Responsible AI Is Good Business — Featuring Wiebke Apitzsch

1
00:00:00,000 --> 00:00:07,560
Yeah, welcome back to the ANSI 65L, the podcast where we explore Microsoft, AI, Cloud, Business,

2
00:00:07,560 --> 00:00:11,760
Transformation and the people sharing the future of technology.

3
00:00:11,760 --> 00:00:13,280
Today's guest brings a perspective.

4
00:00:13,280 --> 00:00:16,960
We don't often hear in AI discussions.

5
00:00:16,960 --> 00:00:23,280
While many conversations focus on bigger models, more agents, faster automation, she asks

6
00:00:23,280 --> 00:00:24,280
gifts and questions.

7
00:00:24,280 --> 00:00:25,480
Should we build it?

8
00:00:25,480 --> 00:00:31,680
People can people trust it. How do we make AI actively useful inside organizations?

9
00:00:31,680 --> 00:00:39,320
Joining us today is we've got a bit, well, founder of AI Impact, AI strategy, advisor,

10
00:00:39,320 --> 00:00:44,320
executive coach, keynote speaker and someone working right at the intersection of technology,

11
00:00:44,320 --> 00:00:50,360
philosophy, ethics and business strategy with experience from top tier, consulting, including

12
00:00:50,360 --> 00:00:57,640
BCG, excluding coaching at in-seat, ongoing philosophy studies and deep practical working,

13
00:00:57,640 --> 00:01:01,920
helping organization adopting generative AI responsible.

14
00:01:01,920 --> 00:01:05,360
She helps leadership teams move from AI hype to AI impact.

15
00:01:05,360 --> 00:01:11,240
Today we will explore responsibly AI, Microsoft co-pilot, AI agents, company culture, trust,

16
00:01:11,240 --> 00:01:14,360
ethics, leaderships and the future of work, and much more.

17
00:01:14,360 --> 00:01:17,360
We've got come to the ANSI 65L as M.

18
00:01:17,360 --> 00:01:18,360
Hi.

19
00:01:18,360 --> 00:01:21,200
That was a very long list of items to cover today.

20
00:01:21,200 --> 00:01:24,440
Looking forward and let's see how far we can get.

21
00:01:24,440 --> 00:01:29,760
Yeah, before we get really deep down into the topics, can you tell us a little bit about

22
00:01:29,760 --> 00:01:37,000
you and why you are fascinated from technology?

23
00:01:37,000 --> 00:01:43,320
Working in technology was never my plan.

24
00:01:43,320 --> 00:01:45,760
After school I didn't really have a clear idea of what to do.

25
00:01:45,760 --> 00:01:51,120
When I went into the hospitality industry and I think what I took from there is customer

26
00:01:51,120 --> 00:01:52,120
first.

27
00:01:52,120 --> 00:01:55,360
I think you don't learn it as well anywhere else.

28
00:01:55,360 --> 00:02:04,120
Then I started to stumble into my first of the three topics, technology, when I worked

29
00:02:04,120 --> 00:02:12,240
with little hotel databases to do CRM and moved to BCG where I joined a very specific team

30
00:02:12,240 --> 00:02:18,600
which was building little prototypes and models for those very big clients and I travelled

31
00:02:18,600 --> 00:02:24,440
a lot and I coded a little not very well but I understood how it works.

32
00:02:24,440 --> 00:02:32,640
Also I was able to see a little bit of the strategy consulting as well, how do people

33
00:02:32,640 --> 00:02:34,640
set up strategies, how do we move forward?

34
00:02:34,640 --> 00:02:38,720
I was part of the team who built up BCGX.

35
00:02:38,720 --> 00:02:45,640
That was more strategic work and one other thing I did in between was to look a little bit

36
00:02:45,640 --> 00:02:46,640
into philosophy.

37
00:02:46,640 --> 00:02:53,280
I did study half a year at the LMU and then I found out that I am not very good at driving

38
00:02:53,280 --> 00:02:58,000
cars and I thought taxi drivers are not for me so I stopped doing that and went to business

39
00:02:58,000 --> 00:02:59,480
administration.

40
00:02:59,480 --> 00:03:02,440
These three topics followed me all my life.

41
00:03:02,440 --> 00:03:08,680
Technology, philosophy and strategy because you have to understand what it actually is.

42
00:03:08,680 --> 00:03:13,480
You have to think where it should go to and then you have to execute.

43
00:03:13,480 --> 00:03:17,400
With one of those is missing, you don't get anywhere.

44
00:03:17,400 --> 00:03:23,360
Over the years I learned and I practiced and I did projects and now I really feel it's

45
00:03:23,360 --> 00:03:28,480
the moment to talk about it more because I changed so many things that we have to go back

46
00:03:28,480 --> 00:03:32,240
to the basics to make good decisions.

47
00:03:32,240 --> 00:03:39,280
How is the consulting part influencing your thinking about your thinking today?

48
00:03:39,280 --> 00:03:47,680
I think these three topics that I mentioned are very different and business administration

49
00:03:47,680 --> 00:03:54,960
and consulting is from a logic perspective the simplest because we calculate with plus

50
00:03:54,960 --> 00:04:00,000
and minus and we have like there's one euro and I need to get 120 back so that it's a

51
00:04:00,000 --> 00:04:04,640
good case and if we only get 90 cents back or nothing then it's a bad case so from that perspective

52
00:04:04,640 --> 00:04:12,720
it's very very simple and but it is very complex in terms of stakeholder management.

53
00:04:12,720 --> 00:04:19,680
There's a lot of emotion, various fears because while in technology you can do a deep analysis

54
00:04:19,680 --> 00:04:25,320
and find out that this model is the best today in business you are already good if you win

55
00:04:25,320 --> 00:04:29,720
at 51% of the cases.

56
00:04:29,720 --> 00:04:36,200
Some people who work in how you call market chairs and trading and all that that's what they

57
00:04:36,200 --> 00:04:37,200
do.

58
00:04:37,200 --> 00:04:43,280
They are 21% right and that makes them billionaires.

59
00:04:43,280 --> 00:04:50,320
They are come to coaching and the executing and the being fast and making both decisions

60
00:04:50,320 --> 00:04:56,120
even though you don't know and it's really hard to integrate but I think if you build good

61
00:04:56,120 --> 00:04:57,560
teams then you can do it.

62
00:04:57,560 --> 00:05:05,040
So just to make it one more in a very short sentence the consulting forces you to make both

63
00:05:05,040 --> 00:05:11,480
decisions fast and to make sure you execute because as long as you are in a theory you're

64
00:05:11,480 --> 00:05:15,760
in academia and you don't make money and we have to make money because else you don't have

65
00:05:15,760 --> 00:05:18,760
companies.

66
00:05:18,760 --> 00:05:23,160
This is what I learned from that and this is what always gives some pressure to my otherwise

67
00:05:23,160 --> 00:05:25,840
more bold thinking.

68
00:05:25,840 --> 00:05:29,920
Yeah and now you are the founder of AI Intect.

69
00:05:29,920 --> 00:05:35,200
Can you tell a little bit about your company and why you found it?

70
00:05:35,200 --> 00:05:44,520
So we are three managing directors and we are very very different and the reason why

71
00:05:44,520 --> 00:05:51,400
I wanted to work on my own company is because I always work very hard but then sometimes

72
00:05:51,400 --> 00:05:58,480
you do it for someone else and I felt that it would be better to do it with getting the

73
00:05:58,480 --> 00:06:06,240
impact for extra hours, not so much the hour itself but the impact, the whole impact.

74
00:06:06,240 --> 00:06:16,080
And with my two co-founders I think it's very interesting but also very valuable mix because

75
00:06:16,080 --> 00:06:23,280
you heard a lot about myself now but then there's Rutga who used to be a professor for physics

76
00:06:23,280 --> 00:06:31,360
and he is very deep in mathematics and logic and models and it took us quite a while to

77
00:06:31,360 --> 00:06:39,720
be able to have a conversation because frankly sometimes I just didn't get it but now that

78
00:06:39,720 --> 00:06:47,640
I started to study philosophy again that became much easier because mathematics and philosophy

79
00:06:47,640 --> 00:06:53,600
it seemed to be very similar and with big ideas like the knowledge machine of light and

80
00:06:53,600 --> 00:06:59,640
a sense of one, we have a common ground and we cannot develop models together that will

81
00:06:59,640 --> 00:07:05,600
actually work in real life so but it's very complex and then we have Felix who is from

82
00:07:05,600 --> 00:07:14,840
the IT area not only, he also is very good at working with clients and thinking into the future

83
00:07:14,840 --> 00:07:19,440
but he also knows all these basics like how do I implement it?

84
00:07:19,440 --> 00:07:21,640
What is the right database to use?

85
00:07:21,640 --> 00:07:28,760
What do we do and how do we exchange data from A to Z?

86
00:07:28,760 --> 00:07:35,520
So that's a combination and I think together we can really make a difference because

87
00:07:35,520 --> 00:07:41,240
only if you think all three dimensions together you can really create projects that deliver

88
00:07:41,240 --> 00:07:46,240
what you expected in the time that you estimated.

89
00:07:46,240 --> 00:07:50,080
Yeah, that's interesting.

90
00:07:50,080 --> 00:07:56,880
By the co-teams though you have all these different parts there.

91
00:07:56,880 --> 00:08:05,920
I think a little bit companies or everybody says they are doing AI, every product has

92
00:08:05,920 --> 00:08:12,280
this label AI but yeah, I see it on my coffee machine.

93
00:08:12,280 --> 00:08:21,560
It's not great and any value for me that's AI inside but yeah, why is AI adoption still

94
00:08:21,560 --> 00:08:24,360
so difficult?

95
00:08:24,360 --> 00:08:33,280
I think the expectation and what the tool can actually deliver, they differ and that's

96
00:08:33,280 --> 00:08:34,280
for many reasons.

97
00:08:34,280 --> 00:08:38,000
So to be honest everyone goes like let's do AI and then people talk to us and we always

98
00:08:38,000 --> 00:08:40,400
go like can we avoid AI?

99
00:08:40,400 --> 00:08:46,840
Do we really have to use that or can we maybe solve that in a different way?

100
00:08:46,840 --> 00:08:50,200
Because how does it press the best?

101
00:08:50,200 --> 00:08:58,040
So I think many years we were building very smart models like machine learning, predictions

102
00:08:58,040 --> 00:09:04,560
and so on and they did what they were supposed to do but most people didn't understand what

103
00:09:04,560 --> 00:09:10,440
happened so they didn't trust the system and we had a problem with data quality but that's

104
00:09:10,440 --> 00:09:11,880
what everyone says.

105
00:09:11,880 --> 00:09:17,880
So now we have Genie I and suddenly it became useful for sea level positions.

106
00:09:17,880 --> 00:09:24,200
So they realized okay if this really improves my email writing then it must be able to do

107
00:09:24,200 --> 00:09:29,120
everything and that is not true.

108
00:09:29,120 --> 00:09:33,800
So there is this question about how do I train the data and how can I apply it, how

109
00:09:33,800 --> 00:09:39,640
should it get access to my own information because the impression is that it can really

110
00:09:39,640 --> 00:09:46,720
write a very cool essay about Kant so it should be able to write a short essay about my

111
00:09:46,720 --> 00:09:51,280
cups that I am producing but of course it doesn't have the data, it doesn't have the background

112
00:09:51,280 --> 00:09:57,160
information, we have to build up a secure system with their confidential stuff in there.

113
00:09:57,160 --> 00:10:04,320
So we have a very enthusiastic start and then we have a very heavy low because people then

114
00:10:04,320 --> 00:10:09,680
realize it's not as easy as I thought and this doesn't work like this and they stop.

115
00:10:09,680 --> 00:10:14,680
I think that's the one thing and the other thing is really a more philosophical question

116
00:10:14,680 --> 00:10:20,720
because after a workshop people are often very enthusiastic and they feel like it's

117
00:10:20,720 --> 00:10:26,440
start, let's do AI first and let's do all of this automated and then when they come home

118
00:10:26,440 --> 00:10:37,200
they feel more like what does it mean for my company, for my colleagues, for the future.

119
00:10:37,200 --> 00:10:44,600
If I apply this cold bot to all of my processes, what will my anchors just say when I arrive

120
00:10:44,600 --> 00:10:46,800
in heaven someday?

121
00:10:46,800 --> 00:10:54,280
And then they hesitate, it's not really that they actively say no but it's more like an

122
00:10:54,280 --> 00:10:59,240
unclear gut feeling, what am I doing to my company?

123
00:10:59,240 --> 00:11:05,040
And I think this keeps many people from execution.

124
00:11:05,040 --> 00:11:12,000
That's why we also offer this ethics guideline and introduction workshops with some

125
00:11:12,000 --> 00:11:16,800
philosophy thinking and that's why we asked the question, who do we want to be with AI going

126
00:11:16,800 --> 00:11:17,800
forward?

127
00:11:17,800 --> 00:11:25,240
And what is my personal view on how this company should operate with Gen AI?

128
00:11:25,240 --> 00:11:29,880
And the big question that we have to ask there is what is mankind?

129
00:11:29,880 --> 00:11:34,440
Who are humans in this company?

130
00:11:34,440 --> 00:11:39,400
And what activities should be only for humans?

131
00:11:39,400 --> 00:11:44,880
Not less if AI could maybe from a technological perspective cover some of it or not.

132
00:11:44,880 --> 00:11:51,760
So once we dice them out, it typically becomes very easy to automate everything else because

133
00:11:51,760 --> 00:11:54,040
the very is gone.

134
00:11:54,040 --> 00:12:00,640
And yeah, and yeah, I can bring it, I don't know, should I go a bit deeper in the philosophy

135
00:12:00,640 --> 00:12:07,440
behind it or is that something new for me, so I would really interested.

136
00:12:07,440 --> 00:12:20,800
So one I really like to cite for this is Kant with his thoughts about dignity and price.

137
00:12:20,800 --> 00:12:29,000
So he says high level, I'm not citing exactly, but he says everything either has a pride or

138
00:12:29,000 --> 00:12:32,760
dignity and everything that has a price can be automated.

139
00:12:32,760 --> 00:12:40,360
So I think that is very logical to us, yeah, if I'm just running through, I don't know, automated

140
00:12:40,360 --> 00:12:45,280
emails with notifications like tonight, the server will be updated or what.

141
00:12:45,280 --> 00:12:51,280
There is no value in having a real person sending that email to 2,000 people, right?

142
00:12:51,280 --> 00:12:59,560
And we don't feel that it is in any way disrespecting us as humans to get such an update

143
00:12:59,560 --> 00:13:00,560
email.

144
00:13:00,560 --> 00:13:07,880
It has a price, it is just a thing that someone has to do to notify me, but there is no personality,

145
00:13:07,880 --> 00:13:09,680
no emotion involved.

146
00:13:09,680 --> 00:13:16,720
And then there's other things where we interact with humans and of course, they also deliver

147
00:13:16,720 --> 00:13:18,680
something with price with value.

148
00:13:18,680 --> 00:13:19,680
Yeah.

149
00:13:19,680 --> 00:13:23,600
So for example, they call me to say that there's maybe a problem with my bank account.

150
00:13:23,600 --> 00:13:26,080
So there's something I would pay for and that's fine.

151
00:13:26,080 --> 00:13:28,700
But I always have to treat the human as a human.

152
00:13:28,700 --> 00:13:32,140
I have to respect his or her dignity.

153
00:13:32,140 --> 00:13:38,980
And by doing so in both directions, we build up a relation and that is something with a

154
00:13:38,980 --> 00:13:40,700
very big value.

155
00:13:40,700 --> 00:13:48,300
And so if I'm going back to business administration, the two most costly things are finding clients

156
00:13:48,300 --> 00:13:52,020
and keeping employees or finding employees, right?

157
00:13:52,020 --> 00:13:53,580
These are the two things.

158
00:13:53,580 --> 00:13:57,500
So if I don't have employees and I don't have clients, then I don't have a company.

159
00:13:57,500 --> 00:14:04,940
So that makes it very clear that this is essential to make sure that every area where I build a

160
00:14:04,940 --> 00:14:09,940
relationship, I have to really protect them.

161
00:14:09,940 --> 00:14:18,460
Does mean that I cannot use AI in these processes, but I have to use them as supportive elements.

162
00:14:18,460 --> 00:14:26,700
Like, if I had said, you know what, I'm really busy whether it's very nice, I will just send

163
00:14:26,700 --> 00:14:31,460
my avatar to do this podcast with you.

164
00:14:31,460 --> 00:14:36,580
And maybe if it's very well trained and I prepared it super good with all my knowledge and

165
00:14:36,580 --> 00:14:38,780
all my personal views, whatever.

166
00:14:38,780 --> 00:14:43,700
Maybe in theory, it would perform as good or maybe much better than I do, but the moment

167
00:14:43,700 --> 00:14:48,260
you find out that I sent my avatar, you would not be happy.

168
00:14:48,260 --> 00:14:55,260
So I am pretty sure that you would not post the podcast and you would be super angry.

169
00:14:55,260 --> 00:15:00,140
But for some reason, yeah, and that's, I think that's the interesting point.

170
00:15:00,140 --> 00:15:07,860
For some reason, you prefer the real life Deepge with her, you know, spell wording issues,

171
00:15:07,860 --> 00:15:12,300
you know, sometimes searching for the right word to say it and maybe saying, ah, we have

172
00:15:12,300 --> 00:15:13,300
to redo this.

173
00:15:13,300 --> 00:15:14,300
I made a mistake.

174
00:15:14,300 --> 00:15:20,780
So you prefer this versus having maybe the perfect bot saying the things I say.

175
00:15:20,780 --> 00:15:27,620
And that is why it is because for some reason, the real human matters.

176
00:15:27,620 --> 00:15:32,580
Yeah, we want a real interaction, and everyone listening to this podcast would also say, hey,

177
00:15:32,580 --> 00:15:37,020
I'm not going to listen to a bot, you know, spreading some kind of more or less, maybe

178
00:15:37,020 --> 00:15:39,620
good and maybe stupid content.

179
00:15:39,620 --> 00:15:42,820
I will listen to a human being sharing experience.

180
00:15:42,820 --> 00:15:45,060
You may get distinction here.

181
00:15:45,060 --> 00:15:50,580
So, and this is what we have to put through all our process thinking when we apply AI,

182
00:15:50,580 --> 00:15:53,260
where do I need a human?

183
00:15:53,260 --> 00:15:55,020
And it is no problem at all.

184
00:15:55,020 --> 00:16:00,540
If I, for instance, had a second screen open and maybe I would sometimes cross check and

185
00:16:00,540 --> 00:16:02,460
say, hey, what was the philosopher's name?

186
00:16:02,460 --> 00:16:04,540
Or can you please bring me my favorite cetacean?

187
00:16:04,540 --> 00:16:07,940
I don't remember what that says, what I was like to say.

188
00:16:07,940 --> 00:16:13,260
No issue with that if I use it as support, but you want the real person saying this.

189
00:16:13,260 --> 00:16:14,260
Yeah.

190
00:16:14,260 --> 00:16:18,940
And I would personally encourage every company to think it through to put out where is

191
00:16:18,940 --> 00:16:22,100
the value of humans in my company?

192
00:16:22,100 --> 00:16:23,620
Protect this.

193
00:16:23,620 --> 00:16:29,100
Support this maybe with AI, but not automated and everything else can be automated with no

194
00:16:29,100 --> 00:16:31,340
bad gut feeling at all.

195
00:16:31,340 --> 00:16:40,100
And will then bring efficiency better results, now higher quality, more automation and

196
00:16:40,100 --> 00:16:45,020
all the good things that we actually want from it.

197
00:16:45,020 --> 00:16:55,740
I think a little bit, yeah, we had also, I think, the change before it's most the industrialization

198
00:16:55,740 --> 00:17:00,980
and there are also this had an impact of philosophy.

199
00:17:00,980 --> 00:17:08,740
So I say, I think one of the famous guys, I'd also like, but Karl Marx and so on.

200
00:17:08,740 --> 00:17:16,100
So did you think also AI has an impact on philosophy?

201
00:17:16,100 --> 00:17:24,140
Well, many philosophers are engaged with, you know, the discussing the role of AI, the

202
00:17:24,140 --> 00:17:32,340
question if it has, if it is conscious, the question of how it can be implemented into

203
00:17:32,340 --> 00:17:33,340
our world.

204
00:17:33,340 --> 00:17:38,700
There's a lot of ethical discussions about how we can actually train these models and

205
00:17:38,700 --> 00:17:48,220
if we are, if how we do it is acceptable and I think the short answers most likely know,

206
00:17:48,220 --> 00:17:49,700
but it's still done.

207
00:17:49,700 --> 00:17:59,740
So there's, there's a lot of thinking and for me, the one big thing is that for thousands

208
00:17:59,740 --> 00:18:07,780
of years, the difference that we made between humans and non-humans overall, yeah, was that

209
00:18:07,780 --> 00:18:15,820
humans are those with a language and the ability to think rationally.

210
00:18:15,820 --> 00:18:23,340
So my dog is super cute and he's very empathetic and fun and loving and all.

211
00:18:23,340 --> 00:18:28,700
And we could maybe discuss about the language because I think science found out that animals

212
00:18:28,700 --> 00:18:36,140
do have their own languages, but they don't have this intellectual language that don't

213
00:18:36,140 --> 00:18:42,660
know, writes philosophical texts about our future or statistics and mathematics and do you

214
00:18:42,660 --> 00:18:43,660
think so?

215
00:18:43,660 --> 00:18:45,860
I think most people would agree that animals don't do it.

216
00:18:45,860 --> 00:18:49,980
Maybe they think it's stupid, but or they don't, they're not able, but they don't do

217
00:18:49,980 --> 00:18:50,980
it.

218
00:18:50,980 --> 00:18:52,980
So this I think we can agree on.

219
00:18:52,980 --> 00:19:00,180
And now this typical distinction becomes problematic because we suddenly have a new kind

220
00:19:00,180 --> 00:19:02,020
of creature.

221
00:19:02,020 --> 00:19:03,020
It is not furry.

222
00:19:03,020 --> 00:19:09,500
It is not sweet and it doesn't have like a nice tail waving, yeah?

223
00:19:09,500 --> 00:19:14,180
But it does something that is at least similar to our thinking.

224
00:19:14,180 --> 00:19:19,620
So again, we can make, we can, our books are written, tons of books are written about is

225
00:19:19,620 --> 00:19:27,460
what Jenny I does, is that something like thinking or is it conscious or not?

226
00:19:27,460 --> 00:19:29,700
And I think that's a very good philosophical question.

227
00:19:29,700 --> 00:19:35,700
It's not very practical question because in essence, yeah, I know a few people who never

228
00:19:35,700 --> 00:19:40,860
shouted at their printers because the printer goes like, I cannot print, I don't have magenta

229
00:19:40,860 --> 00:19:43,380
and you go like, it's black and white, go do it.

230
00:19:43,380 --> 00:19:45,580
And the printer goes like no.

231
00:19:45,580 --> 00:19:54,340
So and we feel then that this printer is a super bad person.

232
00:19:54,340 --> 00:19:59,060
So you make it, you make it a person, you give it a personality, you say my printer is

233
00:19:59,060 --> 00:20:00,060
evil.

234
00:20:00,060 --> 00:20:01,060
Yeah?

235
00:20:01,060 --> 00:20:06,140
So how on earth should standard human beings, yeah?

236
00:20:06,140 --> 00:20:13,260
We shouted printers, make a real life distinction between a system that tells them in a very

237
00:20:13,260 --> 00:20:20,220
empathetic way that they are sorry and not, yeah, humanize it in a way, yeah?

238
00:20:20,220 --> 00:20:24,980
And make this clear distinction like, oh no, it is just, it is just working with statistics

239
00:20:24,980 --> 00:20:27,580
and very good assumptions.

240
00:20:27,580 --> 00:20:31,100
And that's why it wrote me this message and I have to cross check it.

241
00:20:31,100 --> 00:20:37,580
The better it gets, the more we will feel that it is like us, even though it's not, but

242
00:20:37,580 --> 00:20:38,860
you still shout at the printer.

243
00:20:38,860 --> 00:20:44,100
So it doesn't make, it makes a theoretical difference, not a real life difference.

244
00:20:44,100 --> 00:20:51,580
And that makes us so important to create a new way of framing the world and make a simple

245
00:20:51,580 --> 00:20:57,500
distinction why we are still in a cool and thoughtful mood.

246
00:20:57,500 --> 00:21:03,340
So that once we get into the angry mood and maybe into the stress, we don't make mistakes

247
00:21:03,340 --> 00:21:07,980
by making an AI, giving an AI the same role as a human.

248
00:21:07,980 --> 00:21:08,980
Yeah?

249
00:21:08,980 --> 00:21:15,740
And I think that's the big philosophical challenge right now to create this new anthropology

250
00:21:15,740 --> 00:21:17,820
and everything else comes after.

251
00:21:17,820 --> 00:21:22,100
But there's a lot of after as well, or in the same time.

252
00:21:22,100 --> 00:21:25,260
Yeah, that's really, really interesting.

253
00:21:25,260 --> 00:21:34,540
I think one big topic is the ethics and a lot of people are talking about ethics in AI.

254
00:21:34,540 --> 00:21:44,740
So there are a lot of problems, but one of the symbols is what role do human world play

255
00:21:44,740 --> 00:21:45,740
in the future?

256
00:21:45,740 --> 00:21:49,540
I think the one of the biggest questions a lot of people have.

257
00:21:49,540 --> 00:21:52,580
What do you think on this?

258
00:21:52,580 --> 00:21:58,540
Yeah, so I think the role in the future is not an ethical question, but a practical question

259
00:21:58,540 --> 00:21:59,540
right?

260
00:21:59,540 --> 00:22:09,700
Because I know that many, many people call themselves AI ethics experts also.

261
00:22:09,700 --> 00:22:17,780
And most of those people do is to say, I think this is bad or I think this is good.

262
00:22:17,780 --> 00:22:21,780
Which is not, which is an opinion, but not ethical thinking.

263
00:22:21,780 --> 00:22:28,260
So if you were to think about what is a good way or a bad way, so is this acceptable or is

264
00:22:28,260 --> 00:22:32,460
this, and how should we treat it?

265
00:22:32,460 --> 00:22:34,500
How should that, what should we do?

266
00:22:34,500 --> 00:22:35,500
Yeah?

267
00:22:35,500 --> 00:22:40,620
There you have to give a broader, you have to go into a broader scope and you have to distinguish

268
00:22:40,620 --> 00:22:44,980
what model you want to use and if what is good for you and all these definitions, which

269
00:22:44,980 --> 00:22:48,420
is lacking very often.

270
00:22:48,420 --> 00:22:53,540
But this question of what role will it play in the future is a practical question.

271
00:22:53,540 --> 00:22:59,140
So, and there, I think it's, get two options.

272
00:22:59,140 --> 00:23:11,820
Yeah, so one is people who don't, who don't care so much about the overall good in the world,

273
00:23:11,820 --> 00:23:17,820
yeah, can automate many things with AI that should not be automated with AI as a

274
00:23:17,820 --> 00:23:20,820
that's what I described before.

275
00:23:20,820 --> 00:23:25,820
But also those people will run into clear boundaries.

276
00:23:25,820 --> 00:23:27,620
And this is what I explained before,

277
00:23:27,620 --> 00:23:30,740
so if you destroy your client relationship

278
00:23:30,740 --> 00:23:36,100
by sending the CEO of your biggest customer,

279
00:23:36,100 --> 00:23:40,900
a sloppy AI email with maybe even wrong information,

280
00:23:40,900 --> 00:23:45,080
you may use this client and then you don't have a company.

281
00:23:46,020 --> 00:23:50,140
So I think that we will, or I hope,

282
00:23:50,140 --> 00:23:55,140
that we will move into a future where from clear economic,

283
00:23:55,140 --> 00:23:57,820
for clear economic reasons,

284
00:23:57,820 --> 00:24:02,020
we make sure that the transact,

285
00:24:02,020 --> 00:24:05,700
that the relational elements remain relational,

286
00:24:05,700 --> 00:24:08,780
and where humans have to do things that are very hard

287
00:24:08,780 --> 00:24:11,820
or not possible to create with AI.

288
00:24:11,820 --> 00:24:16,820
So for example, to make comprehensive decisions,

289
00:24:16,820 --> 00:24:24,220
to make innovation, like really go into new things.

290
00:24:24,220 --> 00:24:28,420
And the reason why I think it's realistic is

291
00:24:28,420 --> 00:24:31,820
because only humans have actual pain.

292
00:24:31,820 --> 00:24:35,820
So if I want to bring in a business innovation,

293
00:24:35,820 --> 00:24:38,420
I can of course use stochastics and say

294
00:24:38,420 --> 00:24:41,220
what other options are there, and then I can choose one.

295
00:24:41,860 --> 00:24:43,860
But a machine doesn't feel pain.

296
00:24:43,860 --> 00:24:46,460
So the machine does not go outside and say,

297
00:24:46,460 --> 00:24:48,820
this is a cold day, I don't like that.

298
00:24:48,820 --> 00:24:52,660
I need a jacket or the machine doesn't sit and say,

299
00:24:52,660 --> 00:24:55,300
I'm bored, the machine doesn't sit and say,

300
00:24:55,300 --> 00:24:57,460
I'm hungry, yeah, or,

301
00:24:57,460 --> 00:25:02,460
why do I have to gain weight if I eat fast food?

302
00:25:02,460 --> 00:25:04,260
Yeah, I'm not happy with that.

303
00:25:04,260 --> 00:25:06,740
So this kind of impulse is something

304
00:25:06,740 --> 00:25:09,500
that really remains humans because the machine is happy,

305
00:25:09,500 --> 00:25:10,340
yeah?

306
00:25:10,340 --> 00:25:15,340
And if nothing happens, you can put things into their base code

307
00:25:15,340 --> 00:25:19,860
to search for something,

308
00:25:19,860 --> 00:25:21,820
but this still comes from the human to say,

309
00:25:21,820 --> 00:25:24,260
I want you to always be kind or so.

310
00:25:24,260 --> 00:25:28,900
Bringing impulses will remain human,

311
00:25:28,900 --> 00:25:31,340
to execute, to check, to test,

312
00:25:31,340 --> 00:25:34,980
also to test from a stochastic perspective,

313
00:25:34,980 --> 00:25:36,940
like what newsletter works better,

314
00:25:36,940 --> 00:25:40,220
to sell my product, this can be done by AI.

315
00:25:40,220 --> 00:25:42,940
And to build up real relationships,

316
00:25:42,940 --> 00:25:45,900
to make sure someone buys from your company,

317
00:25:45,900 --> 00:25:48,940
even though others exist, remains human,

318
00:25:48,940 --> 00:25:53,140
to then execute tasks like copying things from A to B,

319
00:25:53,140 --> 00:25:57,500
and making sure the newsletter goes out every week,

320
00:25:57,500 --> 00:25:59,820
maybe even following up like PMO task,

321
00:25:59,820 --> 00:26:02,260
this could go to AI, but to,

322
00:26:02,260 --> 00:26:05,340
then we cite,

323
00:26:05,340 --> 00:26:06,780
do we want to do this or not?

324
00:26:06,780 --> 00:26:09,820
Do I purchase this piece of land?

325
00:26:09,820 --> 00:26:12,380
Do I build up a new piece?

326
00:26:12,380 --> 00:26:15,420
This is something that will also remain human

327
00:26:15,420 --> 00:26:19,820
because in the end, there is one person that has the money

328
00:26:19,820 --> 00:26:22,660
and this person in the end will want to decide

329
00:26:22,660 --> 00:26:25,020
what to use the money for, yeah?

330
00:26:25,020 --> 00:26:28,180
So I think if you follow this route of thinking,

331
00:26:28,180 --> 00:26:30,980
you can come to a more or less realistic version

332
00:26:30,980 --> 00:26:33,660
of the future without complete dystopia.

333
00:26:33,660 --> 00:26:39,420
- Yeah, I'm gonna move the money here.

334
00:26:39,420 --> 00:26:44,420
I have tried to give one of my agents his own money,

335
00:26:44,420 --> 00:26:47,940
but he doesn't spend it very well.

336
00:26:47,940 --> 00:26:48,780
(laughs)

337
00:26:48,780 --> 00:26:50,140
He did it for all.

338
00:26:50,140 --> 00:26:53,060
Yeah, I have to try it again.

339
00:26:53,060 --> 00:26:58,060
But yeah, I think a big topic is autonomous AI,

340
00:26:58,060 --> 00:27:02,780
how autonomous should AI be,

341
00:27:02,780 --> 00:27:07,780
and is there anything where we have the man in the loop

342
00:27:08,180 --> 00:27:09,700
or a person in the loop?

343
00:27:09,700 --> 00:27:13,500
And yeah, I think this is my first question.

344
00:27:13,500 --> 00:27:15,420
(laughs)

345
00:27:15,420 --> 00:27:18,100
- Yeah, I think this autonomous AI,

346
00:27:18,100 --> 00:27:25,340
I don't know where the value would be to have it, yeah?

347
00:27:25,340 --> 00:27:31,620
And as you just said, so you gave a certain amount

348
00:27:31,620 --> 00:27:36,180
of money to your AI and you decided

349
00:27:36,180 --> 00:27:38,820
that this is kind of risk money.

350
00:27:38,820 --> 00:27:43,180
And so it sounded as if it was gone, can happen, yeah?

351
00:27:43,180 --> 00:27:45,580
It can happen that you make a lot of money with that,

352
00:27:45,580 --> 00:27:50,500
but we also sometimes invest in risky shares.

353
00:27:50,500 --> 00:27:52,140
And if you're lucky, we get it back.

354
00:27:52,140 --> 00:27:54,020
If you're unlucky, we don't.

355
00:27:54,020 --> 00:27:57,900
And there's some kind of advantage sometimes

356
00:27:57,900 --> 00:28:02,900
to do it like this because agents are not so emotional,

357
00:28:02,980 --> 00:28:07,460
so if you give them a clear advice to always sell if X happens,

358
00:28:07,460 --> 00:28:09,580
then they will do it and not be like a human,

359
00:28:09,580 --> 00:28:11,980
as it's all, but maybe it comes back, yeah?

360
00:28:11,980 --> 00:28:14,940
So that's why from a statistical perspective,

361
00:28:14,940 --> 00:28:18,820
it could be a benefit to let go in certain areas

362
00:28:18,820 --> 00:28:21,740
and give clear instructions how to act

363
00:28:21,740 --> 00:28:24,700
and then rely on the thing to act like this.

364
00:28:24,700 --> 00:28:29,860
But so why would you want an autonomous AI?

365
00:28:29,860 --> 00:28:33,420
So you want maybe to improve decisions

366
00:28:33,420 --> 00:28:39,420
and you can do that for example, if you have insurance companies

367
00:28:39,420 --> 00:28:42,860
who have to pay out certain amount of money

368
00:28:42,860 --> 00:28:44,740
if something happens.

369
00:28:44,740 --> 00:28:49,740
And then if the case occurs, your customer will send in a report

370
00:28:49,740 --> 00:28:52,980
and say, hey, I want 200 euros back

371
00:28:52,980 --> 00:28:56,540
because I don't know my dog broke someone's,

372
00:28:56,540 --> 00:28:59,500
I don't know, stereo system by running through,

373
00:28:59,500 --> 00:29:00,700
anything like this.

374
00:29:00,700 --> 00:29:04,420
So then you can give the AI an autonomy

375
00:29:04,420 --> 00:29:07,140
to decide within certain ranges

376
00:29:07,140 --> 00:29:08,740
and then what you should definitely do

377
00:29:08,740 --> 00:29:12,540
in what people also do is to add a tip,

378
00:29:12,540 --> 00:29:16,780
like a standard fraud detection system

379
00:29:16,780 --> 00:29:21,780
but see if once these decisions are made by the AI,

380
00:29:21,780 --> 00:29:25,100
if they're a pattern, trans up, if someone goes like,

381
00:29:25,100 --> 00:29:28,620
oh, now my dog is breaking stereo systems every week, yeah?

382
00:29:28,620 --> 00:29:31,580
So this would be then cutched by the fraud detection system

383
00:29:31,580 --> 00:29:35,060
and by doing so we can likely save a lot of money,

384
00:29:35,060 --> 00:29:38,340
be more efficient every now and then we pay too much.

385
00:29:38,340 --> 00:29:41,220
That's okay because the overall saving is better.

386
00:29:41,220 --> 00:29:44,260
And as you maybe put a threshold to say 200 euros,

387
00:29:44,260 --> 00:29:46,900
we know it's not gonna break the system,

388
00:29:46,900 --> 00:29:48,260
it's really directional.

389
00:29:48,260 --> 00:29:52,820
And yeah, and the really bad guys

390
00:29:52,820 --> 00:29:54,500
you cover the fraud detection.

391
00:29:54,500 --> 00:29:56,380
So it works autonomous in a way,

392
00:29:56,380 --> 00:29:58,420
but it works autonomous in a way

393
00:29:58,420 --> 00:30:02,540
that we would let a call center agent make decisions.

394
00:30:02,540 --> 00:30:04,820
And we would also check on them

395
00:30:04,820 --> 00:30:06,620
if there's someone who starts to make a deal

396
00:30:06,620 --> 00:30:09,300
with his best friends and, you know,

397
00:30:09,300 --> 00:30:10,620
so it's not really autonomous

398
00:30:10,620 --> 00:30:13,860
because we set the boundaries, yeah?

399
00:30:13,860 --> 00:30:15,700
And I think in that sense,

400
00:30:15,700 --> 00:30:20,700
it can really be useful to create kind of autonomous systems

401
00:30:20,700 --> 00:30:23,740
that we either control with statistics and fraud detection

402
00:30:23,740 --> 00:30:27,220
or whatever, or with humans,

403
00:30:27,220 --> 00:30:32,220
whom we actively involve into the work process.

404
00:30:32,220 --> 00:30:35,700
So I always call this like a Coca-Cola test, yeah?

405
00:30:35,700 --> 00:30:38,260
At least most developers can really relate to it.

406
00:30:38,260 --> 00:30:43,740
Because our most developers have no idea

407
00:30:43,740 --> 00:30:46,860
how to brew real Coca-Cola or a fraud-cola

408
00:30:46,860 --> 00:30:48,820
or whatever they prefer, yeah?

409
00:30:48,820 --> 00:30:52,300
But it only takes them one sip or even just, you know,

410
00:30:52,300 --> 00:30:55,500
the nose to say, that's the real stuff

411
00:30:55,500 --> 00:30:58,420
or are you trying to poison me, yeah?

412
00:30:58,420 --> 00:31:03,420
So if we create workflows that give people

413
00:31:03,420 --> 00:31:08,660
the opportunity to test the results of the AI in such a way, yeah?

414
00:31:08,660 --> 00:31:11,900
So you have a look and without really tracking back

415
00:31:11,900 --> 00:31:15,620
the entire process, you know that this is a reasonable result

416
00:31:15,620 --> 00:31:17,700
or it is definitely wrong.

417
00:31:17,700 --> 00:31:22,140
And you can keep control of the autonomous work in between.

418
00:31:22,140 --> 00:31:25,340
You have a person who can really make the call

419
00:31:25,340 --> 00:31:26,340
is this correct or not?

420
00:31:26,340 --> 00:31:28,820
And if it's not, then you have to do what you always did.

421
00:31:28,820 --> 00:31:31,260
And if it's correct, you can move on.

422
00:31:31,260 --> 00:31:33,580
Again, this is autonomous in the middle,

423
00:31:33,580 --> 00:31:36,460
but not in the decision or in the sign-off.

424
00:31:36,460 --> 00:31:38,660
And in this way, I think it makes sense.

425
00:31:38,660 --> 00:31:40,900
I really can't think of any reason

426
00:31:40,900 --> 00:31:44,340
why I would want an AI to make totally free decisions

427
00:31:44,340 --> 00:31:46,220
with my money and maybe running into that,

428
00:31:46,220 --> 00:31:47,340
the risk is too high.

429
00:31:47,340 --> 00:31:52,060
I have a past, a future, and a today, yeah?

430
00:31:52,060 --> 00:31:53,540
And I want a future

431
00:31:53,540 --> 00:31:58,540
and I want to be able to create it, yeah?

432
00:31:58,540 --> 00:32:01,980
And I can't, if I let agents just do anything

433
00:32:01,980 --> 00:32:04,060
that I become passive and I don't like that.

434
00:32:04,060 --> 00:32:12,060
- Yeah, I think when we think about business,

435
00:32:12,060 --> 00:32:17,260
what see you when you have on your client side,

436
00:32:17,260 --> 00:32:22,260
what are the, what's problem or, or,

437
00:32:22,340 --> 00:32:25,620
I don't know what to write more, what are the problem?

438
00:32:25,620 --> 00:32:29,820
What problems are companies struggling with AI most?

439
00:32:29,820 --> 00:32:33,260
- You mean like when they use it?

440
00:32:33,260 --> 00:32:36,980
- Yeah, when they start with the adoption,

441
00:32:36,980 --> 00:32:39,140
the thing AI in the company.

442
00:32:39,140 --> 00:32:44,180
- Yeah, I think that a big problem is

443
00:32:44,180 --> 00:32:47,420
that everyone is driving companies crazy.

444
00:32:47,420 --> 00:32:48,260
Yeah.

445
00:32:49,460 --> 00:32:52,660
There's so much buzz about what model to use,

446
00:32:52,660 --> 00:32:54,780
what is the newest trend?

447
00:32:54,780 --> 00:32:59,780
How is the best picture creating,

448
00:32:59,780 --> 00:33:04,540
LLM going and when will they release something new?

449
00:33:04,540 --> 00:33:09,220
So people get totally lost in trying to follow trends.

450
00:33:09,220 --> 00:33:15,060
And the sad thing is that from a strategy perspective,

451
00:33:15,060 --> 00:33:17,620
nothing really changed, yeah?

452
00:33:17,620 --> 00:33:19,420
So from a business perspective,

453
00:33:19,420 --> 00:33:22,660
I still have the problem that people have since Stone Age,

454
00:33:22,660 --> 00:33:23,700
yeah?

455
00:33:23,700 --> 00:33:25,460
The problem that we have since Stone Age,

456
00:33:25,460 --> 00:33:29,820
or maybe a little later is I have something here,

457
00:33:29,820 --> 00:33:32,660
an item that cost me 90 cent to produce.

458
00:33:32,660 --> 00:33:37,260
So I have to sell it for 110 so that in this year,

459
00:33:37,260 --> 00:33:39,780
I can still survive and, you know,

460
00:33:39,780 --> 00:33:41,900
maybe grow my business a little bit.

461
00:33:41,900 --> 00:33:44,780
The problem did not change at all.

462
00:33:44,780 --> 00:33:47,300
So the best thing that companies can do

463
00:33:47,300 --> 00:33:51,180
is to say, hey, let me review my strategy and my numbers

464
00:33:51,180 --> 00:33:55,100
as always and find out where my problems are

465
00:33:55,100 --> 00:33:58,620
and where my opportunities are as always.

466
00:33:58,620 --> 00:34:00,980
And only then I asked the question, okay,

467
00:34:00,980 --> 00:34:03,260
maybe there are some things that last year,

468
00:34:03,260 --> 00:34:08,260
I think I couldn't solve, but now having a new tool, Genie I,

469
00:34:08,260 --> 00:34:11,060
maybe it's not possible to do that in an efficient way.

470
00:34:11,060 --> 00:34:13,180
My colleague, Rutka always says,

471
00:34:13,180 --> 00:34:15,340
from a technology perspective,

472
00:34:15,340 --> 00:34:19,220
most questions that people give us can we solve

473
00:34:19,220 --> 00:34:22,540
with Genie I tools that are three years old?

474
00:34:22,540 --> 00:34:23,580
Yeah.

475
00:34:23,580 --> 00:34:26,140
We have not, not even, we are not even close

476
00:34:26,140 --> 00:34:30,300
to solving all the problems that we can solve with technology

477
00:34:30,300 --> 00:34:35,140
with four year old technology or even older, yeah?

478
00:34:35,140 --> 00:34:37,460
It's not the problem.

479
00:34:37,460 --> 00:34:42,460
So the business problem has not changed.

480
00:34:42,460 --> 00:34:45,180
The way we can find out what the business problem is,

481
00:34:45,180 --> 00:34:47,860
did not change, people are completely confused

482
00:34:47,860 --> 00:34:51,060
and looking into the stars and trying to find out

483
00:34:51,060 --> 00:34:54,340
what the newest models can do, but they don't, you need them.

484
00:34:54,340 --> 00:35:00,300
Typically, typical problems that are easier to solve

485
00:35:00,300 --> 00:35:05,060
with Genie I, are text-based, are in areas

486
00:35:05,060 --> 00:35:09,380
that are not GDPR relevant.

487
00:35:09,380 --> 00:35:13,820
So anything that's ore that have reasonable data size

488
00:35:13,820 --> 00:35:17,220
and can be run with local models on local service, yeah?

489
00:35:17,220 --> 00:35:21,060
That's the high level advice, not for if that helps,

490
00:35:21,060 --> 00:35:24,580
but other than that, just keep to business

491
00:35:24,580 --> 00:35:27,420
and then just see what problem needs to be solved

492
00:35:27,420 --> 00:35:29,620
and then we see if it's Genie I or not, yeah?

493
00:35:29,620 --> 00:35:35,940
Above our road, would you say, actually, in implementation,

494
00:35:35,940 --> 00:35:40,940
AI, is it a technology problem or is it a culture problem?

495
00:35:40,940 --> 00:35:42,740
It's a culture problem.

496
00:35:42,740 --> 00:35:47,660
Yeah, I said, I think very often it's first of all,

497
00:35:47,660 --> 00:35:52,020
it's a, who do we want to be with AI problem?

498
00:35:52,020 --> 00:35:54,660
Because that's where the hesitation comes from.

499
00:35:54,660 --> 00:35:57,460
So once you're solved it and have kind of an ethical guidelines

500
00:35:57,460 --> 00:35:59,500
for these things you want to do with AI and these things,

501
00:35:59,500 --> 00:36:03,740
you don't want to do with AI, then you have kind of a calm moment

502
00:36:03,740 --> 00:36:06,420
or a moment to breathe because you made this decision

503
00:36:06,420 --> 00:36:10,220
and technology and change and everything comes after, yeah?

504
00:36:10,220 --> 00:36:14,100
So and sometimes you don't even need change

505
00:36:14,100 --> 00:36:18,220
and technology is easy because if you create a process

506
00:36:18,220 --> 00:36:22,860
that simply cuts out a piece of work and gives it back

507
00:36:22,860 --> 00:36:27,540
in a better structured and maybe easier to solve a way,

508
00:36:27,540 --> 00:36:29,660
you don't need change, yeah?

509
00:36:29,660 --> 00:36:33,740
I did it with, so we have one of our clients

510
00:36:33,740 --> 00:36:36,220
was willing to go through the newspaper with us

511
00:36:36,220 --> 00:36:38,940
and we did a very small process there.

512
00:36:38,940 --> 00:36:43,060
It was about having a very long text to describe a book

513
00:36:43,060 --> 00:36:45,700
and they have catalogues where they sell books.

514
00:36:45,700 --> 00:36:52,380
So the task was to shorten this text and so that it fits

515
00:36:52,380 --> 00:36:56,780
the catalog, the size and then it should be printed.

516
00:36:56,780 --> 00:37:00,300
So what people before us tried to do was to create

517
00:37:00,300 --> 00:37:03,580
the full catalog with Gen AI, but that to a problem

518
00:37:03,580 --> 00:37:07,220
that people had a look and then they liked the text somehow,

519
00:37:07,220 --> 00:37:09,860
but I wanted to do small changes or they were insecure

520
00:37:09,860 --> 00:37:12,100
if the text is accurate or not and so on.

521
00:37:12,100 --> 00:37:16,100
And so they just deleted the whole thing and started new.

522
00:37:16,100 --> 00:37:20,420
So what we did is really think about how the process is good.

523
00:37:20,420 --> 00:37:24,300
So what we did is we stayed in the same program

524
00:37:24,300 --> 00:37:26,020
that people used to do before.

525
00:37:26,020 --> 00:37:28,780
So it's in Adobe Project.

526
00:37:28,780 --> 00:37:33,260
And we put in the long text, we shorten it with AI

527
00:37:33,260 --> 00:37:37,100
and if the target was 100 digits, you made it 120,

528
00:37:37,100 --> 00:37:41,900
yeah, so by default, you made sure people had to change something.

529
00:37:41,900 --> 00:37:44,860
And it's easier to shorten something with 100 digits

530
00:37:44,860 --> 00:37:47,220
than 3000 because you don't have to read the whole thing

531
00:37:47,220 --> 00:37:49,900
and move through your head like what can I put in it?

532
00:37:49,900 --> 00:37:52,900
So we got the pattern on how they did it in the past.

533
00:37:52,900 --> 00:37:54,100
You put this in.

534
00:37:54,100 --> 00:37:58,580
We also gave them the full text as a reference if they wanted to.

535
00:37:58,580 --> 00:38:01,420
So now all they have to do is click through,

536
00:38:01,420 --> 00:38:04,660
cut out 20 digits where they would like to do that

537
00:38:04,660 --> 00:38:06,180
and move on.

538
00:38:06,180 --> 00:38:08,660
And this process didn't really need any change

539
00:38:08,660 --> 00:38:10,660
because they didn't have to change.

540
00:38:10,660 --> 00:38:15,540
The only thing that changed for them is they had to read 120 digits

541
00:38:15,540 --> 00:38:16,740
rather than 3000.

542
00:38:16,740 --> 00:38:20,820
But if they wanted to, they could still do it for quality reasons.

543
00:38:20,820 --> 00:38:23,460
And from a technology perspective, yes, you need,

544
00:38:23,460 --> 00:38:28,980
I think this needed like 10 days of development time over all.

545
00:38:28,980 --> 00:38:31,820
But it was not anything where we doubted

546
00:38:31,820 --> 00:38:33,100
that technology could cover it.

547
00:38:33,100 --> 00:38:36,340
So everything like adjusting text, shortening text,

548
00:38:36,340 --> 00:38:39,980
making it longer, it can be done.

549
00:38:39,980 --> 00:38:43,100
No problem if you put the target first,

550
00:38:43,100 --> 00:38:45,220
if you think through how the processor should look like,

551
00:38:45,220 --> 00:38:49,260
if then you implemented technology is not this threshold.

552
00:38:49,260 --> 00:38:51,660
It's good thinking is this threshold, yeah?

553
00:38:51,660 --> 00:38:56,300
- Yeah, we have also the topic.

554
00:38:56,300 --> 00:39:01,300
I often heard it's the problem with buyers or hallucinations.

555
00:39:01,300 --> 00:39:06,140
And yeah, I think what did you think about this,

556
00:39:06,140 --> 00:39:11,140
how can companies or people work with this problem?

557
00:39:11,140 --> 00:39:19,740
- So it becomes a problem if you're trying to automate

558
00:39:19,740 --> 00:39:22,660
to the fullest potential.

559
00:39:22,660 --> 00:39:23,660
Yeah?

560
00:39:23,660 --> 00:39:29,140
So in my little example, where do hallucinations typically come from?

561
00:39:29,140 --> 00:39:33,460
Yeah, I think the word is really, really bad for the effect

562
00:39:33,460 --> 00:39:38,460
because it is not hallucinations are not completely random,

563
00:39:38,460 --> 00:39:42,260
yeah?

564
00:39:42,260 --> 00:39:43,660
But they come from two reasons.

565
00:39:43,660 --> 00:39:48,660
Either the data is not there and then as any good student,

566
00:39:48,660 --> 00:39:53,140
it solves it in the best possible way.

567
00:39:53,140 --> 00:39:58,020
Or the data has mistakes, or those three, I'm sorry,

568
00:39:58,020 --> 00:39:59,980
either just mistakes in the data,

569
00:39:59,980 --> 00:40:03,340
or you don't have the data or your prompt is not good enough.

570
00:40:03,340 --> 00:40:04,340
Yeah?

571
00:40:04,340 --> 00:40:06,380
That's the typical reasons that comes from.

572
00:40:06,380 --> 00:40:09,140
So if you set up the process in a way

573
00:40:09,140 --> 00:40:13,620
that a knowledgeable human has a look at the final result

574
00:40:13,620 --> 00:40:17,340
in a way that it cannot skip it, yeah?

575
00:40:17,340 --> 00:40:19,380
You have to avoid that people can become lazy

576
00:40:19,380 --> 00:40:21,340
and just say yes, yes, yes.

577
00:40:21,340 --> 00:40:25,340
That's why we put in this 120 words instead of 100

578
00:40:25,340 --> 00:40:29,620
because it enforces the human corrective.

579
00:40:29,620 --> 00:40:31,140
Yeah.

580
00:40:31,140 --> 00:40:34,260
If you do that, you don't have a problem with hallucinations

581
00:40:34,260 --> 00:40:38,780
because you have maybe 200 texts that are good.

582
00:40:38,780 --> 00:40:41,220
And one in which my human goes in and says,

583
00:40:41,220 --> 00:40:42,700
"Hey, there's a picture of a thriller."

584
00:40:42,700 --> 00:40:44,380
And the long text says there's a thriller

585
00:40:44,380 --> 00:40:45,660
and here in the short version,

586
00:40:45,660 --> 00:40:47,660
there's only some romantic bullshit.

587
00:40:47,660 --> 00:40:50,940
Yeah, that is definitely wrong.

588
00:40:50,940 --> 00:40:53,940
And then you go like, yeah, but last year,

589
00:40:53,940 --> 00:40:56,580
I would have had to write 100 articles

590
00:40:56,580 --> 00:40:59,820
and now I only have to write one that I now trapped.

591
00:40:59,820 --> 00:41:02,620
So in this process, hallucinations are not a problem

592
00:41:02,620 --> 00:41:04,300
because we check them.

593
00:41:04,300 --> 00:41:06,260
And this is how you should work with Gen AI.

594
00:41:06,260 --> 00:41:11,460
You should not try to do something that's statistically

595
00:41:11,460 --> 00:41:15,060
impossible, which is to bring a stochastic model

596
00:41:15,060 --> 00:41:16,980
to 100% accuracy.

597
00:41:16,980 --> 00:41:20,100
If not going to happen with this technology,

598
00:41:20,100 --> 00:41:23,140
but you have to take the system as it is

599
00:41:23,140 --> 00:41:27,140
and build your processes so that this shortcoming

600
00:41:27,140 --> 00:41:30,700
of this model is traced down and that's how we do it.

601
00:41:30,700 --> 00:41:38,740
And sometimes I think also human have had my eyes.

602
00:41:38,740 --> 00:41:39,900
I think so.

603
00:41:39,900 --> 00:41:45,140
So, well, there that's, yeah, I think,

604
00:41:45,140 --> 00:41:47,340
actually in the world, all the people think

605
00:41:47,340 --> 00:41:50,220
the most evil company in the world is Palantir

606
00:41:50,220 --> 00:41:56,020
and they build a system where they can detect

607
00:41:56,020 --> 00:41:59,660
where the next criminal act and who are...

608
00:41:59,660 --> 00:42:09,220
Most times, who is the victim, who is the evil guy.

609
00:42:09,220 --> 00:42:11,220
Yeah.

610
00:42:11,220 --> 00:42:13,820
And it's off the track.

611
00:42:13,820 --> 00:42:15,820
It's a black people.

612
00:42:15,820 --> 00:42:21,620
So, then people say that the system is racist.

613
00:42:21,620 --> 00:42:25,380
But did you think the system is racist?

614
00:42:25,380 --> 00:42:33,620
Or is it human bias because we don't like to hear something like that?

615
00:42:33,620 --> 00:42:41,020
Yeah. So, I think there's many elements to your question.

616
00:42:41,020 --> 00:42:51,820
So, the issue with determining where the next criminal activity

617
00:42:51,820 --> 00:42:56,620
will take place is of course that you criminalize people

618
00:42:56,620 --> 00:42:59,620
who have not done anything wrong.

619
00:42:59,620 --> 00:43:08,020
And that is, that this will even create more...

620
00:43:10,020 --> 00:43:15,820
that this sometimes creating the bias that it has in the system in real life.

621
00:43:15,820 --> 00:43:16,820
Yeah.

622
00:43:16,820 --> 00:43:22,820
Because let's assume that you have a city and then there's one area with more criminals

623
00:43:22,820 --> 00:43:25,020
and then one area with less criminals.

624
00:43:25,020 --> 00:43:28,220
So, of course, you send police more to the area with more criminals

625
00:43:28,220 --> 00:43:29,420
which you actually have to do.

626
00:43:29,420 --> 00:43:32,020
That's okay, but even more intense.

627
00:43:32,020 --> 00:43:38,820
So, now you will maybe even control every habitant of the criminal area

628
00:43:38,820 --> 00:43:42,620
and find that 3% more because you never detect everything

629
00:43:42,620 --> 00:43:46,420
and you will not control anyone in the less criminal area

630
00:43:46,420 --> 00:43:51,620
which will even make this bias higher because a few people who do strange stuff there

631
00:43:51,620 --> 00:43:53,620
will not be traced down.

632
00:43:53,620 --> 00:44:04,620
And you enforce this issue and then you always get the problem of coincidence and correlation.

633
00:44:04,620 --> 00:44:11,220
So, even if maybe the...

634
00:44:11,220 --> 00:44:14,220
That's my personal hypothesis.

635
00:44:14,220 --> 00:44:20,220
If the root cause for criminal behavior is poverty, yeah.

636
00:44:20,220 --> 00:44:30,220
But there is a correlation of poverty and say a certain area where you come from.

637
00:44:30,220 --> 00:44:36,820
Then it happens very easily that people assume that the root cause was where they come from

638
00:44:36,820 --> 00:44:40,620
and the real root cause of poverty is not seen.

639
00:44:40,620 --> 00:44:46,620
And so you put in measures that make life even harder for those people

640
00:44:46,620 --> 00:44:49,820
who put them even deeper into poverty, who will even...

641
00:44:49,820 --> 00:44:53,820
He will then become more criminal because they have no way...

642
00:44:53,820 --> 00:44:55,820
No other way to survive.

643
00:44:55,820 --> 00:45:01,420
So this is kind of the palantir effect from my perspective, yeah.

644
00:45:01,420 --> 00:45:03,420
And there's more.

645
00:45:03,420 --> 00:45:08,420
But if you think about it in a different way, I can also really help

646
00:45:08,420 --> 00:45:15,620
because yes, every human being has biases and I like to bring my own example.

647
00:45:15,620 --> 00:45:21,420
So there was at work many years back, there was one female that I have to say it.

648
00:45:21,420 --> 00:45:23,220
I really hated her, yeah.

649
00:45:23,220 --> 00:45:28,820
So the way that person would breathe could make me angry, yeah.

650
00:45:28,820 --> 00:45:32,820
And this is absolutely irrational, completely unfair.

651
00:45:32,820 --> 00:45:38,820
Yeah, I'm the bad person, you know, her, but I cannot change this emotion.

652
00:45:38,820 --> 00:45:46,820
Yeah, so I can tell you for sure that if I was the CHRO of a large company

653
00:45:46,820 --> 00:45:54,420
and I would check pictures of applicants, no female that looked similar to that

654
00:45:54,420 --> 00:45:57,220
other female would ever get a job in this company.

655
00:45:57,220 --> 00:46:03,420
And I even think it is completely fair to say that no female being similar

656
00:46:03,420 --> 00:46:06,820
to that female would ever be my personal assistant, yeah.

657
00:46:06,820 --> 00:46:10,820
Because no matter how competent if there's no personal, no personal fit,

658
00:46:10,820 --> 00:46:12,820
then you can also just leave it.

659
00:46:12,820 --> 00:46:14,820
Yeah, it would be a nightmare for both of us.

660
00:46:14,820 --> 00:46:19,620
But if you are in such a leading position and you would never in person work

661
00:46:19,620 --> 00:46:24,820
with that other person, then your bias is a problem because it's absolutely unfair.

662
00:46:24,820 --> 00:46:25,820
Yeah.

663
00:46:25,820 --> 00:46:31,820
And here, Jenny, I can help a lot because it only sees the data that it actually sees.

664
00:46:31,820 --> 00:46:38,820
So if I build up a system where I cut out all pictures and all, you know,

665
00:46:38,820 --> 00:46:42,020
all hints towards someone is living and so on.

666
00:46:42,020 --> 00:46:47,020
So I clean up this data in this way before and then I match it to positions.

667
00:46:47,020 --> 00:46:52,020
And then I ask my system to say, hey, look from the data you have, yeah,

668
00:46:52,020 --> 00:46:57,020
who should take that role and could be that it would tell me like,

669
00:46:57,020 --> 00:47:07,020
"I just wondering, yeah, this candidate is best qualified and super smart

670
00:47:07,020 --> 00:47:10,020
and lives in the right area and ta-da.

671
00:47:10,020 --> 00:47:17,020
Why did you do this, missor? And I would maybe realize that maybe I was biased.

672
00:47:17,020 --> 00:47:21,020
Maybe I should just stay out of the process and let everyone else decide, yeah.

673
00:47:21,020 --> 00:47:25,020
So maybe my gut feeling is right, but maybe not most likely not.

674
00:47:25,020 --> 00:47:28,020
And I think this is really helpful because you can do that.

675
00:47:28,020 --> 00:47:31,020
People cannot forget, but AI can forget.

676
00:47:31,020 --> 00:47:37,020
So if you realize where your bias, you can make it forget and then you can make it

677
00:47:37,020 --> 00:47:41,020
help you to make good decisions.

678
00:47:41,020 --> 00:47:46,020
Yeah, I think, or there, there, it's on, actually, in some big metal.

679
00:47:46,020 --> 00:47:49,020
It's the HTS, the application, tellers and systems.

680
00:47:49,020 --> 00:47:54,020
A lot of people now writing, AI writes a lot of,

681
00:47:54,020 --> 00:47:59,020
reshians and fight against the systems from the companies.

682
00:47:59,020 --> 00:48:02,020
I think this will, yeah, it will be hard.

683
00:48:02,020 --> 00:48:09,020
I have seen it also. I've written my normal ZB all the time and I see after,

684
00:48:09,020 --> 00:48:15,020
I don't know, 20 minutes, I get to refuse nights at three o'clock and then I think,

685
00:48:15,020 --> 00:48:19,020
okay, you're something, I don't know,

686
00:48:19,020 --> 00:48:25,020
a lot of companies work at this time in the human resources departments.

687
00:48:25,020 --> 00:48:32,020
I have, I've checked and then I have, I've written all my ZB's like what,

688
00:48:32,020 --> 00:48:39,020
what they have inside, they are, they are the keywords and how they describe

689
00:48:39,020 --> 00:48:46,020
the company philosophy and then I see, okay, I get more invites to this.

690
00:48:46,020 --> 00:48:53,020
So, yeah, I think, yeah, that's a good example for, or for the bias.

691
00:48:53,020 --> 00:49:00,020
But, I think a little bit, a lot of people are also confused about, yeah,

692
00:49:00,020 --> 00:49:08,020
I think, security with AI, it can expose, I don't know,

693
00:49:08,020 --> 00:49:16,020
some data that will not be, how did you think about these processes to secure AI

694
00:49:16,020 --> 00:49:22,020
and who should be in charge for this?

695
00:49:22,020 --> 00:49:29,020
Yeah, I think it's, sometimes really, you know, how you say it's frightening.

696
00:49:29,020 --> 00:49:34,020
If you ask the TedBots that you use more often about, you know,

697
00:49:34,020 --> 00:49:39,020
what kind of personality you have or what it thinks you,

698
00:49:39,020 --> 00:49:45,020
it should stop doing or whatever, because they have become more cautious

699
00:49:45,020 --> 00:49:47,020
and testing it every now and then.

700
00:49:47,020 --> 00:49:52,020
Half a year back, they would tell you what psychological problems you have

701
00:49:52,020 --> 00:50:00,020
and now they've really been, you know, rolling back and they become more cautious of assuming

702
00:50:00,020 --> 00:50:03,020
who you might be or what your psychological profile could be.

703
00:50:03,020 --> 00:50:10,020
So, I think that's, it's interesting because it's definitely on the surface only.

704
00:50:10,020 --> 00:50:16,020
So, they made their security prompts better.

705
00:50:16,020 --> 00:50:23,020
So, the tools don't share so much, but it doesn't mean that they don't evaluate the patterns.

706
00:50:23,020 --> 00:50:30,020
And so, this is something that is simply there and we have to be aware that now

707
00:50:30,020 --> 00:50:34,020
if I maybe have a fight with my boss and a big corporation, yeah,

708
00:50:34,020 --> 00:50:39,020
someone could just take all the press interviews and all the emails that you have

709
00:50:39,020 --> 00:50:44,020
and put it into a TedBot and say, okay, how should I best deal with that person?

710
00:50:44,020 --> 00:50:49,020
So, we can manipulate much better.

711
00:50:49,020 --> 00:50:52,020
And I'm when you play it better, yeah.

712
00:50:52,020 --> 00:51:04,020
So, if someone takes the effort, I think the issue about data privacy is also absolutely underestimated.

713
00:51:04,020 --> 00:51:11,020
Because, look, if I'm, for example, producing just, I don't know, bubble gum in Germany, yeah.

714
00:51:11,020 --> 00:51:18,020
So, and maybe I'm just a small private company with no intention to go big, yeah.

715
00:51:18,020 --> 00:51:26,020
And if then I put my financial statement into a TedDBG, open world and train my model, yeah.

716
00:51:26,020 --> 00:51:32,020
Maybe it just doesn't matter because if I, I would say like, hey, if anyone's got a question,

717
00:51:32,020 --> 00:51:37,020
I do pay my taxes in a very, you know, good way and I'm trying to do my best.

718
00:51:37,020 --> 00:51:42,020
If I find out, find out, I can also print it for you, yeah.

719
00:51:42,020 --> 00:51:52,020
If that is what you feel, then there will not be any big result because also no, no ruler of a larger company

720
00:51:52,020 --> 00:51:55,020
would want to read your financial statement, yeah.

721
00:51:55,020 --> 00:51:58,020
And so, how should they get it?

722
00:51:58,020 --> 00:52:04,020
And but as soon as we are in either more relevant areas, where, for example,

723
00:52:04,020 --> 00:52:13,020
it could be interesting for a US or Chinese or Russian company to find out about your patent, your pending patterns.

724
00:52:13,020 --> 00:52:20,020
We have to be aware that they don't even have to get their hands on my actual data to find out what I'm doing

725
00:52:20,020 --> 00:52:27,020
because these networks work in a way that on the one hand, I said my question and on the other hand, the answer is matched.

726
00:52:27,020 --> 00:52:34,020
So I can read from my answer that is created in my system, what your question is, right?

727
00:52:34,020 --> 00:52:42,020
And I cannot really repeat it fully, but I had a longer discussion with an ethical hacker some weeks back.

728
00:52:42,020 --> 00:52:51,020
And I was so, I expected that things would be bad, but I was still super shocked, yeah, because I didn't consider this relation.

729
00:52:51,020 --> 00:53:07,020
And I think for, for really confidential data, such as defense or so, the only way you can grant data security is to have your own machine in your basement with very thick walls

730
00:53:07,020 --> 00:53:13,020
and some other security things that you have to do and then just run a local model and that's it.

731
00:53:13,020 --> 00:53:19,020
Because if someone wants to access the data, they can access the data for stop, yeah.

732
00:53:19,020 --> 00:53:27,020
And everything else, just how high do I build the wall so that it's maybe too much effort for people to climb it?

733
00:53:27,020 --> 00:53:30,020
Yeah, this is how I think about it.

734
00:53:30,020 --> 00:53:32,020
Who cares?

735
00:53:32,020 --> 00:53:34,020
What if I lose it?

736
00:53:34,020 --> 00:53:44,020
And if I building marketing at the ads with Jenny, I then I think, okay, worst case, they hack up book certification in my hotel, yeah.

737
00:53:44,020 --> 00:53:51,020
And you hands out, so then whatever, yeah, maybe he buys my product, maybe not, I couldn't kill us.

738
00:53:51,020 --> 00:54:03,020
For everything where I don't have this kind of an attitude, I have to consider how, who would care and how, and then I can think about using Jenny, I or not, and my security measures.

739
00:54:03,020 --> 00:54:10,020
And of course legal instructions, but I think that's clear, so I'm not mentioning it.

740
00:54:10,020 --> 00:54:33,020
Yeah, I think one topic is trust, so who should, should be the state aware of the AI, should the companies, they build it, should the company use it, should be the user, who should be aware of the AI.

741
00:54:33,020 --> 00:54:36,020
Yeah, which responsible.

742
00:54:36,020 --> 00:54:41,020
And so there's two words, right, responsible in trust.

743
00:54:41,020 --> 00:54:50,020
So trust, from my opinion, trust comes from a repeated behavior that is to my liking.

744
00:54:50,020 --> 00:55:05,020
So if I'm, many times walking outside my house on the lawn, yeah, I trust that the ground will hold me, you know, and that it will feel in a certain way.

745
00:55:05,020 --> 00:55:26,020
And trust that it will be like this the next time again, so, and this, that's a very, very, very, very basic description of it. But for example, if I have a friend, and I get to know that person, then people have this, and humans have this ability to give a little bit of trust in advance.

746
00:55:26,020 --> 00:55:40,020
So we say, yeah, I'm willing to trust you without proof. I will test it, yeah. And the more often we interact and the more often the behavior is as expected and good.

747
00:55:40,020 --> 00:55:44,020
The more I will say a really trust this person. Yeah, so.

748
00:55:44,020 --> 00:56:00,020
And that goes for business if I am interacting with the company. And every time I pay money to do phone calls. And I pay the money and I can do the phone calls, then I start to trust the company.

749
00:56:00,020 --> 00:56:04,020
And doesn't it is not necessarily.

750
00:56:04,020 --> 00:56:19,020
And then it is not a person, because I can trust an institution, a friend or a soil. And this trust can go very, very fast. If suddenly I pay my money, but I cannot do phone calls.

751
00:56:19,020 --> 00:56:31,020
And then comes the question, who is responsible? Yeah, who is responsible? Because responsibility can only come from someone who can explain why it happens.

752
00:56:31,020 --> 00:56:37,020
And who can be accountable. So give me my money back, for example. Yeah. So.

753
00:56:37,020 --> 00:56:49,020
And this can only be a human because it can also be a company, but in this company, people have done to make the decision that they explain me.

754
00:56:49,020 --> 00:56:56,020
I don't know, our network was down and so I can explain network was done. So I can take this.

755
00:56:56,020 --> 00:57:16,020
And I can explain what happened. And I give you your money back. Here's two euros for the day you couldn't take the phone call or 50 euros because maybe I miss an important important appointment somehow so they can make these decisions and they can reimburse me or explain all my trust is completely gone can also happen.

756
00:57:16,020 --> 00:57:29,020
And what is interesting about that is that we have some lawsuits already and where companies wanted to put this responsibility to the provider of the chatbot.

757
00:57:29,020 --> 00:57:43,020
And and a from a legal perspective, it was denied. So they said if we don't care if you work with sloppy chatbot providers or bad staff or if you are, you know, incapable yourself.

758
00:57:43,020 --> 00:57:56,020
But this is within your responsibility. So you have to pay it. And more importantly, people also refuse to explain to accept that apology. They say, I'm your customer.

759
00:57:56,020 --> 00:58:12,020
And as I am your customer, I want you to fix it. And maybe you can pass the charge to the chatbot creating system. But I don't care about it. Your reputation is at stake and not the chatbot providers because it's you that I have the relationship with.

760
00:58:12,020 --> 00:58:24,020
And so I think this responsibility thing is something we really have to consider as business owners. Yeah. And people don't let us get away with it. Someone else is for.

761
00:58:24,020 --> 00:58:35,020
Yeah. So, wow, there's also an awesome episode. Oh, you're talking about one hour. So, yeah, yeah, it's really interesting and something new.

762
00:58:35,020 --> 00:58:45,020
I never have talked about because I more have to talk about technology a lot. And it's interesting to think about. Yeah.

763
00:58:45,020 --> 00:58:56,020
I have a different look on all all this. So yeah, then I will say thank you for for your time and staying with me here. And yeah, give all your perspectives. And yeah, thank you.

764
00:58:56,020 --> 00:58:59,020
And have a nice day.

765
00:58:59,020 --> 00:59:01,020
Yeah, thank you.

766
00:59:01,020 --> 00:59:02,020
Bye.

767
00:59:02,020 --> 00:59:03,020
Bye.

768
00:59:03,020 --> 00:59:32,000
[END PLAYBACK]