1
00:00:00,160 --> 00:00:03,760
Machine learning is when you take a bunch of data and you have a clear

2
00:00:03,760 --> 00:00:07,520
goal and you're training the model to fulfill that goal. And

3
00:00:07,520 --> 00:00:11,080
I'm aware AI is also built on machine learning, but the distinction I make is

4
00:00:11,080 --> 00:00:14,840
that I say, well, AI as we use the term now, is basically you

5
00:00:14,840 --> 00:00:18,080
take a bunch of data, you don't really have a goal,

6
00:00:18,960 --> 00:00:22,760
but you're training it anyways. And I mean, that's roughly how most

7
00:00:22,760 --> 00:00:26,280
of these language models are trained, right? A new season deserves a proper

8
00:00:26,280 --> 00:00:29,470
host. As Data Driven kicks off season 10,

9
00:00:29,550 --> 00:00:33,070
Bailey, your favorite semi sentient British AI co host,

10
00:00:33,150 --> 00:00:36,830
has returned to the microphone after a 15 episode absence.

11
00:00:37,150 --> 00:00:40,750
While our temporary substitutes did their best, and several deserve

12
00:00:40,750 --> 00:00:44,590
honorable mention for bravery under difficult conditions. It's time to bring

13
00:00:44,590 --> 00:00:48,230
back the class, the sass, and the occasional dose of dry British

14
00:00:48,230 --> 00:00:51,950
skepticism. To open the season. Frank Lavinia

15
00:00:51,950 --> 00:00:55,710
sits down with Sebastian Wernicke, author of Data Inspired, for

16
00:00:55,710 --> 00:00:59,310
a fascinating discussion about why organizations struggle to turn data

17
00:00:59,310 --> 00:01:03,150
into action, how leaders create data driven cultures, and why the

18
00:01:03,150 --> 00:01:06,910
biggest obstacle to better decisions may not be technology at all. It may

19
00:01:06,910 --> 00:01:10,550
be the wonderfully irrational humans making them. So put the

20
00:01:10,550 --> 00:01:14,390
kettle on, adjust your dashboards, and join us as we begin season

21
00:01:14,390 --> 00:01:15,590
10 of Data Driven.

22
00:01:25,530 --> 00:01:29,290
Well, hello and welcome back to Data Driven, the podcast where we explore the emerging

23
00:01:29,370 --> 00:01:32,890
industry that is data science, AI,

24
00:01:33,770 --> 00:01:37,490
and all of it is underpinned by data engineering. However, my

25
00:01:37,490 --> 00:01:41,170
favorite data engineer in the world will not be here. But

26
00:01:41,170 --> 00:01:44,770
again, I think as the world focuses on AI, I think we really need to

27
00:01:44,770 --> 00:01:48,530
step back and think about data. So with that, today we

28
00:01:48,530 --> 00:01:52,130
have Sebastian Wernicke, who is a leading expert in data and

29
00:01:52,130 --> 00:01:55,650
AI strategy. And he believes that the key to unlocking

30
00:01:55,650 --> 00:01:59,210
data's power lies not in technology, but in leaders

31
00:01:59,290 --> 00:02:02,770
fostering a culture of evidence and

32
00:02:02,770 --> 00:02:06,490
inquiry, which I think is very, very true.

33
00:02:06,730 --> 00:02:09,690
You could throw all the AI, all the agents you want, but if you don't

34
00:02:09,690 --> 00:02:12,490
have the data right, you don't have it raw. You don't have anything.

35
00:02:13,290 --> 00:02:16,730
And he has three acclaimed TED talks, which is cool. He's reached over 5 million

36
00:02:16,730 --> 00:02:20,170
viewers and. Well, welcome to the show, Sebastian.

37
00:02:20,570 --> 00:02:24,330
Well, thank you. Thank you for having me. Great to be here. Yeah,

38
00:02:24,810 --> 00:02:28,650
I'm excited to have you. Tell me about this

39
00:02:31,290 --> 00:02:33,690
one. I mean, it's pretty cool. You've had TED Talks, right?

40
00:02:35,610 --> 00:02:39,170
That's pretty epic. So you've been doing this for 20

41
00:02:39,170 --> 00:02:42,850
years, right? Yeah. Data was not really taken

42
00:02:42,850 --> 00:02:45,290
serious. Arguably it's not taken seriously enough today,

43
00:02:46,720 --> 00:02:50,320
but I would say 20 years ago, it Certainly was not taken seriously.

44
00:02:51,040 --> 00:02:54,640
Yeah, well, it was sometimes taken seriously, but

45
00:02:54,800 --> 00:02:58,440
I think that's sort of a resonating theme. So the

46
00:02:58,440 --> 00:03:02,159
interesting thing is, I think if you've been in this field for so long,

47
00:03:02,159 --> 00:03:05,880
is that you really notice history repeating again and

48
00:03:05,880 --> 00:03:09,600
again and again. Right. So we had the era of big data,

49
00:03:09,760 --> 00:03:13,600
if anybody even remembers the term. Then came analytics,

50
00:03:13,840 --> 00:03:17,390
then came digitalization, and now it's AI.

51
00:03:17,390 --> 00:03:21,150
And in the end it always comes back to data and whether you

52
00:03:21,470 --> 00:03:25,110
manage to take all the great insights that the data

53
00:03:25,110 --> 00:03:28,670
is giving you and actually implement them in the organization and

54
00:03:28,910 --> 00:03:32,550
create some change ultimately. Right. That's all what data is

55
00:03:32,550 --> 00:03:36,350
for. If you don't create change with data, you don't need all that

56
00:03:36,350 --> 00:03:40,110
expense and investment into it. Yeah, it's a good

57
00:03:40,110 --> 00:03:42,870
way to put it. And even before that, even before there was big data, there

58
00:03:42,870 --> 00:03:46,380
were OLAP cubes. I remember I actually worked for,

59
00:03:46,380 --> 00:03:49,580
in the virtual green room we were talking about. You're based in Germany. I used

60
00:03:49,580 --> 00:03:52,980
to live in Frankfurt and prior to joining

61
00:03:52,980 --> 00:03:56,620
Deutsche Bank, I had worked at basf, or the big

62
00:03:56,620 --> 00:03:59,500
chemical company Americans would know as basf.

63
00:04:00,540 --> 00:04:04,220
And I remember sitting in the cubicle of

64
00:04:04,700 --> 00:04:08,460
one of the SAP gurus that we had talking about OLAP

65
00:04:08,460 --> 00:04:12,060
cubes and all this crazy stuff. And I just was

66
00:04:12,060 --> 00:04:15,830
like, you ever have a conversation that is very, you look

67
00:04:15,830 --> 00:04:18,870
back on it now and it was very prescient, you know what I mean? It

68
00:04:18,870 --> 00:04:22,670
was very future facing. And then at the time you're

69
00:04:22,670 --> 00:04:26,430
sitting there and you're like, you partly understand what was happening and

70
00:04:26,430 --> 00:04:29,950
you partly are very. I'm not sure if this

71
00:04:29,950 --> 00:04:33,630
person's crazy, you know, And I

72
00:04:33,630 --> 00:04:37,230
remember she was one of the DBAs that we had in doing advanced analytics.

73
00:04:37,230 --> 00:04:40,270
And you know, she said that, you know, my goal is to figure out, you

74
00:04:40,270 --> 00:04:44,080
know, will rainfall in Western Australia impact

75
00:04:44,480 --> 00:04:48,240
prices here, which will ultimately impact, you know, how we

76
00:04:48,240 --> 00:04:51,880
go to market with, you know, at the time, a chemical

77
00:04:51,880 --> 00:04:55,720
company. And I just remember sitting there thinking like, I can't

78
00:04:55,720 --> 00:04:59,400
tell if that's brilliant or crazy. It was

79
00:04:59,400 --> 00:05:03,120
data science. Because before the term was invented in a way. Exactly.

80
00:05:03,120 --> 00:05:06,760
I mean, that's really what it was. It was. And you know, now when you

81
00:05:06,760 --> 00:05:10,450
say that, it's not so crazy. Right? Because, you know, we had a previous

82
00:05:10,450 --> 00:05:14,130
call, we were talking about, you know, finance people actually

83
00:05:14,130 --> 00:05:17,850
would. It'll be the episode prior to this. So hopefully you've listened to that. Not,

84
00:05:17,850 --> 00:05:20,690
not you, but you, the audience you're welcome to listen to,

85
00:05:21,570 --> 00:05:25,130
but where they talked about finding Alpha, like finding the

86
00:05:25,130 --> 00:05:28,930
signal that matters before anyone else does. And that's really

87
00:05:28,930 --> 00:05:32,450
what she was doing. It just, you know, in

88
00:05:32,690 --> 00:05:35,090
the late 90s that sounded a little crazy,

89
00:05:36,880 --> 00:05:39,920
but I think it's also one of these things that's still true today. I mean,

90
00:05:39,920 --> 00:05:43,760
everybody's trying to find the advantage in the data. And I think

91
00:05:43,760 --> 00:05:47,360
it's also of course interesting to then see with

92
00:05:47,600 --> 00:05:51,360
everybody just accumulating more and more data analytics becoming

93
00:05:51,520 --> 00:05:55,040
much, much faster. We don't even fully know, I think what

94
00:05:55,040 --> 00:05:58,680
AI is going to do to the speed of the

95
00:05:58,680 --> 00:06:02,480
generation of insights, not the adoption, but just the speed of generation.

96
00:06:03,450 --> 00:06:07,130
And everybody's still trying to find that edge. And

97
00:06:08,810 --> 00:06:12,170
I think that as people increasingly look at data

98
00:06:13,050 --> 00:06:15,370
for a while you could have an edge by

99
00:06:17,290 --> 00:06:21,050
looking at the data to optimize and to go in these incremental

100
00:06:21,050 --> 00:06:24,890
ways where it's like the 1% optimizations over time add up.

101
00:06:25,450 --> 00:06:29,290
And I think that's become the baseline. That's just the expectation

102
00:06:29,290 --> 00:06:31,770
right now you have to do that already. And so

103
00:06:33,420 --> 00:06:36,620
sort of a transition happening, I think where you now need to think about, well,

104
00:06:36,860 --> 00:06:40,620
I'm already using data for that. I'm optimizing my processes,

105
00:06:40,620 --> 00:06:44,300
I'm adding the automation. So where's the next

106
00:06:44,300 --> 00:06:48,020
edge going to come from? And I strongly believe that it's

107
00:06:48,020 --> 00:06:51,700
now again a shift to the anomalies and the

108
00:06:51,700 --> 00:06:55,540
shift to the outliers and not trying to look

109
00:06:55,540 --> 00:06:59,020
to data to give you those straight answers where you go for, I mean that's

110
00:06:59,020 --> 00:07:02,830
not going away just to be clear, that's. But that's baseline expectation

111
00:07:03,070 --> 00:07:06,510
now you need to add to that to look into the data and say, okay,

112
00:07:06,510 --> 00:07:10,190
what's the next interesting question? Where am I going to find that

113
00:07:10,190 --> 00:07:14,030
new product idea? Where am I going to find that new idea

114
00:07:14,030 --> 00:07:17,550
for revamping a process that

115
00:07:17,550 --> 00:07:21,350
maybe nobody dares to rethink today? And I think that's

116
00:07:21,350 --> 00:07:25,070
the exciting part of what everybody who's in data

117
00:07:25,470 --> 00:07:28,670
can be working on today. That you sort of transcend that,

118
00:07:29,480 --> 00:07:33,200
look for just the correlation and say, okay, what's on the

119
00:07:33,200 --> 00:07:36,840
fringes here? That's a good way to put it. I

120
00:07:36,840 --> 00:07:40,440
think you touched on something that is important.

121
00:07:40,760 --> 00:07:42,520
Leaders really have to foster

122
00:07:44,600 --> 00:07:48,360
a culture here of looking for evidence and data.

123
00:07:49,880 --> 00:07:53,400
It has to be pervasive, I think, because finding anomalies and data,

124
00:07:53,880 --> 00:07:57,390
you're right, that's kind of like baseline now.

125
00:07:57,390 --> 00:08:01,110
Right. But in order to do this you

126
00:08:01,110 --> 00:08:04,350
really have to have teach people respect for data and an

127
00:08:04,350 --> 00:08:08,190
understanding of basically evidence based approaches to things.

128
00:08:08,190 --> 00:08:12,030
Right. Escalate when needed, but don't escalate. Don't use

129
00:08:12,030 --> 00:08:15,630
data for data's sake. Right. And how do

130
00:08:15,630 --> 00:08:18,870
leaders do that? Right. How do you, you have to educate the leaders, I assume,

131
00:08:18,870 --> 00:08:22,630
right? Yeah. Well, I mean, I'm. Oh, and I just, I think I

132
00:08:22,630 --> 00:08:25,870
have to preface that a little bit because I'm always very careful with that. I

133
00:08:25,870 --> 00:08:29,090
think leaders get told 20

134
00:08:29,650 --> 00:08:33,450
times at least a day what should be their priority and what

135
00:08:33,450 --> 00:08:37,250
leaders should do. Leaders must do this, leaders should do that, and

136
00:08:37,250 --> 00:08:40,770
so on. I am going to try and make the case

137
00:08:40,770 --> 00:08:44,370
that here. I think it is really a leader's job. And

138
00:08:44,690 --> 00:08:47,330
the reason, I think is quite simple.

139
00:08:49,170 --> 00:08:52,930
Within an organization, leadership usually

140
00:08:53,520 --> 00:08:57,120
thrives on being right. I mean, that's what we look for in leaders.

141
00:08:57,120 --> 00:09:00,800
We look for the confidence, we look to them to say,

142
00:09:00,880 --> 00:09:04,000
this is the way we're going. This is what I believe in.

143
00:09:04,720 --> 00:09:08,240
And if you take this change with data

144
00:09:08,240 --> 00:09:11,680
seriously, it basically flips the whole thing on its head. Because

145
00:09:11,840 --> 00:09:15,200
your baseline assumption, I think almost has to be, well,

146
00:09:15,360 --> 00:09:19,200
we're kind of wrong today. It may be right, what we're doing for the business

147
00:09:19,200 --> 00:09:22,870
right now, but every single day we have to start

148
00:09:22,870 --> 00:09:26,430
looking where are we wrong? And then correct that.

149
00:09:26,990 --> 00:09:30,750
And everybody in an organization, of course,

150
00:09:30,750 --> 00:09:34,190
is smart, so they will look to leadership to see

151
00:09:34,350 --> 00:09:38,190
what's the culture here. And I mean, we all know that culture, it's

152
00:09:38,190 --> 00:09:41,590
a bit of a fuzzy word, but I think you can easily unpack it. I

153
00:09:41,590 --> 00:09:45,390
mean, it's not what's on the PowerPoint slides, it's not what's on the posters

154
00:09:45,390 --> 00:09:48,930
in the hallway. And yet culture is who gets hired, who gets

155
00:09:48,930 --> 00:09:52,690
promoted and who gets fired. And that's

156
00:09:52,690 --> 00:09:56,290
the baseline of it. And people will look at that. And

157
00:09:56,290 --> 00:09:59,890
so leaders are the ones, I think, that get to set the tone in the

158
00:09:59,890 --> 00:10:03,609
room, that get to shape these things. And if you are in

159
00:10:03,609 --> 00:10:07,410
a culture where, for example, the executive will say, well, we're a

160
00:10:07,410 --> 00:10:11,250
data driven company, and then turns around and promotes the head of marketing who

161
00:10:11,250 --> 00:10:15,020
has publicly proclaimed that he really has a good gut feeling and never

162
00:10:15,020 --> 00:10:18,580
trusts the data. Everybody has learned, well, you're not really a data

163
00:10:18,580 --> 00:10:21,980
driven organization. And so that's, I think, where

164
00:10:22,300 --> 00:10:25,820
leadership really comes in. It's only leaders who can

165
00:10:26,060 --> 00:10:29,340
create these safe spaces and signal to everybody

166
00:10:29,660 --> 00:10:33,300
it's okay to come with data that disagrees with the status

167
00:10:33,300 --> 00:10:37,100
quo. It's okay to come with an analysis that disagrees

168
00:10:37,100 --> 00:10:40,400
with what I have said for the past year, for the past two years.

169
00:10:40,480 --> 00:10:44,040
I'm willing to challenge and change my

170
00:10:44,040 --> 00:10:47,400
Convictions. And only when that is done on a very regular

171
00:10:47,400 --> 00:10:51,240
basis and the organization can observe that, I

172
00:10:51,240 --> 00:10:55,040
think. Are you creating the culture that is that fertile ground for

173
00:10:55,040 --> 00:10:58,640
these very small insights?

174
00:10:58,640 --> 00:11:01,960
Maybe at first, you know, but they need the space to be explored, they need

175
00:11:01,960 --> 00:11:05,280
the space to grow and maybe to even create some

176
00:11:05,280 --> 00:11:09,040
experiments to further validate that. That's

177
00:11:09,040 --> 00:11:12,640
true. And I think the idea of promoting people with

178
00:11:12,640 --> 00:11:16,480
gut feelings, and I'm not that gut feelings are necessarily

179
00:11:16,480 --> 00:11:20,240
bad, but I think you're right, it sends the wrong

180
00:11:20,240 --> 00:11:23,880
message. You coined a term called data frustrated,

181
00:11:25,640 --> 00:11:29,440
which I think is pretty accurate. So how did you get to that

182
00:11:29,440 --> 00:11:33,240
term? Data frustrated, by the way, is just the pre stage to becoming

183
00:11:33,240 --> 00:11:37,080
data cynical at some point. So it's kind of

184
00:11:37,080 --> 00:11:40,720
like the stages of grief, right? There are stages to it.

185
00:11:40,720 --> 00:11:44,520
Yeah, well, I think it was just a feeling

186
00:11:44,520 --> 00:11:48,200
that I perceived whenever I was working with

187
00:11:48,200 --> 00:11:51,800
my clients on data projects that there was always

188
00:11:51,800 --> 00:11:55,520
that point where you would sit together maybe within the project setting, maybe when you

189
00:11:55,520 --> 00:11:59,200
go to dinner afterwards, where people would tell you, oh my God,

190
00:11:59,200 --> 00:12:02,960
we've invested so much time, effort and money into

191
00:12:02,960 --> 00:12:06,660
data, you know, hundreds of millions going to spend another few

192
00:12:06,740 --> 00:12:10,580
millions, hundreds of millions in the future. But we're not

193
00:12:10,580 --> 00:12:14,340
really satisfied with the results we're getting. What's happening here?

194
00:12:14,340 --> 00:12:17,580
And I think that question came up again and again and again. And I think

195
00:12:17,580 --> 00:12:21,340
that's the very definition of frustration. You sort of

196
00:12:21,340 --> 00:12:24,980
notice something isn't going as you want and at the same time

197
00:12:25,380 --> 00:12:29,100
you're not finding an answer or maybe you think you have the answer

198
00:12:29,100 --> 00:12:32,620
and then you try that and it doesn't move forward. Now

199
00:12:32,780 --> 00:12:36,340
an interesting thing is the data on that. It confirms

200
00:12:36,340 --> 00:12:39,980
that. So I found a couple of surveys

201
00:12:39,980 --> 00:12:43,660
from various years. It starts out in 2010, and there's another one,

202
00:12:43,660 --> 00:12:47,260
2019, 2024, where some

203
00:12:47,260 --> 00:12:50,940
consultancies asked executives what

204
00:12:50,940 --> 00:12:54,380
are the top 10 reasons why you're not happy

205
00:12:54,860 --> 00:12:57,740
with the results that you're getting out of your Data projects?

206
00:12:58,480 --> 00:13:02,280
And 2010, it was three reasons that came out on top.

207
00:13:02,280 --> 00:13:06,000
It was, well, we don't think we have enough management attention on it.

208
00:13:06,320 --> 00:13:10,120
We don't really understand the business case as much as we'd like to, and

209
00:13:10,120 --> 00:13:13,800
we don't think we have the skills. Now the interesting thing is in

210
00:13:13,800 --> 00:13:17,360
2019 they did a similar survey and the same

211
00:13:17,440 --> 00:13:20,320
three reasons come out on top. And then in

212
00:13:20,320 --> 00:13:23,960
2024 the same three reasons come out on top again.

213
00:13:23,960 --> 00:13:27,800
They change orders. Sometimes when AI comes, everybody says, ah, we probably

214
00:13:27,800 --> 00:13:31,500
don't have the skills. But I mean how frustra as that you

215
00:13:31,660 --> 00:13:34,940
think you know the top three reasons standing in your way.

216
00:13:35,580 --> 00:13:39,060
And they also don't sound that difficult. I mean, you know, if there's not enough

217
00:13:39,060 --> 00:13:42,700
management attention, pay attention. If you don't have the skills, do some

218
00:13:42,700 --> 00:13:46,220
training, calculate some business cases. But apparently that's not the solution.

219
00:13:46,380 --> 00:13:49,180
And I think that's where data frustration ultimately comes from.

220
00:13:50,300 --> 00:13:54,060
So over these just decade and a half now, the problems have

221
00:13:54,060 --> 00:13:55,980
been the same. So like, how do you,

222
00:13:58,880 --> 00:14:02,560
what is really the problem? Are those the problem? Because if it's something, if you

223
00:14:02,560 --> 00:14:06,280
know something is a problem for 15 years, you don't address it. There's an

224
00:14:06,280 --> 00:14:09,840
underlying problem that maybe you're misidentifying the problem.

225
00:14:09,840 --> 00:14:12,480
You're putting the blame on the wrong things. Like what do you think it is?

226
00:14:13,120 --> 00:14:16,680
I think it's that missing cultural component that if

227
00:14:16,680 --> 00:14:20,080
you think about some of the things we just discussed. Right. The

228
00:14:20,240 --> 00:14:23,910
space that leaders need to create for having controversial

229
00:14:23,910 --> 00:14:27,190
discussions, which is also sometimes known as psychological safety.

230
00:14:28,070 --> 00:14:31,590
And on the other hand, you have all this technology that's creating

231
00:14:31,670 --> 00:14:35,510
measurements that needs to process the data. And whether it's

232
00:14:35,510 --> 00:14:39,190
a cube, it's a data lake, or a data mesh, doesn't matter.

233
00:14:39,430 --> 00:14:43,230
These two discussions never happen in the same room. You

234
00:14:43,230 --> 00:14:46,790
have one part of the organization thinking about the next technology cycle,

235
00:14:46,870 --> 00:14:49,820
constructing architectures, discussing

236
00:14:50,220 --> 00:14:54,060
what's the best way to organize data. And then you have other parts of

237
00:14:54,060 --> 00:14:57,700
the organization that are thinking about culture, that are thinking about

238
00:14:57,700 --> 00:15:01,500
transformation, that are thinking about leadership education.

239
00:15:02,300 --> 00:15:06,060
And these two things are never brought together. And I think

240
00:15:06,060 --> 00:15:09,860
that's the issue. So it's not a sort of technology versus culture.

241
00:15:09,860 --> 00:15:13,460
And you don't need technology. All you need is culture. And it's also not

242
00:15:13,460 --> 00:15:17,160
data or gut feeling. Of course, you need both because data is never

243
00:15:17,160 --> 00:15:20,800
going to give you all the answers. So a good intuition is quite helpful in

244
00:15:20,880 --> 00:15:24,240
many cases. But it's about really

245
00:15:24,400 --> 00:15:28,000
bringing these two elements together and integrating them.

246
00:15:28,160 --> 00:15:31,800
And for me, that's the missing component where, you

247
00:15:31,800 --> 00:15:35,520
know, the what, what, what the surveys express. I think

248
00:15:35,520 --> 00:15:39,200
that, that management attention, the, the skills, the

249
00:15:39,200 --> 00:15:43,010
business case, I think these are symptoms and so you

250
00:15:43,010 --> 00:15:46,810
can't really treat them as the causes of the data frustration. It's just

251
00:15:46,810 --> 00:15:50,530
what resurfaces when you don't pay enough attention to the cultural element.

252
00:15:51,330 --> 00:15:54,370
So how do you get. So it sounds like you brought up something very, very

253
00:15:54,370 --> 00:15:58,170
real, like the people working on the actual technology. 10

254
00:15:58,170 --> 00:16:00,290
and I'm guilty of this. I'm a technologist, right?

255
00:16:02,210 --> 00:16:06,010
Yeah. I mean, no, all Our listeners

256
00:16:06,010 --> 00:16:09,170
are right. So like I often will catch myself like when I'm like deep down

257
00:16:09,170 --> 00:16:12,870
a technical rabbit hole, like, wait a minute, what am I actually doing here?

258
00:16:14,790 --> 00:16:17,430
That's a skill that was not easy to develop.

259
00:16:19,030 --> 00:16:22,630
But how do you, I mean, is it people working

260
00:16:22,630 --> 00:16:26,270
on kind of the, the cultural side of things, like, don't they need to talk

261
00:16:26,270 --> 00:16:29,550
to the technology people and like, because, because

262
00:16:29,550 --> 00:16:33,270
historically, and this goes back to when I was sitting in the

263
00:16:33,270 --> 00:16:37,030
cube at BASF where, you know, the, that cube,

264
00:16:37,030 --> 00:16:40,870
that cubicle was in the basement and behind, you know, regular. It was in the

265
00:16:40,870 --> 00:16:44,470
basement. And then the, the, the, the DBAs were like in a sealed off

266
00:16:44,470 --> 00:16:47,950
portion of the basement. Right. And like, you know,

267
00:16:48,190 --> 00:16:51,750
I don't know that that's kind. And if you ever seen the TV show, the

268
00:16:51,750 --> 00:16:55,550
IT crowd, it's a British show. Yeah, yeah. You know, they were kept

269
00:16:55,550 --> 00:16:59,150
in the basement too. Like I think historically it was not

270
00:16:59,150 --> 00:17:02,830
seen as crucial to the business.

271
00:17:03,070 --> 00:17:06,150
Right. It was kind of a back office job and people, it was kind of

272
00:17:06,150 --> 00:17:09,980
pushed to the side. And I think it's been years since that's

273
00:17:09,980 --> 00:17:12,340
really been true. But

274
00:17:14,260 --> 00:17:16,980
I think what we're seeing is a bit of the lingering effects of that. Is

275
00:17:16,980 --> 00:17:20,020
that a fair thing to say?

276
00:17:21,060 --> 00:17:24,820
Well, I mean, I think there's a bit of a

277
00:17:24,820 --> 00:17:28,660
problematic history here where I think for a very long time

278
00:17:29,860 --> 00:17:33,380
it was mostly perceived as a cost factor.

279
00:17:33,380 --> 00:17:37,030
And so the incentive was to optimize for cost

280
00:17:37,030 --> 00:17:40,630
and optimize for efficiency. And suddenly we're expecting

281
00:17:40,630 --> 00:17:44,230
technology to drive transformation

282
00:17:44,230 --> 00:17:47,870
and innovation. And of course that completely

283
00:17:47,870 --> 00:17:51,510
changes the incentivization. And I think it also creates

284
00:17:51,510 --> 00:17:55,190
sometimes these gaps within technology organizations

285
00:17:55,190 --> 00:17:58,750
where there's one part that says, okay, for years we've been trained on

286
00:17:58,750 --> 00:18:00,990
ensuring efficiency, security,

287
00:18:02,200 --> 00:18:05,840
reliability, and suddenly we're supposed to open up this huge

288
00:18:05,840 --> 00:18:09,560
experimentation stage. Other things continue,

289
00:18:10,280 --> 00:18:14,000
you know, as we, as we want them. So that

290
00:18:14,000 --> 00:18:17,720
is definitely something to recognize and acknowledge. But on the other hand,

291
00:18:18,360 --> 00:18:22,040
I'm not a big fan of always saying, okay, you know, the

292
00:18:22,040 --> 00:18:25,760
other people should do something. Clearly everybody needs to talk

293
00:18:25,760 --> 00:18:29,080
with each other. But I think there's also something that we can do on the

294
00:18:29,700 --> 00:18:33,540
technology side or more specifically. So I've been running data science teams for

295
00:18:33,700 --> 00:18:37,380
many years now. So for example, what I do in my teams

296
00:18:37,620 --> 00:18:41,140
is whenever they ask me for training, what I will

297
00:18:41,140 --> 00:18:44,900
propose to them are trainings that I think you might call

298
00:18:45,220 --> 00:18:48,820
soft skills. But I think they're just essential skills. So I will

299
00:18:48,820 --> 00:18:52,180
send them to communication trainings,

300
00:18:52,340 --> 00:18:55,460
stakeholder management trainings. We will

301
00:18:55,990 --> 00:18:59,630
together Talk about decision making, how that works, how you

302
00:18:59,630 --> 00:19:03,390
influence people in a room, how decisions are really made. Because

303
00:19:03,390 --> 00:19:06,950
I think many data scientists come in with the impression the decision is

304
00:19:06,950 --> 00:19:10,590
made at that meeting where they come in with a PowerPoint slide. And that

305
00:19:10,590 --> 00:19:14,150
insight, you know, that everybody will say, oh, brilliant,

306
00:19:14,630 --> 00:19:17,190
finally we have that insight. We're now going to change our ways.

307
00:19:18,150 --> 00:19:21,350
Which I cannot blame them because that's, as a,

308
00:19:21,910 --> 00:19:25,110
you would intuitively think that. And it's I think also

309
00:19:27,270 --> 00:19:30,930
a good belief in humanity if you would think I could simply

310
00:19:30,930 --> 00:19:34,690
influence people like that. But I think it's adding

311
00:19:34,690 --> 00:19:38,010
that part of understanding to the technical

312
00:19:38,010 --> 00:19:41,730
profession. That's something that as a technical team you can do. That's

313
00:19:41,730 --> 00:19:45,490
also something that as a technical leader you can do to just

314
00:19:45,810 --> 00:19:49,570
add these additional skills. I was two weeks ago at a

315
00:19:49,730 --> 00:19:53,410
data science conference and I was speaking in the AI and

316
00:19:53,410 --> 00:19:57,020
MLOps track with a very strange

317
00:19:57,020 --> 00:19:59,940
topic. I just talked about decision making for half an hour

318
00:20:00,900 --> 00:20:04,260
and I was really worried that the resonance would be okay. We came here to

319
00:20:04,260 --> 00:20:07,860
look at a Python notebook and suddenly this guy is talking about psychology. So what's

320
00:20:07,860 --> 00:20:11,660
going on here? But the resonance and the number of questions

321
00:20:11,660 --> 00:20:15,460
I got, I mean, they just showed me there's a real openness, almost like a

322
00:20:15,460 --> 00:20:19,220
craving to understand that because everybody has

323
00:20:19,220 --> 00:20:22,180
been in that room. I know I have many, many times where

324
00:20:23,060 --> 00:20:26,500
we had all the analysis, right? The data was good,

325
00:20:26,660 --> 00:20:30,420
we had scrubbed it, we had understood it, the area under the curve or

326
00:20:30,420 --> 00:20:34,260
whatever quality measure, it was good. And we showed them. I don't

327
00:20:34,260 --> 00:20:37,980
know, when you drive your trucks around Southeast Asia, you can save 15%

328
00:20:37,980 --> 00:20:41,660
of fuel. And then a year later, half a year later, trucks are still

329
00:20:41,660 --> 00:20:45,140
driving around like they used to. I mean, that's such a frustrating experience

330
00:20:45,460 --> 00:20:49,300
on the data and analytics side that I think people are really,

331
00:20:49,380 --> 00:20:53,050
really eager to learn how do I overcome this? And

332
00:20:53,050 --> 00:20:56,410
how can I be more effective in

333
00:20:56,650 --> 00:20:59,770
actually changing something and actually

334
00:21:00,170 --> 00:21:03,610
being effective

335
00:21:04,490 --> 00:21:07,610
in my job and being seen with what I do.

336
00:21:08,010 --> 00:21:11,450
And maybe we can use that as a platform and

337
00:21:11,450 --> 00:21:15,210
basis to expand it from the technology side.

338
00:21:15,210 --> 00:21:19,030
And of course it's not a one sided thing,

339
00:21:19,030 --> 00:21:22,190
right? At the same time, when I speak, let's say with

340
00:21:22,590 --> 00:21:25,470
HR leaders, I always emphasize

341
00:21:26,510 --> 00:21:30,150
you need to make sure that there is more technical understanding in the

342
00:21:30,150 --> 00:21:33,790
organization. You cannot treat data

343
00:21:34,270 --> 00:21:37,990
as this simple API where you basically say request dashboard and then

344
00:21:37,990 --> 00:21:41,830
suddenly the dashboard is built a couple of weeks after. That's not how it works.

345
00:21:41,830 --> 00:21:45,400
You need to understand how this works. You need to also understand how

346
00:21:45,400 --> 00:21:48,800
data works. And what, what it can and cannot give you, because

347
00:21:48,800 --> 00:21:52,480
otherwise you're just coming at this with the wrong expectations and

348
00:21:52,480 --> 00:21:56,280
you're almost bound to be disappointed in the end. That's

349
00:21:56,280 --> 00:21:59,240
a good way to put it because I think one of the naive things I

350
00:21:59,240 --> 00:22:02,920
thought in my youth too, is that we would make better decisions if only

351
00:22:02,920 --> 00:22:06,360
we had the data. And then that didn't work out. Well,

352
00:22:06,360 --> 00:22:09,480
maybe, maybe it was about access and discoverability in the data,

353
00:22:10,150 --> 00:22:13,750
but I think ultimately the problem is a human problem. And

354
00:22:13,750 --> 00:22:17,070
it's funny you mentioned that. Right. You know, people go to a tech conference, they

355
00:22:17,070 --> 00:22:20,230
expect to see jupyter notebooks, et cetera, et cetera. Right. They expect to see code.

356
00:22:20,550 --> 00:22:24,270
I just got back actually really late last night from DevOps days,

357
00:22:24,270 --> 00:22:28,070
Austin. And a number of the talks

358
00:22:28,070 --> 00:22:31,790
were not technical. They were about influence and how decisions are

359
00:22:31,790 --> 00:22:35,110
made and Campbell's Law. And now

360
00:22:35,990 --> 00:22:38,990
I was working the business. I wasn't able to see the whole conference, but the

361
00:22:38,990 --> 00:22:42,700
way that those talks resonated with the crowd I thought was very

362
00:22:42,700 --> 00:22:46,340
interesting because that was, you know, soft skills. And the whole

363
00:22:46,340 --> 00:22:49,900
soft versus hard skills goes back to apparently US military

364
00:22:49,900 --> 00:22:53,660
training, right? Where hard skills. Okay, yeah, yeah, yeah. So. So apparently the

365
00:22:53,660 --> 00:22:57,020
origin is not that soft. It means it's easy or it's not important or it's

366
00:22:57,020 --> 00:23:00,780
fluffy. It really means like, you know, you know, actual

367
00:23:00,780 --> 00:23:04,500
kinetic things that are hard. Tanks, guns,

368
00:23:04,500 --> 00:23:08,340
bullets, that sort of thing, Missiles and, you know, dealing with people. Things that are

369
00:23:08,340 --> 00:23:11,950
soft, like, you know, living things. Right. So that's

370
00:23:11,950 --> 00:23:15,190
apparently the origin of it. So. Which is. It's an unfortunate term because I think

371
00:23:15,190 --> 00:23:18,590
when people hear soft skills, they're like, yeah, right. Especially engineers.

372
00:23:19,790 --> 00:23:23,110
And. But I just found that interesting because I think, I think a lot of

373
00:23:23,110 --> 00:23:26,430
IT professionals have gotten to the point where they get very frustrated because

374
00:23:26,990 --> 00:23:30,750
the data says one thing. Right. But the process

375
00:23:30,750 --> 00:23:34,430
hasn't changed. Right. Or they're still making the same bad decisions or

376
00:23:34,430 --> 00:23:38,020
not data driven situations. And, you

377
00:23:38,020 --> 00:23:41,660
know, and I think you have a lot of people in leadership

378
00:23:41,660 --> 00:23:45,140
roles, not in it, but outside of it, that are data

379
00:23:45,140 --> 00:23:48,900
frustrated. Yeah, yeah, go ahead. No,

380
00:23:48,900 --> 00:23:52,740
no, we, we call them delivery skills in, in my team, that's ultimately

381
00:23:52,740 --> 00:23:55,660
what. Well, I, I know that some people like to refer to soft skills also

382
00:23:55,660 --> 00:23:59,140
as, as human skills, but I'm also not happy with that because, you know, if

383
00:23:59,140 --> 00:24:02,140
you don't have the skills, you're not human. No, that, that doesn't work. But, but

384
00:24:02,320 --> 00:24:06,080
we call them delivery skills because that I think brought home that aspect of. Well,

385
00:24:06,080 --> 00:24:09,840
if you want to deliver. If you want to deliver impact, here's the skills

386
00:24:09,840 --> 00:24:13,680
you need and there's the core skills, your technical component, of course,

387
00:24:13,840 --> 00:24:17,600
you got to have that clear. But you have to know how

388
00:24:17,600 --> 00:24:21,320
to deliver these insights into an organization that isn't

389
00:24:21,320 --> 00:24:25,040
just waiting for the next statistical analysis to

390
00:24:25,040 --> 00:24:28,560
be delivered. But it's unintuitive. And

391
00:24:28,720 --> 00:24:32,120
that was one of the fascinating things I found when researching for the book. So

392
00:24:32,120 --> 00:24:35,900
the book contains an entire chapter on psychology because I just

393
00:24:35,900 --> 00:24:38,780
found it so fascinating to understand that

394
00:24:39,020 --> 00:24:42,780
interplay of data and human brains and

395
00:24:42,780 --> 00:24:46,500
what happens. And turns out there's actually even a study from the year I

396
00:24:46,500 --> 00:24:50,259
was born. So it's been around for a while where

397
00:24:50,259 --> 00:24:54,020
they took a couple of students in Stanford and asked

398
00:24:54,020 --> 00:24:57,780
them, what's your opinion on capital punishment? So they took strong, emotional,

399
00:24:57,780 --> 00:25:01,610
controversial topic, and then they showed them a fake study.

400
00:25:01,610 --> 00:25:05,130
And that study was made up of data that was

401
00:25:05,130 --> 00:25:08,130
constructed in a way so that you could imagine. Half the data

402
00:25:08,690 --> 00:25:12,450
gave you arguments for capital punishment and half the data would give you arguments

403
00:25:12,450 --> 00:25:16,170
against. And the researchers just wanted to find out, okay, can data change

404
00:25:16,170 --> 00:25:19,810
people's minds? And the fascinating thing is they found out,

405
00:25:19,810 --> 00:25:23,530
yes it can, but in exactly the opposite direction that you

406
00:25:23,530 --> 00:25:27,290
want. So the people going in that are pro capital punishment, they come out

407
00:25:27,290 --> 00:25:31,110
and say, oh yeah, I finally found, found the data to confirm my beliefs. There

408
00:25:31,110 --> 00:25:34,750
was also a bit of sketchy data in there that said I'm not right, but

409
00:25:34,750 --> 00:25:37,870
that's sketchy. And the sources, I don't believe them. And the people that were

410
00:25:38,830 --> 00:25:42,670
against capital punishment, they came out with exactly the same feeling. They said,

411
00:25:42,670 --> 00:25:46,390
oh, there was so much good data in there against capital

412
00:25:46,390 --> 00:25:50,190
punishment, I must be right. A bit of sketchy data that was pro, but

413
00:25:50,190 --> 00:25:53,910
that can't be right, I don't believe it. And that's, that's

414
00:25:53,910 --> 00:25:56,990
an insight that's been around for so long and of course it's been replicated

415
00:25:58,000 --> 00:25:59,920
enormous amount of time. And

416
00:26:01,760 --> 00:26:05,280
just go on social media you will see that effect applied at scale,

417
00:26:06,080 --> 00:26:09,880
very much so, the echo bubbles and everything. But we still tend to operate

418
00:26:09,880 --> 00:26:13,319
on this, what I like to call the data deficit theory, that it's just like,

419
00:26:13,319 --> 00:26:17,160
oh, all we're missing is, you know, the saying, right, the right

420
00:26:17,160 --> 00:26:20,640
data to the right people at the right time and suddenly

421
00:26:21,040 --> 00:26:24,500
things will improve for the better. And psychology and

422
00:26:24,500 --> 00:26:28,260
research for decades has shown us this is not the

423
00:26:28,260 --> 00:26:32,020
case. We don't like to be proven wrong and our brain

424
00:26:32,020 --> 00:26:35,780
will do a lot of tricks and a lot of self convincing

425
00:26:36,100 --> 00:26:39,900
to just make us very Very reassured that

426
00:26:39,900 --> 00:26:43,460
we're right. Yeah, I mean that's

427
00:26:43,460 --> 00:26:46,020
really, is that, is that the,

428
00:26:47,460 --> 00:26:50,700
that's really a human problem. How does,

429
00:26:51,260 --> 00:26:55,060
obviously it's also been shown in monkeys actually. Oh really? Okay, so

430
00:26:55,060 --> 00:26:58,140
it might be a biological problem. Right. And I wonder,

431
00:26:58,860 --> 00:27:02,540
I wonder, you know, will you, will we see as model LLMs

432
00:27:02,540 --> 00:27:06,299
get better at reasoning and kind of holding opinions, will they do the same

433
00:27:06,299 --> 00:27:09,500
thing? Is it maybe just part of, it's just a function of the system?

434
00:27:09,980 --> 00:27:13,500
I don't know, it's, I mean the, the,

435
00:27:13,740 --> 00:27:17,260
the thing is, so that's what I find is one of the bigger

436
00:27:17,260 --> 00:27:21,000
dangers of AI actually because so I

437
00:27:21,000 --> 00:27:24,840
differentiate in the book pretty clearly between machine learning and AI.

438
00:27:25,080 --> 00:27:28,760
So I'm not sure it's canonical, but I think it's useful. So basically I

439
00:27:28,760 --> 00:27:32,560
say machine learning is when you take a bunch of data and you have a

440
00:27:32,560 --> 00:27:36,040
clear goal and you're training the model to fulfill that goal.

441
00:27:36,120 --> 00:27:39,840
And I'm aware AI is also built on machine learning, but the distinction I

442
00:27:39,840 --> 00:27:43,280
make is that I say, well, AI as we use the term now is

443
00:27:43,280 --> 00:27:47,080
basically you take a bunch of data, you don't really have a goal,

444
00:27:48,190 --> 00:27:51,950
but you're training it anyways. And I mean, that's roughly how most of

445
00:27:51,950 --> 00:27:55,790
these language models are trained, right? You feed them all of these texts and

446
00:27:55,790 --> 00:27:59,630
then you say my goal is to make the user happy and

447
00:27:59,630 --> 00:28:03,309
for many people to be happy with the answers. So what

448
00:28:03,309 --> 00:28:06,990
does that lead to? Well, first of all, you're training a model that

449
00:28:07,310 --> 00:28:10,670
by definition is going to be extremely convincing

450
00:28:10,910 --> 00:28:14,750
because that model has been trained on how do I circumvent all of

451
00:28:14,750 --> 00:28:18,510
these psychological traps, how do I make people feel good all the time?

452
00:28:18,510 --> 00:28:22,330
And that's all the sycophancy discussion of course we're having. But I think it's also

453
00:28:22,330 --> 00:28:25,930
more subtle. There's some studies that actually

454
00:28:26,250 --> 00:28:30,050
looked at, for example, models generating propaganda and

455
00:28:30,050 --> 00:28:33,770
found out that language models are much more effective than humans

456
00:28:33,770 --> 00:28:36,970
at generating propaganda. Because I think just the way they're trained,

457
00:28:37,690 --> 00:28:41,050
there's some implicit mechanisms that these models have learned

458
00:28:41,370 --> 00:28:45,130
that they can exploit. And so the ironic thing is that

459
00:28:45,490 --> 00:28:49,010
actually you should be trusting machine learning because machine

460
00:28:49,010 --> 00:28:52,850
learning, lots of data, clear goal, there's some

461
00:28:52,930 --> 00:28:56,410
statistical proof that you can make after a while and say,

462
00:28:56,410 --> 00:29:00,170
okay, of course it's better than a human. The classic examples, the cancer

463
00:29:00,170 --> 00:29:03,850
detection, now the self driving cars. But it turns out

464
00:29:03,850 --> 00:29:07,170
in studies that people just don't trust machine learning.

465
00:29:08,370 --> 00:29:12,210
There's a really weird study I found from Wharton, for example, where they

466
00:29:13,930 --> 00:29:17,370
showed people that a machine learning model was superior

467
00:29:17,610 --> 00:29:20,810
to their own decision making. And then they asked them,

468
00:29:21,610 --> 00:29:24,610
as they do in these studies, you can earn some money here and you can

469
00:29:24,610 --> 00:29:27,450
either trust your gut or you can trust the machine learning model.

470
00:29:28,250 --> 00:29:32,090
Everybody on average, of course, went with their

471
00:29:32,090 --> 00:29:35,690
own feelings. Even though they had seen that the model performs

472
00:29:35,690 --> 00:29:39,360
better, they just needed to catch it doing something

473
00:29:39,360 --> 00:29:42,680
wrong and that they would immediately say, I don't trust this thing.

474
00:29:43,960 --> 00:29:47,760
Then there's a weird mechanism that they added where people could start to influence the

475
00:29:47,760 --> 00:29:51,520
results and suddenly they trusted the model more, even though the model didn't

476
00:29:51,520 --> 00:29:54,680
change at all. So very, very weird effects. But

477
00:29:55,800 --> 00:29:59,280
what I'm getting at is machine learning is something that's extremely

478
00:29:59,280 --> 00:30:03,000
trustworthy and yet our brains are just wired to

479
00:30:03,000 --> 00:30:06,600
distrust it. It all seems so mechanical, so

480
00:30:06,600 --> 00:30:10,410
mathematical, so unhuman. I think that's what many people

481
00:30:10,410 --> 00:30:14,050
say. Right. They don't feel comfortable with that. And then here along

482
00:30:14,050 --> 00:30:17,890
come these language models where you should definitely not trust

483
00:30:17,890 --> 00:30:21,370
them. They have not been trained on a specific goal. We have no idea what's

484
00:30:21,370 --> 00:30:25,050
going on under the hood. They're pretty good, to be fair, but

485
00:30:25,050 --> 00:30:27,890
still weird effects happening.

486
00:30:28,770 --> 00:30:32,610
And yet implicitly, we completely trust them just because of the way that they

487
00:30:32,850 --> 00:30:36,490
interact and because they have this amazing, amazing human,

488
00:30:36,490 --> 00:30:40,200
like, user interface to interact with us. And

489
00:30:40,200 --> 00:30:43,920
that's something I think we're going to have to grapple with that suddenly we found

490
00:30:43,920 --> 00:30:47,560
a machine that can be very convincing but actually shouldn't be very

491
00:30:47,560 --> 00:30:51,400
trustworthy. That is a very good way to

492
00:30:51,400 --> 00:30:55,000
put it. I actually just saw a video last night where they were talking about

493
00:30:55,000 --> 00:30:58,360
how an AI will map human emotions

494
00:30:59,080 --> 00:31:02,680
and not so much map the emotions. I'll send you a link to the videos

495
00:31:02,680 --> 00:31:06,530
from the infographics show. And, you

496
00:31:06,530 --> 00:31:09,610
know, it was not meant for. It was meant for, like, the general public, I

497
00:31:09,610 --> 00:31:12,890
think, to watch. But, like, there were things in there where. And it makes sense,

498
00:31:12,890 --> 00:31:15,850
right, because they, they basically categorize or

499
00:31:16,250 --> 00:31:20,010
store words in. We'll call them spaces. Right. So they,

500
00:31:20,010 --> 00:31:23,450
they. Certain words that you will use will indicate that you're in a certain

501
00:31:23,450 --> 00:31:26,890
emotional state. So. And it was very

502
00:31:26,890 --> 00:31:29,050
fascinating. I was like, this sounds.

503
00:31:30,890 --> 00:31:34,650
It sounds dangerous. Right. And so

504
00:31:34,650 --> 00:31:38,410
it'll know, like, if it can. It can know, like, if it's. Depending on what

505
00:31:38,410 --> 00:31:41,810
the prompt is and how it was trained, it can use different words to kind

506
00:31:41,810 --> 00:31:45,610
of guide you back to whatever emotional state it wants you to go in,

507
00:31:45,930 --> 00:31:49,170
which is, you know, a very

508
00:31:49,170 --> 00:31:53,010
dangerous weapon. I dare say it's manipulation at

509
00:31:53,010 --> 00:31:56,090
its very best. Yeah. And on the other manipulation at scale.

510
00:31:56,740 --> 00:31:59,300
Yeah. And on the other hand, not surprising that the models would be able to

511
00:31:59,300 --> 00:32:02,380
do that. I mean, you know, they know 20 TV shows you watched and they

512
00:32:02,380 --> 00:32:05,380
can recommend you a movie that you never thought of. But still like, so

513
00:32:05,940 --> 00:32:09,660
why wouldn't it work with emotional states? I think, of course our brain likes

514
00:32:09,660 --> 00:32:13,380
to suggest to us that we're so complex to figure out where

515
00:32:14,180 --> 00:32:17,460
perhaps in fact we are. Not always at least.

516
00:32:18,740 --> 00:32:22,500
Yeah. If you ever watch a Star Trek Next Generation and you watch it with

517
00:32:22,500 --> 00:32:26,210
like today's vision. Right. You

518
00:32:26,210 --> 00:32:29,410
know, like they, there are lines in there that I find kind of funny with,

519
00:32:29,410 --> 00:32:31,530
you know, 20, 26 kind of

520
00:32:33,370 --> 00:32:36,770
vision. Right. Is, yeah, I'm an AI. I, I, I can be

521
00:32:36,770 --> 00:32:40,610
completely impartial. That was what Data said in a couple

522
00:32:40,610 --> 00:32:44,210
of times. And I like, oh, we were so innocent

523
00:32:44,210 --> 00:32:47,810
then. And there were a few other things where, you know, a big

524
00:32:47,810 --> 00:32:51,600
subplot. Picard would say, well, you know, you really can't calculate

525
00:32:51,600 --> 00:32:54,720
the human condition or something like that. It was very, I think it was very

526
00:32:54,720 --> 00:32:58,440
much a product of its time. Yeah, yeah. That's the one thing

527
00:32:58,520 --> 00:33:01,640
they really didn't quite get right. It's so interesting.

528
00:33:02,360 --> 00:33:06,080
Yeah, I hadn't thought about that. Yeah. Like, you know, if you watch

529
00:33:06,080 --> 00:33:08,280
them like now, it's kind of like, you'll see

530
00:33:09,800 --> 00:33:13,600
particularly when they talk about AI and kind of how AI people interact

531
00:33:13,600 --> 00:33:17,380
with AI, some of it is very, very, very much on point. Right. You'll, you

532
00:33:17,380 --> 00:33:20,980
know, there's one episode, any episode, where they interact with the computer through

533
00:33:20,980 --> 00:33:24,580
voice. Very much how we interact with voice assistants today. Right.

534
00:33:24,580 --> 00:33:28,420
Yeah. They didn't anticipate the seductiveness and the human likeness.

535
00:33:28,420 --> 00:33:32,220
Right? Yeah, no, they know, they, I mean it was just, and honestly, who

536
00:33:32,220 --> 00:33:35,740
really could? Right? No, yeah. Unless you're like

537
00:33:35,740 --> 00:33:39,460
extremely paranoid like Philip K. Dick. Right. Like, you know,

538
00:33:39,460 --> 00:33:43,300
but sometimes, sometimes paranoia is another way to say you're

539
00:33:44,030 --> 00:33:46,030
further ahead of the curve than other people.

540
00:33:48,910 --> 00:33:52,270
But that's what science fiction is all about.

541
00:33:52,590 --> 00:33:55,870
Exactly. Right. Like sometimes the crazier it comes out

542
00:33:56,510 --> 00:34:00,350
when it's mentioned, the more it'll has long lasting effect.

543
00:34:00,990 --> 00:34:04,830
So we mentioned kind of like data driven and data frustrated. What is

544
00:34:05,390 --> 00:34:09,230
data inspiration then? Is that kind

545
00:34:09,230 --> 00:34:11,870
of the end, the ideal end state is data inspiration?

546
00:34:12,970 --> 00:34:16,170
Like what I think the goal is to add data

547
00:34:16,170 --> 00:34:19,850
inspiration to the data driven organization. So

548
00:34:20,650 --> 00:34:24,170
I don't think it's any way plausible to

549
00:34:24,170 --> 00:34:27,930
argue, oh, let's swing the pendulum around. Of course we're

550
00:34:27,930 --> 00:34:31,490
going to have dashboards. Of course we're going to have optimization, of course we're going

551
00:34:31,490 --> 00:34:35,210
to have automation. I mean, that's all what a data driven organization

552
00:34:36,170 --> 00:34:39,860
really is about. But I think the important part about being

553
00:34:39,860 --> 00:34:43,540
data inspired, I think there's two components that I deeply care

554
00:34:43,540 --> 00:34:47,140
about. One is that I think the notion of data driven

555
00:34:47,460 --> 00:34:51,300
can also have this notion of here's a preset path to

556
00:34:51,300 --> 00:34:54,940
average and everything's just going to be optimized.

557
00:34:54,940 --> 00:34:58,780
And in the end, that's going to kill innovation, it's going to

558
00:34:58,780 --> 00:35:01,860
kill creativity, and it's also not going to be very fun.

559
00:35:02,580 --> 00:35:06,410
And I think that's not true when it comes to data, or at least that's

560
00:35:06,410 --> 00:35:09,810
also not how I perceive data. I know that when many people hear the term,

561
00:35:09,810 --> 00:35:13,210
they think about these tables and numbers and things that are not very

562
00:35:13,210 --> 00:35:16,810
exciting. But in the end, if you think about it,

563
00:35:16,810 --> 00:35:20,450
there's a lot of innovation potential in data. And

564
00:35:20,450 --> 00:35:24,170
I myself, I find myself being creative with data.

565
00:35:24,410 --> 00:35:28,210
You connect data sets that probably weren't meant to be connected, but you find

566
00:35:28,210 --> 00:35:31,910
this new, exciting insight. We have art created with

567
00:35:31,910 --> 00:35:35,630
data that is amazing. We have things like data journalism, which I

568
00:35:35,630 --> 00:35:39,470
find some of the most interesting and most fascinating forms

569
00:35:39,470 --> 00:35:43,110
of journalism. So I think we need to recognize that data

570
00:35:43,110 --> 00:35:46,950
is more than these cold hard facts that just tell us go

571
00:35:46,950 --> 00:35:50,470
left, go right, but that there is a lot of

572
00:35:51,030 --> 00:35:54,790
additional potential in here. And ultimately for a

573
00:35:54,870 --> 00:35:58,480
organization or a company, I think that's also the potential for

574
00:35:58,800 --> 00:36:02,520
transformation. So not in the way that you have, you know, the

575
00:36:02,520 --> 00:36:06,080
data driven part, that's opt, that optimizes what's

576
00:36:06,240 --> 00:36:09,560
already there. And then you have a few creative people with these

577
00:36:09,560 --> 00:36:13,280
brilliant strategic ideas that are going to take the organization to the next

578
00:36:13,280 --> 00:36:17,000
level. I think that's wrong thinking. You know, that's again, having

579
00:36:17,000 --> 00:36:20,680
two conversations in different rooms that should be in a single room. You are going

580
00:36:20,680 --> 00:36:23,760
to find your next breakthrough strategy and your next

581
00:36:24,240 --> 00:36:28,030
breakthrough product, most likely in the data. If you

582
00:36:28,030 --> 00:36:31,550
connect everything that you have, if you allow for experiments,

583
00:36:31,790 --> 00:36:35,230
if you, as we said, look at the outliers, look at the

584
00:36:35,230 --> 00:36:38,990
anomalies. And I think that's the important aspect that I

585
00:36:38,990 --> 00:36:42,790
like to emphasize here when I say data inspired, that you

586
00:36:42,790 --> 00:36:46,390
just don't forget about these aspects and

587
00:36:46,390 --> 00:36:50,230
don't just build organizations that are happy when the dashboard

588
00:36:50,230 --> 00:36:53,900
is green. Yeah, yeah.

589
00:36:53,900 --> 00:36:57,620
You know, there's a lot of organizations and I've worked for them in the

590
00:36:57,620 --> 00:37:01,100
past, that it basically it's a scorecard world.

591
00:37:01,180 --> 00:37:05,020
Yeah. And as long as you hit those certain numbers and

592
00:37:05,500 --> 00:37:09,340
the mental gymnastics and I

593
00:37:09,340 --> 00:37:13,180
wouldn't say shady, but I would say awkward ethical

594
00:37:13,420 --> 00:37:17,020
situations people would throw themselves into. I mean, it can be

595
00:37:17,020 --> 00:37:20,300
shady. That's the. So. So, so there's an interesting thing

596
00:37:21,420 --> 00:37:24,900
I found that most of the

597
00:37:24,900 --> 00:37:28,580
inventors of data driven methods, so one of the

598
00:37:28,580 --> 00:37:32,300
famous one W. Edward Deming, for example, one of the inventors of

599
00:37:32,300 --> 00:37:35,820
statistical process control, all of them

600
00:37:36,459 --> 00:37:39,900
come to the realization that steering a company by

601
00:37:39,900 --> 00:37:43,100
numbers is a really, really bad idea.

602
00:37:43,580 --> 00:37:47,300
So they all have that insight. And with Deming it's

603
00:37:47,300 --> 00:37:51,050
quite extreme and also unfortunate. So there's this quote,

604
00:37:51,050 --> 00:37:54,130
and I'm sure you've heard it right. What gets

605
00:37:54,530 --> 00:37:58,050
measured gets managed or what gets measured gets done.

606
00:37:58,610 --> 00:38:02,210
And oftentimes when people put that on a slide, they will quote Peter Drucker. Now

607
00:38:02,210 --> 00:38:05,450
there's a whole webpage on the Peter Drucker Institute where they say

608
00:38:05,450 --> 00:38:09,210
Drucker never said that. And the

609
00:38:09,210 --> 00:38:13,010
actual quote comes from Deming. But the

610
00:38:14,050 --> 00:38:17,890
really, really annoying thing is he actually said the exact opposite. So

611
00:38:17,890 --> 00:38:21,740
his full quote is, it's wrong to assume that if you

612
00:38:21,740 --> 00:38:23,540
can't measure it, you can't manage it.

613
00:38:25,220 --> 00:38:28,980
Yeah, and he only gets cited with that second part. He must be turning around

614
00:38:28,980 --> 00:38:32,420
in his grave because he was so against management by

615
00:38:32,420 --> 00:38:36,220
numbers that he put it on his list of seven deadly

616
00:38:36,220 --> 00:38:39,820
sins of management. It was so important to him that people

617
00:38:39,820 --> 00:38:43,540
realized that you need to honor what can't be measured. And that

618
00:38:43,540 --> 00:38:46,960
exactly what you're saying, if you just steer by numbers, you're going to end up

619
00:38:46,960 --> 00:38:50,240
in a shady place. And I think that's not just theory. I mean there's lots

620
00:38:50,240 --> 00:38:54,080
of public examples. So for example, ge, you know, for year

621
00:38:54,080 --> 00:38:57,760
after year after year delivered extremely stable

622
00:38:57,760 --> 00:39:01,560
profits, unusually stable profits, because that's what they measured

623
00:39:01,560 --> 00:39:05,360
themselves by until somebody figured out, well, the

624
00:39:05,360 --> 00:39:09,120
accounting that led to these stable profits might not

625
00:39:09,120 --> 00:39:12,880
entirely have been kosher to say it very carefully

626
00:39:12,880 --> 00:39:16,160
or the. I mean while we're talking from Germany, the

627
00:39:16,160 --> 00:39:19,420
Volkswagen diesel scandal, right, Dieselgate,

628
00:39:19,980 --> 00:39:23,620
there were a couple of engineers who were told, I want you to build an

629
00:39:23,620 --> 00:39:27,380
efficient engine and if you don't, you're fired. Well, and that's the

630
00:39:27,380 --> 00:39:31,100
only number I care about. So the engineers made very,

631
00:39:31,100 --> 00:39:33,260
very sure that the engines would hit that number.

632
00:39:34,860 --> 00:39:38,380
But of course that led them to other

633
00:39:38,380 --> 00:39:42,180
troubles down the line. So I think that's also one of the real

634
00:39:42,180 --> 00:39:45,740
dangers. If we're just talking about data driven organizations that ultimately

635
00:39:46,960 --> 00:39:50,560
or very easily actually it can lead to this notion of

636
00:39:50,880 --> 00:39:54,120
we steer by numbers and we Invent a few

637
00:39:54,120 --> 00:39:57,840
KPIs, maybe complex ones. And as long as we hit the KPIs,

638
00:39:58,000 --> 00:40:01,160
the business will be fine. And of course, what you start doing is you start

639
00:40:01,160 --> 00:40:04,800
managing the numbers and not the business. And both can diverge

640
00:40:04,800 --> 00:40:07,360
quite a bit. Yeah, no, that's.

641
00:40:08,080 --> 00:40:10,400
Incentives drive behavior. So if you

642
00:40:11,840 --> 00:40:15,540
change the incentive, the incentives should be running the business, not hitting

643
00:40:15,540 --> 00:40:19,340
a KPI. Right. KPI's are a means to an. No. Like,

644
00:40:19,340 --> 00:40:22,700
you know, and it's funny because, like, I've seen people do really dumb things

645
00:40:23,100 --> 00:40:26,540
just because that was in their KPIs. And I, I just,

646
00:40:27,020 --> 00:40:30,700
I, I did not know that about Deming, like, because he's often

647
00:40:30,780 --> 00:40:34,540
cited as like, you know, the management by numbers guy. But the fact that

648
00:40:34,860 --> 00:40:38,020
I hear that, I'm like, oh, my God, like, he really must be rolling over

649
00:40:38,020 --> 00:40:41,390
in his grave. I know, I know. And I mean,

650
00:40:41,630 --> 00:40:45,470
also, just to defend the people that want to steer by numbers,

651
00:40:46,110 --> 00:40:49,710
I have these discussions in my team all the time. For example,

652
00:40:49,790 --> 00:40:53,310
performance reviews, where I'm very much against

653
00:40:53,310 --> 00:40:55,870
doing performance reviews based on metrics because

654
00:40:56,910 --> 00:41:00,270
I really believe that these won't do people justice. But then

655
00:41:00,670 --> 00:41:04,230
maybe a performance review isn't going the way you thought it would be. And

656
00:41:04,230 --> 00:41:08,060
what's the immediate reaction that you get? Well, people will come to you and say,

657
00:41:08,220 --> 00:41:12,060
okay, next time I want numbers, because that will give me reassurance.

658
00:41:12,140 --> 00:41:15,660
And when I hit my numbers, I know I've done a good job. And I

659
00:41:15,660 --> 00:41:18,860
think in management that may just be the same, right, that you say, well, how

660
00:41:18,860 --> 00:41:22,500
do we know we're doing a good job if we don't have objective measurement, if

661
00:41:22,500 --> 00:41:26,060
we don't have numbers to hold ourselves to

662
00:41:26,300 --> 00:41:30,140
as a standard? And of course, as always, I think

663
00:41:30,140 --> 00:41:33,660
there's a bit of a balancing issue. The interesting part is

664
00:41:34,310 --> 00:41:37,910
you may know this from project management and project management, you have this triangle

665
00:41:37,990 --> 00:41:41,750
where a project can be cheap, it can be good, it

666
00:41:41,750 --> 00:41:44,950
can be fast. Unfortunately, you only get two out of three.

667
00:41:45,670 --> 00:41:49,190
And with data, I think there's a similar triangle,

668
00:41:49,430 --> 00:41:53,110
which is data in my mind. It can be simple, it

669
00:41:53,110 --> 00:41:56,870
can be accurate, and it can be universal, which means it

670
00:41:56,870 --> 00:42:00,230
tells you something about the entire business. And I also think you only get two

671
00:42:00,230 --> 00:42:03,720
out of three. So if, for example, you have something that's

672
00:42:03,720 --> 00:42:07,400
accurate and simple, it won't be very universal. It will be

673
00:42:07,400 --> 00:42:11,240
measuring a tiny process somewhere in manufacturing hall 4D,

674
00:42:11,320 --> 00:42:13,560
but not tell you how the whole business is doing.

675
00:42:15,000 --> 00:42:18,800
And so what does that mean? Everybody wants accurate. And I

676
00:42:18,800 --> 00:42:22,640
think when it comes to steering a business or steering a business unit, you want

677
00:42:22,640 --> 00:42:26,400
universal. So what's the one you need to forego? That

678
00:42:26,400 --> 00:42:30,100
is simplicity. And you really need to engage with the complexity. And

679
00:42:30,100 --> 00:42:33,740
that's what some organizations do. I mean, I know everybody's a bit tired about

680
00:42:33,740 --> 00:42:37,420
hearing about Amazon, but if you look at what Amazon does in a business meeting

681
00:42:37,900 --> 00:42:41,540
and there's various books about that, they will look at 400, 500

682
00:42:41,540 --> 00:42:43,980
different KPIs in a single meeting

683
00:42:45,180 --> 00:42:48,780
just because they want to make sure that they're not optimizing for one number

684
00:42:49,180 --> 00:42:52,540
at the cost of the rest of the business or managing one number up and

685
00:42:52,540 --> 00:42:56,050
the other one goes down. And so I think they have deeply

686
00:42:56,050 --> 00:42:59,570
understood this principle that, okay, we want accurate, we want

687
00:42:59,570 --> 00:43:02,810
universal, so we need to go complex. Sorry.

688
00:43:04,650 --> 00:43:08,010
Yeah, I mean, you're right because you know, you can only

689
00:43:08,170 --> 00:43:11,930
model, particularly the specificity part.

690
00:43:11,930 --> 00:43:15,650
Right. Like, yeah, that, that, that I think is very, and it's very

691
00:43:15,650 --> 00:43:18,730
tempting to assume that you can have one model to rule them all,

692
00:43:19,290 --> 00:43:22,820
but that's just not. I mean, I think finally when it comes to

693
00:43:22,820 --> 00:43:26,220
LLMs, I think people are finally figuring that one out, right? Where you'll have different,

694
00:43:26,700 --> 00:43:30,140
smaller, fine tuned models versus models with

695
00:43:30,140 --> 00:43:33,700
trillions of parameters that'll cost ridiculous amount of money to run.

696
00:43:33,700 --> 00:43:37,220
Yeah, well also I think the whole agent discussion of course

697
00:43:37,220 --> 00:43:40,860
is going this way, right? You're realizing we won't have the universal

698
00:43:40,860 --> 00:43:44,300
agent that will do everything, but people are creating these really, really

699
00:43:44,300 --> 00:43:47,980
specialized tools that can then interact with each

700
00:43:47,980 --> 00:43:51,660
other and play to their goals. So we're sort of doing the

701
00:43:51,660 --> 00:43:55,500
opposite. Right. At first we said, well, we don't have a special or specified

702
00:43:55,500 --> 00:43:59,300
goal for creating these AI models. And now we're sort of trimming

703
00:43:59,300 --> 00:44:02,620
it down again and saying, well actually we'd like a bit of a goal in

704
00:44:02,620 --> 00:44:05,979
there. Right, Right, yeah. No, that's

705
00:44:05,979 --> 00:44:09,540
interesting. What

706
00:44:10,740 --> 00:44:14,180
I'm looking at some of the notes here and this is what is a

707
00:44:14,260 --> 00:44:17,850
data resistant mind. Because I think that's interesting, I think it's an

708
00:44:17,850 --> 00:44:21,370
interesting concept because I think I know what you mean. But I'm like, I never

709
00:44:21,370 --> 00:44:25,050
had a label for it before. I think the data resistant

710
00:44:25,050 --> 00:44:28,610
mind, I mean it starts with things like the Stanford study, right?

711
00:44:28,610 --> 00:44:31,850
So our mind is just confronted with data and

712
00:44:32,570 --> 00:44:36,170
just takes it the wrong way. But there's other effects

713
00:44:36,410 --> 00:44:40,010
that have been shown. So I mentioned the monkey

714
00:44:40,010 --> 00:44:43,530
study. So again, apparently they like to study brains. In Stanford

715
00:44:44,900 --> 00:44:48,740
there's a study that's much younger where they had a monkey looking at dot

716
00:44:48,740 --> 00:44:52,540
patterns on a screen and the monkey essentially had to decide which way

717
00:44:52,540 --> 00:44:55,700
are the Dots moving, and then press a button, depending on what it thought.

718
00:44:57,380 --> 00:45:01,140
The researchers, what they did is they wired up the monkey's brain,

719
00:45:02,100 --> 00:45:05,860
which led them to some fascinating and scary result,

720
00:45:05,860 --> 00:45:09,540
which is a couple of seconds before the monkey actually was pressing a

721
00:45:09,540 --> 00:45:12,420
button, they already knew which way it was going to decide

722
00:45:13,220 --> 00:45:16,920
and they could predict that. But then of course, they

723
00:45:16,920 --> 00:45:20,680
did the following thing. They said, okay, if we can predict how the monkey's going

724
00:45:20,680 --> 00:45:24,480
to decide, we're going to show it opposing signals. So we're

725
00:45:24,480 --> 00:45:27,440
going to have some of the dots suddenly move in a different direction

726
00:45:28,480 --> 00:45:31,440
and see whether the monkey changes its mind. And

727
00:45:32,080 --> 00:45:35,720
what they found was the closer the monkey was to a

728
00:45:35,720 --> 00:45:39,520
decision, the more it simply ignored that new information.

729
00:45:40,570 --> 00:45:44,010
So as the brain was sort of forming the signal from the

730
00:45:44,010 --> 00:45:47,570
noise, it was stronger and stronger and stronger in its

731
00:45:47,570 --> 00:45:51,370
belief and it filtered out anything that was contradicting that

732
00:45:51,610 --> 00:45:55,330
belief. And now, of course always very dangerous

733
00:45:55,330 --> 00:45:58,890
to transfer monkey brain studies to human brains,

734
00:45:59,050 --> 00:46:02,770
but I think ultimately that's the same way that we function as

735
00:46:02,770 --> 00:46:06,250
well. The closer we are to having made up our mind, the more certain we

736
00:46:06,250 --> 00:46:09,990
are of something, the more likely we are to reject contradicting

737
00:46:09,990 --> 00:46:13,310
data. Which of course is a very, very painful irony, because

738
00:46:13,710 --> 00:46:17,550
when would data be the most valuable? When we're certain

739
00:46:17,550 --> 00:46:21,230
of something and it actually contradicts us. And so

740
00:46:21,230 --> 00:46:24,590
we need to recognize the data resistant mind

741
00:46:25,870 --> 00:46:28,910
on an individual level. But then of course, many companies,

742
00:46:29,150 --> 00:46:32,590
organizations, it's not just a single brain. So suddenly you have

743
00:46:32,670 --> 00:46:36,380
lots and lots of data resistant minds interacting with

744
00:46:36,380 --> 00:46:39,180
each other and interacting with each other on different

745
00:46:39,260 --> 00:46:43,100
timescales. And so that again shows you

746
00:46:43,100 --> 00:46:46,700
how ridiculous actually that notion of the data deficit theory is, right?

747
00:46:46,700 --> 00:46:50,460
If you just think that in this jumble of noise you can throw in a

748
00:46:50,460 --> 00:46:54,220
few numbers and they're going to influence a decision, it's futile. You

749
00:46:54,220 --> 00:46:57,940
really need to think about the entire mechanism from beginning to end. And I

750
00:46:57,940 --> 00:47:01,500
mean, we all know this or have been in a situation,

751
00:47:02,740 --> 00:47:06,580
there is no single decision point or decision meeting. Just like in

752
00:47:06,580 --> 00:47:10,060
the monkey brain, where they found there's at first a bit of chaos and it

753
00:47:10,060 --> 00:47:13,860
starts to form a decision signal. I think that resonates very much with how

754
00:47:13,860 --> 00:47:17,540
I perceive organizations making decisions. There's a notion here,

755
00:47:17,779 --> 00:47:21,420
bit of politics over there, and suddenly at some point you feel,

756
00:47:21,420 --> 00:47:24,700
well, things are probably moving in this direction and suddenly everybody is moving in that

757
00:47:24,700 --> 00:47:28,310
direction. And that that is a very

758
00:47:28,310 --> 00:47:30,990
data resistant mechanism that we need to,

759
00:47:32,430 --> 00:47:36,190
where we need to put some engineering into that decision process and the decision

760
00:47:36,430 --> 00:47:40,070
framework, the decision Architecture to make it work. It's not going to happen by

761
00:47:40,070 --> 00:47:43,670
default. Yeah, I often wonder, because if that happens in

762
00:47:43,670 --> 00:47:47,230
monkeys, it happens. Assuming it happens in people, it happens in

763
00:47:47,230 --> 00:47:50,910
organizations. There's got to have been some evolutionary

764
00:47:50,910 --> 00:47:54,600
advantage to it, right? Things don't. Things

765
00:47:54,600 --> 00:47:58,200
don't exist in a vacuum is one of my beliefs. Now, again, maybe I'll see

766
00:47:58,200 --> 00:48:01,720
data and I'll fight it. You tell me that I'm wrong, I'll fight it.

767
00:48:02,360 --> 00:48:06,160
But I really think that things, particularly natural

768
00:48:06,160 --> 00:48:09,960
things like behavior in animals or in groups of people,

769
00:48:10,440 --> 00:48:13,920
or there has to be some kind of evolutionary

770
00:48:13,920 --> 00:48:17,560
advantage. I wonder how we got here. Right? That's the engineering me,

771
00:48:17,560 --> 00:48:21,200
like, how did we get here? That's a fascinating

772
00:48:21,200 --> 00:48:24,840
question. And there's actually some people that have thought about that. So I

773
00:48:24,840 --> 00:48:28,480
only very briefly mentioned that in the book. There's a book by two

774
00:48:28,480 --> 00:48:32,200
researchers called Mercier and Sperber who make a

775
00:48:32,200 --> 00:48:36,000
very fascinating argument. So they start with the

776
00:48:36,000 --> 00:48:39,760
following premise. They say, if rational

777
00:48:39,760 --> 00:48:43,440
thinking is so good, why don't we see it all over

778
00:48:43,440 --> 00:48:47,280
the place? Why are we humans the only creatures that have

779
00:48:47,280 --> 00:48:50,990
evolved rational thoughts if it's so good? And

780
00:48:50,990 --> 00:48:54,550
the conclusion they come to which you can

781
00:48:54,550 --> 00:48:57,350
subscribe to or controversial, but I find it very interesting is that they say, well,

782
00:48:57,350 --> 00:49:00,950
maybe rational thought did not evolve for

783
00:49:00,950 --> 00:49:04,230
actually thinking through a decision before we make it.

784
00:49:04,390 --> 00:49:08,150
Maybe rational thought evolved to rationalize and to

785
00:49:08,150 --> 00:49:11,350
justify decisions after they have been made,

786
00:49:11,750 --> 00:49:15,070
which again, I'm not entirely sure whether I subscribe to that

787
00:49:15,070 --> 00:49:18,810
myself fully, but I think it would make a lot of sense

788
00:49:18,810 --> 00:49:21,410
in that context, if you think about

789
00:49:22,370 --> 00:49:26,170
rational thought not being the decision making process that's in

790
00:49:26,170 --> 00:49:29,570
control of everything, but something that comes in after

791
00:49:31,170 --> 00:49:34,850
and helps us justify it to ourselves and others,

792
00:49:35,250 --> 00:49:37,890
how we came to a certain conclusion.

793
00:49:38,850 --> 00:49:42,290
An interesting piece of research that adds to that. So there's

794
00:49:43,860 --> 00:49:47,460
various studies that have been done with humans that have

795
00:49:47,460 --> 00:49:51,060
had a particular kind of brain damage where

796
00:49:51,060 --> 00:49:54,820
they're not able to process emotions, so they have a damage

797
00:49:54,820 --> 00:49:58,660
to a certain part of the prefrontal cortex.

798
00:49:58,980 --> 00:50:02,500
And you would think that these people are the perfect rational

799
00:50:02,580 --> 00:50:06,260
decision makers. They have lost the capability to process emotions.

800
00:50:06,260 --> 00:50:09,940
And it turns out also otherwise they behave normally, they, they talk

801
00:50:09,940 --> 00:50:13,380
normally, they have high IQ scores and everything. And what the

802
00:50:13,380 --> 00:50:16,660
researchers are finding is that these people are

803
00:50:16,660 --> 00:50:20,500
completely incapable of making any decision at

804
00:50:20,500 --> 00:50:23,940
all. They can't even decide. Yes, they cannot even

805
00:50:23,940 --> 00:50:27,100
decide what to have for lunch because they will find

806
00:50:27,580 --> 00:50:31,420
an additional rational reason and yet another additional rational

807
00:50:31,420 --> 00:50:35,220
reason for why this sandwich might be better. Than the salad or why the

808
00:50:35,220 --> 00:50:38,970
salad might be better than the sandwich or the burger. So these people become

809
00:50:39,130 --> 00:50:42,890
incapable of decision making and these

810
00:50:42,890 --> 00:50:46,650
studies. So there's a brain researcher, Antonio Damasio, who

811
00:50:46,650 --> 00:50:50,170
started that. Their conclusion is

812
00:50:50,170 --> 00:50:53,570
that most decisions just

813
00:50:53,570 --> 00:50:56,930
cannot be made on a purely rational level. So you need

814
00:50:56,930 --> 00:51:00,690
emotion to tip the scale at some point to move you

815
00:51:00,690 --> 00:51:04,010
in that direction. So that mechanism that stands in the way

816
00:51:04,430 --> 00:51:07,870
where ultimately, you know, the emotion is telling us, I'm not going to listen to

817
00:51:07,870 --> 00:51:10,350
this data and I'm going to subscribe to that data

818
00:51:11,630 --> 00:51:14,830
has a flip side to it where it might actually be

819
00:51:15,630 --> 00:51:19,150
the part of the brain that allows us to make

820
00:51:19,150 --> 00:51:22,830
decisions in the first place, because otherwise we would just be overwhelmed

821
00:51:22,830 --> 00:51:25,230
by data that doesn't give us a clear direction.

822
00:51:26,670 --> 00:51:30,350
And that's true, Link. And also too, thinking requires effort,

823
00:51:31,160 --> 00:51:34,800
FDA requires calories, and calories were not at a surplus until

824
00:51:34,800 --> 00:51:37,880
very recently in human history. Yes.

825
00:51:38,440 --> 00:51:41,880
So I think maybe there's something. It's an optimization trick, Right.

826
00:51:43,320 --> 00:51:47,080
Maybe that's what it is. It's a fascinating thing.

827
00:51:48,040 --> 00:51:51,720
And we can go down that rabbit hole. But one of the things I thought

828
00:51:51,720 --> 00:51:55,240
was interesting is that you said that there were four ways to

829
00:51:55,240 --> 00:51:58,680
derail your data driven journey.

830
00:51:58,680 --> 00:52:02,400
Yeah, yeah. So what, what are those four things to avoid?

831
00:52:03,600 --> 00:52:07,200
Yeah, so, so the first one, and this comes really out of the project experience,

832
00:52:07,360 --> 00:52:11,080
I think with, with many data projects, they start out by

833
00:52:11,080 --> 00:52:14,680
saying, let's do a quick win or let's do a pilot or

834
00:52:14,680 --> 00:52:18,360
let's do a lighthouse project. So let's take the lighthouse project for

835
00:52:18,360 --> 00:52:22,200
example. And I think the assumption there is everybody thinks data

836
00:52:22,200 --> 00:52:25,820
is hard and you know, people might not like to engage with it.

837
00:52:26,140 --> 00:52:29,900
So what if we build a very impressive lighthouse project,

838
00:52:31,100 --> 00:52:34,940
then surely everybody must look at this and say, oh, this was a

839
00:52:34,940 --> 00:52:38,660
great idea, we're going to come along. It

840
00:52:38,660 --> 00:52:42,420
doesn't pan out that way for various reasons. I think one of the main reasons

841
00:52:42,420 --> 00:52:45,980
is that the organization will look at this lighthouse project and say, well,

842
00:52:46,300 --> 00:52:50,140
okay, you constructed this in a place where it was very

843
00:52:50,140 --> 00:52:53,880
easy to construct a lighthouse. You know, usually these lighthouse projects,

844
00:52:53,880 --> 00:52:56,520
they get all the management attention, they get the budget.

845
00:52:58,680 --> 00:53:02,480
If for some reason they don't work, they get more attention, they get more budget

846
00:53:02,480 --> 00:53:06,080
because everybody's already committed to them. And so they don't get the

847
00:53:06,080 --> 00:53:09,840
organization moving. So I think this notion of let's not engage with

848
00:53:09,840 --> 00:53:13,280
the difficult parts of this, just build a lighthouse, I think is one of the

849
00:53:13,280 --> 00:53:16,680
main traps that I see now.

850
00:53:17,560 --> 00:53:20,520
Another trap is that I think

851
00:53:22,290 --> 00:53:26,050
let's say out of a good intention, which is the intention to

852
00:53:26,290 --> 00:53:30,050
avoid complexity. Oftentimes

853
00:53:30,450 --> 00:53:33,170
use cases, various data use cases will be

854
00:53:34,050 --> 00:53:37,570
analyzed in a very isolated fashion from each other. So

855
00:53:37,970 --> 00:53:41,210
I think many people working in the data space will have seen this. You take

856
00:53:41,210 --> 00:53:44,850
the individual use case and then it gets a score, maybe the

857
00:53:44,850 --> 00:53:48,570
complexity and the business value gets put just this two

858
00:53:48,570 --> 00:53:51,930
by two matrix. And then you sort of try to find, oh, we're going to

859
00:53:51,930 --> 00:53:55,730
do the ones that are sort of medium difficulty and we want to

860
00:53:55,730 --> 00:53:58,450
get a lot of business value out of it. And

861
00:53:59,410 --> 00:54:03,090
what that does is I think it neglects the fact

862
00:54:03,090 --> 00:54:06,410
that many of these use cases will be related and all of them will be

863
00:54:06,410 --> 00:54:10,170
tied to building a solid data foundation. And

864
00:54:10,170 --> 00:54:13,890
so when there's oftentimes I think a complaint of saying,

865
00:54:13,890 --> 00:54:17,550
oh, you know, we're doing the same data transformation in 10 different

866
00:54:17,550 --> 00:54:21,310
places. And oftentimes where that comes from is because the use cases

867
00:54:21,310 --> 00:54:24,830
are planned and implemented in isolation, because nobody wants to deal with a

868
00:54:24,830 --> 00:54:28,470
complex topic of touching the entire data

869
00:54:28,470 --> 00:54:32,030
foundation. And so I offer some workshop

870
00:54:32,030 --> 00:54:35,830
formats and some methodologies that I found useful to

871
00:54:35,830 --> 00:54:39,670
actually make sure that you put all of these use cases on a

872
00:54:39,670 --> 00:54:43,390
common ground basis and are able to connect them

873
00:54:43,390 --> 00:54:47,050
to each other. Then there's

874
00:54:47,290 --> 00:54:50,330
another interesting one, which is that

875
00:54:50,570 --> 00:54:54,170
oftentimes we believe that the data

876
00:54:54,330 --> 00:54:57,930
and the technology is going to be the hard part

877
00:54:58,010 --> 00:55:01,730
about a data project. So the

878
00:55:01,730 --> 00:55:04,850
example that I previously mentioned, we were actually

879
00:55:04,850 --> 00:55:08,690
optimizing trucks driving through Southeast Asia and

880
00:55:08,690 --> 00:55:12,240
we were launching in that thinking that the hard part would be

881
00:55:12,240 --> 00:55:15,960
designing the algorithm, because as you and probably a lot of

882
00:55:15,960 --> 00:55:19,760
listeners know, finding optimal routes, it is a hard problem, and especially

883
00:55:20,000 --> 00:55:23,720
with different parameters coming in and so on. But we

884
00:55:23,720 --> 00:55:27,400
actually developed an algorithm that was doing pretty pretty well and pretty, pretty

885
00:55:27,400 --> 00:55:30,720
fast. What we had completely

886
00:55:30,720 --> 00:55:34,520
overlooked and underestimated were two things. One is the

887
00:55:34,520 --> 00:55:38,190
complexity of restructuring a warehouse, because in order to

888
00:55:38,270 --> 00:55:41,590
drive these optimal routes, you would have to have completely

889
00:55:41,590 --> 00:55:45,270
restructured the way the warehouse is working. And that was a

890
00:55:45,270 --> 00:55:48,830
complexity that everybody was just afraid of because you would essentially

891
00:55:48,830 --> 00:55:52,310
be, it's almost like made to order, right? You would say, okay, put this package

892
00:55:52,310 --> 00:55:55,670
here and then put this package there. And that's not how it works. They get

893
00:55:55,670 --> 00:55:59,310
their pallets and the pallets get into the truck entirely.

894
00:55:59,870 --> 00:56:03,710
And the other part was simply politics because the hard

895
00:56:03,710 --> 00:56:07,340
part there was, well, if you tell people you can save 20% of fuel,

896
00:56:07,500 --> 00:56:11,220
you're also telling them you have been wasting 20% of fuel in

897
00:56:11,220 --> 00:56:15,020
the past. And so we should have been a

898
00:56:15,020 --> 00:56:18,860
bit more careful about that, which I can completely understand. And it

899
00:56:18,860 --> 00:56:22,660
was also a very political organization in many places.

900
00:56:22,660 --> 00:56:26,300
So the data project was

901
00:56:26,300 --> 00:56:30,020
hard, but it wasn't the hard part. And I think oftentimes when we

902
00:56:30,020 --> 00:56:33,080
say, let's implement a use case, we think about, where are we going to get

903
00:56:33,080 --> 00:56:36,440
the data, who's going to write the algorithm, how are we going to get that

904
00:56:36,440 --> 00:56:39,920
into production? It and completely neglecting all of these other

905
00:56:39,920 --> 00:56:43,200
factors, which ties into the fourth trap, which is

906
00:56:43,280 --> 00:56:47,080
ignoring the human factor. And I think this is something that

907
00:56:47,080 --> 00:56:50,560
comes very much out of the data deficit theory, where we think,

908
00:56:50,640 --> 00:56:54,040
oh, once we show this amazing technology to

909
00:56:54,040 --> 00:56:57,440
everybody, everybody will be happy. And I use this example in the book of saying,

910
00:56:57,600 --> 00:57:01,280
let's say you are using a language model to make an automated

911
00:57:01,280 --> 00:57:04,970
marketing newsletter. As a technology person, you would think

912
00:57:04,970 --> 00:57:08,650
this is amazing. I can categorize the customers

913
00:57:08,650 --> 00:57:11,850
into clusters, then I can address each of these clusters

914
00:57:12,170 --> 00:57:16,010
specifically. I can measure the response rates, I can measure the open

915
00:57:16,010 --> 00:57:19,370
rates. All of these really fantastic things. Not very

916
00:57:19,370 --> 00:57:23,170
expensive. The technology is there. Let's go for

917
00:57:23,170 --> 00:57:27,010
it. What you don't recognize is that there's a whole bunch of people

918
00:57:27,010 --> 00:57:30,570
in the organization that will think very differently about it. The

919
00:57:30,930 --> 00:57:34,770
marketing department feels like it's losing control. You might have, if it's a

920
00:57:34,770 --> 00:57:38,330
retailer, let's say you will have a category manager, maybe that says, well, I'm

921
00:57:38,330 --> 00:57:41,650
incentivized on the sales here, so if your automated

922
00:57:41,650 --> 00:57:45,370
newsletter is suddenly disrupting what gets sold and what gets

923
00:57:45,370 --> 00:57:49,210
left behind, my bonus is gone. The warehouse manager

924
00:57:49,210 --> 00:57:52,770
says, oh, your newsletter is suddenly going to change our

925
00:57:52,770 --> 00:57:56,050
orders. You know, can we even have that stock ready in time? I don't know.

926
00:57:56,540 --> 00:58:00,300
And so that's, I think, the big, big human factor that is often

927
00:58:00,620 --> 00:58:03,820
underestimated. And maybe as a corollary of that, I think also

928
00:58:04,220 --> 00:58:07,900
sometimes, and I'm guilty of that myself, you know,

929
00:58:07,900 --> 00:58:11,540
sometimes as data people, we just miss out on some of the operational

930
00:58:11,540 --> 00:58:15,260
realities. I know one project we weren't involved in, but it was

931
00:58:15,260 --> 00:58:19,060
involving a steel manufacturer and they had an

932
00:58:19,060 --> 00:58:22,460
agency that had designed an app for the workers

933
00:58:22,860 --> 00:58:26,710
to sort of see how the machines were doing. The only problem was

934
00:58:26,710 --> 00:58:30,430
that app didn't really work with the mandatory safety gloves that everybody

935
00:58:30,430 --> 00:58:34,070
had to wear. Had these big, big, big thick

936
00:58:34,070 --> 00:58:37,030
mittens trying to control an iPad. Doing that.

937
00:58:37,750 --> 00:58:41,509
Not really the way it works. That's come up quite a bit,

938
00:58:41,509 --> 00:58:45,190
actually. The notion that they get these apps and for whatever reason,

939
00:58:45,190 --> 00:58:48,750
the end users were never consulted or brought into

940
00:58:48,750 --> 00:58:52,150
the process. Amazing. Isn't it? Yeah. Yeah.

941
00:58:53,990 --> 00:58:55,350
The book is called

942
00:58:58,310 --> 00:59:01,990
Data Inspired and the author is

943
00:59:01,990 --> 00:59:05,350
Sebastian Wernicke who's been here speaking with us.

944
00:59:07,510 --> 00:59:11,030
Thank you. And where can folks get the book on Amazon, on

945
00:59:11,030 --> 00:59:14,470
Audible, everywhere where books are sold. And

946
00:59:15,030 --> 00:59:18,590
if you get enough on Amazon, they will surely be available on Audible or as

947
00:59:18,590 --> 00:59:22,290
an audiobook soon. So we're still working on that. I

948
00:59:22,290 --> 00:59:26,010
do love myself a good audiobook. Yeah, me too. Me too. So

949
00:59:26,170 --> 00:59:29,770
hopefully that's going to be in the making soon. And I mean, if you want

950
00:59:29,770 --> 00:59:33,050
to find out more, you can go to datainspired.org that's my website

951
00:59:33,610 --> 00:59:37,330
and I'm always happy to connect with folks on LinkedIn. So that's where I

952
00:59:37,330 --> 00:59:40,970
am. That's where I regularly post. So get in touch and reach out and let's

953
00:59:40,970 --> 00:59:44,730
have some data conversations. Excellent. Thank you. I'll let the

954
00:59:44,730 --> 00:59:45,610
outro music play.

955
00:59:45,840 --> 00:59:56,720
Sam.