1
00:00:00,420 --> 00:00:05,910
[Music]

2
00:00:11,200 --> 00:00:15,679
Thanks for coming. I'm Lily. Um, we'll

3
00:00:14,320 --> 00:00:18,000
talk about some falsehoods about

4
00:00:15,679 --> 00:00:20,880
reality. I chose this topic. I can't

5
00:00:18,000 --> 00:00:23,119
remember why. Um, and you know, I've

6
00:00:20,880 --> 00:00:25,199
been kicking myself for it, but as as we

7
00:00:23,119 --> 00:00:27,680
do Oh,

8
00:00:25,199 --> 00:00:31,039
and then I did that.

9
00:00:27,680 --> 00:00:35,200
We can edit this out in post, right?

10
00:00:31,039 --> 00:00:37,280
Anyway, hi, I'm Lily. Um,

11
00:00:35,200 --> 00:00:39,600
I am an information security engineer.

12
00:00:37,280 --> 00:00:41,280
I'm an exhistorian. I'm a synthesizer

13
00:00:39,600 --> 00:00:43,360
enthusiast. And I am a professionally

14
00:00:41,280 --> 00:00:45,920
anxious human.

15
00:00:43,360 --> 00:00:47,280
I like working at the edges of systems

16
00:00:45,920 --> 00:00:49,039
because, in my opinion, the most

17
00:00:47,280 --> 00:00:51,280
interesting things happen there. It's

18
00:00:49,039 --> 00:00:52,960
also where things tend to break down and

19
00:00:51,280 --> 00:00:54,879
where you can find the things that

20
00:00:52,960 --> 00:00:56,879
break. And every time there is another

21
00:00:54,879 --> 00:00:58,640
cycle of move fast and break things in

22
00:00:56,879 --> 00:01:00,719
the security space, we see a lot of the

23
00:00:58,640 --> 00:01:03,199
broken things. You may have noticed that

24
00:01:00,719 --> 00:01:05,600
we are in a cycle like that right now.

25
00:01:03,199 --> 00:01:08,320
And that means I'm seeing a lot of the

26
00:01:05,600 --> 00:01:10,159
broken aspects of projects and products

27
00:01:08,320 --> 00:01:12,880
and and what a good portion of them stem

28
00:01:10,159 --> 00:01:15,520
from in my view is a belief held by

29
00:01:12,880 --> 00:01:17,200
someone central to that work that the

30
00:01:15,520 --> 00:01:18,960
information that feeds and drives their

31
00:01:17,200 --> 00:01:20,880
systems is real in some kind of

32
00:01:18,960 --> 00:01:22,960
objective way. that their thing is

33
00:01:20,880 --> 00:01:24,640
naturally going to handle the complexity

34
00:01:22,960 --> 00:01:27,119
of the real world with computers because

35
00:01:24,640 --> 00:01:30,720
computers are very complicated and

36
00:01:27,119 --> 00:01:33,280
reality is simple.

37
00:01:30,720 --> 00:01:34,640
The core issue here is that reality is

38
00:01:33,280 --> 00:01:37,040
not literally possible to capture

39
00:01:34,640 --> 00:01:38,720
objectively, but we need to represent

40
00:01:37,040 --> 00:01:41,840
reality to our computers in order to do

41
00:01:38,720 --> 00:01:43,520
our jobs. So phrases like the map is not

42
00:01:41,840 --> 00:01:45,439
the territory, they're common, they're

43
00:01:43,520 --> 00:01:48,479
quotable because they come up so often

44
00:01:45,439 --> 00:01:50,799
in our work. And put simply, most of the

45
00:01:48,479 --> 00:01:52,880
work we do with with computers has some

46
00:01:50,799 --> 00:01:54,560
kind of relationship to things happening

47
00:01:52,880 --> 00:01:56,240
in pieces of the world somewhere along

48
00:01:54,560 --> 00:01:58,640
the line. And in order to deal with

49
00:01:56,240 --> 00:02:00,560
that, we have to find a way to represent

50
00:01:58,640 --> 00:02:02,079
the world to the computer in order to

51
00:02:00,560 --> 00:02:03,600
make sense of it. So we have to make

52
00:02:02,079 --> 00:02:05,520
decisions to turn those things into data

53
00:02:03,600 --> 00:02:07,200
sets and models and abstract systems

54
00:02:05,520 --> 00:02:09,520
that can be used to stand in for the

55
00:02:07,200 --> 00:02:11,520
things that we want to talk about and

56
00:02:09,520 --> 00:02:13,840
help us represent them and make new

57
00:02:11,520 --> 00:02:16,480
observations and test hypotheses about

58
00:02:13,840 --> 00:02:18,800
whatever this thing is.

59
00:02:16,480 --> 00:02:20,640
The falsehoods programmers believe in

60
00:02:18,800 --> 00:02:22,959
the title of this talk, falsehoods

61
00:02:20,640 --> 00:02:25,520
programmers believe about reality. Um,

62
00:02:22,959 --> 00:02:27,599
this is a longunning format for lists

63
00:02:25,520 --> 00:02:29,760
about things that seem simple to deal

64
00:02:27,599 --> 00:02:31,360
with on the surface and that get more

65
00:02:29,760 --> 00:02:33,680
complicated the more that you dig into

66
00:02:31,360 --> 00:02:35,920
them. There are many excellent lists and

67
00:02:33,680 --> 00:02:38,959
talks as well about how hard it is to

68
00:02:35,920 --> 00:02:43,040
handle every single edge case for names

69
00:02:38,959 --> 00:02:45,280
and time and geography and gender and

70
00:02:43,040 --> 00:02:47,519
email addresses and phone numbers and

71
00:02:45,280 --> 00:02:51,519
you know the list goes on. You can find

72
00:02:47,519 --> 00:02:53,680
a list of lists at that link.

73
00:02:51,519 --> 00:02:55,280
And this title is also sort of kind of a

74
00:02:53,680 --> 00:02:56,879
little bit clickbay to get you here

75
00:02:55,280 --> 00:02:59,840
because I suspect that a lot of you will

76
00:02:56,879 --> 00:03:01,280
know most of or parts of or some of all

77
00:02:59,840 --> 00:03:03,360
of this sort of stuff I'm going to talk

78
00:03:01,280 --> 00:03:06,159
about already. And if you haven't yet

79
00:03:03,360 --> 00:03:09,040
run into it, you will the longer that

80
00:03:06,159 --> 00:03:11,200
you go on doing this kind of work. But

81
00:03:09,040 --> 00:03:13,360
this talk is not just for the people who

82
00:03:11,200 --> 00:03:15,040
are in this room or the people who are

83
00:03:13,360 --> 00:03:16,959
watching it online live or in the

84
00:03:15,040 --> 00:03:19,120
future. It's also for the people who

85
00:03:16,959 --> 00:03:20,400
would never have chosen to watch this

86
00:03:19,120 --> 00:03:21,920
talk.

87
00:03:20,400 --> 00:03:23,920
It's to give all of you who did choose

88
00:03:21,920 --> 00:03:26,159
to watch this talk the more tools, all

89
00:03:23,920 --> 00:03:27,680
these tools that you can use to have

90
00:03:26,159 --> 00:03:29,840
these kinds of conversations with the

91
00:03:27,680 --> 00:03:32,080
people in your lives in both personal

92
00:03:29,840 --> 00:03:34,400
and professional capacities who have

93
00:03:32,080 --> 00:03:36,560
fallen into the idea that we can capture

94
00:03:34,400 --> 00:03:38,319
reality objectively and they're going

95
00:03:36,560 --> 00:03:41,280
ahead and building things based on this

96
00:03:38,319 --> 00:03:43,760
premise. And because reality is slippery

97
00:03:41,280 --> 00:03:46,400
and brains are weird, um sometimes the

98
00:03:43,760 --> 00:03:48,959
person who needs that reminder is you or

99
00:03:46,400 --> 00:03:52,440
me. We forget some of this in small ways

100
00:03:48,959 --> 00:03:52,440
all the time.

101
00:03:56,640 --> 00:04:00,879
So although the map is not the territory

102
00:03:58,959 --> 00:04:03,040
was a phrase that was coined, it it was

103
00:04:00,879 --> 00:04:06,000
a phrase that was coined in the 1930s,

104
00:04:03,040 --> 00:04:09,120
but it's only one expression of a very

105
00:04:06,000 --> 00:04:11,439
old observation. It's not by any means a

106
00:04:09,120 --> 00:04:13,599
new idea.

107
00:04:11,439 --> 00:04:14,959
It's there's been so much philosophy and

108
00:04:13,599 --> 00:04:16,479
literature and all kinds of other stuff

109
00:04:14,959 --> 00:04:17,680
about it that if I started to list it,

110
00:04:16,479 --> 00:04:20,400
we wouldn't be doing anything else for

111
00:04:17,680 --> 00:04:22,479
the rest of this time slot. I will have

112
00:04:20,400 --> 00:04:24,000
to gloss over things like semiotics,

113
00:04:22,479 --> 00:04:26,320
which I am just as disappointed about as

114
00:04:24,000 --> 00:04:28,160
you are.

115
00:04:26,320 --> 00:04:29,440
So, here are some works you can dip into

116
00:04:28,160 --> 00:04:31,520
later if you feel like going further

117
00:04:29,440 --> 00:04:34,080
down this rabbit hole.

118
00:04:31,520 --> 00:04:35,840
Anyway, this map territory thing has

119
00:04:34,080 --> 00:04:38,320
become increasingly loadbearing as we

120
00:04:35,840 --> 00:04:39,600
have made things more complex and as the

121
00:04:38,320 --> 00:04:41,199
pieces of the world that we're dealing

122
00:04:39,600 --> 00:04:43,040
with in our data get more and more

123
00:04:41,199 --> 00:04:45,040
nuanced. So, back in the time of the

124
00:04:43,040 --> 00:04:46,960
ancient Greeks, they had started to

125
00:04:45,040 --> 00:04:49,199
tease out some of these ideas, but they

126
00:04:46,960 --> 00:04:50,880
were often the kinds of musings that you

127
00:04:49,199 --> 00:04:52,639
would get at parties from people who had

128
00:04:50,880 --> 00:04:55,199
ingested some really interesting stuff

129
00:04:52,639 --> 00:04:57,280
most of the time. By the time we got to

130
00:04:55,199 --> 00:04:58,880
the industrial revolution and humanity

131
00:04:57,280 --> 00:05:00,960
started to try making and organizing

132
00:04:58,880 --> 00:05:04,080
things in a common way at some kind of

133
00:05:00,960 --> 00:05:06,000
scale, these little reality maps became

134
00:05:04,080 --> 00:05:08,800
way more important because they were

135
00:05:06,000 --> 00:05:11,039
necessary for anything to be organized

136
00:05:08,800 --> 00:05:12,320
or consistent when you were doing

137
00:05:11,039 --> 00:05:13,680
anything at the scale of like larger

138
00:05:12,320 --> 00:05:16,800
than the number of people who could fit

139
00:05:13,680 --> 00:05:18,160
in a town hall at the one time. And by

140
00:05:16,800 --> 00:05:20,400
the time that we had gotten to the point

141
00:05:18,160 --> 00:05:21,919
of making computers about it, we had

142
00:05:20,400 --> 00:05:23,440
made things nice and complicated for

143
00:05:21,919 --> 00:05:25,759
ourselves. We'd invented the concept of

144
00:05:23,440 --> 00:05:27,840
middle managers and our species was

145
00:05:25,759 --> 00:05:30,000
capable of screwing up a whole lot of

146
00:05:27,840 --> 00:05:31,840
things way faster.

147
00:05:30,000 --> 00:05:34,960
But in our current moment, we have

148
00:05:31,840 --> 00:05:36,560
dialed it up even further than even the

149
00:05:34,960 --> 00:05:38,960
ancient Greeks would ever have believed.

150
00:05:36,560 --> 00:05:42,000
We've taken perfectly good databases and

151
00:05:38,960 --> 00:05:43,919
we have given them anxiety.

152
00:05:42,000 --> 00:05:45,199
We have built large language models on

153
00:05:43,919 --> 00:05:47,039
the back of literally everything on the

154
00:05:45,199 --> 00:05:49,199
entire open web and quite a few things

155
00:05:47,039 --> 00:05:52,080
that aren't. We've developed algorithms

156
00:05:49,199 --> 00:05:54,160
that make decisions about people. um and

157
00:05:52,080 --> 00:05:55,840
what they will see, what they will hear

158
00:05:54,160 --> 00:05:57,919
as they go about their days. And we've

159
00:05:55,840 --> 00:05:59,759
based them on, you know, mass analytics

160
00:05:57,919 --> 00:06:01,600
gathering and physical senses and where

161
00:05:59,759 --> 00:06:05,280
you bought lunch yesterday and who you

162
00:06:01,600 --> 00:06:07,199
went to school with. So, we need to

163
00:06:05,280 --> 00:06:09,199
continue to make this observation and

164
00:06:07,199 --> 00:06:11,199
especially now because we are beyond big

165
00:06:09,199 --> 00:06:13,600
data and we're into the realm of sort of

166
00:06:11,199 --> 00:06:15,199
weekend Bernies and big data around to

167
00:06:13,600 --> 00:06:17,600
make decisions about the world and the

168
00:06:15,199 --> 00:06:18,800
people and the things in it.

169
00:06:17,600 --> 00:06:20,240
So, there are terms like, you know,

170
00:06:18,800 --> 00:06:22,000
digital twins and world models and

171
00:06:20,240 --> 00:06:25,280
ground truth. They get thrown around in

172
00:06:22,000 --> 00:06:28,880
so many places. And these terms have the

173
00:06:25,280 --> 00:06:30,400
effect of carrying on this idea that we

174
00:06:28,880 --> 00:06:32,560
are able to measure everything

175
00:06:30,400 --> 00:06:35,680
objectively and then probably make money

176
00:06:32,560 --> 00:06:38,000
from it. We've taken all of these maps

177
00:06:35,680 --> 00:06:39,440
from philosophy to industry to

178
00:06:38,000 --> 00:06:41,120
individually shaping the information

179
00:06:39,440 --> 00:06:43,280
that people receive on a daily basis.

180
00:06:41,120 --> 00:06:44,960
And this whole set of things is still

181
00:06:43,280 --> 00:06:47,520
not actually the territory. Not even

182
00:06:44,960 --> 00:06:49,520
close. Like just ask an LLM how many L's

183
00:06:47,520 --> 00:06:52,319
there are in surveillance capitalism.

184
00:06:49,520 --> 00:06:55,720
and then like ask it again if it's

185
00:06:52,319 --> 00:06:55,720
really sure.

186
00:06:59,360 --> 00:07:02,240
So let's talk about some of the thinking

187
00:07:00,880 --> 00:07:04,319
traps that we fall into every now and

188
00:07:02,240 --> 00:07:06,080
then. I've broken some of these

189
00:07:04,319 --> 00:07:08,639
fallacies up into a few statements that

190
00:07:06,080 --> 00:07:10,720
we can talk through like any structure

191
00:07:08,639 --> 00:07:12,160
we put over reality. The points I want

192
00:07:10,720 --> 00:07:13,280
to make as we talk about each of these

193
00:07:12,160 --> 00:07:14,560
statements are going to overlap with

194
00:07:13,280 --> 00:07:17,520
each other a little bit contradict each

195
00:07:14,560 --> 00:07:19,919
other in some places. um you know but

196
00:07:17,520 --> 00:07:21,199
having worked on this talk for so long I

197
00:07:19,919 --> 00:07:23,759
consider all of that to be kind of part

198
00:07:21,199 --> 00:07:27,840
of the fun that comes with the territory

199
00:07:23,759 --> 00:07:29,680
so to speak. So here we go. Surely the

200
00:07:27,840 --> 00:07:32,080
more data we have the more perfectly we

201
00:07:29,680 --> 00:07:34,880
can represent reality.

202
00:07:32,080 --> 00:07:36,160
So to an extent this is true. If you

203
00:07:34,880 --> 00:07:37,840
have one data source you have one view

204
00:07:36,160 --> 00:07:40,080
of the world. If you have two you have

205
00:07:37,840 --> 00:07:42,080
two views that allow you to approximate

206
00:07:40,080 --> 00:07:44,479
something more general about the world.

207
00:07:42,080 --> 00:07:46,000
And if you have 500 or 10 billion

208
00:07:44,479 --> 00:07:48,400
sources, then you probably end up with

209
00:07:46,000 --> 00:07:50,800
something that is increasingly nuanced,

210
00:07:48,400 --> 00:07:52,720
but it's still not reality. It's a bunch

211
00:07:50,800 --> 00:07:54,720
of different models of slices of reality

212
00:07:52,720 --> 00:07:57,520
with your interpretation of it laid over

213
00:07:54,720 --> 00:07:59,440
the top. Raw data as a term is an

214
00:07:57,520 --> 00:08:00,560
oxymoron. There's also a book with this

215
00:07:59,440 --> 00:08:02,879
title that goes into some of the

216
00:08:00,560 --> 00:08:04,479
examples about how this is so, why it is

217
00:08:02,879 --> 00:08:06,720
that way.

218
00:08:04,479 --> 00:08:09,120
All data is the result of someone's

219
00:08:06,720 --> 00:08:10,960
choices to measure something, how to

220
00:08:09,120 --> 00:08:13,280
measure it, what that measurement stands

221
00:08:10,960 --> 00:08:15,280
for, and so on. Raw data, I feel like,

222
00:08:13,280 --> 00:08:17,520
is more like raw milk, like it's

223
00:08:15,280 --> 00:08:19,199
unprocessed, unpasteurized, it's full of

224
00:08:17,520 --> 00:08:20,639
bugs. It needs some kind of processing

225
00:08:19,199 --> 00:08:22,639
to improve its quality and make it fit

226
00:08:20,639 --> 00:08:24,240
for consumption. There are plenty of

227
00:08:22,639 --> 00:08:26,400
things that cause problems with raw

228
00:08:24,240 --> 00:08:28,400
data, like interference and changing

229
00:08:26,400 --> 00:08:30,080
parameters and faulty sensors and faulty

230
00:08:28,400 --> 00:08:32,320
assumptions.

231
00:08:30,080 --> 00:08:34,240
Quality data sources are what really

232
00:08:32,320 --> 00:08:35,839
make the difference here. Things do tend

233
00:08:34,240 --> 00:08:37,599
to work better when you choose what you

234
00:08:35,839 --> 00:08:39,039
need for the purposes that you need it

235
00:08:37,599 --> 00:08:41,760
for and you build your models and

236
00:08:39,039 --> 00:08:43,440
systems with that purpose in mind. In

237
00:08:41,760 --> 00:08:45,279
the era of scaling everything and

238
00:08:43,440 --> 00:08:47,600
relentless growth and like line go up,

239
00:08:45,279 --> 00:08:51,040
it's almost a default assumption I feel

240
00:08:47,600 --> 00:08:53,279
that you know some places will work with

241
00:08:51,040 --> 00:08:56,160
the view that react that quality can be

242
00:08:53,279 --> 00:08:57,680
coaxed out of a model later and that the

243
00:08:56,160 --> 00:09:01,440
quantity of the sources is what matters

244
00:08:57,680 --> 00:09:03,279
the most and to an extent that works but

245
00:09:01,440 --> 00:09:05,120
it is also very brittle. You have to

246
00:09:03,279 --> 00:09:07,600
work with your model just so and finesse

247
00:09:05,120 --> 00:09:10,320
it just right to get the thing that you

248
00:09:07,600 --> 00:09:11,920
want. Many of us are probably intimately

249
00:09:10,320 --> 00:09:13,920
familiar with how badly this goes when

250
00:09:11,920 --> 00:09:16,240
someone at your workplace decides to

251
00:09:13,920 --> 00:09:17,839
wire up 20 years of SharePoint to an LLM

252
00:09:16,240 --> 00:09:19,279
to finally solve the problem of like

253
00:09:17,839 --> 00:09:21,360
nobody ever being able to find anything

254
00:09:19,279 --> 00:09:23,040
around here. And you know what's

255
00:09:21,360 --> 00:09:25,600
supposed to come out of that at the

256
00:09:23,040 --> 00:09:29,279
other end is insights. And what comes

257
00:09:25,600 --> 00:09:31,120
out is usually like an air horn noise.

258
00:09:29,279 --> 00:09:32,480
And to correct for it, we often end up

259
00:09:31,120 --> 00:09:33,680
having to write prompts that are so

260
00:09:32,480 --> 00:09:35,279
complex that it sounds like you're

261
00:09:33,680 --> 00:09:36,800
trying to outwit a genie by covering

262
00:09:35,279 --> 00:09:38,560
literally all the edge cases so it

263
00:09:36,800 --> 00:09:41,440
doesn't grant your wish in like the most

264
00:09:38,560 --> 00:09:43,120
cursed way possible.

265
00:09:41,440 --> 00:09:45,279
So many talks about this conference have

266
00:09:43,120 --> 00:09:47,760
been about the issues that exists in all

267
00:09:45,279 --> 00:09:49,120
data sets and models

268
00:09:47,760 --> 00:09:51,600
and particularly how Python helped them

269
00:09:49,120 --> 00:09:54,080
work with them. And this is a Python

270
00:09:51,600 --> 00:09:56,240
conference. So cool. So there's uh Ishad

271
00:09:54,080 --> 00:09:58,000
Zaman's talk about continuing about

272
00:09:56,240 --> 00:09:59,600
counting the neurons in mouse brains.

273
00:09:58,000 --> 00:10:01,600
That was a really cool one. Renee

274
00:09:59,600 --> 00:10:03,680
Noble's talk about matching student

275
00:10:01,600 --> 00:10:06,080
names in a register to what the students

276
00:10:03,680 --> 00:10:07,760
write down when they sign in and how

277
00:10:06,080 --> 00:10:09,360
even like multi-round approaches with a

278
00:10:07,760 --> 00:10:11,279
range of techniques and technologies

279
00:10:09,360 --> 00:10:13,519
couldn't match all of the names 100%

280
00:10:11,279 --> 00:10:15,200
correctly. And there was Dave Cole's

281
00:10:13,519 --> 00:10:17,120
talk about finding workable and

282
00:10:15,200 --> 00:10:18,480
realistic road trip routes for electric

283
00:10:17,120 --> 00:10:21,800
vehicles that accounted for all of the

284
00:10:18,480 --> 00:10:21,800
broken charges.

285
00:10:22,000 --> 00:10:26,959
The way um

286
00:10:24,800 --> 00:10:28,320
the way that many AI product companies

287
00:10:26,959 --> 00:10:30,160
market and encourage users to think

288
00:10:28,320 --> 00:10:32,640
about large language models is I think a

289
00:10:30,160 --> 00:10:34,399
pretty big example of this fallacy. It's

290
00:10:32,640 --> 00:10:36,640
not really the fault of the LLMs. The

291
00:10:34,399 --> 00:10:38,720
technology itself is not something that

292
00:10:36,640 --> 00:10:40,480
should be about being a mirror of the

293
00:10:38,720 --> 00:10:43,360
world. But it is very tempting and very

294
00:10:40,480 --> 00:10:44,560
profitable to fall into the idea that it

295
00:10:43,360 --> 00:10:45,760
should be so. Especially when you're

296
00:10:44,560 --> 00:10:47,600
told it's this general purpose

297
00:10:45,760 --> 00:10:49,920
everything machine. It's what I call the

298
00:10:47,600 --> 00:10:51,519
sparkle emoji effect. You put all of the

299
00:10:49,920 --> 00:10:53,920
LLM back features behind a little button

300
00:10:51,519 --> 00:10:56,320
with a sparkle emoji on it and that

301
00:10:53,920 --> 00:10:58,240
encourages the idea that it's magic and

302
00:10:56,320 --> 00:11:00,399
it's got a side dose of like it's very

303
00:10:58,240 --> 00:11:02,959
complicated. Don't think about it. It'll

304
00:11:00,399 --> 00:11:05,920
just work TM.

305
00:11:02,959 --> 00:11:07,440
And it's actually really important in

306
00:11:05,920 --> 00:11:09,200
deciding when to use an LLM

307
00:11:07,440 --> 00:11:10,560
appropriately that the users know the

308
00:11:09,200 --> 00:11:11,920
basics about what it actually does,

309
00:11:10,560 --> 00:11:13,120
which is mathematically predicting and

310
00:11:11,920 --> 00:11:14,959
outputting tokens based on the tokens

311
00:11:13,120 --> 00:11:17,279
that were input and what it doesn't do,

312
00:11:14,959 --> 00:11:19,279
which is think and reason in some kind

313
00:11:17,279 --> 00:11:21,279
of like objective and neutral fashion or

314
00:11:19,279 --> 00:11:23,600
at all.

315
00:11:21,279 --> 00:11:25,680
It feels to me like the root of this

316
00:11:23,600 --> 00:11:27,600
idea is like check it out. We've scraped

317
00:11:25,680 --> 00:11:29,200
all of the data from the entire internet

318
00:11:27,600 --> 00:11:31,279
and all of human everything and we are

319
00:11:29,200 --> 00:11:33,680
regurgitating it back to you. But that

320
00:11:31,279 --> 00:11:35,040
isn't actually everything. It's all

321
00:11:33,680 --> 00:11:37,200
written data that has been made

322
00:11:35,040 --> 00:11:38,560
available in a digital form which favors

323
00:11:37,200 --> 00:11:40,240
the written languages and styles and

324
00:11:38,560 --> 00:11:42,880
topics that are used by the groups of

325
00:11:40,240 --> 00:11:45,200
people who have the most access and the

326
00:11:42,880 --> 00:11:46,800
earliest access to the internet and to

327
00:11:45,200 --> 00:11:49,120
publication technologies in all of the

328
00:11:46,800 --> 00:11:50,880
years gone by. That is like a small

329
00:11:49,120 --> 00:11:52,959
slice of perspectives when we're talking

330
00:11:50,880 --> 00:11:55,440
about the scale of humanity. But there's

331
00:11:52,959 --> 00:11:57,200
also a small slice of what language is.

332
00:11:55,440 --> 00:11:59,680
We've had written language for maybe

333
00:11:57,200 --> 00:12:02,000
6,000 years, which is nothing compared

334
00:11:59,680 --> 00:12:03,680
to like the 100,000ish years that we

335
00:12:02,000 --> 00:12:06,800
have had language in general, like as a

336
00:12:03,680 --> 00:12:09,200
species. Written language can seem like

337
00:12:06,800 --> 00:12:11,680
the most important thing today. But

338
00:12:09,200 --> 00:12:13,920
language itself is also gestures and

339
00:12:11,680 --> 00:12:16,320
tone and posture and facial expression

340
00:12:13,920 --> 00:12:17,760
and timing. And in an over in an

341
00:12:16,320 --> 00:12:21,200
overwhelming number of cases, it is

342
00:12:17,760 --> 00:12:23,839
spoken and it is interactive. It assumes

343
00:12:21,200 --> 00:12:26,160
communities and societies and others at

344
00:12:23,839 --> 00:12:28,000
the other end. It is so bound up with

345
00:12:26,160 --> 00:12:29,920
our species that it has impacted the

346
00:12:28,000 --> 00:12:31,920
evolution of our hands and our faces and

347
00:12:29,920 --> 00:12:34,320
our larynxes and our limbs compared to

348
00:12:31,920 --> 00:12:35,839
other species. Um, you should really

349
00:12:34,320 --> 00:12:38,320
check out Steven Levenson's work on

350
00:12:35,839 --> 00:12:39,839
this. He's a linguist. He spent over 40

351
00:12:38,320 --> 00:12:41,440
years researching the intersection of

352
00:12:39,839 --> 00:12:45,279
linguistics and evolutionary biology.

353
00:12:41,440 --> 00:12:47,440
It's really cool. Anyway, when we take

354
00:12:45,279 --> 00:12:49,120
just the internet accessible written

355
00:12:47,440 --> 00:12:51,200
parts of some languages and we turn it

356
00:12:49,120 --> 00:12:53,279
into a chatbot and it is marketed like

357
00:12:51,200 --> 00:12:55,120
that as the basis of all like objective

358
00:12:53,279 --> 00:12:57,920
human knowledge and expression and

359
00:12:55,120 --> 00:12:59,600
everything, it really isn't. It is very

360
00:12:57,920 --> 00:13:01,360
good at natural language processing and

361
00:12:59,600 --> 00:13:03,279
like semantic search and it can be made

362
00:13:01,360 --> 00:13:06,880
to be partially reliable performing a

363
00:13:03,279 --> 00:13:09,680
subset of tasks if you hold it like that

364
00:13:06,880 --> 00:13:11,360
and don't use too many run-on sentences

365
00:13:09,680 --> 00:13:12,880
or low resource languages and you know

366
00:13:11,360 --> 00:13:15,279
maybe give it all the answers before you

367
00:13:12,880 --> 00:13:17,920
start. But it is not a perfect

368
00:13:15,279 --> 00:13:20,240
representation of reality and is also

369
00:13:17,920 --> 00:13:21,920
not a therapist or a romantic partner or

370
00:13:20,240 --> 00:13:23,680
an intern.

371
00:13:21,920 --> 00:13:25,040
can't even run a drive-thru because

372
00:13:23,680 --> 00:13:26,320
ingesting a bunch of training manuals on

373
00:13:25,040 --> 00:13:27,920
how to run a drive-through is not the

374
00:13:26,320 --> 00:13:29,200
same as having the experience of being a

375
00:13:27,920 --> 00:13:30,959
teenage human being who knows that the

376
00:13:29,200 --> 00:13:32,720
audio quality sucks and the people

377
00:13:30,959 --> 00:13:34,560
placing the orders at the other end are

378
00:13:32,720 --> 00:13:36,000
probably tired and distracted and

379
00:13:34,560 --> 00:13:38,079
occasionally likely to mess with you for

380
00:13:36,000 --> 00:13:40,880
like no good reason. It is hard to

381
00:13:38,079 --> 00:13:42,079
measure that experience in data.

382
00:13:40,880 --> 00:13:43,920
And yeah, this is another one of these

383
00:13:42,079 --> 00:13:45,600
fallacies about data that assumes

384
00:13:43,920 --> 00:13:48,079
everything is measurable when that is

385
00:13:45,600 --> 00:13:51,440
not at all true. How do you measure

386
00:13:48,079 --> 00:13:53,519
justice or security or love? These

387
00:13:51,440 --> 00:13:55,200
things are really important, but they

388
00:13:53,519 --> 00:13:58,320
are not measurable. You end up having to

389
00:13:55,200 --> 00:13:59,760
use things as proxies for other things.

390
00:13:58,320 --> 00:14:01,120
Sometimes we get so caught up in the

391
00:13:59,760 --> 00:14:03,040
proxy that we forget that it isn't

392
00:14:01,120 --> 00:14:05,600
actually measuring the thing itself. So

393
00:14:03,040 --> 00:14:07,199
like speaking as a security person, the

394
00:14:05,600 --> 00:14:09,040
number of vulnerabilities that you patch

395
00:14:07,199 --> 00:14:10,320
in your app's code, it doesn't

396
00:14:09,040 --> 00:14:12,880
necessarily mean that your app is like

397
00:14:10,320 --> 00:14:14,880
more secure than it was before. It might

398
00:14:12,880 --> 00:14:16,560
still be perfectly possible for the app

399
00:14:14,880 --> 00:14:18,160
to be breached because a core

400
00:14:16,560 --> 00:14:19,440
developer's GitHub password was fished

401
00:14:18,160 --> 00:14:21,040
and the attackers can now control your

402
00:14:19,440 --> 00:14:23,360
build servers. It's all down to what we

403
00:14:21,040 --> 00:14:26,480
choose to observe or what we think we

404
00:14:23,360 --> 00:14:29,040
observe because reality doesn't look the

405
00:14:26,480 --> 00:14:31,440
same to everyone.

406
00:14:29,040 --> 00:14:33,519
For a start, not all of us can see the

407
00:14:31,440 --> 00:14:37,600
same color spectrum. Not all of us have

408
00:14:33,519 --> 00:14:40,800
vision. We all have we also have like

409
00:14:37,600 --> 00:14:42,639
opinions and histories that that shape

410
00:14:40,800 --> 00:14:44,959
us and what we think is good and what

411
00:14:42,639 --> 00:14:46,480
isn't. And those views can change

412
00:14:44,959 --> 00:14:48,160
constantly. Some of you watching this

413
00:14:46,480 --> 00:14:50,399
talk probably disagree with me about

414
00:14:48,160 --> 00:14:52,000
what I'm saying or how I'm saying it.

415
00:14:50,399 --> 00:14:53,920
Some of you might be bored and some of

416
00:14:52,000 --> 00:14:57,040
you might be fascinated or distracted by

417
00:14:53,920 --> 00:15:01,199
social media or by the ineffable nature

418
00:14:57,040 --> 00:15:02,639
of the universe, which fair enough.

419
00:15:01,199 --> 00:15:04,000
But we often fall into this trap of

420
00:15:02,639 --> 00:15:06,320
thinking that there's only one reality

421
00:15:04,000 --> 00:15:08,639
experienced by everyone because it is

422
00:15:06,320 --> 00:15:10,720
more or less a defense mechanism. It's

423
00:15:08,639 --> 00:15:12,320
not actually good for our brains to be

424
00:15:10,720 --> 00:15:14,320
consistently aware of the subjective

425
00:15:12,320 --> 00:15:15,440
nature of reality for like every waking

426
00:15:14,320 --> 00:15:17,120
moment.

427
00:15:15,440 --> 00:15:19,279
There's an amount that it can be

428
00:15:17,120 --> 00:15:20,959
liberating to realize that a lot of

429
00:15:19,279 --> 00:15:22,880
nonsense like I'm really bad at

430
00:15:20,959 --> 00:15:25,199
painting. That's a matter of perspective

431
00:15:22,880 --> 00:15:28,000
and taste and experience, but it makes

432
00:15:25,199 --> 00:15:30,079
it very hard to participate in wider

433
00:15:28,000 --> 00:15:32,639
functional systems with others or like

434
00:15:30,079 --> 00:15:34,639
to buy groceries or merge in traffic or

435
00:15:32,639 --> 00:15:36,560
hold a conversation.

436
00:15:34,639 --> 00:15:38,079
We need consensus to get anything done

437
00:15:36,560 --> 00:15:40,079
on a cooperative level. So, we

438
00:15:38,079 --> 00:15:43,440
approximate constantly and this is

439
00:15:40,079 --> 00:15:46,079
usually mostly good enough. Take web

440
00:15:43,440 --> 00:15:48,160
standards. There is nothing to stop you

441
00:15:46,079 --> 00:15:49,600
from writing a web app that uses get

442
00:15:48,160 --> 00:15:52,399
requests to make changes to the

443
00:15:49,600 --> 00:15:55,279
application state. Like it's really not

444
00:15:52,399 --> 00:15:56,880
advisable to do that. But in my time as

445
00:15:55,279 --> 00:15:58,639
a penetration tester, I have seen so

446
00:15:56,880 --> 00:16:00,480
many apps that are built this way. And

447
00:15:58,639 --> 00:16:02,240
it works, but not all of the other

448
00:16:00,480 --> 00:16:03,920
systems make the same assumptions about

449
00:16:02,240 --> 00:16:07,600
how that's going to work. And it gets

450
00:16:03,920 --> 00:16:09,440
weird really fast. And this is all human

451
00:16:07,600 --> 00:16:11,600
understanding. That's how it works.

452
00:16:09,440 --> 00:16:13,839
We've decided that words in a language

453
00:16:11,600 --> 00:16:15,040
represent things and ideas and mostly we

454
00:16:13,839 --> 00:16:17,040
share those understandings in order to

455
00:16:15,040 --> 00:16:19,680
communicate. For example, it's also the

456
00:16:17,040 --> 00:16:22,639
foundation of all science. So one paper,

457
00:16:19,680 --> 00:16:24,399
no matter how well researched or

458
00:16:22,639 --> 00:16:27,360
thoroughly documented, is just an

459
00:16:24,399 --> 00:16:29,519
expression of one experience. But if we

460
00:16:27,360 --> 00:16:31,440
do enough experience experiments in

461
00:16:29,519 --> 00:16:32,720
enough ways, we gather enough

462
00:16:31,440 --> 00:16:34,560
perspectives over time, we put them

463
00:16:32,720 --> 00:16:36,959
together, we discuss them robustly and

464
00:16:34,560 --> 00:16:38,560
widely among a diverse community, and we

465
00:16:36,959 --> 00:16:40,800
constantly search for new input, we

466
00:16:38,560 --> 00:16:42,480
approach like scientific consensus,

467
00:16:40,800 --> 00:16:44,320
which tells us what is more likely to be

468
00:16:42,480 --> 00:16:46,399
happening in the world around us and is

469
00:16:44,320 --> 00:16:49,199
a pretty solid base to build additional

470
00:16:46,399 --> 00:16:51,360
things on top of. Being aware of shared

471
00:16:49,199 --> 00:16:53,600
consensus becomes important when we're

472
00:16:51,360 --> 00:16:55,199
in charge of making decisions and or you

473
00:16:53,600 --> 00:16:58,160
know have the ability to influence

474
00:16:55,199 --> 00:17:00,000
decisions about how a reality a slice of

475
00:16:58,160 --> 00:17:04,040
reality will be interpreted for a data

476
00:17:00,000 --> 00:17:04,040
set or a model or a system.

477
00:17:05,520 --> 00:17:10,000
Our roles as observers also mean that we

478
00:17:08,160 --> 00:17:11,839
become describers and we bring our

479
00:17:10,000 --> 00:17:14,000
biases along whether we are aware of it

480
00:17:11,839 --> 00:17:16,240
or not. We are biased in the very act of

481
00:17:14,000 --> 00:17:18,720
choosing what to observe. We then make

482
00:17:16,240 --> 00:17:20,959
choices about how to describe the world

483
00:17:18,720 --> 00:17:22,480
to our model or our data set which will

484
00:17:20,959 --> 00:17:24,799
reflect our experiences and our

485
00:17:22,480 --> 00:17:27,120
education and our beliefs in some way or

486
00:17:24,799 --> 00:17:29,039
another. There is nothing too small that

487
00:17:27,120 --> 00:17:30,559
isn't subject to some kind of opinion.

488
00:17:29,039 --> 00:17:32,240
Anytime you change a column name at

489
00:17:30,559 --> 00:17:35,200
someone work, someone at work gets like

490
00:17:32,240 --> 00:17:36,880
mad about it. You've experienced this.

491
00:17:35,200 --> 00:17:38,640
There is no data set or model system

492
00:17:36,880 --> 00:17:40,799
that is not biased. You cannot remove

493
00:17:38,640 --> 00:17:42,960
bias. You can only continually be open

494
00:17:40,799 --> 00:17:45,520
to more quality input and see it as like

495
00:17:42,960 --> 00:17:48,320
an evolving consensus over time. Which

496
00:17:45,520 --> 00:17:49,919
leads us to the issue of categorization.

497
00:17:48,320 --> 00:17:51,440
There was a really spicy lightning talk

498
00:17:49,919 --> 00:17:53,120
at the conference yesterday about like

499
00:17:51,440 --> 00:17:56,400
phoggenetic bracketing. Every

500
00:17:53,120 --> 00:18:01,640
terrestrial vertebrae is a fish.

501
00:17:56,400 --> 00:18:01,640
So is is a hot dog a sandwich?

502
00:18:03,120 --> 00:18:08,720
No.

503
00:18:05,679 --> 00:18:10,240
A jaffers dumplings. Good question.

504
00:18:08,720 --> 00:18:11,600
I can see some nodding heads about the

505
00:18:10,240 --> 00:18:13,120
hot dogs and I've seen a lot of people

506
00:18:11,600 --> 00:18:14,799
shouting no as well. You're all really

507
00:18:13,120 --> 00:18:20,440
confident.

508
00:18:14,799 --> 00:18:20,440
All of you. Is cereal a soup?

509
00:18:21,919 --> 00:18:24,919
Cool.

510
00:18:26,320 --> 00:18:32,240
Reality has like no inherent structure.

511
00:18:29,840 --> 00:18:34,480
Any structure that we put on reality is

512
00:18:32,240 --> 00:18:36,799
artificial and it's imposed. And these

513
00:18:34,480 --> 00:18:38,320
structures can be really helpful. Like

514
00:18:36,799 --> 00:18:39,919
you can decide what you want to choose

515
00:18:38,320 --> 00:18:41,440
on a menu for example or where you would

516
00:18:39,919 --> 00:18:43,760
find things in a supermarket. But they

517
00:18:41,440 --> 00:18:46,640
are still made up and there will always

518
00:18:43,760 --> 00:18:48,320
be outliers and overlaps and gray areas

519
00:18:46,640 --> 00:18:49,760
and debates and I'd be very happy to get

520
00:18:48,320 --> 00:18:51,520
into some debates later because those

521
00:18:49,760 --> 00:18:52,720
are fun.

522
00:18:51,520 --> 00:18:54,960
Categories though they're really

523
00:18:52,720 --> 00:18:56,960
powerful and like all models they carry

524
00:18:54,960 --> 00:18:58,559
the power to shape the world and people

525
00:18:56,960 --> 00:19:00,240
use categories quite validly as

526
00:18:58,559 --> 00:19:02,480
shorthands for many things. As we said

527
00:19:00,240 --> 00:19:03,919
before we have to sort of like delegate

528
00:19:02,480 --> 00:19:06,640
delegate and trust each other and

529
00:19:03,919 --> 00:19:08,960
operate by some kind of consensus to get

530
00:19:06,640 --> 00:19:11,520
through life. But categories are

531
00:19:08,960 --> 00:19:13,840
arbitrary and sticky and political. And

532
00:19:11,520 --> 00:19:16,320
at their best, they are a way to

533
00:19:13,840 --> 00:19:18,080
understand something about the world. At

534
00:19:16,320 --> 00:19:20,400
their worst, they lead to oppressions

535
00:19:18,080 --> 00:19:21,919
and human rights abuses. And when people

536
00:19:20,400 --> 00:19:24,080
mistake them for objective truth,

537
00:19:21,919 --> 00:19:25,919
especially when people use it as a

538
00:19:24,080 --> 00:19:27,919
weapon to control the world and force it

539
00:19:25,919 --> 00:19:29,440
to fit that descriptor, it gets really

540
00:19:27,919 --> 00:19:30,799
bad.

541
00:19:29,440 --> 00:19:32,400
The anthropologist and political

542
00:19:30,799 --> 00:19:34,799
scientist James C. Scott wrote about

543
00:19:32,400 --> 00:19:36,240
this very extensively in his 1998 book,

544
00:19:34,799 --> 00:19:37,919
Seeing Like a State, which goes really

545
00:19:36,240 --> 00:19:39,679
deeply into how governments, companies,

546
00:19:37,919 --> 00:19:42,000
and other groups have attempted to

547
00:19:39,679 --> 00:19:43,760
impose categories on the world at scale.

548
00:19:42,000 --> 00:19:45,520
The goal, whether consciously or

549
00:19:43,760 --> 00:19:48,160
unconsciously expressed, is to make the

550
00:19:45,520 --> 00:19:49,520
world legible. I think it was

551
00:19:48,160 --> 00:19:51,360
Jordi Miller put it in his lightning

552
00:19:49,520 --> 00:19:53,440
talk yesterday, you alter reality to

553
00:19:51,360 --> 00:19:55,440
match the software.

554
00:19:53,440 --> 00:19:57,120
This makes a lot of things very easy to

555
00:19:55,440 --> 00:19:58,720
manage, much easier to manage,

556
00:19:57,120 --> 00:20:00,640
especially when it comes to deciding how

557
00:19:58,720 --> 00:20:02,720
to allocate resources or plan for the

558
00:20:00,640 --> 00:20:04,160
future at the scale of a country, how to

559
00:20:02,720 --> 00:20:06,240
ensure that everyone in that country has

560
00:20:04,160 --> 00:20:08,000
the best chance to receive basic medical

561
00:20:06,240 --> 00:20:09,600
care and an education and how to react

562
00:20:08,000 --> 00:20:11,760
and respond effectively and efficiently

563
00:20:09,600 --> 00:20:14,160
to disasters. But it doesn't always

564
00:20:11,760 --> 00:20:16,320
necessarily improve things. In making

565
00:20:14,160 --> 00:20:19,200
the world legible to a government or to

566
00:20:16,320 --> 00:20:20,799
a computer, it flattens and erases local

567
00:20:19,200 --> 00:20:22,480
knowledge and complexity. complexity

568
00:20:20,799 --> 00:20:25,200
that was often doing really important

569
00:20:22,480 --> 00:20:27,200
work in the context where it came up and

570
00:20:25,200 --> 00:20:29,039
it's often really hard to understand and

571
00:20:27,200 --> 00:20:30,799
measure outside of that context. So it

572
00:20:29,039 --> 00:20:32,480
gets discarded as unimportant or like

573
00:20:30,799 --> 00:20:34,080
annoying.

574
00:20:32,480 --> 00:20:36,320
Scott gives a lot of examples in his

575
00:20:34,080 --> 00:20:38,960
book including one about scientific

576
00:20:36,320 --> 00:20:42,080
forestry in late 18 late 18th century

577
00:20:38,960 --> 00:20:43,600
Prussia and Saxony.

578
00:20:42,080 --> 00:20:46,320
There were these diverse old growth

579
00:20:43,600 --> 00:20:48,720
forests in these areas and been used for

580
00:20:46,320 --> 00:20:51,440
centuries by the local communities from

581
00:20:48,720 --> 00:20:54,720
everything to medicine from everything

582
00:20:51,440 --> 00:20:56,240
from medicines to like animal fod to

583
00:20:54,720 --> 00:20:57,760
building materials and food like they

584
00:20:56,240 --> 00:20:59,520
were hunting grounds the whole thing.

585
00:20:57,760 --> 00:21:01,840
But when the state's wood supply started

586
00:20:59,520 --> 00:21:03,280
to decline, the government decided to

587
00:21:01,840 --> 00:21:05,440
systematically replace the forest with

588
00:21:03,280 --> 00:21:06,799
like the ordered rows of single species

589
00:21:05,440 --> 00:21:09,679
timber plantations. I think it was like

590
00:21:06,799 --> 00:21:12,960
a Norway spruce. And the state could now

591
00:21:09,679 --> 00:21:14,720
read the forest from like tables in an

592
00:21:12,960 --> 00:21:16,640
office because they planted them

593
00:21:14,720 --> 00:21:18,799
specifically in those ways. They got rid

594
00:21:16,640 --> 00:21:20,400
of the old one. This also enabled them

595
00:21:18,799 --> 00:21:23,280
to predict things like yields and

596
00:21:20,400 --> 00:21:25,440
revenues with really precise

597
00:21:23,280 --> 00:21:28,799
calculations. And for about 80 years,

598
00:21:25,440 --> 00:21:30,880
this approach was extremely successful.

599
00:21:28,799 --> 00:21:33,280
These plantations produced like uniform

600
00:21:30,880 --> 00:21:34,880
highquality timber with unprecedented

601
00:21:33,280 --> 00:21:36,799
efficiency. But once the first lot of

602
00:21:34,880 --> 00:21:39,200
trees had been harvested and the second

603
00:21:36,799 --> 00:21:42,960
rotation had been established, it became

604
00:21:39,200 --> 00:21:45,200
really clear that these like whole

605
00:21:42,960 --> 00:21:46,720
simplified forests were just collapsing.

606
00:21:45,200 --> 00:21:49,039
The second lot were not growing as well

607
00:21:46,720 --> 00:21:50,880
as the first lot did because the complex

608
00:21:49,039 --> 00:21:52,960
ecological relationships between the

609
00:21:50,880 --> 00:21:54,640
local communities and the diversity of

610
00:21:52,960 --> 00:21:56,880
the plant and animal species and the old

611
00:21:54,640 --> 00:21:59,440
growth forest had been sustaining the

612
00:21:56,880 --> 00:22:01,039
soil health and the nutrient cycling and

613
00:21:59,440 --> 00:22:03,360
pest resistance and all that sort of

614
00:22:01,039 --> 00:22:05,679
stuff. The new forests, especially that

615
00:22:03,360 --> 00:22:09,120
first one, had been living off what the

616
00:22:05,679 --> 00:22:12,240
old forests had built up over centuries.

617
00:22:09,120 --> 00:22:14,799
And once that inheritance like ran out,

618
00:22:12,240 --> 00:22:16,480
timber yields crashed, the forests died.

619
00:22:14,799 --> 00:22:18,559
So what looked like irrelevant

620
00:22:16,480 --> 00:22:20,159
complexity to state planners was

621
00:22:18,559 --> 00:22:22,880
actually the foundation that made the

622
00:22:20,159 --> 00:22:24,720
whole system work. It's not just forests

623
00:22:22,880 --> 00:22:26,400
though. It when powerful groups create

624
00:22:24,720 --> 00:22:28,480
categories. And when there are strong

625
00:22:26,400 --> 00:22:30,400
incentives to fit into those categories

626
00:22:28,480 --> 00:22:32,400
or serious disadvantages for standing

627
00:22:30,400 --> 00:22:34,240
out, people in places start to reshape

628
00:22:32,400 --> 00:22:35,919
themselves to match the description.

629
00:22:34,240 --> 00:22:37,360
Like there's a particular kind of

630
00:22:35,919 --> 00:22:38,640
aesthetic. It's, you know, industrial

631
00:22:37,360 --> 00:22:40,240
lighting and reclaimed wood and

632
00:22:38,640 --> 00:22:42,080
minimalist furniture. You know, the one

633
00:22:40,240 --> 00:22:43,520
it's um probably a couple of Mona

634
00:22:42,080 --> 00:22:47,039
plants, the whole thing. It's it's

635
00:22:43,520 --> 00:22:50,320
coffee shops and Airbnbs and, you know,

636
00:22:47,039 --> 00:22:53,520
all of that co-working spaces.

637
00:22:50,320 --> 00:22:55,120
Um, the Verge journalist Kyle Chika

638
00:22:53,520 --> 00:22:56,960
decided to call this Airspace. And this

639
00:22:55,120 --> 00:23:00,000
is back in 2016 when it was already like

640
00:22:56,960 --> 00:23:01,679
a this thing. It's everywhere. Why? And

641
00:23:00,000 --> 00:23:04,720
he needed this name for this sort of

642
00:23:01,679 --> 00:23:07,440
homogenized interior design vibe because

643
00:23:04,720 --> 00:23:09,039
it had been popping up in his view all

644
00:23:07,440 --> 00:23:11,760
over the planet without any connection

645
00:23:09,039 --> 00:23:14,880
to local context. There was it looked

646
00:23:11,760 --> 00:23:18,880
the same in Barcelona as it did in New

647
00:23:14,880 --> 00:23:21,840
York as it did in Istanbul as it did in

648
00:23:18,880 --> 00:23:23,919
Shanghai like everywhere. Um and he

649
00:23:21,840 --> 00:23:25,840
found that this aesthetic once he

650
00:23:23,919 --> 00:23:27,360
started to dig into it was pretty much

651
00:23:25,840 --> 00:23:29,600
down to platform based algorithmic

652
00:23:27,360 --> 00:23:31,520
trends. These kind of spaces photograph

653
00:23:29,600 --> 00:23:33,200
really well for Instagram and they

654
00:23:31,520 --> 00:23:34,720
appeal to the demographic that uses

655
00:23:33,200 --> 00:23:36,720
these kinds of platforms and they get

656
00:23:34,720 --> 00:23:38,400
recommended by the same apps that guide

657
00:23:36,720 --> 00:23:39,919
people to discover places to visit and

658
00:23:38,400 --> 00:23:41,520
so more people made the things look like

659
00:23:39,919 --> 00:23:43,919
that so that they would also get

660
00:23:41,520 --> 00:23:46,000
favored. Spotify's genre categories work

661
00:23:43,919 --> 00:23:47,840
in a similar way for music. Spotify will

662
00:23:46,000 --> 00:23:49,280
commission and generate its own music

663
00:23:47,840 --> 00:23:51,600
specifically to fit the vibe of a

664
00:23:49,280 --> 00:23:53,600
particular playlist. But as the

665
00:23:51,600 --> 00:23:55,520
journalist Liz P documented in her

666
00:23:53,600 --> 00:23:57,120
recent book Mood Machine, the platform

667
00:23:55,520 --> 00:23:58,559
also uses their artist analytic

668
00:23:57,120 --> 00:24:00,960
dashboards. the ones that the artists

669
00:23:58,559 --> 00:24:02,640
themselves are using to nudge the actual

670
00:24:00,960 --> 00:24:04,799
musicians to write and market their

671
00:24:02,640 --> 00:24:06,880
music to fit more closely into these

672
00:24:04,799 --> 00:24:08,400
categories with the aim of increasing

673
00:24:06,880 --> 00:24:10,480
their chances of ending up on these

674
00:24:08,400 --> 00:24:12,480
playlists which has become essential for

675
00:24:10,480 --> 00:24:15,600
any kind of visibility and therefore

676
00:24:12,480 --> 00:24:17,600
income on that platform. So if we make

677
00:24:15,600 --> 00:24:21,960
reality fit this we get sort of like

678
00:24:17,600 --> 00:24:21,960
cultural equivalent of beige.

679
00:24:24,720 --> 00:24:30,320
This one is the main trap. It's not

680
00:24:26,960 --> 00:24:32,480
actually useful or desirable to do this.

681
00:24:30,320 --> 00:24:34,320
For one thing, we'll be here literally

682
00:24:32,480 --> 00:24:36,080
forever if we try. As I said earlier,

683
00:24:34,320 --> 00:24:37,360
reality has no inherent structure.

684
00:24:36,080 --> 00:24:39,440
There's only the structures that we give

685
00:24:37,360 --> 00:24:41,279
it. This mindset also leads to

686
00:24:39,440 --> 00:24:42,799
overcapture of data, the indiscriminate

687
00:24:41,279 --> 00:24:44,400
inclusion of lower quality and

688
00:24:42,799 --> 00:24:46,240
extraneous data, bad decisions,

689
00:24:44,400 --> 00:24:48,960
inflexible systems, surveillance,

690
00:24:46,240 --> 00:24:50,480
capitalism. For developers, this can

691
00:24:48,960 --> 00:24:52,240
also be really paralyzing. I've

692
00:24:50,480 --> 00:24:54,240
definitely started projects only to get

693
00:24:52,240 --> 00:24:55,840
completely hung up on trying to design a

694
00:24:54,240 --> 00:24:58,080
system that will gracefully handle as

695
00:24:55,840 --> 00:24:59,919
many of my potential future use cases as

696
00:24:58,080 --> 00:25:01,600
possible. And sometimes I get so stuck

697
00:24:59,919 --> 00:25:03,520
there that I never end up building

698
00:25:01,600 --> 00:25:05,039
anything.

699
00:25:03,520 --> 00:25:06,400
This doesn't mean that it is useless to

700
00:25:05,039 --> 00:25:07,840
make models. We just have to be aware

701
00:25:06,400 --> 00:25:09,520
that what we are doing is making a

702
00:25:07,840 --> 00:25:11,760
representation of the world for our

703
00:25:09,520 --> 00:25:14,080
specific purpose. And once you get

704
00:25:11,760 --> 00:25:16,640
there, that can be pretty liberating.

705
00:25:14,080 --> 00:25:18,640
Once I remember that this thing only has

706
00:25:16,640 --> 00:25:20,159
to work for my purpose right now and I

707
00:25:18,640 --> 00:25:22,240
take the trouble to articulate what the

708
00:25:20,159 --> 00:25:24,080
purpose actually is, I am able to move

709
00:25:22,240 --> 00:25:25,600
more confidently and I'm much less

710
00:25:24,080 --> 00:25:27,679
precious about changing things later if

711
00:25:25,600 --> 00:25:29,039
my current tools don't fit the job. So

712
00:25:27,679 --> 00:25:31,039
let's talk about abstractions. We get

713
00:25:29,039 --> 00:25:34,559
into some principles for how to use them

714
00:25:31,039 --> 00:25:37,120
mindfully and well.

715
00:25:34,559 --> 00:25:39,840
Abstractions are tools, not mirrors.

716
00:25:37,120 --> 00:25:42,400
Your tools are there to be useful, not

717
00:25:39,840 --> 00:25:44,159
perfect. As Bill Kent says, he wrote a

718
00:25:42,400 --> 00:25:45,279
book called Data and Reality. The second

719
00:25:44,159 --> 00:25:48,159
edition is the one you should really

720
00:25:45,279 --> 00:25:51,200
check out. It's 1978, I think, and it is

721
00:25:48,159 --> 00:25:53,760
very good. Um, there is an important

722
00:25:51,200 --> 00:25:55,679
difference between truth and utility. We

723
00:25:53,760 --> 00:25:56,720
want things that are useful, at least in

724
00:25:55,679 --> 00:25:58,960
this business, otherwise we'd be

725
00:25:56,720 --> 00:26:01,760
philosophers and artists.

726
00:25:58,960 --> 00:26:03,600
We are only working with computers. We

727
00:26:01,760 --> 00:26:04,960
have a task to do with what we're

728
00:26:03,600 --> 00:26:07,120
coding. And if we don't, we should ask

729
00:26:04,960 --> 00:26:11,480
ourselves what we're doing. We need to

730
00:26:07,120 --> 00:26:11,480
leave the hard stuff to the artists.

731
00:26:12,000 --> 00:26:16,400
We should design with context and with

732
00:26:13,760 --> 00:26:19,120
purpose. Every abstraction should start

733
00:26:16,400 --> 00:26:21,039
a clear start with a clear sense of why

734
00:26:19,120 --> 00:26:23,360
you are building it and what you need it

735
00:26:21,039 --> 00:26:25,679
to do. Thinking in systems by Dana

736
00:26:23,360 --> 00:26:27,279
Meadows is full of so many quotes that I

737
00:26:25,679 --> 00:26:29,600
could have put in many places in this

738
00:26:27,279 --> 00:26:32,000
talk. Um but I've restrained myself to

739
00:26:29,600 --> 00:26:33,600
this one. There are no separate systems.

740
00:26:32,000 --> 00:26:35,679
The world is a continuum. Where to draw

741
00:26:33,600 --> 00:26:37,520
a boundary around a system depends on

742
00:26:35,679 --> 00:26:39,520
the purpose of the discussion and the

743
00:26:37,520 --> 00:26:40,880
questions that we want to ask. This

744
00:26:39,520 --> 00:26:42,640
means being explicit about your

745
00:26:40,880 --> 00:26:44,559
constraints. What problem are you

746
00:26:42,640 --> 00:26:47,039
solving for whom? Under what conditions?

747
00:26:44,559 --> 00:26:48,640
What are you optimizing for? The same

748
00:26:47,039 --> 00:26:49,919
piece of reality might need completely

749
00:26:48,640 --> 00:26:51,440
different abstractions depending on

750
00:26:49,919 --> 00:26:53,520
whether you're trying to track inventory

751
00:26:51,440 --> 00:26:55,760
or understand user behavior or plan a

752
00:26:53,520 --> 00:26:57,840
budget. Designed for your actual

753
00:26:55,760 --> 00:27:00,000
purpose, not for some imagined perfect

754
00:26:57,840 --> 00:27:02,960
system.

755
00:27:00,000 --> 00:27:05,600
And abstractions have consequences. This

756
00:27:02,960 --> 00:27:08,000
doesn't mean that we magically avoid the

757
00:27:05,600 --> 00:27:09,360
inherent issues with measuring the

758
00:27:08,000 --> 00:27:10,480
world. Just because we acknowledge that

759
00:27:09,360 --> 00:27:11,760
we're working with abstractions of

760
00:27:10,480 --> 00:27:13,200
reality, that doesn't mean that they

761
00:27:11,760 --> 00:27:15,120
don't have consequences. We've covered

762
00:27:13,200 --> 00:27:17,279
this off. The choices we make about what

763
00:27:15,120 --> 00:27:19,120
we measure and how we name it have an

764
00:27:17,279 --> 00:27:21,760
impact on the world. There are both

765
00:27:19,120 --> 00:27:24,559
technical decisions and also sometimes

766
00:27:21,760 --> 00:27:26,320
decisions that affect people's lives and

767
00:27:24,559 --> 00:27:28,320
opportunities and access to resources,

768
00:27:26,320 --> 00:27:29,520
which means we have to be thoughtful. We

769
00:27:28,320 --> 00:27:31,840
should ask when we're designing our

770
00:27:29,520 --> 00:27:34,400
tools, who benefits from this design?

771
00:27:31,840 --> 00:27:36,559
who might be excluded or misrepresented,

772
00:27:34,400 --> 00:27:39,520
what behavior is this design likely to

773
00:27:36,559 --> 00:27:41,039
encourage? And sometimes the answer is

774
00:27:39,520 --> 00:27:42,720
it's fine. We're just trying to track

775
00:27:41,039 --> 00:27:44,960
what laptops in our inventory need

776
00:27:42,720 --> 00:27:47,760
repairs. But sometimes it is bigger than

777
00:27:44,960 --> 00:27:49,360
that. So yeah, we're we're never going

778
00:27:47,760 --> 00:27:50,799
to be able to build perfect

779
00:27:49,360 --> 00:27:52,400
representations of reality. We're going

780
00:27:50,799 --> 00:27:54,320
to keep running into edge cases and

781
00:27:52,400 --> 00:27:55,760
categories that don't quite fit and data

782
00:27:54,320 --> 00:27:58,159
that tells us conflicting things about

783
00:27:55,760 --> 00:28:00,640
the world. And I am really here to say

784
00:27:58,159 --> 00:28:02,240
basically that that's that's okay. Once

785
00:28:00,640 --> 00:28:03,679
we stop aiming to capture reality

786
00:28:02,240 --> 00:28:05,919
completely, we can start building things

787
00:28:03,679 --> 00:28:07,919
that are useful. We can focus on solving

788
00:28:05,919 --> 00:28:09,279
problems instead of getting stuck in

789
00:28:07,919 --> 00:28:11,200
debates about whether a hot dog is a

790
00:28:09,279 --> 00:28:13,600
sandwich, unless that is your thing. I'm

791
00:28:11,200 --> 00:28:15,200
not here to judge. We can build

792
00:28:13,600 --> 00:28:16,720
abstractions that work well enough for

793
00:28:15,200 --> 00:28:18,240
their purposes, that acknowledge their

794
00:28:16,720 --> 00:28:20,720
limitations, that treat the people who

795
00:28:18,240 --> 00:28:23,440
have to live with them with respect. The

796
00:28:20,720 --> 00:28:25,919
map isn't the territory, but a decent

797
00:28:23,440 --> 00:28:28,399
map built mindfully and with a purpose

798
00:28:25,919 --> 00:28:29,380
can get you where you need to go.

799
00:28:28,399 --> 00:28:39,949
Thanks,

800
00:28:29,380 --> 00:28:39,949
[Applause]