08.400 --> 01:03:13.119
<v Speaker 4>think it's maths problems. The others weren't, and they did

1032
01:03:13.159 --> 01:03:15.840
<v Speaker 4>a taste. So what they found was that the kids

1033
01:03:15.880 --> 01:03:19.480
<v Speaker 4>that had chat boats did better than the kids that

1034
01:03:19.599 --> 01:03:23.760
<v Speaker 4>didn't have. Then they performed another taste a few weeks later,

1035
01:03:24.239 --> 01:03:27.840
<v Speaker 4>and they found that the kids that used chat boats

1036
01:03:28.239 --> 01:03:32.559
<v Speaker 4>performed way worse than the kids that didn't have. So

1037
01:03:32.679 --> 01:03:38.159
<v Speaker 4>people are realizing that these systems inhibit learning. Of course,

1038
01:03:38.199 --> 01:03:41.880
<v Speaker 4>you know, education is not just information dissemination, the teacher

1039
01:03:41.960 --> 01:03:45.360
<v Speaker 4>going into class and just like taining the students facts,

1040
01:03:45.599 --> 01:03:48.519
<v Speaker 4>but rather it's an interaction. It's a two way street,

1041
01:03:48.639 --> 01:03:52.840
<v Speaker 4>both for the student and the teacher, developing the skills,

1042
01:03:53.360 --> 01:03:58.840
<v Speaker 4>especially critical skills to analyze and to decipher fact from fiction.

1043
01:03:59.480 --> 01:04:02.599
<v Speaker 4>You know, pormon for misinformation and so on. And when

1044
01:04:02.639 --> 01:04:07.559
<v Speaker 4>you use AI chatbots without knowing their limitations, you tend

1045
01:04:07.599 --> 01:04:10.920
<v Speaker 4>to kind of trust the output, you tend to treat

1046
01:04:10.960 --> 01:04:14.280
<v Speaker 4>it as facts. But also it inhibits your learning. It

1047
01:04:14.400 --> 01:04:18.159
<v Speaker 4>inhibits your critical skills. And if you don't have the

1048
01:04:18.280 --> 01:04:21.280
<v Speaker 4>knowledge to begin wish to verify the answer, you have

1049
01:04:21.440 --> 01:04:23.840
<v Speaker 4>no way of knowing what you have, what you are

1050
01:04:23.880 --> 01:04:27.119
<v Speaker 4>getting is correct or incorrect. So in the long term,

1051
01:04:27.199 --> 01:04:30.519
<v Speaker 4>studies are coming out to show that they might seem

1052
01:04:30.880 --> 01:04:33.760
<v Speaker 4>helpful in the immediate term, but in the long term,

1053
01:04:34.039 --> 01:04:37.559
<v Speaker 4>these chatbots might be inhibiting the learning process.

1054
01:04:38.639 --> 01:04:40.559
<v Speaker 2>Last listener question and I know my husband has this

1055
01:04:40.639 --> 01:04:43.599
<v Speaker 2>question too. DV and C share rample and Chelsea and

1056
01:04:43.639 --> 01:04:47.039
<v Speaker 2>her doctor Charlie want to know. Chelsea asked, why does

1057
01:04:47.119 --> 01:04:50.360
<v Speaker 2>AI do better when you threaten it? Is that ethical?

1058
01:04:50.400 --> 01:04:52.239
<v Speaker 2>Because it doesn't feel like a good president to set

1059
01:04:52.280 --> 01:04:54.719
<v Speaker 2>in any part of life. VNC asked, is it weird

1060
01:04:54.719 --> 01:04:56.360
<v Speaker 2>that I feel the need to say please and thank you?

1061
01:04:56.360 --> 01:04:59.559
<v Speaker 2>We're talking to chatbots. Will the AI overlords be nicer

1062
01:04:59.599 --> 01:05:02.239
<v Speaker 2>to me when they take over? I assume all of

1063
01:05:02.239 --> 01:05:05.559
<v Speaker 2>my conversations will be logged for eternity and Jarett, my

1064
01:05:05.639 --> 01:05:09.039
<v Speaker 2>husband also doesn't use chat ChiPT very much, but when

1065
01:05:09.039 --> 01:05:11.760
<v Speaker 2>it came out, he was trying to teach it to

1066
01:05:11.840 --> 01:05:14.840
<v Speaker 2>be civil and I was like, boy, I don't think

1067
01:05:14.840 --> 01:05:15.480
<v Speaker 2>that's gonna work.

1068
01:05:16.400 --> 01:05:16.599
<v Speaker 4>Yeah.

1069
01:05:16.800 --> 01:05:18.480
<v Speaker 2>Do manners matter?

1070
01:05:18.639 --> 01:05:18.800
<v Speaker 7>Yeah?

1071
01:05:18.880 --> 01:05:22.719
<v Speaker 4>Yeah, yeah yeah. So for models like chat chipet, they

1072
01:05:22.800 --> 01:05:25.280
<v Speaker 4>have what they call a knowledge cut off dates, so

1073
01:05:25.599 --> 01:05:29.159
<v Speaker 4>the training data, the you know, your interaction won't really

1074
01:05:29.239 --> 01:05:32.519
<v Speaker 4>input into into the learning system of the model. It's

1075
01:05:32.559 --> 01:05:35.079
<v Speaker 4>the training data set for I think Chad Chipet that

1076
01:05:35.280 --> 01:05:37.800
<v Speaker 4>ends I think around twenty twenty one or twenty twenty two.

1077
01:05:37.840 --> 01:05:40.880
<v Speaker 4>So Jadechipt, for example, can't give you a cohorrent answer

1078
01:05:40.960 --> 01:05:42.960
<v Speaker 4>for any event that has happened recently.

1079
01:05:43.440 --> 01:05:46.800
<v Speaker 2>So different models are using data from different timelines. They

1080
01:05:46.840 --> 01:05:49.239
<v Speaker 2>have to collect it and clean it and process it first,

1081
01:05:49.239 --> 01:05:51.880
<v Speaker 2>so it's not as real time as I thought it was,

1082
01:05:52.000 --> 01:05:53.320
<v Speaker 2>or some people might expect.

1083
01:05:53.599 --> 01:05:57.679
<v Speaker 4>For I for kind of when you speak to it,

1084
01:05:57.800 --> 01:05:59.599
<v Speaker 4>was it aggressively cratly?

1085
01:05:59.760 --> 01:06:03.280
<v Speaker 2>Yeah, threateningly? Why does it do better? Threaten it?

1086
01:06:03.400 --> 01:06:05.880
<v Speaker 4>So it's the first time I'm hearing this, so I

1087
01:06:05.920 --> 01:06:08.079
<v Speaker 4>should I should try it out. I should check it

1088
01:06:08.119 --> 01:06:12.280
<v Speaker 4>out and see if that also happens to me. But yeah,

1089
01:06:12.280 --> 01:06:13.920
<v Speaker 4>it's it's the first time I'm hearing it.

1090
01:06:14.239 --> 01:06:16.840
<v Speaker 2>Okay, let's hit the books for this. Specifically a twenty

1091
01:06:16.880 --> 01:06:20.639
<v Speaker 2>twenty four study about how to interact with large language models,

1092
01:06:20.639 --> 01:06:24.719
<v Speaker 2>and it's titled should We Respect lms Across Lingual Study

1093
01:06:24.760 --> 01:06:28.559
<v Speaker 2>on the Influence of prompt Politeness on LM performance. So

1094
01:06:28.639 --> 01:06:33.000
<v Speaker 2>this abstract explains that they did testing in English, Japanese,

1095
01:06:33.079 --> 01:06:37.880
<v Speaker 2>and Chinese language models and that as the politeness level descends,

1096
01:06:38.239 --> 01:06:43.239
<v Speaker 2>the answers generated get shorter. However, on the far side

1097
01:06:43.400 --> 01:06:47.719
<v Speaker 2>of the rude scale, impolite prompts often results in poor performance,

1098
01:06:48.000 --> 01:06:52.280
<v Speaker 2>but overly polite language does not guarantee better outcomes, and

1099
01:06:52.320 --> 01:06:55.840
<v Speaker 2>the best politeness level is different according to the language.

1100
01:06:56.119 --> 01:06:59.039
<v Speaker 2>And they say that this suggests that lllms not only

1101
01:06:59.159 --> 01:07:02.519
<v Speaker 2>reflect human being behavior, but are also influenced by language,

1102
01:07:02.559 --> 01:07:06.239
<v Speaker 2>particularly in different cultural contexts. So what about the future.

1103
01:07:06.519 --> 01:07:10.079
<v Speaker 2>The researchers say that it is conjectured that GBT four,

1104
01:07:10.280 --> 01:07:14.639
<v Speaker 2>being a superior model, might prioritize the task itself and

1105
01:07:14.800 --> 01:07:19.000
<v Speaker 2>effectively control its tendency to quote argue at a low

1106
01:07:19.079 --> 01:07:23.559
<v Speaker 2>politeness level, so as it matures, it just won't engage.

1107
01:07:23.639 --> 01:07:27.239
<v Speaker 2>It's like this AI new generation has been the therapy

1108
01:07:27.440 --> 01:07:29.119
<v Speaker 2>if it were a person, which it's not.

1109
01:07:29.519 --> 01:07:32.920
<v Speaker 4>And as to the AI overlords, again, I mean, it's

1110
01:07:33.000 --> 01:07:37.519
<v Speaker 4>just models. It's just you know, data sets in algorithms,

1111
01:07:37.639 --> 01:07:41.840
<v Speaker 4>and you know connection of networks. Okay, there is no

1112
01:07:42.400 --> 01:07:47.639
<v Speaker 4>kind of all knowing godlike or seeing AI. But of

1113
01:07:47.679 --> 01:07:50.440
<v Speaker 4>course you know, the people that are running AI companies

1114
01:07:50.880 --> 01:07:53.679
<v Speaker 4>come close to that because they have access to the data,

1115
01:07:53.760 --> 01:07:56.880
<v Speaker 4>because they have access to the algorithms, so you might

1116
01:07:56.920 --> 01:08:00.239
<v Speaker 4>worry about those people. You know, using your data has

1117
01:08:00.320 --> 01:08:07.599
<v Speaker 4>like almost invisible but very nuanced downstream impact. So in

1118
01:08:07.639 --> 01:08:12.000
<v Speaker 4>the US, for example, authorities as you are forcing companies

1119
01:08:12.079 --> 01:08:15.599
<v Speaker 4>like Meta to give up data so that authorities are

1120
01:08:15.679 --> 01:08:19.600
<v Speaker 4>hunting down women that had abortions. For example, in areas

1121
01:08:19.640 --> 01:08:24.279
<v Speaker 4>where abortion is prohibited. Law enforcement is working with Amazon

1122
01:08:24.359 --> 01:08:28.720
<v Speaker 4>for example, for was it Amazon bail ring bail ring?

1123
01:08:29.279 --> 01:08:33.520
<v Speaker 2>Oh right, this is that camera enabled an Amazon owned

1124
01:08:33.800 --> 01:08:34.520
<v Speaker 2>ring door bell.

1125
01:08:34.960 --> 01:08:41.079
<v Speaker 4>So law enforcement like ice uses that kind of data

1126
01:08:41.119 --> 01:08:44.720
<v Speaker 4>from those two to do something to even deport people.

1127
01:08:44.840 --> 01:08:48.319
<v Speaker 4>So what we should worry about is not really AI overlords,

1128
01:08:48.359 --> 01:08:52.079
<v Speaker 4>but this company is working with powerful entities to really

1129
01:08:52.760 --> 01:08:56.079
<v Speaker 4>kind of identify people that might be in trouble with

1130
01:08:56.239 --> 01:08:59.359
<v Speaker 4>the law or that might be doing you know that

1131
01:08:59.359 --> 01:09:02.880
<v Speaker 4>that Violet the know because that data gives them access,

1132
01:09:02.960 --> 01:09:06.960
<v Speaker 4>gives them the knowledge about the whereabouts in the interactions

1133
01:09:07.000 --> 01:09:08.279
<v Speaker 4>and the activities of people.

1134
01:09:08.560 --> 01:09:12.039
<v Speaker 2>Yeah. And previously, according to a twenty twenty Newsweek article

1135
01:09:12.239 --> 01:09:16.199
<v Speaker 2>titled police are monitoring Black Lives Matter protests with ring

1136
01:09:16.279 --> 01:09:20.079
<v Speaker 2>doorbell data and drones, activists say it's reported that Amazon

1137
01:09:20.159 --> 01:09:23.399
<v Speaker 2>Ring has video sharing partnerships with more than thirteen hundred

1138
01:09:23.680 --> 01:09:28.199
<v Speaker 2>law enforcement agencies across the US. However, in January twenty

1139
01:09:28.239 --> 01:09:31.239
<v Speaker 2>twenty four, Ring said that it would stop letting police

1140
01:09:31.239 --> 01:09:35.079
<v Speaker 2>departments request and receive user's footage on its app. Now,

1141
01:09:35.119 --> 01:09:38.960
<v Speaker 2>on the flip side, some Ring doorbell owners are posting

1142
01:09:39.079 --> 01:09:42.079
<v Speaker 2>on the Ring Neighbors app when ice rates are going

1143
01:09:42.119 --> 01:09:44.840
<v Speaker 2>down locally and they're alerting their community. Now, Ring of

1144
01:09:44.880 --> 01:09:47.600
<v Speaker 2>course notes that those are user generated posts and has

1145
01:09:47.600 --> 01:09:50.560
<v Speaker 2>nothing to do with them. Whether or not they'll censor

1146
01:09:50.600 --> 01:09:54.680
<v Speaker 2>those user generated posts is like anyone's guess, Hey, let's

1147
01:09:54.760 --> 01:09:58.000
<v Speaker 2>take a welcome departure from reality for sex? Shall we

1148
01:09:58.199 --> 01:09:59.359
<v Speaker 2>do any movies get it right?

1149
01:09:59.439 --> 01:09:59.520
<v Speaker 7>Like?

1150
01:09:59.560 --> 01:10:00.920
<v Speaker 2>Does it make get it right?

1151
01:10:01.119 --> 01:10:01.359
<v Speaker 4>Any?

1152
01:10:01.479 --> 01:10:04.880
<v Speaker 2>Does AI? That old Spielberg movie? Does anyone actually get

1153
01:10:04.920 --> 01:10:07.640
<v Speaker 2>it AI right? Or does that drive you absolutely insane

1154
01:10:07.640 --> 01:10:08.199
<v Speaker 2>to watch TV?

1155
01:10:08.640 --> 01:10:11.600
<v Speaker 4>So I love science fiction actually, like The Matrix is

1156
01:10:11.640 --> 01:10:12.800
<v Speaker 4>one of my favorite movies.

1157
01:10:13.159 --> 01:10:14.079
<v Speaker 2>I knew it, I knew it.

1158
01:10:14.159 --> 01:10:17.680
<v Speaker 4>It's a good one. But also that's like, that's nowhere

1159
01:10:17.800 --> 01:10:20.239
<v Speaker 4>you have to treat science fiction as science fiction for

1160
01:10:20.279 --> 01:10:23.560
<v Speaker 4>the sake of like it's really some really good science

1161
01:10:23.640 --> 01:10:28.600
<v Speaker 4>fiction really brings you into a world that you couldn't

1162
01:10:28.600 --> 01:10:31.680
<v Speaker 4>even envision. So I love that element about science fiction.

1163
01:10:32.159 --> 01:10:35.680
<v Speaker 4>But a lot of these like Robot Uprising, Terminator like

1164
01:10:36.800 --> 01:10:40.720
<v Speaker 4>movies are really just for entertainment. There is nothing that

1165
01:10:41.119 --> 01:10:43.800
<v Speaker 4>can be extrapolated and said, oh, this could happen to

1166
01:10:43.840 --> 01:10:47.880
<v Speaker 4>really AI. But you have kind of very nuanced sci

1167
01:10:47.880 --> 01:10:52.399
<v Speaker 4>fi movies. That's nail it. So you have Continuum. It's

1168
01:10:52.439 --> 01:10:55.119
<v Speaker 4>not a movie, it's a series. It wasn't on Netflix

1169
01:10:55.159 --> 01:10:55.880
<v Speaker 4>a wild but a.

1170
01:10:55.800 --> 01:10:59.680
<v Speaker 2>Few years side for what for what you're going to do?

1171
01:11:00.439 --> 01:11:03.880
<v Speaker 4>So what happens here is that you know, as AI

1172
01:11:03.960 --> 01:11:08.079
<v Speaker 4>companies become powerful, they take over government and they become

1173
01:11:09.000 --> 01:11:13.199
<v Speaker 4>you know, the bodies that really govern society. So that

1174
01:11:13.279 --> 01:11:16.920
<v Speaker 4>kind of sci fi is very close to reality than

1175
01:11:17.039 --> 01:11:18.840
<v Speaker 4>you know, Terminator like movies. Yeah.

1176
01:11:19.079 --> 01:11:20.039
<v Speaker 2>How about Black Mirror?

1177
01:11:20.600 --> 01:11:21.800
<v Speaker 4>Oh black Mirror.

1178
01:11:21.520 --> 01:11:22.000
<v Speaker 2>Is so good?

1179
01:11:22.119 --> 01:11:26.239
<v Speaker 4>I mean Black Mirror. There are some things that are like, eh, Now,

1180
01:11:26.439 --> 01:11:29.199
<v Speaker 4>when Black Mirror came out, it was just like wow,

1181
01:11:29.239 --> 01:11:31.039
<v Speaker 4>this could happen. And now I was like, oh that

1182
01:11:31.119 --> 01:11:34.960
<v Speaker 4>has happened, or it's like oh yeah this is you know,

1183
01:11:35.039 --> 01:11:37.199
<v Speaker 4>this is what's happening? Is this in that government?

1184
01:11:37.279 --> 01:11:41.640
<v Speaker 2>So yeah, wow, yep. And the last two questions I

1185
01:11:41.680 --> 01:11:45.159
<v Speaker 2>always ask her is worst and best. I guess your

1186
01:11:45.520 --> 01:11:48.319
<v Speaker 2>your most loathed thing about AI and your favorite thing

1187
01:11:48.560 --> 01:11:51.119
<v Speaker 2>about it. I guess we've talked a lot about cautionary

1188
01:11:51.199 --> 01:11:53.359
<v Speaker 2>but like in terms of what you do or in

1189
01:11:53.520 --> 01:11:55.640
<v Speaker 2>terms of your job, worst and best thing.

1190
01:11:56.079 --> 01:11:59.600
<v Speaker 4>Yeah, so the worst thing is really just the hype.

1191
01:12:00.159 --> 01:12:02.680
<v Speaker 4>As a researcher, I have my own research agen, but

1192
01:12:02.920 --> 01:12:06.079
<v Speaker 4>the hype is so destructive. You see something that's not

1193
01:12:06.199 --> 01:12:10.680
<v Speaker 4>true being disseminated, going viral, and you know, as an expert,

1194
01:12:10.760 --> 01:12:12.960
<v Speaker 4>it's really traveling, so you have to stop what you

1195
01:12:13.000 --> 01:12:15.439
<v Speaker 4>are doing and do some work to kind of correct

1196
01:12:15.479 --> 01:12:18.840
<v Speaker 4>it or at least that tempt. So yeah, a lot

1197
01:12:18.880 --> 01:12:24.319
<v Speaker 4>of the hype really is what really gets on minor

1198
01:12:24.520 --> 01:12:27.319
<v Speaker 4>and it becomes also a problem in terms of getting

1199
01:12:27.399 --> 01:12:30.479
<v Speaker 4>my own work done. But what excites me about AI

1200
01:12:30.760 --> 01:12:35.600
<v Speaker 4>is I'm still extremely optimistic about AI, But unfortunately a

1201
01:12:35.600 --> 01:12:38.920
<v Speaker 4>lot of the AI I get excited about is not

1202
01:12:39.119 --> 01:12:45.439
<v Speaker 4>something that results in you know, massive profits. So you know,

1203
01:12:45.600 --> 01:12:50.199
<v Speaker 4>using AI for disaster mapping, using AI for soil heads monitoring,

1204
01:12:50.239 --> 01:12:52.560
<v Speaker 4>and so on, these are things that really excite me,

1205
01:12:52.680 --> 01:12:56.079
<v Speaker 4>but there is no monetary value in developing AI for

1206
01:12:56.479 --> 01:12:58.720
<v Speaker 4>these system So these are the things that really get

1207
01:12:58.760 --> 01:13:02.119
<v Speaker 4>me excited, that really make me feel like, Wow, this

1208
01:13:02.239 --> 01:13:05.399
<v Speaker 4>is powerful tool that we can use to actually do

1209
01:13:05.520 --> 01:13:06.640
<v Speaker 4>some good in the world.

1210
01:13:06.800 --> 01:13:07.239
<v Speaker 1>Yeah.

1211
01:13:07.279 --> 01:13:09.479
<v Speaker 2>We could make sure that everyone is fed and has

1212
01:13:09.520 --> 01:13:12.760
<v Speaker 2>healthcare and that resources are allocated in a way that's fair.

1213
01:13:13.520 --> 01:13:15.960
<v Speaker 2>And we just don't because of money.

1214
01:13:16.159 --> 01:13:17.640
<v Speaker 4>Yeah, because it doesn't make you money.

1215
01:13:17.880 --> 01:13:20.520
<v Speaker 2>Yeah, which is I think once again, money is the

1216
01:13:20.560 --> 01:13:21.079
<v Speaker 2>root of allie.

1217
01:13:21.119 --> 01:13:23.159
<v Speaker 4>Yes, yeah, yeah, yeah.

1218
01:13:23.159 --> 01:13:24.800
<v Speaker 2>Thank you so much for doing this. This has been

1219
01:13:24.840 --> 01:13:27.399
<v Speaker 2>so illuminating and it's great to talk to someone who

1220
01:13:27.439 --> 01:13:28.479
<v Speaker 2>knows their shit about it.

1221
01:13:29.520 --> 01:13:31.880
<v Speaker 4>Thank you so much for having me. I really enjoyed

1222
01:13:31.880 --> 01:13:33.880
<v Speaker 4>our conversations.

1223
01:13:35.039 --> 01:13:39.560
<v Speaker 2>So ask real people, real not smart and important questions,

1224
01:13:39.560 --> 01:13:42.119
<v Speaker 2>because how else are we supposed to learn anything. So

1225
01:13:42.199 --> 01:13:44.720
<v Speaker 2>thank you so much to Trinity College's doctor Rohney for

1226
01:13:44.760 --> 01:13:47.359
<v Speaker 2>sitting down with me and making the trip to Ireland

1227
01:13:47.600 --> 01:13:50.520
<v Speaker 2>so eventful. I loved this talk and you can find

1228
01:13:50.560 --> 01:13:52.239
<v Speaker 2>links to her and her work in the show notes,

1229
01:13:52.279 --> 01:13:54.159
<v Speaker 2>as well as to the Cause of the week. We

1230
01:13:54.239 --> 01:13:58.600
<v Speaker 2>are at ologies on Meta owned Instagram and on Blue Sky,

1231
01:13:58.960 --> 01:14:02.079
<v Speaker 2>and I'm giving my data as ali Ward with just

1232
01:14:02.159 --> 01:14:04.720
<v Speaker 2>one l on both, and our website has links to

1233
01:14:04.720 --> 01:14:06.600
<v Speaker 2>all the studies we talked about and that link is

1234
01:14:06.640 --> 01:14:08.760
<v Speaker 2>in the show notes. If you're looking to become a patron,

1235
01:14:08.800 --> 01:14:10.600
<v Speaker 2>you can go to patreon dot com slash Ologies and

1236
01:14:10.640 --> 01:14:13.279
<v Speaker 2>you can join up there. If you need shorter, kid

1237
01:14:13.319 --> 01:14:16.359
<v Speaker 2>friendly versions of Ologies episodes, we have them for free

1238
01:14:16.399 --> 01:14:19.720
<v Speaker 2>in their own feed. Just look for Smologies that's also

1239
01:14:19.800 --> 01:14:21.720
<v Speaker 2>linked to the show notes. Please spread the word on that,

1240
01:14:22.119 --> 01:14:25.159
<v Speaker 2>and we have Ologies merch at ologiesmerch dot com. Thank

1241
01:14:25.159 --> 01:14:28.159
<v Speaker 2>you to Aaron Talbert for adminting the Ologies podcast Facebook group.

1242
01:14:28.239 --> 01:14:31.960
<v Speaker 2>Aveline Mallick does our professional human made transcripts. Kelly R.

1243
01:14:32.000 --> 01:14:34.680
<v Speaker 2>Dwyerd is the website. Noel Dilworth is our Flesh and

1244
01:14:34.720 --> 01:14:38.800
<v Speaker 2>Blood scheduling producer, Human Organism. Susan Hale managing directs the

1245
01:14:38.800 --> 01:14:42.119
<v Speaker 2>whole show. A live editor Jake Chafe helps put it together,

1246
01:14:42.479 --> 01:14:46.119
<v Speaker 2>and the Connective Tissue lead editor is Mercedes Maitland of

1247
01:14:46.159 --> 01:14:49.239
<v Speaker 2>Maitland Audio and Nick Thorbern made the theme music using

1248
01:14:49.279 --> 01:14:52.159
<v Speaker 2>his brain and ears and fingers. And if you stick

1249
01:14:52.159 --> 01:14:53.399
<v Speaker 2>around to the end of the show, I tell you

1250
01:14:53.399 --> 01:14:55.920
<v Speaker 2>a secret, and this week it's two. One is that

1251
01:14:55.960 --> 01:14:58.279
<v Speaker 2>I think I'm going to be shooting something next week

1252
01:14:58.319 --> 01:14:59.920
<v Speaker 2>and I want to tell patrons about it. For a

1253
01:15:00.119 --> 01:15:02.600
<v Speaker 2>but also i'll do some posting on social media if

1254
01:15:02.600 --> 01:15:05.039
<v Speaker 2>and when it happens, I'm really excited. I don't mean

1255
01:15:05.079 --> 01:15:07.840
<v Speaker 2>to be secretive, but just send good vibes. Next week

1256
01:15:07.880 --> 01:15:09.800
<v Speaker 2>i'll tell you as a secret effort happens. And the

1257
01:15:09.800 --> 01:15:12.800
<v Speaker 2>other secret is, before I went to Ireland, I got

1258
01:15:12.840 --> 01:15:16.119
<v Speaker 2>a couple of those film cameras disposed because it's like, ooh,

1259
01:15:16.159 --> 01:15:18.640
<v Speaker 2>what is this a film in this? And I took

1260
01:15:18.680 --> 01:15:21.640
<v Speaker 2>all the pictures and I haven't gotten them developed yet,

1261
01:15:21.720 --> 01:15:23.920
<v Speaker 2>and I kind of feel like the longer you wait

1262
01:15:23.960 --> 01:15:26.319
<v Speaker 2>to get them developed, the more you'll like them. And

1263
01:15:26.399 --> 01:15:28.920
<v Speaker 2>so I don't know what the appropriate amount of time

1264
01:15:29.039 --> 01:15:32.920
<v Speaker 2>is to forget about this disposable camera and then get

1265
01:15:32.920 --> 01:15:35.560
<v Speaker 2>it developed. If it should be like a couple more months,

1266
01:15:35.680 --> 01:15:37.800
<v Speaker 2>or if I should get it developed in a year. And

1267
01:15:37.880 --> 01:15:40.640
<v Speaker 2>so now I just have this disposable camera in my

1268
01:15:40.680 --> 01:15:45.039
<v Speaker 2>backpack and I don't know how long I should I

1269
01:15:45.079 --> 01:15:47.000
<v Speaker 2>also don't know where to get it developed, if I'm

1270
01:15:47.039 --> 01:15:49.199
<v Speaker 2>being honest. But anyway, if anyone has thoughts about that,

1271
01:15:49.239 --> 01:15:52.399
<v Speaker 2>feel free to advise me. That is a very analog

1272
01:15:52.520 --> 01:15:55.560
<v Speaker 2>update here for me, all right, fucking please do not

1273
01:15:55.920 --> 01:16:01.039
<v Speaker 2>use chat GPT to write papers, or illustrate anything important.

1274
01:16:01.479 --> 01:16:07.159
<v Speaker 2>Hire an illustrator if you can. Illustrators, writers, artists, musicians,

1275
01:16:07.439 --> 01:16:10.560
<v Speaker 2>please let them live. They are alive, okay, be good.

1276
01:16:10.439 --> 01:16:24.199
<v Speaker 7>For by pacodermatology, homology or doo zoology, lithology, technology, meteorology, pertology, anthology, seriology, elinology.

1277
01:16:30.000 --> 01:16:33.680
<v Speaker 2>We marveled at our own magnificence as we gave birth

1278
01:16:34.239 --> 01:16:35.039
<v Speaker 2>to a r