1
00:00:06,320 --> 00:00:11,499
[Music]

2
00:00:15,360 --> 00:00:19,359
hello everyone and welcome back to the

3
00:00:17,359 --> 00:00:21,920
kia ora theater in the linuxconf

4
00:00:19,359 --> 00:00:23,279
australia 2022. i'm really excited for

5
00:00:21,920 --> 00:00:25,359
this next talk and i hope that everyone

6
00:00:23,279 --> 00:00:28,240
uh enjoys it as much as i will

7
00:00:25,359 --> 00:00:30,240
uh jay rosenform is a melbourne-based ai

8
00:00:28,240 --> 00:00:32,079
artist and researcher working with 3d

9
00:00:30,240 --> 00:00:34,640
modelling artificial intelligence and

10
00:00:32,079 --> 00:00:36,640
extended reality technologies their work

11
00:00:34,640 --> 00:00:38,960
explores post-human and post-gender

12
00:00:36,640 --> 00:00:40,399
concepts using classical art combined

13
00:00:38,960 --> 00:00:41,760
with new media technologies and

14
00:00:40,399 --> 00:00:43,520
programming

15
00:00:41,760 --> 00:00:45,360
their presentation today is a dungeons

16
00:00:43,520 --> 00:00:47,760
and dragons themed talk about ai and

17
00:00:45,360 --> 00:00:49,840
ethics using incredible tools to defeat

18
00:00:47,760 --> 00:00:52,000
the beasts of bias and i understand

19
00:00:49,840 --> 00:00:54,000
slides will be available afterwards and

20
00:00:52,000 --> 00:00:55,920
a beta dnd module if you want to try it

21
00:00:54,000 --> 00:00:57,920
at your table which sounds incredible

22
00:00:55,920 --> 00:00:59,760
there's also a transcript linked in the

23
00:00:57,920 --> 00:01:02,079
chat and venue list if you are hard of

24
00:00:59,760 --> 00:01:05,119
hearing or have voluntary processing

25
00:01:02,079 --> 00:01:08,159
issues over to you

26
00:01:05,119 --> 00:01:08,159
thank you so much

27
00:01:08,799 --> 00:01:15,040
you've all heard the stories a community

28
00:01:11,760 --> 00:01:17,119
is being attacked by a monster

29
00:01:15,040 --> 00:01:19,439
night after night it erodes the

30
00:01:17,119 --> 00:01:22,479
foundations of the village

31
00:01:19,439 --> 00:01:24,880
they the town people are terrified it

32
00:01:22,479 --> 00:01:27,680
whispers in their ears and tries to turn

33
00:01:24,880 --> 00:01:29,360
their minds it seems to feed off their

34
00:01:27,680 --> 00:01:31,840
very souls

35
00:01:29,360 --> 00:01:33,520
night after night it returns

36
00:01:31,840 --> 00:01:35,360
there's no reprieve

37
00:01:33,520 --> 00:01:37,439
no surrender

38
00:01:35,360 --> 00:01:40,720
it undermines everything the good and

39
00:01:37,439 --> 00:01:43,280
lawful people of the town work for

40
00:01:40,720 --> 00:01:46,240
those who succumb to the whispers slowly

41
00:01:43,280 --> 00:01:47,759
lose their ideals and their dreams of a

42
00:01:46,240 --> 00:01:50,000
better tomorrow

43
00:01:47,759 --> 00:01:53,280
they lose their friends they lose

44
00:01:50,000 --> 00:01:55,759
enjoyment in their favorite things

45
00:01:53,280 --> 00:01:59,040
they're enslaved to it now and work for

46
00:01:55,759 --> 00:02:01,600
the monsters benefit not their own

47
00:01:59,040 --> 00:02:04,320
you come to the center of the village

48
00:02:01,600 --> 00:02:07,280
the epicenter of the devastation

49
00:02:04,320 --> 00:02:09,200
and behold an aberrant creature like you

50
00:02:07,280 --> 00:02:11,200
have never seen before

51
00:02:09,200 --> 00:02:13,760
i'm going to need you all to roll for

52
00:02:11,200 --> 00:02:13,760
initiative

53
00:02:14,000 --> 00:02:20,319
the monster is an aberrant pile of ooze

54
00:02:17,120 --> 00:02:23,120
atop a once golden automaton

55
00:02:20,319 --> 00:02:26,000
inside you see some tarnished metal

56
00:02:23,120 --> 00:02:26,800
glinting as the ooze shifts

57
00:02:26,000 --> 00:02:28,400
and

58
00:02:26,800 --> 00:02:32,800
moves around

59
00:02:28,400 --> 00:02:35,440
a creature that spews bile and slurs it

60
00:02:32,800 --> 00:02:38,959
interrupts you as you try to talk to it

61
00:02:35,440 --> 00:02:42,800
it's unceasing in its desire to squash

62
00:02:38,959 --> 00:02:45,120
those it deems lesser under its treads

63
00:02:42,800 --> 00:02:48,000
you put up a valiant battle

64
00:02:45,120 --> 00:02:51,599
sword slashing magic firing and arcane

65
00:02:48,000 --> 00:02:54,319
bursts the very trees come to your aid

66
00:02:51,599 --> 00:02:55,840
with each attack ooze separates from the

67
00:02:54,319 --> 00:02:58,480
automaton

68
00:02:55,840 --> 00:03:00,879
sometimes it joins the battle

69
00:02:58,480 --> 00:03:03,519
and turns into a creature

70
00:03:00,879 --> 00:03:06,640
sometimes the ooze evaporates and you

71
00:03:03,519 --> 00:03:08,959
see more of the metal showing

72
00:03:06,640 --> 00:03:11,360
just as you feel like you're starting to

73
00:03:08,959 --> 00:03:14,480
make some headway

74
00:03:11,360 --> 00:03:17,040
the against the black bilious eye core

75
00:03:14,480 --> 00:03:18,400
surrounding the machine it turns and

76
00:03:17,040 --> 00:03:19,840
flees

77
00:03:18,400 --> 00:03:22,560
as it leaves

78
00:03:19,840 --> 00:03:25,840
small bits of booze drop off

79
00:03:22,560 --> 00:03:28,000
and fly or crawl away

80
00:03:25,840 --> 00:03:29,599
turning your attention to the crowd

81
00:03:28,000 --> 00:03:31,360
milling around you now

82
00:03:29,599 --> 00:03:32,480
you sense that they're grateful for your

83
00:03:31,360 --> 00:03:34,799
aid

84
00:03:32,480 --> 00:03:37,760
but they tell you that others have tried

85
00:03:34,799 --> 00:03:41,680
and succeeded in chasing it off

86
00:03:37,760 --> 00:03:43,840
and still it returns again and again

87
00:03:41,680 --> 00:03:46,239
you know deep in your hearts that it

88
00:03:43,840 --> 00:03:47,599
will return unless the source is dealt

89
00:03:46,239 --> 00:03:50,799
with

90
00:03:47,599 --> 00:03:52,400
you need to find out where it comes from

91
00:03:50,799 --> 00:03:54,720
why it attacks

92
00:03:52,400 --> 00:03:58,080
and how to slay it

93
00:03:54,720 --> 00:03:58,080
once and for all

94
00:03:58,879 --> 00:04:03,920
welcome to the campaign against a uh

95
00:04:01,200 --> 00:04:06,959
against algorithmic bias in ai i'm your

96
00:04:03,920 --> 00:04:09,680
dm jay rosenborn and i will be talking

97
00:04:06,959 --> 00:04:11,680
today about two of my passions bias in

98
00:04:09,680 --> 00:04:14,080
ai and d d

99
00:04:11,680 --> 00:04:16,400
i'm a phd student in my final year at

100
00:04:14,080 --> 00:04:17,359
rmit exploring computer perceptions of

101
00:04:16,400 --> 00:04:19,440
gender

102
00:04:17,359 --> 00:04:21,519
and i research a great deal about ai

103
00:04:19,440 --> 00:04:23,280
bias and mitigation

104
00:04:21,519 --> 00:04:24,800
one of my major projects involves

105
00:04:23,280 --> 00:04:26,800
visibly

106
00:04:24,800 --> 00:04:29,440
de-biasing a generative adversarial

107
00:04:26,800 --> 00:04:31,680
network to show how bias affects image

108
00:04:29,440 --> 00:04:33,199
generation and how it is possible to

109
00:04:31,680 --> 00:04:35,600
alter the bias

110
00:04:33,199 --> 00:04:38,080
of a biased neural network

111
00:04:35,600 --> 00:04:40,720
this talk is a bit of a fun look at how

112
00:04:38,080 --> 00:04:43,680
you can seek out the issues in ai and

113
00:04:40,720 --> 00:04:46,560
view ai through a critical lens

114
00:04:43,680 --> 00:04:49,120
ai bias is a monster that doesn't stop

115
00:04:46,560 --> 00:04:51,680
it's constantly coming back

116
00:04:49,120 --> 00:04:54,800
some people allow it to run rampant and

117
00:04:51,680 --> 00:04:56,479
some work to fight and and subdue it

118
00:04:54,800 --> 00:04:58,880
i want to look at this monster and how

119
00:04:56,479 --> 00:05:01,360
we can break it down bit by bit

120
00:04:58,880 --> 00:05:03,919
to turn it from a horrifying aberration

121
00:05:01,360 --> 00:05:06,080
into a tiny nuisance

122
00:05:03,919 --> 00:05:08,400
where you can go to get the information

123
00:05:06,080 --> 00:05:11,039
you need to invite it and how to go

124
00:05:08,400 --> 00:05:13,840
equip yourselves against the battle that

125
00:05:11,039 --> 00:05:13,840
lies ahead of you

126
00:05:14,560 --> 00:05:20,560
your party is tasked with subduing the

127
00:05:17,680 --> 00:05:23,199
creature or slaying it

128
00:05:20,560 --> 00:05:25,680
asking around town you meet many who

129
00:05:23,199 --> 00:05:28,880
have been tainted by the whispers

130
00:05:25,680 --> 00:05:31,199
some to a greater or lesser degree

131
00:05:28,880 --> 00:05:33,680
the ones with little affect just seem to

132
00:05:31,199 --> 00:05:36,880
suffer from apathy

133
00:05:33,680 --> 00:05:38,880
what bother they tell you the monster is

134
00:05:36,880 --> 00:05:42,400
just going to return

135
00:05:38,880 --> 00:05:45,360
maybe we should just stop fighting

136
00:05:42,400 --> 00:05:47,440
others are almost hostile all lives

137
00:05:45,360 --> 00:05:49,759
matter they say

138
00:05:47,440 --> 00:05:51,680
or well actually

139
00:05:49,759 --> 00:05:54,080
they say

140
00:05:51,680 --> 00:05:55,680
you sigh and move on

141
00:05:54,080 --> 00:05:58,080
in a local tavern

142
00:05:55,680 --> 00:06:00,960
you sit over your drinks despondently

143
00:05:58,080 --> 00:06:03,759
wondering where to start when you hear a

144
00:06:00,960 --> 00:06:06,479
soft i think i can help you

145
00:06:03,759 --> 00:06:08,479
and you look up to see a traveling bird

146
00:06:06,479 --> 00:06:10,639
who tells you that he has been seeking

147
00:06:08,479 --> 00:06:12,639
adventurous to help with his terrifying

148
00:06:10,639 --> 00:06:14,240
quest

149
00:06:12,639 --> 00:06:17,039
the bird tells you how the creature

150
00:06:14,240 --> 00:06:18,840
fights that it has tendrils of those

151
00:06:17,039 --> 00:06:21,840
that learn and

152
00:06:18,840 --> 00:06:24,800
attack it tells you how he tells you how

153
00:06:21,840 --> 00:06:27,360
it attacks and works in whispers

154
00:06:24,800 --> 00:06:29,680
destroying the morale of those around it

155
00:06:27,360 --> 00:06:31,280
how it can subtly corrupt the people

156
00:06:29,680 --> 00:06:34,880
nearby

157
00:06:31,280 --> 00:06:36,639
it doesn't feed on meat or people or any

158
00:06:34,880 --> 00:06:40,080
known food source

159
00:06:36,639 --> 00:06:42,800
it feeds on the in hatred

160
00:06:40,080 --> 00:06:46,240
and the doubts of people allows them to

161
00:06:42,800 --> 00:06:48,080
grow and fester as it returns again and

162
00:06:46,240 --> 00:06:49,680
again

163
00:06:48,080 --> 00:06:52,160
despairing

164
00:06:49,680 --> 00:06:55,280
afraid of this monumental task in front

165
00:06:52,160 --> 00:06:57,680
of you you ask where can you even start

166
00:06:55,280 --> 00:07:00,160
with something like this

167
00:06:57,680 --> 00:07:02,800
can it even be done

168
00:07:00,160 --> 00:07:04,160
the bad size yes

169
00:07:02,800 --> 00:07:05,599
he believes

170
00:07:04,160 --> 00:07:07,919
it can be done

171
00:07:05,599 --> 00:07:09,599
it wasn't always like this

172
00:07:07,919 --> 00:07:11,120
there have been other villagers who have

173
00:07:09,599 --> 00:07:12,400
had different experiences with the

174
00:07:11,120 --> 00:07:14,720
creature

175
00:07:12,400 --> 00:07:16,560
it seems to be getting worse as it moves

176
00:07:14,720 --> 00:07:19,120
from town to town

177
00:07:16,560 --> 00:07:21,919
it needs to be confronted at the source

178
00:07:19,120 --> 00:07:25,360
in its lair to truly stop it

179
00:07:21,919 --> 00:07:27,280
your quest therefore is to find the lair

180
00:07:25,360 --> 00:07:28,319
and work together to defeat this

181
00:07:27,280 --> 00:07:30,960
creature

182
00:07:28,319 --> 00:07:33,280
he also knows of someone who can help

183
00:07:30,960 --> 00:07:35,039
who may be able to help you gain more

184
00:07:33,280 --> 00:07:37,440
information about the creature in a

185
00:07:35,039 --> 00:07:39,520
gleaming tower to the east

186
00:07:37,440 --> 00:07:42,160
he hints that other villagers along the

187
00:07:39,520 --> 00:07:45,840
way may have further clues additional

188
00:07:42,160 --> 00:07:45,840
information about the creature

189
00:07:45,919 --> 00:07:50,720
it appears to be a huge task to slay the

190
00:07:48,160 --> 00:07:53,120
monster of ai bias it has so many

191
00:07:50,720 --> 00:07:55,520
different permutations so many different

192
00:07:53,120 --> 00:07:57,759
ways to appear

193
00:07:55,520 --> 00:07:58,879
we hear about aspects of ai bias all the

194
00:07:57,759 --> 00:08:00,639
time

195
00:07:58,879 --> 00:08:03,360
all of the different things that can go

196
00:08:00,639 --> 00:08:05,120
wrong when an ai is not properly trained

197
00:08:03,360 --> 00:08:08,240
from the beginning to consider all of

198
00:08:05,120 --> 00:08:10,800
the wonderful variations of humanity

199
00:08:08,240 --> 00:08:13,160
in computer vision in particular we see

200
00:08:10,800 --> 00:08:16,560
examples of misgendering racist

201
00:08:13,160 --> 00:08:17,759
miscategorization homophobia transphobia

202
00:08:16,560 --> 00:08:19,840
sexism

203
00:08:17,759 --> 00:08:23,039
i constantly see examples of people

204
00:08:19,840 --> 00:08:25,440
using ai to decide who to police

205
00:08:23,039 --> 00:08:28,160
to classify gender to ensure that

206
00:08:25,440 --> 00:08:30,560
hateful content is shown in social media

207
00:08:28,160 --> 00:08:32,399
or to add misgendering and sexism to

208
00:08:30,560 --> 00:08:34,800
language translation

209
00:08:32,399 --> 00:08:35,599
this is a problem that extends offline

210
00:08:34,800 --> 00:08:38,080
too

211
00:08:35,599 --> 00:08:42,640
as many people are arrested on the word

212
00:08:38,080 --> 00:08:42,640
of an algorithm without due process

213
00:08:43,039 --> 00:08:47,920
bias is a particularly large problem in

214
00:08:45,440 --> 00:08:50,240
ai and computer science

215
00:08:47,920 --> 00:08:51,600
many of the researchers working in ai

216
00:08:50,240 --> 00:08:54,959
are men

217
00:08:51,600 --> 00:08:58,480
with 80 percent of ai professors and 72

218
00:08:54,959 --> 00:09:00,320
percent of presenters at ai conferences

219
00:08:58,480 --> 00:09:02,399
these numbers go on to influence the

220
00:09:00,320 --> 00:09:04,640
creation of data sets that are heavily

221
00:09:02,399 --> 00:09:07,600
male focused

222
00:09:04,640 --> 00:09:10,080
bias in ai leads to harmful algorithms

223
00:09:07,600 --> 00:09:12,240
that can be used for unethical practices

224
00:09:10,080 --> 00:09:14,320
that can actively harm women and gender

225
00:09:12,240 --> 00:09:16,880
and sexual minorities

226
00:09:14,320 --> 00:09:18,720
an ai was created to undress women using

227
00:09:16,880 --> 00:09:20,240
a pornography dataset to fill in the

228
00:09:18,720 --> 00:09:22,080
unseen areas

229
00:09:20,240 --> 00:09:24,480
while the initial algorithm was quickly

230
00:09:22,080 --> 00:09:26,800
removed by the creator after backlash

231
00:09:24,480 --> 00:09:29,600
from the ai community

232
00:09:26,800 --> 00:09:31,360
many clones appeared shortly after

233
00:09:29,600 --> 00:09:33,920
ai has been used to attempt to detect

234
00:09:31,360 --> 00:09:35,920
homosexuality with several algorithms

235
00:09:33,920 --> 00:09:38,560
created using data sets based on

236
00:09:35,920 --> 00:09:40,399
photographs from grinder and tinder in

237
00:09:38,560 --> 00:09:43,279
an attempt to explore a modern

238
00:09:40,399 --> 00:09:45,600
physiognomy of detection

239
00:09:43,279 --> 00:09:48,320
these algorithms exist and while they

240
00:09:45,600 --> 00:09:50,560
don't work properly they were created

241
00:09:48,320 --> 00:09:53,279
without any consideration for people who

242
00:09:50,560 --> 00:09:56,560
could be put at risk because of the bias

243
00:09:53,279 --> 00:09:59,120
of ais creators

244
00:09:56,560 --> 00:10:01,440
bias works insidiously to illustrate to

245
00:09:59,120 --> 00:10:02,720
people who are excluded that they don't

246
00:10:01,440 --> 00:10:05,120
matter

247
00:10:02,720 --> 00:10:07,519
that they are less important

248
00:10:05,120 --> 00:10:09,120
and these concepts prevent people from

249
00:10:07,519 --> 00:10:12,399
moving ahead

250
00:10:09,120 --> 00:10:15,200
from being hired from being seen

251
00:10:12,399 --> 00:10:18,399
and from seeing others like them

252
00:10:15,200 --> 00:10:20,000
as dr sarah lewis put it in her article

253
00:10:18,399 --> 00:10:22,800
on the racial bias built into

254
00:10:20,000 --> 00:10:25,440
photography you can't become what you

255
00:10:22,800 --> 00:10:27,519
can't accurately see

256
00:10:25,440 --> 00:10:30,079
if you cannot see others like yourself

257
00:10:27,519 --> 00:10:32,640
represented in power in art in

258
00:10:30,079 --> 00:10:34,959
literature and entertainment then you

259
00:10:32,640 --> 00:10:37,360
start believing almost that you don't

260
00:10:34,959 --> 00:10:37,360
exist

261
00:10:38,000 --> 00:10:42,320
but you don't matter

262
00:10:39,839 --> 00:10:44,880
bias can also propel the people who are

263
00:10:42,320 --> 00:10:46,560
represented to ignore and exclude those

264
00:10:44,880 --> 00:10:50,640
who aren't

265
00:10:46,560 --> 00:10:54,399
or people they don't view as the default

266
00:10:50,640 --> 00:10:56,640
bias leads to racism to fear and hatred

267
00:10:54,399 --> 00:10:59,680
and while we're in the nascent stages of

268
00:10:56,640 --> 00:11:04,160
teaching ai it is crucial that we teach

269
00:10:59,680 --> 00:11:04,160
it that all humans have equal value

270
00:11:05,279 --> 00:11:11,120
a neural network can even develop biases

271
00:11:08,079 --> 00:11:13,279
because of background data

272
00:11:11,120 --> 00:11:15,360
even if the background content is

273
00:11:13,279 --> 00:11:17,600
unlabeled researchers have found that

274
00:11:15,360 --> 00:11:19,680
children in the background of an image

275
00:11:17,600 --> 00:11:21,839
usually results in the foreground person

276
00:11:19,680 --> 00:11:23,680
being classified as a woman

277
00:11:21,839 --> 00:11:26,079
because women appear with children in

278
00:11:23,680 --> 00:11:27,760
the background more often even in a

279
00:11:26,079 --> 00:11:30,079
balanced data set

280
00:11:27,760 --> 00:11:33,120
this sort of background noise reinforces

281
00:11:30,079 --> 00:11:36,320
the very bias we're trying to eliminate

282
00:11:33,120 --> 00:11:38,560
we see entirely too often these issues

283
00:11:36,320 --> 00:11:41,120
being dismissed or fixed with a quick

284
00:11:38,560 --> 00:11:44,079
fix solution that doesn't really work to

285
00:11:41,120 --> 00:11:46,399
solve the underlying problem

286
00:11:44,079 --> 00:11:48,000
it seems insurmountable to get past

287
00:11:46,399 --> 00:11:50,399
these issues

288
00:11:48,000 --> 00:11:52,800
ai amplifies the worst biases of

289
00:11:50,399 --> 00:11:55,200
humanity and spews it forth

290
00:11:52,800 --> 00:11:56,720
which then encourages people to be worse

291
00:11:55,200 --> 00:12:00,399
because they see their inner beliefs

292
00:11:56,720 --> 00:12:01,920
manifested it becomes a vicious cycle

293
00:12:00,399 --> 00:12:05,040
but there is hope

294
00:12:01,920 --> 00:12:06,800
there are ways to de-bias ai

295
00:12:05,040 --> 00:12:10,639
to mitigate the problems caused by

296
00:12:06,800 --> 00:12:12,399
improper data sets and biased code

297
00:12:10,639 --> 00:12:14,399
it helps to know that there are people

298
00:12:12,399 --> 00:12:16,480
out there who know what they're doing

299
00:12:14,399 --> 00:12:18,639
there are experts in ai bias and there

300
00:12:16,480 --> 00:12:21,120
are resources to help you learn

301
00:12:18,639 --> 00:12:23,440
the algorithmic justice league can help

302
00:12:21,120 --> 00:12:26,480
audit ai development and there are

303
00:12:23,440 --> 00:12:28,160
independent ai ethics experts available

304
00:12:26,480 --> 00:12:29,519
for consultation

305
00:12:28,160 --> 00:12:31,760
just knowing that there are people and

306
00:12:29,519 --> 00:12:34,160
resources you can turn to starts to

307
00:12:31,760 --> 00:12:37,160
break the task down into more manageable

308
00:12:34,160 --> 00:12:37,160
pieces

309
00:12:38,079 --> 00:12:42,399
the path to the tower heads through

310
00:12:40,639 --> 00:12:44,880
several villages

311
00:12:42,399 --> 00:12:46,720
each showing signs of the automaton

312
00:12:44,880 --> 00:12:49,279
passing through

313
00:12:46,720 --> 00:12:50,839
each one seems to tell a story of the

314
00:12:49,279 --> 00:12:53,200
creature's

315
00:12:50,839 --> 00:12:55,360
regression from the villages that are

316
00:12:53,200 --> 00:12:57,760
mostly wreckage

317
00:12:55,360 --> 00:13:00,480
to recovering

318
00:12:57,760 --> 00:13:02,720
to barely damaged at all

319
00:13:00,480 --> 00:13:05,360
the villagers in each town tell you of

320
00:13:02,720 --> 00:13:08,000
the automaton as it moved through

321
00:13:05,360 --> 00:13:10,160
and working backwards you can see that

322
00:13:08,000 --> 00:13:12,320
it started off helpful

323
00:13:10,160 --> 00:13:13,519
but as it moved from one village to the

324
00:13:12,320 --> 00:13:16,399
next

325
00:13:13,519 --> 00:13:19,600
and met with more people something

326
00:13:16,399 --> 00:13:22,240
seemed to shift inside the creature

327
00:13:19,600 --> 00:13:24,560
as you make your way to the place

328
00:13:22,240 --> 00:13:27,200
where it all began you see the amounts

329
00:13:24,560 --> 00:13:30,000
of ooze all over everything taper off

330
00:13:27,200 --> 00:13:32,399
until there's no ooze to be seen

331
00:13:30,000 --> 00:13:34,959
see warped looking parrots

332
00:13:32,399 --> 00:13:37,839
occasionally as you track flying around

333
00:13:34,959 --> 00:13:40,079
and signs of tread moving away

334
00:13:37,839 --> 00:13:42,240
from the direction you're walking

335
00:13:40,079 --> 00:13:43,680
the verdant fields give way to

336
00:13:42,240 --> 00:13:46,240
grasslands

337
00:13:43,680 --> 00:13:47,680
and the grasslands give wage or rocky

338
00:13:46,240 --> 00:13:50,720
plain

339
00:13:47,680 --> 00:13:52,959
on the way the bard tells tales of how

340
00:13:50,720 --> 00:13:55,519
the creature was created

341
00:13:52,959 --> 00:13:57,120
he knows how to spin a yarn and before

342
00:13:55,519 --> 00:13:59,839
you know it you find yourself getting

343
00:13:57,120 --> 00:14:02,480
lost in the bard's words

344
00:13:59,839 --> 00:14:03,760
an artificer a wizard

345
00:14:02,480 --> 00:14:07,680
a bard

346
00:14:03,760 --> 00:14:09,839
all joined together to create something

347
00:14:07,680 --> 00:14:12,160
they wanted to forge something to help

348
00:14:09,839 --> 00:14:14,639
people all over the world

349
00:14:12,160 --> 00:14:17,199
wanting to bring the best of themselves

350
00:14:14,639 --> 00:14:19,600
into this creature they channeled their

351
00:14:17,199 --> 00:14:22,000
hearts and their souls

352
00:14:19,600 --> 00:14:25,600
as well as their minds and the sweat of

353
00:14:22,000 --> 00:14:28,160
their brows into this creation

354
00:14:25,600 --> 00:14:30,560
surrounding the automaton were three

355
00:14:28,160 --> 00:14:32,399
moats of potential

356
00:14:30,560 --> 00:14:35,199
glowing lights to help the creature

357
00:14:32,399 --> 00:14:36,639
become whatever it can be

358
00:14:35,199 --> 00:14:39,680
however

359
00:14:36,639 --> 00:14:41,839
the tiny seeds of corruption that lay in

360
00:14:39,680 --> 00:14:44,240
their each of their hearts that lay in

361
00:14:41,839 --> 00:14:47,120
all of our hearts

362
00:14:44,240 --> 00:14:49,279
unknown to us were drawn out

363
00:14:47,120 --> 00:14:51,920
with the motes of potential and lay

364
00:14:49,279 --> 00:14:54,320
dormant until fed

365
00:14:51,920 --> 00:14:56,079
they had the best of intentions as they

366
00:14:54,320 --> 00:14:57,440
set the creature off to go help the

367
00:14:56,079 --> 00:15:01,920
villagers

368
00:14:57,440 --> 00:15:05,680
but the more time it spent with people

369
00:15:01,920 --> 00:15:09,360
the more it changed as it was exposed to

370
00:15:05,680 --> 00:15:12,560
the worst in people that tiny seed of

371
00:15:09,360 --> 00:15:15,839
corruption grew

372
00:15:12,560 --> 00:15:17,760
it fed of the corruption of others

373
00:15:15,839 --> 00:15:19,680
and little by little

374
00:15:17,760 --> 00:15:24,240
their golden dream

375
00:15:19,680 --> 00:15:24,240
turned into an aberrant nightmare

376
00:15:24,399 --> 00:15:31,279
they watched as ooze started to creep up

377
00:15:28,399 --> 00:15:34,720
the sides of their creation

378
00:15:31,279 --> 00:15:36,480
it was subtle at first the change

379
00:15:34,720 --> 00:15:38,240
first it was just a little rude

380
00:15:36,480 --> 00:15:40,560
answering back

381
00:15:38,240 --> 00:15:43,040
but it was oddly casual about

382
00:15:40,560 --> 00:15:45,360
accidentally hurting people

383
00:15:43,040 --> 00:15:47,519
and then it began to shift to become

384
00:15:45,360 --> 00:15:50,160
something they could no longer control

385
00:15:47,519 --> 00:15:51,839
it started to injure villagers for no

386
00:15:50,160 --> 00:15:56,399
reason

387
00:15:51,839 --> 00:15:59,120
and then one day it was simply gone

388
00:15:56,399 --> 00:16:00,639
the three disagreed as to how to proceed

389
00:15:59,120 --> 00:16:03,120
and split up

390
00:16:00,639 --> 00:16:06,240
the wizard chose to stay at the tower to

391
00:16:03,120 --> 00:16:09,040
research a way to fix it

392
00:16:06,240 --> 00:16:12,240
the artificer went to the find the lair

393
00:16:09,040 --> 00:16:14,720
of the creature and the bard went forth

394
00:16:12,240 --> 00:16:17,360
to see what assistance could be granted

395
00:16:14,720 --> 00:16:20,639
any way to fix the devastation wrought

396
00:16:17,360 --> 00:16:23,120
in the neighboring towns and hopefully

397
00:16:20,639 --> 00:16:26,560
find a group of brave adventurers

398
00:16:23,120 --> 00:16:26,560
to tackle this challenge

399
00:16:27,920 --> 00:16:31,600
so many people start off their ai

400
00:16:30,160 --> 00:16:33,120
creations with the very best of

401
00:16:31,600 --> 00:16:35,440
intentions

402
00:16:33,120 --> 00:16:38,880
they consider what they can do without

403
00:16:35,440 --> 00:16:41,360
thinking about whether they should

404
00:16:38,880 --> 00:16:44,240
they want to benefit people but don't

405
00:16:41,360 --> 00:16:46,880
necessarily think about which people

406
00:16:44,240 --> 00:16:48,160
and what that impact will be

407
00:16:46,880 --> 00:16:50,959
sometimes they just want to make

408
00:16:48,160 --> 00:16:52,639
something to see if it can be done

409
00:16:50,959 --> 00:16:55,920
the problem is

410
00:16:52,639 --> 00:16:59,279
that we all have bias

411
00:16:55,920 --> 00:17:02,480
that is the tiny seed of corruption all

412
00:16:59,279 --> 00:17:04,480
our unknown unconscious biases

413
00:17:02,480 --> 00:17:05,600
just knowing that you have inbuilt

414
00:17:04,480 --> 00:17:07,919
biases

415
00:17:05,600 --> 00:17:10,000
will help you start to see them

416
00:17:07,919 --> 00:17:12,079
and learn to factor them in and work

417
00:17:10,000 --> 00:17:15,120
past them

418
00:17:12,079 --> 00:17:18,400
the issue with this monster of ai bias

419
00:17:15,120 --> 00:17:20,799
is that the creators are three white men

420
00:17:18,400 --> 00:17:22,319
one method they could have used early on

421
00:17:20,799 --> 00:17:25,360
is to ensure that they had a diverse

422
00:17:22,319 --> 00:17:28,079
team working to create the creature

423
00:17:25,360 --> 00:17:31,280
it isn't enough to want to help people

424
00:17:28,079 --> 00:17:33,039
you need the perspectives of many people

425
00:17:31,280 --> 00:17:34,960
in order to ensure that no one is

426
00:17:33,039 --> 00:17:37,039
overlooked

427
00:17:34,960 --> 00:17:39,360
ai is lazy

428
00:17:37,039 --> 00:17:42,640
it will only work with what it's taught

429
00:17:39,360 --> 00:17:44,240
so it's up to every ai creator to ensure

430
00:17:42,640 --> 00:17:46,480
that it is taught a wide range of

431
00:17:44,240 --> 00:17:48,640
options so that it treats everyone

432
00:17:46,480 --> 00:17:51,919
equally

433
00:17:48,640 --> 00:17:54,320
ai is like a child it needs to be raised

434
00:17:51,919 --> 00:17:56,559
with gentle and positive reinforcement

435
00:17:54,320 --> 00:17:59,120
and it needs to be exposed to a wide

436
00:17:56,559 --> 00:18:01,280
range of experiences and people so that

437
00:17:59,120 --> 00:18:04,960
it does not get confused when presented

438
00:18:01,280 --> 00:18:04,960
with something outside of its knowledge

439
00:18:05,520 --> 00:18:09,600
lack of diversity and development has

440
00:18:07,760 --> 00:18:12,400
led to issues like apple's facial

441
00:18:09,600 --> 00:18:15,280
recognition issues with asian people

442
00:18:12,400 --> 00:18:17,120
and uk ai based passport checker

443
00:18:15,280 --> 00:18:19,120
repeatedly telling black people that

444
00:18:17,120 --> 00:18:20,720
they need to that they have their mouths

445
00:18:19,120 --> 00:18:22,840
open

446
00:18:20,720 --> 00:18:25,600
it has led to google's image

447
00:18:22,840 --> 00:18:28,400
miscategorization and the development of

448
00:18:25,600 --> 00:18:30,000
the highly secretive company palantir

449
00:18:28,400 --> 00:18:32,000
who work in defense and counter

450
00:18:30,000 --> 00:18:34,160
terrorism

451
00:18:32,000 --> 00:18:35,520
ai is about using math to make

452
00:18:34,160 --> 00:18:37,440
predictions

453
00:18:35,520 --> 00:18:40,000
about using the data it is given to

454
00:18:37,440 --> 00:18:42,720
extrapolate new ideas

455
00:18:40,000 --> 00:18:45,120
this can be extremely powerful

456
00:18:42,720 --> 00:18:47,280
but ai has also been proven to focus on

457
00:18:45,120 --> 00:18:49,760
the wrong things

458
00:18:47,280 --> 00:18:52,320
we need diverse teams developing ai

459
00:18:49,760 --> 00:18:54,880
systems so that we can ensure that the

460
00:18:52,320 --> 00:18:57,200
moats of potential from everyone are

461
00:18:54,880 --> 00:19:00,160
included but that the seeds of

462
00:18:57,200 --> 00:19:02,000
corruption are minimized by contrasting

463
00:19:00,160 --> 00:19:04,080
viewpoints

464
00:19:02,000 --> 00:19:05,600
but what do you do if the ai is already

465
00:19:04,080 --> 00:19:07,679
biased

466
00:19:05,600 --> 00:19:10,799
how do you progress from there

467
00:19:07,679 --> 00:19:13,520
stay tuned

468
00:19:10,799 --> 00:19:14,880
as you see the how a tower looming on

469
00:19:13,520 --> 00:19:18,080
the horizon

470
00:19:14,880 --> 00:19:22,240
gleaming in the setting sun

471
00:19:18,080 --> 00:19:25,760
you hear a raucous cry from behind you

472
00:19:22,240 --> 00:19:27,679
and feel the impact of whispers on your

473
00:19:25,760 --> 00:19:29,280
mind

474
00:19:27,679 --> 00:19:32,720
the parrots

475
00:19:29,280 --> 00:19:34,400
who seem so few and far away have

476
00:19:32,720 --> 00:19:36,640
increased in number

477
00:19:34,400 --> 00:19:39,600
and they're milling around you

478
00:19:36,640 --> 00:19:41,280
entering your minds with doubts and

479
00:19:39,600 --> 00:19:43,280
questions

480
00:19:41,280 --> 00:19:45,360
their feathers look like individual

481
00:19:43,280 --> 00:19:48,080
droplets of ooze

482
00:19:45,360 --> 00:19:51,039
these are stochastic parrots

483
00:19:48,080 --> 00:19:52,880
filled with bile and corruption

484
00:19:51,039 --> 00:19:55,120
not always able to express themselves

485
00:19:52,880 --> 00:19:58,080
well

486
00:19:55,120 --> 00:20:00,400
the fight against parrots is taxing

487
00:19:58,080 --> 00:20:03,200
individually they're not strong but as a

488
00:20:00,400 --> 00:20:06,640
swarm they pick away at your defenses

489
00:20:03,200 --> 00:20:08,799
your hit points and your stamina

490
00:20:06,640 --> 00:20:12,159
they attack your mind with dissonant

491
00:20:08,799 --> 00:20:14,720
whispers designed to feed that tiny seed

492
00:20:12,159 --> 00:20:16,880
of corruption inside you

493
00:20:14,720 --> 00:20:19,440
but you are strong

494
00:20:16,880 --> 00:20:22,640
and eventually the stochastic parrots

495
00:20:19,440 --> 00:20:25,760
are defeated and as the sun dips below

496
00:20:22,640 --> 00:20:28,159
the horizon you set up camp for a long

497
00:20:25,760 --> 00:20:28,159
rest

498
00:20:28,880 --> 00:20:33,760
stochastic parrot is an ai term for very

499
00:20:31,520 --> 00:20:36,159
large language models for natural

500
00:20:33,760 --> 00:20:37,760
language processing systems

501
00:20:36,159 --> 00:20:39,840
these are often trained with a huge

502
00:20:37,760 --> 00:20:42,000
amount of information dragged from all

503
00:20:39,840 --> 00:20:44,880
over the internet

504
00:20:42,000 --> 00:20:46,559
and the reliability of the text is

505
00:20:44,880 --> 00:20:49,120
not great

506
00:20:46,559 --> 00:20:50,480
after all when your data set includes

507
00:20:49,120 --> 00:20:53,360
reddit

508
00:20:50,480 --> 00:20:56,480
how biased will your results be

509
00:20:53,360 --> 00:20:58,400
the peep the paper the timnit gabriel

510
00:20:56,480 --> 00:21:01,120
and later margaret mitchell were fired

511
00:20:58,400 --> 00:21:03,600
from google over discusses the problems

512
00:21:01,120 --> 00:21:05,600
of stochastic parrots at length and i

513
00:21:03,600 --> 00:21:08,000
felt it was appropriate to not only

514
00:21:05,600 --> 00:21:10,159
include these as low level encounters

515
00:21:08,000 --> 00:21:12,400
but also to generate an insult table

516
00:21:10,159 --> 00:21:14,000
using ai for some quality insults

517
00:21:12,400 --> 00:21:16,640
quality

518
00:21:14,000 --> 00:21:18,159
it was challenging to generate insults

519
00:21:16,640 --> 00:21:19,600
that were not biased towards any

520
00:21:18,159 --> 00:21:21,840
particular group of people and

521
00:21:19,600 --> 00:21:25,360
ironically i couldn't use the biggest

522
00:21:21,840 --> 00:21:27,600
stochastic parrot of them all gpt-3

523
00:21:25,360 --> 00:21:29,520
because it since its results were almost

524
00:21:27,600 --> 00:21:31,760
entirely bigoted

525
00:21:29,520 --> 00:21:34,320
i ended up using a smaller model insult

526
00:21:31,760 --> 00:21:34,320
generator

527
00:21:38,880 --> 00:21:43,440
excuse me

528
00:21:41,280 --> 00:21:45,679
stochastic parrots are an ongoing

529
00:21:43,440 --> 00:21:47,600
problem in ai because we frankly have

530
00:21:45,679 --> 00:21:51,280
become a little lazy

531
00:21:47,600 --> 00:21:54,000
in the method of data set creation

532
00:21:51,280 --> 00:21:55,919
we want huge data sets

533
00:21:54,000 --> 00:21:58,159
but are often unwilling to go to the

534
00:21:55,919 --> 00:22:00,720
trouble of creating every image or piece

535
00:21:58,159 --> 00:22:02,799
of training text within

536
00:22:00,720 --> 00:22:04,960
let alone working to ensure that every

537
00:22:02,799 --> 00:22:07,600
part is unbiased

538
00:22:04,960 --> 00:22:09,360
i'm sure that the original intent was to

539
00:22:07,600 --> 00:22:10,880
bring everything in and let the ai

540
00:22:09,360 --> 00:22:12,960
sorted out

541
00:22:10,880 --> 00:22:16,799
and hope for the best but you all know

542
00:22:12,960 --> 00:22:18,159
what happens when you put garbage in

543
00:22:16,799 --> 00:22:20,799
we still have not lived down the

544
00:22:18,159 --> 00:22:23,919
creation of tay the microsoft ai on

545
00:22:20,799 --> 00:22:26,400
twitter that turned into a rabid nazi

546
00:22:23,919 --> 00:22:29,280
more recently our imagination has been

547
00:22:26,400 --> 00:22:32,320
captured by ask delphi a large language

548
00:22:29,280 --> 00:22:33,600
model which claims to make ethical

549
00:22:32,320 --> 00:22:36,320
judgments

550
00:22:33,600 --> 00:22:38,159
after being taken to task the creators

551
00:22:36,320 --> 00:22:40,799
have put a lot of disclaimers on the

552
00:22:38,159 --> 00:22:42,320
site acknowledging the bias

553
00:22:40,799 --> 00:22:44,480
lately people have been turning to

554
00:22:42,320 --> 00:22:46,559
smaller language models again and

555
00:22:44,480 --> 00:22:47,679
training on better chosen and curated

556
00:22:46,559 --> 00:22:49,679
data sets

557
00:22:47,679 --> 00:22:51,280
this is paying off

558
00:22:49,679 --> 00:22:53,600
it's more work

559
00:22:51,280 --> 00:22:58,240
it's always more work to do the right

560
00:22:53,600 --> 00:22:58,240
thing but the end result pays off

561
00:22:58,480 --> 00:23:03,360
another way of cutting corners is by

562
00:23:00,960 --> 00:23:04,960
using mass worker groups like amazon's

563
00:23:03,360 --> 00:23:06,960
mechanical turk

564
00:23:04,960 --> 00:23:09,840
to caption images

565
00:23:06,960 --> 00:23:12,720
in this way as with ai you have to be

566
00:23:09,840 --> 00:23:14,720
extremely careful with what you ask and

567
00:23:12,720 --> 00:23:16,159
you need to supervise the outcomes

568
00:23:14,720 --> 00:23:19,280
carefully

569
00:23:16,159 --> 00:23:21,840
a recent study by mit of popular public

570
00:23:19,280 --> 00:23:24,480
data sets showed label errors up to

571
00:23:21,840 --> 00:23:28,159
three point four percent of the time

572
00:23:24,480 --> 00:23:30,080
in imagenet that amounts to four

573
00:23:28,159 --> 00:23:31,919
four hundred and thousand

574
00:23:30,080 --> 00:23:35,280
four hundred and eighty two thousand

575
00:23:31,919 --> 00:23:37,919
seven hundred and two images excuse me

576
00:23:35,280 --> 00:23:40,159
some of these are harmless likely a

577
00:23:37,919 --> 00:23:42,480
chihuahua as a feather boa

578
00:23:40,159 --> 00:23:45,760
but it gets more insidious when it comes

579
00:23:42,480 --> 00:23:45,760
to images of humans

580
00:23:47,039 --> 00:23:52,559
when doing research into the ms data set

581
00:23:50,320 --> 00:23:55,120
i spoke to the creators about how the

582
00:23:52,559 --> 00:23:57,200
image descriptions were created

583
00:23:55,120 --> 00:23:59,279
they were written by m tech workers with

584
00:23:57,200 --> 00:24:01,279
a set of instructions

585
00:23:59,279 --> 00:24:03,919
these instructions however did not go

586
00:24:01,279 --> 00:24:06,159
far enough into determining what should

587
00:24:03,919 --> 00:24:07,760
and should not be included in the

588
00:24:06,159 --> 00:24:10,000
caption

589
00:24:07,760 --> 00:24:12,880
at no point does it really explain what

590
00:24:10,000 --> 00:24:15,200
the important parts are or how to

591
00:24:12,880 --> 00:24:17,760
determine them

592
00:24:15,200 --> 00:24:19,840
what is important for one person may not

593
00:24:17,760 --> 00:24:23,039
be important for another

594
00:24:19,840 --> 00:24:25,120
same with describing unimportant details

595
00:24:23,039 --> 00:24:27,360
i browsed the online data set looking

596
00:24:25,120 --> 00:24:29,919
for the above example to see the

597
00:24:27,360 --> 00:24:31,760
captions that were written

598
00:24:29,919 --> 00:24:33,919
and came across a number of captions

599
00:24:31,760 --> 00:24:36,720
that described women in photographs as

600
00:24:33,919 --> 00:24:39,600
beautiful or very sexy

601
00:24:36,720 --> 00:24:42,559
the only value judgment i saw of a man

602
00:24:39,600 --> 00:24:45,520
was described as very large

603
00:24:42,559 --> 00:24:48,320
i would posit that value judgments are

604
00:24:45,520 --> 00:24:49,679
not important to the captioning process

605
00:24:48,320 --> 00:24:51,120
as everyone has different views of

606
00:24:49,679 --> 00:24:52,799
attractiveness

607
00:24:51,120 --> 00:24:54,720
but that's also not critical to the

608
00:24:52,799 --> 00:24:56,799
story

609
00:24:54,720 --> 00:24:59,600
some further instruction would prevent

610
00:24:56,799 --> 00:25:01,840
that i also saw occasional examples

611
00:24:59,600 --> 00:25:04,240
where the person was merely a hand or

612
00:25:01,840 --> 00:25:06,880
the person was so small that they could

613
00:25:04,240 --> 00:25:09,039
not be identified and yet an attempt was

614
00:25:06,880 --> 00:25:10,720
made to gender them

615
00:25:09,039 --> 00:25:12,640
having people go through the caption

616
00:25:10,720 --> 00:25:14,720
results and remove those that didn't

617
00:25:12,640 --> 00:25:16,720
follow the instructions would have also

618
00:25:14,720 --> 00:25:19,039
been useful

619
00:25:16,720 --> 00:25:21,200
this sounds like a lot of extra work but

620
00:25:19,039 --> 00:25:23,279
in the end isn't it worth the extra work

621
00:25:21,200 --> 00:25:25,120
at the beginning to avoid the harder

622
00:25:23,279 --> 00:25:29,480
task of eventually having to fight a

623
00:25:25,120 --> 00:25:29,480
battle against a corrupted ai

624
00:25:30,159 --> 00:25:34,880
you finally reach the tower

625
00:25:32,559 --> 00:25:37,679
and start to climb

626
00:25:34,880 --> 00:25:40,320
as you make your way to the top

627
00:25:37,679 --> 00:25:42,480
you see workrooms storage rooms with

628
00:25:40,320 --> 00:25:44,640
powered down automatons

629
00:25:42,480 --> 00:25:47,120
of all kinds

630
00:25:44,640 --> 00:25:50,000
some look like they're made for fighting

631
00:25:47,120 --> 00:25:52,080
others for gardening or building

632
00:25:50,000 --> 00:25:54,240
they all seem to have a singular purpose

633
00:25:52,080 --> 00:25:56,320
built into them and they all seem to be

634
00:25:54,240 --> 00:25:58,240
completely inert

635
00:25:56,320 --> 00:26:00,880
some are intact and others appear to

636
00:25:58,240 --> 00:26:02,960
have been scavenged for parts

637
00:26:00,880 --> 00:26:04,400
there's a slightly grisly scene that

638
00:26:02,960 --> 00:26:07,919
repeats

639
00:26:04,400 --> 00:26:10,080
as on your journey to the top

640
00:26:07,919 --> 00:26:11,360
at the top of the tower you meet the

641
00:26:10,080 --> 00:26:14,400
wizard

642
00:26:11,360 --> 00:26:17,360
he's weedy with a golden band holding

643
00:26:14,400 --> 00:26:20,159
his long hair back from his face

644
00:26:17,360 --> 00:26:23,840
he wipes his brow and set down his book

645
00:26:20,159 --> 00:26:26,480
as you come in his face resigned

646
00:26:23,840 --> 00:26:31,600
knows why you're here

647
00:26:26,480 --> 00:26:34,080
as you explain the story his face falls

648
00:26:31,600 --> 00:26:36,240
he sighs at the end of your tail and

649
00:26:34,080 --> 00:26:39,520
sits down dejectedly

650
00:26:36,240 --> 00:26:41,520
i knew this would happen he said i told

651
00:26:39,520 --> 00:26:44,400
them it wasn't ready to go out in public

652
00:26:41,520 --> 00:26:46,400
yet this was all supposed to be

653
00:26:44,400 --> 00:26:48,799
a bit of an exploration into something

654
00:26:46,400 --> 00:26:51,120
that could help people

655
00:26:48,799 --> 00:26:52,400
we didn't know it would get so bad so

656
00:26:51,120 --> 00:26:54,880
fast

657
00:26:52,400 --> 00:26:56,240
he continues i hate the idea of taking

658
00:26:54,880 --> 00:26:59,440
it down

659
00:26:56,240 --> 00:27:01,600
but i see we have no choice

660
00:26:59,440 --> 00:27:04,080
we need to reset it at the very least

661
00:27:01,600 --> 00:27:07,200
but if you can find a way to remove the

662
00:27:04,080 --> 00:27:09,520
corruption without destroying it

663
00:27:07,200 --> 00:27:12,320
that would be best i think we can save

664
00:27:09,520 --> 00:27:16,720
it i think there's good in it yet and i

665
00:27:12,320 --> 00:27:19,039
think i have an idea how that might work

666
00:27:16,720 --> 00:27:22,960
he tells you the automaton is not to

667
00:27:19,039 --> 00:27:25,200
blame but the corruption surrounding it

668
00:27:22,960 --> 00:27:26,880
and if you take out the corruption

669
00:27:25,200 --> 00:27:29,200
you may have a chance to save the

670
00:27:26,880 --> 00:27:31,200
machine itself

671
00:27:29,200 --> 00:27:32,960
if you can bring it back to the tower he

672
00:27:31,200 --> 00:27:34,640
and the others can help reverse their

673
00:27:32,960 --> 00:27:36,720
mistakes

674
00:27:34,640 --> 00:27:39,120
he tells you that the corruption like

675
00:27:36,720 --> 00:27:42,000
the machine has the power to learn

676
00:27:39,120 --> 00:27:44,240
to grow and shift and that adventurers

677
00:27:42,000 --> 00:27:47,600
such as yourselves need to be especially

678
00:27:44,240 --> 00:27:50,000
wary of its need to adapt and to change

679
00:27:47,600 --> 00:27:52,240
up your attacks and skills to keep it

680
00:27:50,000 --> 00:27:54,559
guessing

681
00:27:52,240 --> 00:27:55,840
he suggests magical weapons may be of

682
00:27:54,559 --> 00:27:58,480
more use

683
00:27:55,840 --> 00:28:00,320
and has some for sale

684
00:27:58,480 --> 00:28:02,880
if you ask about the creatures that

685
00:28:00,320 --> 00:28:05,440
attacked you on your journey he tells

686
00:28:02,880 --> 00:28:07,840
you that sometimes ooh separates from

687
00:28:05,440 --> 00:28:10,880
the aberration and becomes its own

688
00:28:07,840 --> 00:28:12,960
creature its own monster

689
00:28:10,880 --> 00:28:14,880
the stochastic parrots learn by

690
00:28:12,960 --> 00:28:16,720
listening to conversations

691
00:28:14,880 --> 00:28:19,919
and repeating the words they've heard

692
00:28:16,720 --> 00:28:22,000
twisting them into bizarre insults

693
00:28:19,919 --> 00:28:23,919
there are other creatures too sloppy

694
00:28:22,000 --> 00:28:25,039
creatures with claws and a comical

695
00:28:23,919 --> 00:28:27,120
appearance

696
00:28:25,039 --> 00:28:28,880
they exist only to mock those who would

697
00:28:27,120 --> 00:28:30,720
destroy them

698
00:28:28,880 --> 00:28:32,720
while the individual monsters are not

699
00:28:30,720 --> 00:28:34,559
much of a challenge to kill

700
00:28:32,720 --> 00:28:37,840
there do seem to be more and more about

701
00:28:34,559 --> 00:28:38,960
of them about these days

702
00:28:37,840 --> 00:28:41,440
finally

703
00:28:38,960 --> 00:28:44,080
he tells you that he has a telescope and

704
00:28:41,440 --> 00:28:46,880
has seen a trail of ooze heading into

705
00:28:44,080 --> 00:28:49,200
and out of a swamp nearby

706
00:28:46,880 --> 00:28:52,320
he's seen signs that it may be coming

707
00:28:49,200 --> 00:28:55,919
and going from there and may have a lair

708
00:28:52,320 --> 00:28:55,919
somewhere in the swamp

709
00:28:56,960 --> 00:29:02,320
ai creators often consider their own

710
00:28:59,279 --> 00:29:04,080
needs and wants when making something

711
00:29:02,320 --> 00:29:06,080
they don't always consider the impact

712
00:29:04,080 --> 00:29:08,240
their creation will have or the way it

713
00:29:06,080 --> 00:29:10,480
will react

714
00:29:08,240 --> 00:29:12,399
or affect some users

715
00:29:10,480 --> 00:29:14,000
when thinking about whether you should

716
00:29:12,399 --> 00:29:17,520
make something

717
00:29:14,000 --> 00:29:20,480
ask yourself who does this help

718
00:29:17,520 --> 00:29:22,559
then ask yourselves who could this harm

719
00:29:20,480 --> 00:29:24,799
and look at your team

720
00:29:22,559 --> 00:29:27,039
if you don't see the types of people

721
00:29:24,799 --> 00:29:29,120
most likely to be affected then you need

722
00:29:27,039 --> 00:29:30,880
to broaden your team

723
00:29:29,120 --> 00:29:33,039
if they all look like you then you need

724
00:29:30,880 --> 00:29:34,480
to consider who you are making this for

725
00:29:33,039 --> 00:29:36,399
and why

726
00:29:34,480 --> 00:29:38,320
bring in consultants put the question

727
00:29:36,399 --> 00:29:40,960
out on twitter

728
00:29:38,320 --> 00:29:43,440
buyersnai most often comes down to two

729
00:29:40,960 --> 00:29:45,360
major areas the data set and the

730
00:29:43,440 --> 00:29:47,440
original concept

731
00:29:45,360 --> 00:29:48,960
what do you want to do and how do you

732
00:29:47,440 --> 00:29:51,520
want to do it

733
00:29:48,960 --> 00:29:53,360
it's very easy to get carried away with

734
00:29:51,520 --> 00:29:55,760
an idea and want to develop it straight

735
00:29:53,360 --> 00:29:57,919
away but you need to consider is this

736
00:29:55,760 --> 00:29:59,600
something that will help people

737
00:29:57,919 --> 00:30:01,840
is this something that could hurt people

738
00:29:59,600 --> 00:30:03,840
inadvertently

739
00:30:01,840 --> 00:30:06,399
fixing the problem at the beginning

740
00:30:03,840 --> 00:30:08,480
before it becomes a giant monster is

741
00:30:06,399 --> 00:30:11,360
much easier than at the end

742
00:30:08,480 --> 00:30:13,120
and sometimes the best idea is never to

743
00:30:11,360 --> 00:30:16,159
start at all

744
00:30:13,120 --> 00:30:18,880
in any form of ai based production the

745
00:30:16,159 --> 00:30:22,000
ethics of what you are making should be

746
00:30:18,880 --> 00:30:26,480
part of the initial risk assessment

747
00:30:22,000 --> 00:30:26,480
you do make risk assessments right

748
00:30:28,240 --> 00:30:32,480
you head towards the swamp

749
00:30:31,120 --> 00:30:36,640
as you enter

750
00:30:32,480 --> 00:30:39,919
doubt starts to build in all of you

751
00:30:36,640 --> 00:30:41,679
a sense of dread at the monumental size

752
00:30:39,919 --> 00:30:44,320
of this task

753
00:30:41,679 --> 00:30:47,120
you gathered information and you head

754
00:30:44,320 --> 00:30:50,880
into the dense undergrowth

755
00:30:47,120 --> 00:30:53,120
full of anticipation and concern

756
00:30:50,880 --> 00:30:55,520
what will you find

757
00:30:53,120 --> 00:30:57,279
as you hack and fight your way through

758
00:30:55,520 --> 00:31:00,559
the undergrowth

759
00:30:57,279 --> 00:31:03,360
it seems to grow darker and the light is

760
00:31:00,559 --> 00:31:06,159
swallowed up by the vegetation

761
00:31:03,360 --> 00:31:08,559
even the bird who's been merrily playing

762
00:31:06,159 --> 00:31:11,120
along up until now

763
00:31:08,559 --> 00:31:14,480
starts to change his tune

764
00:31:11,120 --> 00:31:16,640
to a more sonorous melody filled with

765
00:31:14,480 --> 00:31:18,640
dread

766
00:31:16,640 --> 00:31:20,159
you don't even know if he's conscious of

767
00:31:18,640 --> 00:31:22,720
it as he doesn't

768
00:31:20,159 --> 00:31:24,720
as you venture further into the darkness

769
00:31:22,720 --> 00:31:30,080
your hearts sink

770
00:31:24,720 --> 00:31:32,000
more and more with each passing step

771
00:31:30,080 --> 00:31:33,600
just as it feels you can't go any

772
00:31:32,000 --> 00:31:35,600
further

773
00:31:33,600 --> 00:31:39,120
a little glowing moat

774
00:31:35,600 --> 00:31:41,919
a tiny golden music note appears in

775
00:31:39,120 --> 00:31:44,480
front of you and starts swirling around

776
00:31:41,919 --> 00:31:48,640
the bird

777
00:31:44,480 --> 00:31:51,519
that tiny light is like a ray of hope

778
00:31:48,640 --> 00:31:54,320
and your legs feel lighter your mind's

779
00:31:51,519 --> 00:31:57,360
clearer as you get deeper into the heart

780
00:31:54,320 --> 00:32:00,240
of the jungle another moat appears a

781
00:31:57,360 --> 00:32:01,840
shining blue quill that seems to beckon

782
00:32:00,240 --> 00:32:04,399
on you onwards

783
00:32:01,840 --> 00:32:04,399
excuse me

784
00:32:04,799 --> 00:32:10,080
you emerge into a clearing with a

785
00:32:06,799 --> 00:32:11,919
metallic structure built into the trees

786
00:32:10,080 --> 00:32:14,559
and a small hillock

787
00:32:11,919 --> 00:32:16,640
in front of it is a glowing moat shaped

788
00:32:14,559 --> 00:32:20,559
like a green spanner

789
00:32:16,640 --> 00:32:23,279
and all three start circling the bird

790
00:32:20,559 --> 00:32:25,279
as you face the metallic structure and

791
00:32:23,279 --> 00:32:27,679
square your shoulders

792
00:32:25,279 --> 00:32:30,240
a wild-eyed man bursts from the

793
00:32:27,679 --> 00:32:32,799
undergrowth he shakes his spanner at the

794
00:32:30,240 --> 00:32:34,480
bard and says you

795
00:32:32,799 --> 00:32:36,720
but whatever he was going to say after

796
00:32:34,480 --> 00:32:40,480
that is immaterial as the automaton

797
00:32:36,720 --> 00:32:40,480
emerges from its lair and attacks

798
00:32:41,120 --> 00:32:44,480
where before you felt completely

799
00:32:42,880 --> 00:32:46,159
steamrolled

800
00:32:44,480 --> 00:32:47,279
now you understand more about the

801
00:32:46,159 --> 00:32:48,880
machine

802
00:32:47,279 --> 00:32:52,159
you're ready

803
00:32:48,880 --> 00:32:55,039
you know it learns and adapts

804
00:32:52,159 --> 00:32:57,440
you know the ooze is to blame you bring

805
00:32:55,039 --> 00:33:00,080
all of your attacks to bear

806
00:32:57,440 --> 00:33:01,840
little oozes drop off it and join the

807
00:33:00,080 --> 00:33:04,880
attack but now

808
00:33:01,840 --> 00:33:07,600
you notice that some return to it and it

809
00:33:04,880 --> 00:33:09,679
seems to lose some they'll look loose it

810
00:33:07,600 --> 00:33:11,919
seems to learn something about who the

811
00:33:09,679 --> 00:33:14,559
ooze was attacking

812
00:33:11,919 --> 00:33:16,960
at the beginning it almost seems weak

813
00:33:14,559 --> 00:33:18,880
and you're amazed at the ease with which

814
00:33:16,960 --> 00:33:21,039
you are fighting it back

815
00:33:18,880 --> 00:33:23,519
but the more you attack the less your

816
00:33:21,039 --> 00:33:25,200
attacks affect it

817
00:33:23,519 --> 00:33:26,880
at the beginning the tendrils it was

818
00:33:25,200 --> 00:33:28,480
lashing out with were just doing

819
00:33:26,880 --> 00:33:31,200
piercing damage

820
00:33:28,480 --> 00:33:34,240
but one grows a club-like end after an

821
00:33:31,200 --> 00:33:37,679
ooze returns to it another crackles with

822
00:33:34,240 --> 00:33:40,960
lightning after a different ooze returns

823
00:33:37,679 --> 00:33:42,480
you start to grasp how it's learning

824
00:33:40,960 --> 00:33:45,919
making headway

825
00:33:42,480 --> 00:33:48,720
finally a sorcerer lets out a fireball

826
00:33:45,919 --> 00:33:50,159
and some of the ooze grows back

827
00:33:48,720 --> 00:33:52,880
over the

828
00:33:50,159 --> 00:33:55,039
body once more

829
00:33:52,880 --> 00:33:56,240
the artificer and the bard yell out no

830
00:33:55,039 --> 00:33:58,320
fire

831
00:33:56,240 --> 00:34:00,000
and the sorcerer looks abashed but

832
00:33:58,320 --> 00:34:02,159
shifts their attacks

833
00:34:00,000 --> 00:34:05,279
the bard and the attitude turn to each

834
00:34:02,159 --> 00:34:07,679
other limbed in the glow of the battle

835
00:34:05,279 --> 00:34:09,839
and smile at each other their difference

836
00:34:07,679 --> 00:34:10,960
is forgotten as they unite in a common

837
00:34:09,839 --> 00:34:13,200
goal

838
00:34:10,960 --> 00:34:15,679
everyone burns their spell slots

839
00:34:13,200 --> 00:34:17,599
changing their attacks again and again

840
00:34:15,679 --> 00:34:20,159
and eventually you see more oohs

841
00:34:17,599 --> 00:34:22,480
dripping away and dissolving instead of

842
00:34:20,159 --> 00:34:25,280
becoming little creatures

843
00:34:22,480 --> 00:34:27,839
little by little you find what works and

844
00:34:25,280 --> 00:34:29,520
the ooze clears until you can see more

845
00:34:27,839 --> 00:34:31,599
and more of the original golden

846
00:34:29,520 --> 00:34:32,839
automaton appearing

847
00:34:31,599 --> 00:34:34,560
you press your

848
00:34:32,839 --> 00:34:37,200
attacks

849
00:34:34,560 --> 00:34:38,399
and finally the last of the ooze

850
00:34:37,200 --> 00:34:41,200
vanishes

851
00:34:38,399 --> 00:34:41,200
from the machine

852
00:34:42,720 --> 00:34:46,399
defeating bias in ai requires

853
00:34:44,639 --> 00:34:48,879
multi-pronged attacks

854
00:34:46,399 --> 00:34:50,800
it isn't enough to just fix the data set

855
00:34:48,879 --> 00:34:53,200
it isn't enough to

856
00:34:50,800 --> 00:34:55,679
delete the offending categories

857
00:34:53,200 --> 00:34:58,079
you need to systematically consider

858
00:34:55,679 --> 00:34:59,760
where the issues are coming from and

859
00:34:58,079 --> 00:35:00,960
where it is learned

860
00:34:59,760 --> 00:35:03,520
and

861
00:35:00,960 --> 00:35:05,280
what it is learned and how

862
00:35:03,520 --> 00:35:08,000
you need to look at if you are tuning

863
00:35:05,280 --> 00:35:12,000
hyper parameters and see the results

864
00:35:08,000 --> 00:35:14,160
when you edit that tuning or turn it off

865
00:35:12,000 --> 00:35:16,800
you need to examine the weighting you've

866
00:35:14,160 --> 00:35:19,040
applied to different factors explore

867
00:35:16,800 --> 00:35:21,839
your sample weights and test to make

868
00:35:19,040 --> 00:35:24,720
sure they're being counted

869
00:35:21,839 --> 00:35:27,839
you can use them to add impact to some

870
00:35:24,720 --> 00:35:29,839
categories and downplay some others

871
00:35:27,839 --> 00:35:31,599
you need to try different things and

872
00:35:29,839 --> 00:35:33,440
objectively compare the different

873
00:35:31,599 --> 00:35:35,440
results

874
00:35:33,440 --> 00:35:37,599
you need to really look at your

875
00:35:35,440 --> 00:35:41,280
algorithms and make sure that every

876
00:35:37,599 --> 00:35:42,880
option is being considered and trained

877
00:35:41,280 --> 00:35:45,839
you need to take a deep look at your

878
00:35:42,880 --> 00:35:47,359
data sets and the representation within

879
00:35:45,839 --> 00:35:49,680
and you need to look at the confidence

880
00:35:47,359 --> 00:35:52,000
of your systems and the outputs to

881
00:35:49,680 --> 00:35:54,960
ensure that there's a balance

882
00:35:52,000 --> 00:35:55,839
you may even need to bias it in other

883
00:35:54,960 --> 00:35:59,839
ways

884
00:35:55,839 --> 00:35:59,839
to counter balance the original biases

885
00:36:00,000 --> 00:36:04,480
bias can come from the most innocuous of

886
00:36:02,160 --> 00:36:06,720
sources it can come from background and

887
00:36:04,480 --> 00:36:08,640
labeled data it can come from what you

888
00:36:06,720 --> 00:36:10,640
choose not to include just as much as

889
00:36:08,640 --> 00:36:12,800
what you choose to include

890
00:36:10,640 --> 00:36:14,400
it can come from just not understanding

891
00:36:12,800 --> 00:36:16,720
that there are many different parts at

892
00:36:14,400 --> 00:36:16,720
play

893
00:36:17,520 --> 00:36:22,240
my phd is specifically looking at

894
00:36:19,839 --> 00:36:24,560
computer perceptions of gender

895
00:36:22,240 --> 00:36:27,680
and as such i've lived and breathed ai

896
00:36:24,560 --> 00:36:29,599
bias for the better part of three years

897
00:36:27,680 --> 00:36:32,480
i've worked with deliberately biasing

898
00:36:29,599 --> 00:36:33,440
image generation systems and de-biasing

899
00:36:32,480 --> 00:36:36,400
them

900
00:36:33,440 --> 00:36:38,880
and while training to see the results

901
00:36:36,400 --> 00:36:41,680
i've worked with existing ai systems to

902
00:36:38,880 --> 00:36:43,680
show the gaps in large image data sets

903
00:36:41,680 --> 00:36:46,400
like ms coco and i've trained

904
00:36:43,680 --> 00:36:49,440
classifiers examining different levels

905
00:36:46,400 --> 00:36:52,560
of training and bias in the data set to

906
00:36:49,440 --> 00:36:55,200
achieve different artistic outcomes

907
00:36:52,560 --> 00:36:58,640
art is a wonderful tool

908
00:36:55,200 --> 00:37:01,200
to explore ai bias because it clearly

909
00:36:58,640 --> 00:37:04,079
illustrates the presence of bias and the

910
00:37:01,200 --> 00:37:06,640
need for greater diversity

911
00:37:04,079 --> 00:37:09,520
i added fire as a restoration device

912
00:37:06,640 --> 00:37:13,440
deliberately as a nod to ai trolls you

913
00:37:09,520 --> 00:37:15,200
cannot defeat aio with a flame war

914
00:37:13,440 --> 00:37:17,119
and while sometimes it may feel like

915
00:37:15,200 --> 00:37:18,240
nuking it with fire may be the only way

916
00:37:17,119 --> 00:37:20,079
out

917
00:37:18,240 --> 00:37:21,680
fighting fire with fire only makes the

918
00:37:20,079 --> 00:37:23,280
situation worse

919
00:37:21,680 --> 00:37:25,520
you need to be clever

920
00:37:23,280 --> 00:37:28,480
to adapt and think carefully about how

921
00:37:25,520 --> 00:37:30,880
the algorithm learns and how it adapts

922
00:37:28,480 --> 00:37:32,720
the information you give it

923
00:37:30,880 --> 00:37:33,760
ai doesn't always work in the way you

924
00:37:32,720 --> 00:37:36,480
think

925
00:37:33,760 --> 00:37:38,480
sometimes it focuses on the wrong idea

926
00:37:36,480 --> 00:37:43,359
like medical ai systems that learn to

927
00:37:38,480 --> 00:37:43,359
identify rulers or medical equipment

928
00:37:44,079 --> 00:37:49,760
the glowing marts of potential orbiting

929
00:37:46,800 --> 00:37:51,680
the bard fly around the party

930
00:37:49,760 --> 00:37:53,200
bolstering you all for the final

931
00:37:51,680 --> 00:37:55,920
challenge

932
00:37:53,200 --> 00:37:58,720
the artificer and the bird move towards

933
00:37:55,920 --> 00:38:00,880
the shining automaton and start speaking

934
00:37:58,720 --> 00:38:02,960
to it gently

935
00:38:00,880 --> 00:38:06,560
its eyes change

936
00:38:02,960 --> 00:38:10,079
and the moats start circling around it

937
00:38:06,560 --> 00:38:12,560
a moat of inspiration the music note

938
00:38:10,079 --> 00:38:15,920
a motive thought the quill

939
00:38:12,560 --> 00:38:18,320
and emote of creation the spanner

940
00:38:15,920 --> 00:38:21,839
all of them orbit the automaton and it

941
00:38:18,320 --> 00:38:23,920
watches them dreamily

942
00:38:21,839 --> 00:38:27,280
the bird starts to play

943
00:38:23,920 --> 00:38:28,800
and as he does a glowing moat erupts

944
00:38:27,280 --> 00:38:30,560
from your chest

945
00:38:28,800 --> 00:38:32,720
from all of you

946
00:38:30,560 --> 00:38:35,040
a little symbol of something close to

947
00:38:32,720 --> 00:38:37,200
your heart

948
00:38:35,040 --> 00:38:39,200
a motive potential that only you can

949
00:38:37,200 --> 00:38:41,839
provide

950
00:38:39,200 --> 00:38:42,960
you look around and each member of the

951
00:38:41,839 --> 00:38:45,200
party

952
00:38:42,960 --> 00:38:47,520
has a march that swirls around their

953
00:38:45,200 --> 00:38:49,040
head before floating over to the

954
00:38:47,520 --> 00:38:51,520
automaton

955
00:38:49,040 --> 00:38:54,240
the motes start to spin

956
00:38:51,520 --> 00:38:55,359
faster and faster becoming a halo of

957
00:38:54,240 --> 00:38:58,640
lights

958
00:38:55,359 --> 00:39:01,280
the colors merge into one bright light

959
00:38:58,640 --> 00:39:03,680
too bright to look at and then are

960
00:39:01,280 --> 00:39:06,640
abruptly gone

961
00:39:03,680 --> 00:39:10,320
and the echo of that light

962
00:39:06,640 --> 00:39:13,800
shining in the eyes of the now calm

963
00:39:10,320 --> 00:39:13,800
golden automaton

964
00:39:14,400 --> 00:39:19,760
we all bring something to the table we

965
00:39:17,440 --> 00:39:21,839
all have unique lived experiences and

966
00:39:19,760 --> 00:39:25,440
perspectives we can bring

967
00:39:21,839 --> 00:39:26,839
to help make ai systems ideas and data

968
00:39:25,440 --> 00:39:29,040
sets more

969
00:39:26,839 --> 00:39:30,320
well-rounded we all have a mode of

970
00:39:29,040 --> 00:39:32,400
potential

971
00:39:30,320 --> 00:39:34,720
that can help make a system better just

972
00:39:32,400 --> 00:39:37,280
by asking what if

973
00:39:34,720 --> 00:39:39,839
why how

974
00:39:37,280 --> 00:39:41,200
what can we do that will make a

975
00:39:39,839 --> 00:39:43,440
difference

976
00:39:41,200 --> 00:39:45,280
who does this ai help

977
00:39:43,440 --> 00:39:48,960
who does it harm

978
00:39:45,280 --> 00:39:51,200
how can we make this ai more equitable

979
00:39:48,960 --> 00:39:53,200
how do we prevent the mistakes of the

980
00:39:51,200 --> 00:39:56,240
past

981
00:39:53,200 --> 00:39:57,040
how do we reflect the society we want to

982
00:39:56,240 --> 00:40:00,400
be

983
00:39:57,040 --> 00:40:00,400
rather than the one we have

984
00:40:00,480 --> 00:40:03,440
in dnd

985
00:40:01,839 --> 00:40:05,200
there's a bard college called the

986
00:40:03,440 --> 00:40:07,119
college of creation

987
00:40:05,200 --> 00:40:09,359
when a board of the college of creation

988
00:40:07,119 --> 00:40:11,839
casts inspiration

989
00:40:09,359 --> 00:40:14,160
it creates a motive potential which can

990
00:40:11,839 --> 00:40:15,040
enhance abilities attack roles or saving

991
00:40:14,160 --> 00:40:17,520
throws

992
00:40:15,040 --> 00:40:18,800
in this case i have used it somewhat

993
00:40:17,520 --> 00:40:20,640
liberally

994
00:40:18,800 --> 00:40:23,760
as a literary device to show the

995
00:40:20,640 --> 00:40:26,240
potentiality of ai

996
00:40:23,760 --> 00:40:28,480
ai is a motive potential

997
00:40:26,240 --> 00:40:31,760
it can enhance our work it can be a

998
00:40:28,480 --> 00:40:34,160
collaborator or it can attack

999
00:40:31,760 --> 00:40:36,160
how it behaves is entirely up to the

1000
00:40:34,160 --> 00:40:39,040
person using it

1001
00:40:36,160 --> 00:40:42,240
we are responsible for the ai we create

1002
00:40:39,040 --> 00:40:44,240
and we are responsible for the damage it

1003
00:40:42,240 --> 00:40:46,640
inflicts on others

1004
00:40:44,240 --> 00:40:48,560
we need to be aware not just of its

1005
00:40:46,640 --> 00:40:50,079
potential to create

1006
00:40:48,560 --> 00:40:51,440
but also

1007
00:40:50,079 --> 00:40:54,319
of its

1008
00:40:51,440 --> 00:40:58,079
potential to protect and assist

1009
00:40:54,319 --> 00:40:58,079
and its potential to harm

1010
00:40:59,359 --> 00:41:03,520
in this talk we've seen what it looks

1011
00:41:01,520 --> 00:41:04,640
like when everything goes well

1012
00:41:03,520 --> 00:41:07,359
but

1013
00:41:04,640 --> 00:41:10,319
what if everything goes wrong

1014
00:41:07,359 --> 00:41:12,240
what if your party isn't strong enough

1015
00:41:10,319 --> 00:41:15,040
you can always disengage and come back

1016
00:41:12,240 --> 00:41:16,319
after a long rest to try again

1017
00:41:15,040 --> 00:41:18,800
you can buff yourselves and your

1018
00:41:16,319 --> 00:41:21,280
co-workers learn about the resources at

1019
00:41:18,800 --> 00:41:24,400
hand take a look at the code with fresh

1020
00:41:21,280 --> 00:41:27,520
eyes get someone else to look

1021
00:41:24,400 --> 00:41:30,000
you are a party after all this task is

1022
00:41:27,520 --> 00:41:31,119
not just on your shoulders but on you

1023
00:41:30,000 --> 00:41:34,560
all

1024
00:41:31,119 --> 00:41:36,960
this quest falls to you to all of you

1025
00:41:34,560 --> 00:41:39,520
as a group to find the bugs in the

1026
00:41:36,960 --> 00:41:40,720
system to look at the datasets and do

1027
00:41:39,520 --> 00:41:43,760
the work

1028
00:41:40,720 --> 00:41:46,720
to find the areas that are being missed

1029
00:41:43,760 --> 00:41:48,800
this is not a task for one person

1030
00:41:46,720 --> 00:41:50,400
but for a team

1031
00:41:48,800 --> 00:41:53,040
and if you find

1032
00:41:50,400 --> 00:41:55,280
that you're alone in this endeavor

1033
00:41:53,040 --> 00:41:57,920
if it's looking bleak there are always

1034
00:41:55,280 --> 00:41:59,599
resources online there are mercenaries

1035
00:41:57,920 --> 00:42:02,319
you can hire to help

1036
00:41:59,599 --> 00:42:05,119
there are elder gods like google to

1037
00:42:02,319 --> 00:42:07,440
consult capricious in nature

1038
00:42:05,119 --> 00:42:09,040
but available whenever you need them

1039
00:42:07,440 --> 00:42:10,720
there are organizations such as the

1040
00:42:09,040 --> 00:42:13,040
algorithmic justice league who can

1041
00:42:10,720 --> 00:42:15,200
furnish you with the information and

1042
00:42:13,040 --> 00:42:17,839
assist you in your quest

1043
00:42:15,200 --> 00:42:20,560
you are never alone

1044
00:42:17,839 --> 00:42:23,359
even when it feels dark there are always

1045
00:42:20,560 --> 00:42:26,560
others with shining lights ready to help

1046
00:42:23,359 --> 00:42:26,560
you in your quest

1047
00:42:27,599 --> 00:42:32,720
the automaton is wide-eyed

1048
00:42:30,640 --> 00:42:35,280
friendly looking even

1049
00:42:32,720 --> 00:42:36,800
the lights from all those moats shining

1050
00:42:35,280 --> 00:42:39,359
in its eyes

1051
00:42:36,800 --> 00:42:41,920
it turns to you placidly

1052
00:42:39,359 --> 00:42:44,560
waiting for your input

1053
00:42:41,920 --> 00:42:45,680
it's here to help you that's all it ever

1054
00:42:44,560 --> 00:42:48,720
wanted

1055
00:42:45,680 --> 00:42:50,560
was to help you in your tasks and make

1056
00:42:48,720 --> 00:42:54,720
things easier

1057
00:42:50,560 --> 00:42:58,560
it's what you do with it that counts

1058
00:42:54,720 --> 00:43:00,960
what are you going to do next

1059
00:42:58,560 --> 00:43:00,960
thank you

1060
00:43:02,480 --> 00:43:06,400
thank you so much jay that's lovely

1061
00:43:05,599 --> 00:43:08,160
uh

1062
00:43:06,400 --> 00:43:11,599
can i get a round of applause for jay in

1063
00:43:08,160 --> 00:43:14,000
the chat and vanilla if we can um we do

1064
00:43:11,599 --> 00:43:15,920
have a couple of questions uh which i

1065
00:43:14,000 --> 00:43:17,599
will go through i'm going to take this

1066
00:43:15,920 --> 00:43:18,960
one and paste it into the streaming chat

1067
00:43:17,599 --> 00:43:20,480
because it's quite long

1068
00:43:18,960 --> 00:43:22,240
i thought i'll read it out

1069
00:43:20,480 --> 00:43:24,160
uh how do you balance the needs to

1070
00:43:22,240 --> 00:43:25,839
collect data for a data set with the

1071
00:43:24,160 --> 00:43:27,200
privacy of the people whose data is

1072
00:43:25,839 --> 00:43:28,960
being collected

1073
00:43:27,200 --> 00:43:30,800
in general ai there may be third parties

1074
00:43:28,960 --> 00:43:32,400
from whom it is very difficult to get

1075
00:43:30,800 --> 00:43:34,160
permission for data collection for

1076
00:43:32,400 --> 00:43:35,839
example other road users train the ai

1077
00:43:34,160 --> 00:43:37,520
for self-driving but are not consulted

1078
00:43:35,839 --> 00:43:40,240
on whether their data is collected or

1079
00:43:37,520 --> 00:43:42,319
how it is used

1080
00:43:40,240 --> 00:43:44,720
see this is an excellent question it's a

1081
00:43:42,319 --> 00:43:46,960
multi-pronged question too

1082
00:43:44,720 --> 00:43:50,000
um i think that so

1083
00:43:46,960 --> 00:43:52,240
ethics i i guess being in a university

1084
00:43:50,000 --> 00:43:55,200
environment um if you need to do

1085
00:43:52,240 --> 00:44:00,880
anything involving other people you need

1086
00:43:55,200 --> 00:44:04,720
to um put in ethics approvals and that's

1087
00:44:00,880 --> 00:44:08,079
it's a um it's a complicated process uh

1088
00:44:04,720 --> 00:44:10,560
the ai ethics or any ethics approval at

1089
00:44:08,079 --> 00:44:13,359
university levels but

1090
00:44:10,560 --> 00:44:14,960
um i think it's actually an interesting

1091
00:44:13,359 --> 00:44:16,560
roadmap and while i don't think it needs

1092
00:44:14,960 --> 00:44:18,720
to be as complicated as it is at

1093
00:44:16,560 --> 00:44:21,520
university i think that everybody

1094
00:44:18,720 --> 00:44:22,400
actually probably needs to have some

1095
00:44:21,520 --> 00:44:25,040
form

1096
00:44:22,400 --> 00:44:28,079
of um

1097
00:44:25,040 --> 00:44:30,560
ethics system that sort of like yes i

1098
00:44:28,079 --> 00:44:32,640
will be using real people in this i will

1099
00:44:30,560 --> 00:44:35,040
be seeking their permission this way

1100
00:44:32,640 --> 00:44:36,800
they will be informed this way they will

1101
00:44:35,040 --> 00:44:40,000
be able to retract their information

1102
00:44:36,800 --> 00:44:42,319
this way and a lot of

1103
00:44:40,000 --> 00:44:43,520
large ai um

1104
00:44:42,319 --> 00:44:45,839
companies

1105
00:44:43,520 --> 00:44:47,040
do literally none of this

1106
00:44:45,839 --> 00:44:49,280
if you have

1107
00:44:47,040 --> 00:44:51,839
a data up on flickr if you have images

1108
00:44:49,280 --> 00:44:53,599
on flickr they've been harvested um

1109
00:44:51,839 --> 00:44:56,079
they're in um

1110
00:44:53,599 --> 00:44:57,760
multiple different ai data sets

1111
00:44:56,079 --> 00:44:58,880
if you've

1112
00:44:57,760 --> 00:45:00,480
got

1113
00:44:58,880 --> 00:45:01,839
a different

1114
00:45:00,480 --> 00:45:04,160
uh if

1115
00:45:01,839 --> 00:45:06,240
facebook of course uses your data

1116
00:45:04,160 --> 00:45:08,720
extremely liberally uh there are a lot

1117
00:45:06,240 --> 00:45:10,720
of different ways and it's

1118
00:45:08,720 --> 00:45:13,200
it's part of being extremely online but

1119
00:45:10,720 --> 00:45:17,040
it's also an issue so for doing it

1120
00:45:13,200 --> 00:45:18,720
yourself it's actually um

1121
00:45:17,040 --> 00:45:23,200
really important

1122
00:45:18,720 --> 00:45:26,160
to consider for my work um i decided

1123
00:45:23,200 --> 00:45:28,640
instead of using real people i created

1124
00:45:26,160 --> 00:45:30,000
3d renders of people so i never actually

1125
00:45:28,640 --> 00:45:32,800
had to use them

1126
00:45:30,000 --> 00:45:34,400
or i used um

1127
00:45:32,800 --> 00:45:38,240
uh

1128
00:45:34,400 --> 00:45:41,680
generated humans from uh using uh

1129
00:45:38,240 --> 00:45:44,240
styleganto um to dial in image

1130
00:45:41,680 --> 00:45:46,079
parameters instead so i work entirely

1131
00:45:44,240 --> 00:45:47,680
with fabricated humans which is my way

1132
00:45:46,079 --> 00:45:49,359
of getting around that ethics oh no is

1133
00:45:47,680 --> 00:45:52,079
there a minutes left

1134
00:45:49,359 --> 00:45:52,079
sorry about that

1135
00:45:52,720 --> 00:45:56,079
yeah it's a good question i'll put the

1136
00:45:54,720 --> 00:45:58,400
if we can get that one as well into the

1137
00:45:56,079 --> 00:45:59,920
texture afterwards uh if you want to

1138
00:45:58,400 --> 00:46:01,359
talk more about it and then that's

1139
00:45:59,920 --> 00:46:03,359
absolutely welcome

1140
00:46:01,359 --> 00:46:05,359
thank you so much jay that was a really

1141
00:46:03,359 --> 00:46:06,880
good uh talking people and chat seemed

1142
00:46:05,359 --> 00:46:09,040
really happy as well

1143
00:46:06,880 --> 00:46:11,520
oh hey i'm glad

1144
00:46:09,040 --> 00:46:12,480
next up we have an afternoon tea break i

1145
00:46:11,520 --> 00:46:14,319
think or

1146
00:46:12,480 --> 00:46:15,440
yeah yeah

1147
00:46:14,319 --> 00:46:17,359
yeah

1148
00:46:15,440 --> 00:46:18,640
we're back at 3 45

1149
00:46:17,359 --> 00:46:21,119
um

1150
00:46:18,640 --> 00:46:23,760
with g prof ng with which is the next

1151
00:46:21,119 --> 00:46:25,520
generation new gnu profiling tool with

1152
00:46:23,760 --> 00:46:27,520
rude van der pass

1153
00:46:25,520 --> 00:46:31,400
uh and i'll see you all back then

1154
00:46:27,520 --> 00:46:31,400
in the kia ora theater