23 Mar 2026
40 minutes
40 minutes
Cultural Intelligence Podcast | When AI goes global: Culture, trust and the limits of scale.
Description
Transcript
In this first episode of The Cultural Intelligence Series, host Marine Esquenet sits down with Marina Pantcheva to explore what happens when global growth moves faster than our ability to manage meaning, context and trust.
Marina helps us unpack
- Why culture remains one of the hardest variables for global leaders to master but also the easiest to underestimate
- The real-world risks of one-size-fits-all communication across markets
- What leaders must rethink to ensure their global strategies truly resonate across cultures
Don’t miss Marina’s recipe for cultural confidence, blending curiosity, empathy and local insight into every decision. A powerful start to our series on Cultural Intelligence.
Read More
00:00:00.240 --> 00:00:02.320
Hello everyone, and welcome to the first episode
00:00:02.320 --> 00:00:06.400
of the Cultural Intelligence podcast series by RWS.
00:00:09.600 --> 00:00:12.800
Today we're welcoming Marina Pancheva to talk
00:00:12.800 --> 00:00:15.040
all things about when AI goes global,
00:00:15.919 --> 00:00:18.240
culture, trust, and the limit of scale.
00:00:18.560 --> 00:00:20.960
We are going to cover three main points today.
00:00:21.120 --> 00:00:22.960
The first one is why culture is one
00:00:22.960 --> 00:00:25.520
of the hardest variables for AI to model,
00:00:26.380 --> 00:00:28.940
and also the easiest for leaders to underestimate.
00:00:29.020 --> 00:00:31.580
We're going to talk about the risks connected to
00:00:31.740 --> 00:00:34.580
monocultural AI, and we're also going to talk about
00:00:34.580 --> 00:00:37.900
what leaders need to take into consideration for AI
00:00:37.900 --> 00:00:39.740
to work reliably in the real world.
00:00:40.300 --> 00:00:42.300
Stick around till the end to have
00:00:42.380 --> 00:00:45.780
Marina's secret recipe for enterprise AI to
00:00:45.780 --> 00:00:48.780
remain culturally impactful in the real world.
00:00:55.180 --> 00:00:56.950
Hello everyone, and welcome to our
00:00:56.950 --> 00:01:00.590
Cultural Intelligence, podcast series by RWS.
00:01:00.910 --> 00:01:03.750
Today we're going to talk all things about when
00:01:03.750 --> 00:01:07.110
AI goes global, culture, trust, and maybe a bit
00:01:07.110 --> 00:01:09.790
the limits, as well of scaling it.
00:01:11.280 --> 00:01:15.470
We are welcoming, warmly welcoming Marina Pancheva,
00:01:16.200 --> 00:01:18.940
director of Linguistic AI Services at RWS.
00:01:19.630 --> 00:01:20.630
Marina, welcome.
00:01:20.630 --> 00:01:21.550
Thank you for being here.
00:01:21.550 --> 00:01:22.790
Thank you for having me here.
00:01:22.790 --> 00:01:24.510
I'm very happy to meet you.
00:01:24.910 --> 00:01:26.630
And, it's amazing to have two
00:01:26.630 --> 00:01:28.490
Marinas on a podcast, isn't it?
00:01:29.050 --> 00:01:30.730
I'm sure everyone is going to be very
00:01:30.730 --> 00:01:33.610
happy to hear that we're interviewing you.
00:01:34.870 --> 00:01:36.290
You're just very active in the
00:01:36.290 --> 00:01:38.410
field, like, you're presenting everywhere.
00:01:38.410 --> 00:01:40.570
So, yeah, it's an honor to have you here today, really.
00:01:41.590 --> 00:01:44.850
Do you want to maybe start by introducing
00:01:44.850 --> 00:01:47.050
yourself shortly before we dive right in?
00:01:47.450 --> 00:01:49.290
So, my name is Marina Pancheva.
00:01:49.290 --> 00:01:52.810
I'm Director of Linguistic AI Services, at RWS.
00:01:53.130 --> 00:01:56.780
And I started, as a theoretical linguist.
00:01:56.780 --> 00:01:58.820
So I have an academic background and a love
00:01:58.820 --> 00:02:00.860
for research, which I still keep in me.
00:02:01.180 --> 00:02:04.180
But I joined this industry, localization and
00:02:04.180 --> 00:02:06.780
translation industry, about 10, 12 years ago.
00:02:07.180 --> 00:02:08.460
I think it's more like 10.
00:02:08.699 --> 00:02:10.539
And, yeah, I've been through it all.
00:02:10.539 --> 00:02:15.340
I first started as regular, analog linguistic, services
00:02:15.980 --> 00:02:18.460
and then had a couple of very exciting
00:02:18.780 --> 00:02:20.540
projects where I got to work a lot
00:02:20.540 --> 00:02:25.020
on quality working with translation community until I
00:02:25.020 --> 00:02:27.840
reached the stage when AI became a thing.
00:02:27.840 --> 00:02:32.080
So I started exploring AI and, gradually we formed
00:02:32.080 --> 00:02:35.400
the center of Excellence on AI Research and Development,
00:02:35.960 --> 00:02:40.040
which then branched off into delivery service delivery.
00:02:40.280 --> 00:02:41.839
And there is still the
00:02:41.839 --> 00:02:43.560
branch of research and development.
00:02:44.040 --> 00:02:50.040
So our daily work is focused on, finding ways to
00:02:50.040 --> 00:02:55.560
implement AI in linguistic services in a way that makes
00:02:55.640 --> 00:03:01.300
really sense meaning, implement AI where it really helps, brings
00:03:01.300 --> 00:03:05.300
value and also be aware of the limitations that AI
00:03:05.300 --> 00:03:08.700
has In a way counteracting the hype a bit, knowing,
00:03:08.700 --> 00:03:11.220
you know, where we can use it, where we can't
00:03:11.220 --> 00:03:13.620
use it, and researching all the time.
00:03:13.940 --> 00:03:16.980
As the field is moving at an immense speed.
00:03:17.860 --> 00:03:19.380
I was going to say, what a time to
00:03:19.380 --> 00:03:21.980
be alive when this is your topic of expertise.
00:03:21.980 --> 00:03:23.340
You know, like with all the
00:03:23.340 --> 00:03:25.300
developments and you have a lot.
00:03:25.780 --> 00:03:28.500
Yeah, it's moving so fast. I can imagine.
00:03:28.500 --> 00:03:31.120
Yeah, we even obviously outside of, you know,
00:03:31.120 --> 00:03:34.760
the expertise kind of field and I can't
00:03:34.760 --> 00:03:36.960
imagine how much there is to understand and.
00:03:37.120 --> 00:03:39.200
Yeah, yeah, it's an ever expanding
00:03:39.280 --> 00:03:41.920
kind of, border of knowledge.
00:03:41.920 --> 00:03:43.480
So, you know, the feeling the more you
00:03:43.480 --> 00:03:45.040
know, the more you know, you don't know.
00:03:45.040 --> 00:03:47.320
And it's moving so fast that I'm constantly
00:03:47.320 --> 00:03:49.280
having the feeling that I am catching up,
00:03:49.280 --> 00:03:51.480
catching up, catching up, and I'll never catch
00:03:51.480 --> 00:03:52.960
up with the knowledge that is out there.
00:03:52.960 --> 00:03:54.240
Yeah, I can imagine that.
00:03:54.640 --> 00:03:57.200
All right, let's let's dive in.
00:03:58.430 --> 00:04:02.550
How do you think leaders should define culture, in
00:04:02.550 --> 00:04:08.350
the context of enterprise, AI and global content? Yeah.
00:04:08.350 --> 00:04:11.550
Well, I mean, defining culture, my first instinct
00:04:11.550 --> 00:04:13.670
is let's define culture in the first place. Right.
00:04:13.670 --> 00:04:15.870
So there are different ways to define culture.
00:04:16.390 --> 00:04:19.350
There is the kind of psychological take on what
00:04:19.350 --> 00:04:24.350
culture is, which means like individual person's, behavior of
00:04:24.350 --> 00:04:27.310
a habit, the way they react in different situation.
00:04:27.390 --> 00:04:30.180
But I think the one we which is more
00:04:30.180 --> 00:04:34.260
relevant for us is the anthropological definition of culture
00:04:34.340 --> 00:04:40.140
as being a system of shared values, reaction styles,
00:04:40.140 --> 00:04:42.660
you know, behaviors in different contexts.
00:04:42.660 --> 00:04:44.820
And that's very important because
00:04:46.100 --> 00:04:50.180
these shared values, create predictability
00:04:50.820 --> 00:04:53.380
and the predictability creates trust.
00:04:53.780 --> 00:04:56.660
So this is a very simple explanation for a
00:04:56.660 --> 00:04:59.340
5 year old why two people from the same
00:04:59.340 --> 00:05:02.720
culture, they feel that they can have friendship easier.
00:05:02.720 --> 00:05:03.040
Right?
00:05:03.280 --> 00:05:06.960
Because by us sharing the same culture, I actually can
00:05:06.960 --> 00:05:09.280
predict how you're gonna react to what I do.
00:05:09.440 --> 00:05:10.840
And by me knowing how you're
00:05:10.840 --> 00:05:13.200
gonna react, I can trust you.
00:05:13.200 --> 00:05:15.120
I feel safer if I meet a
00:05:15.120 --> 00:05:17.200
person from a totally alien culture.
00:05:17.280 --> 00:05:20.160
I may say something means something, but
00:05:20.160 --> 00:05:22.320
it gets interpreted in a different way.
00:05:22.320 --> 00:05:26.200
The unpredictability factor rises and thereby also
00:05:26.200 --> 00:05:29.120
the insecurity and the instinct for self
00:05:29.120 --> 00:05:31.280
preservation and all those things, you know,
00:05:31.280 --> 00:05:36.620
that behavior at the very, basic level.
00:05:37.580 --> 00:05:41.740
So shared culture creates predictability,
00:05:41.820 --> 00:05:43.980
which in turn creates trust.
00:05:44.700 --> 00:05:47.740
And this is what, global leaders
00:05:47.740 --> 00:05:49.820
should actually be aware of. Right?
00:05:49.820 --> 00:05:51.980
It's the trust element, of the
00:05:51.980 --> 00:05:53.940
end customer that matters here.
00:05:53.940 --> 00:05:57.460
But thinking about enterprises, I think again,
00:05:57.460 --> 00:06:00.140
we have two levels of culture.
00:06:00.460 --> 00:06:03.520
One is the kind of the company culture or the brand.
00:06:03.520 --> 00:06:06.840
Culture is like the message you want to send to
00:06:06.840 --> 00:06:11.280
the world about your culture, your product, what you deliver.
00:06:11.600 --> 00:06:13.160
As an example, like let's take
00:06:13.160 --> 00:06:15.040
a famous shoe producer, right?
00:06:15.520 --> 00:06:17.639
Their culture and message and
00:06:17.639 --> 00:06:19.840
brand is about individual achievement.
00:06:20.080 --> 00:06:21.360
Just do it, right?
00:06:21.360 --> 00:06:23.080
It's very individualistic.
00:06:23.080 --> 00:06:25.520
So it would speak very easily to,
00:06:26.050 --> 00:06:28.320
end user from a western type of
00:06:28.320 --> 00:06:30.960
culture which is more individualistic, right?
00:06:31.280 --> 00:06:33.350
And this message is the core message.
00:06:33.660 --> 00:06:37.020
It makes the core of the culture of
00:06:37.020 --> 00:06:40.460
the brand, and it has to be preserved.
00:06:40.460 --> 00:06:43.020
Like you cannot change that for different markets.
00:06:43.020 --> 00:06:44.220
You cannot transform.
00:06:44.380 --> 00:06:46.940
Just do it to let us all do it, right?
00:06:47.020 --> 00:06:48.940
It breaks the message and the culture.
00:06:49.020 --> 00:06:51.660
So that core culture must be preserved.
00:06:52.060 --> 00:06:54.900
But then it has to be fine tuned
00:06:54.900 --> 00:06:58.460
for the regional culture of the markets, right?
00:06:58.700 --> 00:07:01.500
So while you're keeping your core message, you may
00:07:01.500 --> 00:07:06.480
have to deliver it differently for a American, for
00:07:06.480 --> 00:07:09.240
the American market and for the Japanese market.
00:07:09.560 --> 00:07:12.680
So while it is just do it and it is,
00:07:13.100 --> 00:07:16.480
placing focus on individual achievements, the way you talk about
00:07:16.480 --> 00:07:19.840
it may have to be changed when this is marketed
00:07:19.840 --> 00:07:23.960
in a more collectivist culture where the values are different
00:07:24.200 --> 00:07:28.240
and it's more like the group achievement than the individual
00:07:28.240 --> 00:07:30.440
achievement of a single person.
00:07:31.080 --> 00:07:34.440
So for global leaders, what is very important is to,
00:07:35.860 --> 00:07:41.340
on the one hand, preserve their core, culture that characterizes
00:07:41.340 --> 00:07:45.820
the brand and on the other hand, fine tune it
00:07:45.820 --> 00:07:50.500
and modify it for the regional variations, but still preserving
00:07:50.500 --> 00:07:52.260
that core message without losing it.
00:07:52.340 --> 00:07:54.100
I always think about it, you know, I
00:07:54.100 --> 00:07:56.100
always try to explain things to my kids.
00:07:56.500 --> 00:07:59.540
So I always think about it like, think Coca Cola.
00:07:59.700 --> 00:08:02.710
It has a taste and it has a, unique taste, right?
00:08:02.710 --> 00:08:06.270
Everybody recognizes Coca Cola, but in fact
00:08:06.270 --> 00:08:10.190
it tastes slightly differently in every market.
00:08:10.190 --> 00:08:15.630
It is sweeter in Asia, it is, a bit sharper in Europe.
00:08:16.110 --> 00:08:17.230
So just do it.
00:08:17.230 --> 00:08:21.870
The Coca Cola style, your culture, brand culture is,
00:08:21.950 --> 00:08:24.830
you know, the beverage and then you change it
00:08:24.830 --> 00:08:28.310
slightly for the different markets to fine tune it
00:08:28.310 --> 00:08:31.510
to the specific taste of the end user.
00:08:31.870 --> 00:08:33.429
Also the cultural taste.
00:08:34.150 --> 00:08:36.870
I always ask myself this question about culture.
00:08:36.870 --> 00:08:38.309
Where does it really end?
00:08:38.870 --> 00:08:43.188
Because a family in itself is a culture, you know, so
00:08:43.188 --> 00:08:46.550
yeah, you have the country, the culture of the country.
00:08:47.110 --> 00:08:48.790
But how far can you take it?
00:08:48.790 --> 00:08:50.670
You know, it's kind of never ending really.
00:08:50.670 --> 00:08:53.350
Obviously you have general standards that apply
00:08:53.350 --> 00:08:55.590
for groups of people in general.
00:08:56.310 --> 00:08:57.920
But can you really.
00:08:58.240 --> 00:09:02.000
It's like nearly Persona per person, per
00:09:02.000 --> 00:09:03.440
person at this point, you know?
00:09:03.760 --> 00:09:05.920
Well, that is a bit the transition from
00:09:05.920 --> 00:09:08.880
the psychological definition of culture, which starts with
00:09:08.880 --> 00:09:12.000
the individual, to the anthropological definition of culture,
00:09:12.000 --> 00:09:14.320
which basically applies to a whole ethnic group.
00:09:14.480 --> 00:09:15.680
But you're totally right.
00:09:15.840 --> 00:09:17.840
Culture is not a discrete thing.
00:09:17.840 --> 00:09:19.080
It's not that if we have a
00:09:19.080 --> 00:09:22.320
border between two countries, suddenly, you know,
00:09:22.320 --> 00:09:25.320
you have completely different behavior. It flows.
00:09:25.630 --> 00:09:26.190
It flows.
00:09:26.430 --> 00:09:28.990
And there is constantly the interaction between different
00:09:28.990 --> 00:09:31.510
cultures which generate new cultures and so on.
00:09:31.510 --> 00:09:32.750
That's the fascinating thing.
00:09:32.750 --> 00:09:34.270
But you'd be surprised how many
00:09:34.270 --> 00:09:36.590
universal things there are about culture.
00:09:36.910 --> 00:09:40.190
Like there are cultural universals and universal
00:09:40.190 --> 00:09:44.190
human values that are shared, across cultures.
00:09:45.090 --> 00:09:47.470
Just like, even, you know,
00:09:47.470 --> 00:09:49.310
body language facial expressions.
00:09:49.550 --> 00:09:52.710
Facial expressions differ in different cultures. Right.
00:09:52.710 --> 00:09:54.910
The way people smile. Nodding.
00:09:54.910 --> 00:09:56.110
You're nodding at me. Right?
00:09:56.110 --> 00:09:59.050
If you were in Bulgaria, that would mean no all the time.
00:09:59.210 --> 00:10:01.130
Because it's reversed. Right.
00:10:01.130 --> 00:10:03.130
In Bulgarian, nodding is negative and
00:10:03.130 --> 00:10:05.690
shaking your head is positive. Right.
00:10:05.690 --> 00:10:08.130
Super confusing. But there are.
00:10:08.130 --> 00:10:10.010
So that's definitely not universal.
00:10:10.090 --> 00:10:12.890
Yeah, but smiling is universal everywhere.
00:10:13.129 --> 00:10:15.650
Now it depends on how much you smile.
00:10:15.650 --> 00:10:18.810
In some cultures, it's not appropriate to show one's
00:10:18.810 --> 00:10:21.650
teeth, so you smile without showing your teeth.
00:10:21.650 --> 00:10:23.690
In other cultures you can laugh
00:10:23.930 --> 00:10:26.410
out loud and it's totally appropriate.
00:10:26.650 --> 00:10:30.590
Sadness, crying these are univers, kind
00:10:30.590 --> 00:10:32.350
of, messages that you can send.
00:10:32.350 --> 00:10:33.750
And it's the same with culture.
00:10:34.530 --> 00:10:35.910
If we have time, we can dig
00:10:35.910 --> 00:10:39.910
into cultural universals which, have been established,
00:10:40.170 --> 00:10:44.310
through a fascinating social and cultural experiment
00:10:44.310 --> 00:10:48.070
called, Moral Machine, where different...
00:10:48.630 --> 00:10:52.230
Which basically asks people from all over the planet.
00:10:52.230 --> 00:10:54.790
There were like 233 countries that were
00:10:54.790 --> 00:10:58.070
surveyed and more than 40 million, answers
00:10:58.070 --> 00:11:01.020
to the survey about moral decisions which,
00:11:01.090 --> 00:11:04.050
which reflect moral values and cultures.
00:11:04.130 --> 00:11:06.850
And there were a couple of universals that
00:11:06.850 --> 00:11:09.890
applied to all, decisions that people make.
00:11:09.890 --> 00:11:12.170
Yeah, we can get into that if there is time.
00:11:12.170 --> 00:11:15.010
I might have a little question that I think
00:11:15.170 --> 00:11:18.770
will lead you to talk again about this later.
00:11:19.090 --> 00:11:19.730
Looking forward.
00:11:22.530 --> 00:11:24.850
How do cultural bias creep,
00:11:25.720 --> 00:11:28.090
into AI models and workflows?
00:11:28.090 --> 00:11:30.130
And why does it happen? Where does it happen?
00:11:32.330 --> 00:11:35.290
Unfortunately, there are many doors and even portals
00:11:35.450 --> 00:11:38.730
through which cultural bias creeps into AI models.
00:11:38.890 --> 00:11:41.770
The one that is biggest, probably
00:11:42.330 --> 00:11:46.569
is, through data, training data.
00:11:46.890 --> 00:11:49.930
So unbalanced data sets where one culture is
00:11:49.930 --> 00:11:54.730
overrepresented, are going to lead to, AI models
00:11:54.890 --> 00:11:58.370
adopting this culture, absorbing it from the training
00:11:58.370 --> 00:12:01.530
and becoming basically a monocultural AI.
00:12:01.690 --> 00:12:04.390
Even even though this model might be multilingual,
00:12:04.550 --> 00:12:08.230
it will still be a monocultural, AI model.
00:12:08.390 --> 00:12:13.190
And in reality, that's what's happening because between 70
00:12:13.190 --> 00:12:16.790
and 95% of training data for the really large
00:12:16.790 --> 00:12:19.120
models, you know, the ones coming out of Silicon
00:12:19.310 --> 00:12:23.350
Valley and other big models is actually English data.
00:12:23.750 --> 00:12:25.510
And the rest is primarily data
00:12:25.510 --> 00:12:28.550
coming from the western digitalized world. Right?
00:12:28.550 --> 00:12:30.870
That would be Spanish and German and so on.
00:12:31.110 --> 00:12:32.910
Minority languages, long tail
00:12:32.910 --> 00:12:34.850
languages are under represented.
00:12:35.090 --> 00:12:38.130
So by necessity AI models actually learn about
00:12:38.290 --> 00:12:43.330
western culture and they absorb the western cultural
00:12:43.490 --> 00:12:45.970
values like the so called weird culture.
00:12:46.050 --> 00:12:48.130
I'm not a big fan of this acronym.
00:12:48.130 --> 00:12:50.930
But it's like Western, educated,
00:12:51.010 --> 00:12:54.050
industrialized, rich, democratic society.
00:12:54.850 --> 00:12:59.850
So data is one of the portals through which bias
00:12:59.850 --> 00:13:04.450
creeps into, and cultural bias creeps into AI models.
00:13:04.930 --> 00:13:07.310
But then there are also other ways.
00:13:07.870 --> 00:13:12.750
One is another source for bias
00:13:12.750 --> 00:13:15.910
and cultural bias comes from the
00:13:15.910 --> 00:13:19.470
humans themselves, human labeling, for example.
00:13:20.270 --> 00:13:23.110
So in the process of training a
00:13:23.110 --> 00:13:25.030
large language model, there are different stages.
00:13:25.030 --> 00:13:28.070
In the first stage it is called pre training.
00:13:28.070 --> 00:13:29.870
So you use massive amounts of data.
00:13:29.950 --> 00:13:31.510
And this data, as we already
00:13:31.510 --> 00:13:34.980
mentioned, is unbalanced data, predominantly English.
00:13:35.540 --> 00:13:37.340
But then in the next stages you actually
00:13:37.340 --> 00:13:40.260
use humans to fine tune the model and
00:13:40.260 --> 00:13:42.900
teach the model to be a helpful assistant
00:13:42.980 --> 00:13:45.780
and respond adequately to questions and so on.
00:13:45.780 --> 00:13:48.580
And there is one stage called reinforcement learning through
00:13:48.580 --> 00:13:54.180
human feedback, where humans need to provide feedback which
00:13:54.260 --> 00:13:57.980
responses from the AI are best to teach.
00:13:57.980 --> 00:13:59.420
The model produce the kind of
00:13:59.420 --> 00:14:01.220
completions that humans like best.
00:14:01.380 --> 00:14:04.090
And that's where a lot of cultural bias comes in.
00:14:04.560 --> 00:14:06.760
Because you as a human, if you are given three
00:14:06.760 --> 00:14:09.840
versions of a prompt completion and you are asked to
00:14:09.840 --> 00:14:13.280
choose the one that you like best, to rank them,
00:14:13.280 --> 00:14:17.400
you are gonna rank the one that is aligned to
00:14:17.400 --> 00:14:21.200
your culture, your beliefs, your values as the best one
00:14:21.200 --> 00:14:22.960
because it aligns with you.
00:14:23.040 --> 00:14:24.480
This comes from the so called
00:14:24.560 --> 00:14:26.640
confirmation bias that humans have.
00:14:27.200 --> 00:14:31.040
We like and we choose and we think that
00:14:31.200 --> 00:14:35.850
those answers, completions, you know, facts that align with
00:14:35.850 --> 00:14:39.330
our own beliefs and understandings and values are correct.
00:14:39.730 --> 00:14:43.970
Unfortunately, this bias, this cognitive bias, is so
00:14:43.970 --> 00:14:47.170
strong that not even intelligence can beat it.
00:14:47.170 --> 00:14:49.290
In fact, there is a very disturbing piece of
00:14:49.290 --> 00:14:52.370
research which shows that the more intelligent a person
00:14:52.450 --> 00:14:56.610
is, the more they are susceptible to confirmation bias.
00:14:56.610 --> 00:14:58.370
Yes, this is bad news, I agree.
00:14:58.610 --> 00:15:01.090
Yes, very bad news.
00:15:01.170 --> 00:15:03.170
But there is a logical explanation to that.
00:15:03.170 --> 00:15:06.530
The more, the more intelligent the person is, the
00:15:06.530 --> 00:15:12.170
more they can interpret evidence in a way that
00:15:12.170 --> 00:15:15.690
it would in the end confirm their preconception. Right?
00:15:15.930 --> 00:15:16.410
Yeah.
00:15:16.570 --> 00:15:18.970
So another source of bias comes from labeling.
00:15:18.970 --> 00:15:21.570
Like sometimes we just need to have label data.
00:15:21.570 --> 00:15:23.810
For example, sentiment labeling.
00:15:23.810 --> 00:15:25.490
Like label these images.
00:15:25.490 --> 00:15:27.610
Is the person here happy or unhappy?
00:15:28.340 --> 00:15:31.690
Or label this message, is it polite or impolite?
00:15:31.850 --> 00:15:34.150
Now a very direct message.
00:15:35.830 --> 00:15:39.110
Say a response to a user flagging an issue, right?
00:15:39.110 --> 00:15:41.370
If it needs to be labeled, a German,
00:15:41.830 --> 00:15:44.310
labeler will say, it's perfect, it's direct.
00:15:44.390 --> 00:15:47.110
It immediately gets to the point. I like it.
00:15:47.350 --> 00:15:50.550
A Japanese person may be like, it's very rude.
00:15:50.550 --> 00:15:52.230
You know, it needs to have this whole
00:15:52.230 --> 00:15:55.390
apologetic preamble, you know, it doesn't use the
00:15:55.390 --> 00:15:58.630
right kind of, doesn't sound respectful enough. Right.
00:15:58.630 --> 00:16:02.510
So if you have predominantly Western labelers or people
00:16:02.510 --> 00:16:06.570
who label Western data, then you're gonna teach the
00:16:06.570 --> 00:16:10.730
model to be very direct, low context kind of
00:16:10.730 --> 00:16:15.490
communication, transactional relationships, and that sort of thing.
00:16:15.490 --> 00:16:17.330
So that's another source for
00:16:17.950 --> 00:16:19.890
cultural bias in AI models.
00:16:19.890 --> 00:16:22.690
And they exhibit it, big time really.
00:16:23.170 --> 00:16:26.930
There's a lot of research benchmarking evaluations
00:16:26.930 --> 00:16:29.410
of language models over and over again.
00:16:29.490 --> 00:16:33.050
They show that large language models from the
00:16:33.050 --> 00:16:36.610
big providers align with Western cultural values.
00:16:36.610 --> 00:16:40.130
I mean, yeah, is there such a
00:16:40.130 --> 00:16:43.250
thing as an unbiased AI model?
00:16:43.410 --> 00:16:45.250
We all bias to some extent.
00:16:45.570 --> 00:16:46.770
Everyone out there, right?
00:16:46.770 --> 00:16:48.410
Like biased based on our life
00:16:48.410 --> 00:16:52.290
experience, our parents, I guess.
00:16:52.290 --> 00:16:54.610
Like, is it, how do we do this?
00:16:54.610 --> 00:16:57.490
Because so an AI model learns from
00:16:57.490 --> 00:17:00.370
the, what we give them, you know.
00:17:00.850 --> 00:17:04.530
So is there anything like an unbiased model?
00:17:05.650 --> 00:17:07.849
Well, bias in a way is in the
00:17:07.849 --> 00:17:10.369
eye of the observer in one sense. Right.
00:17:10.690 --> 00:17:13.210
While I can interpret something as being biased,
00:17:13.210 --> 00:17:15.170
you may say, no, that's totally fine.
00:17:15.250 --> 00:17:18.210
I think bias is very, very culture specific.
00:17:19.869 --> 00:17:23.569
But then again there is objective
00:17:23.569 --> 00:17:26.410
bias, for example gender, racial bias. Right.
00:17:26.410 --> 00:17:27.329
Age bias.
00:17:27.490 --> 00:17:29.880
And a lot of that bias is also historic bias.
00:17:30.350 --> 00:17:34.710
So these are biases that AI creeped
00:17:35.270 --> 00:17:40.230
into AI models, from historical training data,
00:17:40.230 --> 00:17:43.070
for example literature from the 19th century.
00:17:43.710 --> 00:17:48.990
Unfortunately, it is enough to have very few examples
00:17:49.310 --> 00:17:52.990
of bias for a model to learn that bias.
00:17:54.830 --> 00:17:57.390
This also has been explored a lot in
00:17:57.710 --> 00:18:00.990
data poisoning and trying to misalign models.
00:18:02.010 --> 00:18:05.770
Recent research came out which shows that 250,
00:18:07.810 --> 00:18:11.890
poisoned examples are enough to actually make a
00:18:11.890 --> 00:18:14.770
model learn a behavior or learn something that
00:18:14.770 --> 00:18:16.970
is undesired, an undesired feature.
00:18:17.210 --> 00:18:23.530
So it's enough to have 250 data chunks,
00:18:24.500 --> 00:18:27.330
that exhibit a particular type of bias for
00:18:27.330 --> 00:18:29.760
the AI model to actually absorb it.
00:18:29.840 --> 00:18:31.200
It's a shocking finding.
00:18:31.200 --> 00:18:33.440
And what is even more shocking is that it actually,
00:18:33.760 --> 00:18:36.040
it doesn't depend on the size of the model.
00:18:36.040 --> 00:18:38.960
It's not a percentage, it's an absolute number.
00:18:39.200 --> 00:18:42.460
So whether your model will be 6B,
00:18:42.720 --> 00:18:47.200
parameters or 13B parameters, it's about 250.
00:18:47.440 --> 00:18:50.240
Imagine how easy it is to skew
00:18:50.320 --> 00:18:52.880
the behavior of a large language model.
00:18:53.520 --> 00:18:56.160
Yes, very interesting, thank you.
00:18:58.340 --> 00:19:00.580
My next question is a bit more
00:19:00.580 --> 00:19:03.460
language related from, you know, my translator.
00:19:03.700 --> 00:19:05.060
Translator past.
00:19:05.060 --> 00:19:09.420
I guess it seems like it's so easy for
00:19:09.420 --> 00:19:13.419
a model to learn language patterns and, you know,
00:19:13.419 --> 00:19:16.900
understand that, but then it struggles with the culture
00:19:16.900 --> 00:19:19.700
and the context part of things. Why is that?
00:19:20.260 --> 00:19:21.220
Yeah, yeah.
00:19:22.800 --> 00:19:26.330
Let me just maybe start with the learning.
00:19:26.570 --> 00:19:29.130
What does a language model learn when it is language?
00:19:29.130 --> 00:19:31.690
Well, it basically learns statistical patterns.
00:19:32.250 --> 00:19:36.250
So, this is a very, very fancy autocomplete state
00:19:36.250 --> 00:19:38.170
of the art models are more than that because
00:19:38.170 --> 00:19:41.690
they're also trained on reasoning, they're trained on code,
00:19:41.850 --> 00:19:45.050
they are trained to access, they can access tools.
00:19:45.130 --> 00:19:45.530
Right.
00:19:46.730 --> 00:19:51.290
But they still work based on pure statistics.
00:19:51.770 --> 00:19:54.530
And you can extract statistical patterns from
00:19:54.530 --> 00:20:00.070
linguistic, identify relationships, but it stays at
00:20:00.070 --> 00:20:02.150
the level of the linguistic data.
00:20:02.550 --> 00:20:05.110
In order to absorb culture, you need to go
00:20:05.110 --> 00:20:08.670
one level deeper and actually go beyond language.
00:20:08.670 --> 00:20:10.790
Language, it's like a tip of an iceberg.
00:20:10.870 --> 00:20:13.030
Language is 10% of culture.
00:20:13.110 --> 00:20:14.950
The rest is shared history,
00:20:15.590 --> 00:20:19.350
religion, interpersonal relationships, traditions.
00:20:19.670 --> 00:20:22.470
And that is not encoded in language in
00:20:22.470 --> 00:20:25.530
such a way that it's statistical, right? It is.
00:20:25.690 --> 00:20:28.450
You can see at the culture
00:20:28.450 --> 00:20:30.090
through the lens of language.
00:20:30.250 --> 00:20:33.610
And to a certain extent the two are correlated. Right.
00:20:33.610 --> 00:20:36.190
You've probably, studied about the famous Sapir
00:20:36.810 --> 00:20:40.210
Whorf hypothesis which says that language influences
00:20:40.210 --> 00:20:41.810
the way people see the world.
00:20:41.810 --> 00:20:44.810
You know, if you have 1000 like hundred words
00:20:44.810 --> 00:20:48.730
of snow, you know, it's gonna, you know, make
00:20:48.730 --> 00:20:51.170
you perceive snow in a different way than if
00:20:51.170 --> 00:20:53.530
you have words for something else.
00:20:53.850 --> 00:20:56.710
So language does influen thinking to a certain extent.
00:20:56.710 --> 00:20:58.230
I think it's also the other way around.
00:20:58.230 --> 00:20:59.550
Like if you live in a place with
00:20:59.550 --> 00:21:02.070
snow, you're gonna have hundred words about snow.
00:21:02.150 --> 00:21:03.830
And if you don't have snow, then why do
00:21:03.830 --> 00:21:05.590
we need the words in the first place? Right.
00:21:05.670 --> 00:21:06.510
One would be enough.
00:21:06.510 --> 00:21:08.230
Like something white and cold. Yeah.
00:21:08.870 --> 00:21:12.310
So, but when AI learns language,
00:21:12.310 --> 00:21:15.110
it basically extracts the statistical patterns.
00:21:15.350 --> 00:21:16.710
It learns code as well.
00:21:16.710 --> 00:21:18.950
AI is excellent at learning code.
00:21:18.950 --> 00:21:23.440
Does it understand the goal of programming when
00:21:23.440 --> 00:21:26.440
it programs an app, writes the code, which
00:21:26.440 --> 00:21:28.960
is a type of very well regulated language.
00:21:29.120 --> 00:21:31.560
Like does it actually intend to
00:21:31.560 --> 00:21:33.200
create the thing it creates?
00:21:33.280 --> 00:21:35.080
No, it's the same with language.
00:21:35.080 --> 00:21:36.360
So it produces language, but
00:21:36.360 --> 00:21:37.840
does it understand the intent?
00:21:37.840 --> 00:21:39.360
I think the best way to put it
00:21:39.360 --> 00:21:42.840
is like with humans, when we communicate, when
00:21:42.840 --> 00:21:44.960
we read, we read between the lines. Actually.
00:21:45.040 --> 00:21:47.400
Communication is a lot about the things you
00:21:47.400 --> 00:21:49.940
don't see than the things you actually Say
00:21:50.420 --> 00:21:54.020
AI reads the lines, we read between them,
00:21:54.100 --> 00:21:57.140
and it's between the lines where culture hides.
00:21:57.700 --> 00:21:59.060
And how big is this problem?
00:21:59.380 --> 00:22:04.220
When AI is deployed globally, the
00:22:04.220 --> 00:22:05.620
problem is big and growing.
00:22:06.420 --> 00:22:08.540
And the problem has as a root the
00:22:08.540 --> 00:22:10.540
fact that most of the models that are
00:22:10.540 --> 00:22:13.300
being used globally are Silicon Valley models.
00:22:14.190 --> 00:22:16.590
There are some Chinese models, but to a
00:22:16.590 --> 00:22:18.350
certain extent they are trained on the same
00:22:18.350 --> 00:22:21.710
data or they distill mutually outputs.
00:22:21.870 --> 00:22:24.270
So it's Silicon Valley culture.
00:22:25.710 --> 00:22:31.190
And, when these models are deployed globally, then
00:22:31.190 --> 00:22:34.070
in markets where regions where the culture is
00:22:34.070 --> 00:22:38.870
very different, they still produce this, American centric
00:22:38.870 --> 00:22:42.150
or Western centric type of output with the
00:22:42.150 --> 00:22:43.830
values shining through, right?
00:22:43.830 --> 00:22:45.220
The cultural values, the communic
00:22:45.290 --> 00:22:47.210
communication style, and so on.
00:22:47.210 --> 00:22:50.650
So you have probably experienced that whenever you ask
00:22:50.730 --> 00:22:53.410
one of the big models a question, often they
00:22:53.410 --> 00:22:55.850
come back with, this is such an excellent question.
00:22:56.330 --> 00:22:57.330
Genius. Brilliant.
00:22:57.330 --> 00:22:58.690
You know, especially some models
00:22:58.690 --> 00:23:02.730
are really very, exaggerating, right?
00:23:03.050 --> 00:23:06.730
And in my culture this feels like, what did I do wrong?
00:23:07.050 --> 00:23:08.610
What was happening here?
00:23:08.610 --> 00:23:11.010
Like, I just ask a simple question and I
00:23:11.010 --> 00:23:13.320
don't need to be convinced I'm, a genius.
00:23:13.320 --> 00:23:14.960
And it's amazing even if the
00:23:14.960 --> 00:23:16.680
question is really silly, right?
00:23:16.760 --> 00:23:19.480
So for my culture, this sounds annoying.
00:23:19.560 --> 00:23:24.040
This is like you are not, really honest with me here.
00:23:24.040 --> 00:23:25.480
You're trying to please me.
00:23:26.430 --> 00:23:28.120
But I guess in other cultures this is
00:23:28.120 --> 00:23:30.080
just taken as granted and you ignore it.
00:23:30.080 --> 00:23:31.760
Like you don't really take it seriously,
00:23:31.760 --> 00:23:34.120
asking yourself, am I really a genius?
00:23:34.360 --> 00:23:35.720
Yeah, I never thought so.
00:23:36.040 --> 00:23:40.280
So, now on a more serious note, this
00:23:40.280 --> 00:23:42.720
is also dangerous because this leads us to
00:23:42.720 --> 00:23:46.420
this algorithmic monoc culture and cultural flattening.
00:23:46.420 --> 00:23:49.420
I just opened an article, like just before our
00:23:49.420 --> 00:23:52.300
podcast, as I was taking a break, I opened
00:23:52.300 --> 00:23:55.540
an article which was all about language flattening.
00:23:55.620 --> 00:23:56.780
So it's not even only
00:23:56.780 --> 00:23:58.780
culture, even language gets flattened.
00:23:58.780 --> 00:24:00.500
And we are using the same linguistic
00:24:00.500 --> 00:24:03.180
patterns that come from English, even when
00:24:03.180 --> 00:24:05.300
writing in different languages, right?
00:24:05.780 --> 00:24:08.980
So, if that happens in language, just
00:24:08.980 --> 00:24:11.770
think about what happens with culture. Okay?
00:24:11.930 --> 00:24:14.010
So this also leads to digital
00:24:14.250 --> 00:24:15.850
imperialism in a way, right?
00:24:15.850 --> 00:24:20.290
You have this western centered culture, being
00:24:20.290 --> 00:24:22.730
in a way imposed through AI models
00:24:23.050 --> 00:24:26.730
in Africa, in India, in China. Right?
00:24:27.850 --> 00:24:32.490
So, it is definitely not a good development if
00:24:32.490 --> 00:24:36.090
you want to preserve diversity and the cultural richness.
00:24:36.250 --> 00:24:38.930
But there is an even bigger problem, and that is trust.
00:24:38.930 --> 00:24:41.020
As I mentioned mentioned, as I mentioned,
00:24:41.420 --> 00:24:44.060
culture, shared culture is the most important
00:24:44.300 --> 00:24:46.820
factor for building trust, right?
00:24:46.820 --> 00:24:48.980
So if an AI model starts telling you
00:24:48.980 --> 00:24:52.060
how fantastic your silly question was, are you
00:24:52.060 --> 00:24:54.700
actually really gonna trust what it responds. Right.
00:24:54.700 --> 00:24:57.060
It's not that you need to trust AI models blindly.
00:24:57.060 --> 00:25:00.380
I always say fact check, they hallucinate. Right.
00:25:00.380 --> 00:25:04.020
But imagine that you are communicating.
00:25:04.020 --> 00:25:06.740
You're an end customer and you're communicating with
00:25:06.740 --> 00:25:10.420
an AI powered chatbot for user like assistance.
00:25:10.420 --> 00:25:15.160
Say you have an, you want to return your package or
00:25:15.160 --> 00:25:17.160
something happened or you want to know where it's stuck.
00:25:17.160 --> 00:25:17.480
Right.
00:25:17.720 --> 00:25:22.440
And then you start communicating with an AI chatbot
00:25:22.760 --> 00:25:27.000
which just doesn't use the right tone and doesn't
00:25:27.000 --> 00:25:31.360
approach your issue from the right angle.
00:25:31.360 --> 00:25:34.920
It can be too direct or it can be overly polite.
00:25:34.920 --> 00:25:36.120
I had a problem recently.
00:25:36.520 --> 00:25:39.520
I was communicating with a Chinese based chatbot which
00:25:39.520 --> 00:25:42.670
was so polite that I was at some point
00:25:42.670 --> 00:25:45.430
wondering is this thing ever gonna resolve my problem
00:25:45.430 --> 00:25:49.310
or is it just gonna profusely apologize forever? Right.
00:25:49.310 --> 00:25:50.670
In the end I got it resolved.
00:25:50.670 --> 00:25:53.990
It was actually a very well done AI powered chatbot.
00:25:53.990 --> 00:25:55.670
I really enjoyed communicating with it,
00:25:55.910 --> 00:25:58.150
trying to see how it's made. Right.
00:25:58.150 --> 00:26:01.110
I mean I started actually trying to dismantle
00:26:01.110 --> 00:26:04.190
it and say like, does it use tools? Yeah.
00:26:04.190 --> 00:26:06.030
Will it make an agentic call there
00:26:06.030 --> 00:26:07.470
and there like in the end? Yeah.
00:26:07.470 --> 00:26:10.540
It hallucinated twice on numbers but
00:26:10.540 --> 00:26:11.540
then it got it right.
00:26:12.260 --> 00:26:14.580
So back to the topic of culture.
00:26:15.730 --> 00:26:17.460
Yeah, it is a very big problem.
00:26:17.860 --> 00:26:22.300
And the more enterprises adopt AI, the
00:26:22.300 --> 00:26:25.860
more they risk to flatten the way
00:26:25.860 --> 00:26:30.900
they talk to different markets, homogenize culture.
00:26:31.060 --> 00:26:32.420
And I think it's going to be
00:26:32.420 --> 00:26:34.260
a problem in general for humanity.
00:26:35.940 --> 00:26:36.260
Yeah.
00:26:36.260 --> 00:26:38.260
I mean when you say it, I can see that happen.
00:26:38.900 --> 00:26:43.220
Like you, it's just gonna be standardized. Everything.
00:26:43.380 --> 00:26:46.500
In a way, it's a bit scary. Yeah.
00:26:47.140 --> 00:26:48.780
Well, in a way it is a
00:26:48.780 --> 00:26:52.300
natural development of this whole globalization trend.
00:26:52.300 --> 00:26:52.580
Right.
00:26:52.580 --> 00:26:54.700
I mean when you think about it, when I was
00:26:54.700 --> 00:26:56.580
a kid and I would go to a foreign country
00:26:56.580 --> 00:26:59.540
on a holiday, I was so excited because you go
00:26:59.540 --> 00:27:02.940
to a shop and it's completely different stuff. Right.
00:27:02.940 --> 00:27:04.380
And you've never seen that.
00:27:04.380 --> 00:27:08.720
Even nowadays I go to shop in any European countries
00:27:09.440 --> 00:27:12.440
country and it's all the same on the supermarket shelves.
00:27:12.440 --> 00:27:12.720
Right.
00:27:12.720 --> 00:27:14.840
I mean you have these little things like you
00:27:14.840 --> 00:27:17.400
are going to get fish, what is it called?
00:27:17.400 --> 00:27:18.960
Fiske kaker in Norway. Right.
00:27:18.960 --> 00:27:21.280
I mean, I don't know the word in English. Fish.
00:27:22.300 --> 00:27:22.879
Cakes.
00:27:22.880 --> 00:27:24.160
Fish cakes, Yeah.
00:27:24.160 --> 00:27:27.040
I don't think you can buy fish cakes in Spain.
00:27:27.200 --> 00:27:28.880
Pretty sure we can. Yes.
00:27:29.040 --> 00:27:32.000
So there are a few things that have remained local,
00:27:33.010 --> 00:27:35.810
but apart from that it's all very, very standardized.
00:27:35.810 --> 00:27:37.890
So it becomes boring to travel in a way.
00:27:38.130 --> 00:27:40.690
Well similarly, culture gets standardized.
00:27:41.890 --> 00:27:47.570
How should leaders ensure that AI remains culturally
00:27:47.570 --> 00:27:52.450
competent, trustworthy and accountable at scale as well.
00:27:53.160 --> 00:27:56.410
Okay, so I've been thinking about that and I
00:27:56.410 --> 00:27:59.690
think I have a recipe with the main ingredients. Right?
00:27:59.690 --> 00:28:01.090
Step by step. Manual.
00:28:01.330 --> 00:28:02.770
Instruction manual. Simple.
00:28:03.000 --> 00:28:04.600
So let me break this down.
00:28:04.600 --> 00:28:07.720
So leaders, the first thing they need to do
00:28:07.800 --> 00:28:12.840
is actually establish the culture of their brand.
00:28:12.840 --> 00:28:15.280
Like what we were talking about, right? Just do it.
00:28:15.280 --> 00:28:18.600
Like what do you want the message to be?
00:28:18.600 --> 00:28:20.080
Like who are you? Exactly?
00:28:20.080 --> 00:28:21.760
As you put what are your values?
00:28:21.760 --> 00:28:25.880
How do you market your products?
00:28:26.610 --> 00:28:28.680
What do you want people to take from that?
00:28:28.840 --> 00:28:32.720
The main message, the main emotion is well, right?
00:28:32.720 --> 00:28:34.560
And then this is non negotiable.
00:28:34.560 --> 00:28:35.760
This shouldn't change, right?
00:28:35.760 --> 00:28:37.080
As we said, just do it.
00:28:37.080 --> 00:28:39.040
Shouldn't be like let's all do it together, right?
00:28:39.040 --> 00:28:40.160
This is your message.
00:28:40.480 --> 00:28:44.960
So define the core and then start fine tuning
00:28:45.040 --> 00:28:47.719
the recipe, you know, the way the core feels
00:28:47.719 --> 00:28:50.880
and sounds for different regions and markets.
00:28:51.440 --> 00:28:55.360
For that what you need to do is actually benchmark
00:28:55.600 --> 00:28:58.880
identify what are the cultural values of every region.
00:28:59.120 --> 00:29:02.160
I believe not all enterprises have mapped
00:29:02.240 --> 00:29:05.100
out the cultures of the markets.
00:29:05.260 --> 00:29:07.500
So they probably have a lot of data
00:29:07.900 --> 00:29:12.300
about you know, sales, conversion rates, whatever.
00:29:12.620 --> 00:29:14.940
But do you actually as a leader
00:29:15.340 --> 00:29:18.300
have a cultural map of your markets?
00:29:18.460 --> 00:29:22.860
Do you know whether this market is a collectivist type
00:29:22.860 --> 00:29:29.660
of culture which value cohesion and being together, respect for
00:29:29.660 --> 00:29:35.090
the elderly and for example have kind of a, a
00:29:35.170 --> 00:29:37.570
let's say non linear idea of time.
00:29:37.890 --> 00:29:41.930
And that market might be valuing efficiency and directness and
00:29:41.930 --> 00:29:45.010
they are very low context and they have this linear
00:29:45.010 --> 00:29:48.810
plan time and they need to be talked to in
00:29:48.810 --> 00:29:52.610
a very direct, very short and condensed type of way.
00:29:52.850 --> 00:29:55.650
So after you've defined your core culture
00:29:55.650 --> 00:29:58.330
and the values, then go out onto
00:29:58.330 --> 00:30:01.010
the markets and map them culturally.
00:30:01.010 --> 00:30:03.330
For that you need cultural specialists, you need
00:30:03.330 --> 00:30:05.850
people who will tell you on these dimensions.
00:30:05.850 --> 00:30:08.990
And there are various ways like there are
00:30:08.990 --> 00:30:12.230
various studies, you know, of cultural dimensions.
00:30:12.230 --> 00:30:15.070
You have the famous hofstede cultural dimensions.
00:30:15.310 --> 00:30:16.950
Six different dimensions, right?
00:30:16.950 --> 00:30:19.870
You have some frameworks with 12 different dimensions.
00:30:19.870 --> 00:30:22.789
So choose your favorite framework and define
00:30:22.789 --> 00:30:25.150
you know, where each market sits.
00:30:26.110 --> 00:30:27.430
Next thing that needs to be
00:30:27.430 --> 00:30:30.230
done is start testing, right?
00:30:30.230 --> 00:30:31.710
Create prompts because this is
00:30:31.870 --> 00:30:33.790
for AI implementation, right?
00:30:34.110 --> 00:30:38.990
Create prompts that will, will tweak the message, the
00:30:38.990 --> 00:30:42.030
core message, so that it talks to the people
00:30:42.030 --> 00:30:45.550
in the different regions, to their culture, right?
00:30:45.550 --> 00:30:47.510
So create different types of messages.
00:30:47.990 --> 00:30:49.030
Play with prompts.
00:30:49.030 --> 00:30:52.550
Create basically a test set, like a rich test
00:30:52.550 --> 00:30:57.510
set for modifying the core messages so that it
00:30:57.990 --> 00:31:00.790
speaks to the end users in each market.
00:31:01.510 --> 00:31:04.810
And then evaluate, okay again evaluate
00:31:04.810 --> 00:31:07.690
using local experts, cultural experts.
00:31:07.930 --> 00:31:10.890
Cultural, specialists, you know, people who understand,
00:31:11.610 --> 00:31:14.490
gather the feedback and incorporate again into
00:31:14.650 --> 00:31:16.970
your prompt your solution, your workflow, whatever
00:31:16.970 --> 00:31:18.210
you have there, right?
00:31:18.210 --> 00:31:22.930
So that you actually gather the learnings test again until
00:31:22.930 --> 00:31:25.570
you've gotten it right now I bet your next question
00:31:25.570 --> 00:31:27.810
will be like, how do you implement that?
00:31:27.810 --> 00:31:29.970
How do you tweak an AI model, you know,
00:31:29.970 --> 00:31:34.670
to talk to Japanese audience in one way and
00:31:34.670 --> 00:31:36.470
to a German audience in another way?
00:31:37.110 --> 00:31:40.390
Well, there are again different solutions and each
00:31:40.390 --> 00:31:43.670
one of them has different level of invasiveness.
00:31:43.670 --> 00:31:45.110
I call it invasiveness.
00:31:45.350 --> 00:31:46.470
First one, you can try to
00:31:46.470 --> 00:31:48.230
do it through prompting, right?
00:31:48.310 --> 00:31:49.670
You can try to start your
00:31:49.670 --> 00:31:51.550
prompt with this famous role prompting.
00:31:51.550 --> 00:31:54.510
Like you're a Japanese I
00:31:54.510 --> 00:31:57.430
don't know, customer support, agent.
00:31:57.510 --> 00:31:59.510
And you are very polite.
00:31:59.590 --> 00:32:03.750
You always apologize to the user and
00:32:03.750 --> 00:32:07.250
so give the guidelines for the model.
00:32:08.680 --> 00:32:11.090
If that doesn't work, because sometimes this
00:32:11.170 --> 00:32:13.730
role prompting actually doesn't work that well.
00:32:14.450 --> 00:32:17.570
You can go one level more invasive and
00:32:17.570 --> 00:32:19.810
use something that is called soft prompting.
00:32:19.810 --> 00:32:22.769
For that you already need more technical expertise.
00:32:23.490 --> 00:32:25.130
Soft prompting is a very interesting,
00:32:25.130 --> 00:32:26.610
it's not very much talked about.
00:32:26.690 --> 00:32:30.090
Soft prompting is a bit similar to fine tuning
00:32:30.090 --> 00:32:32.690
a model, but you actually don't really change.
00:32:34.180 --> 00:32:35.860
You know, when you fine tune a model,
00:32:35.940 --> 00:32:38.420
you go under the hood and you basically
00:32:39.500 --> 00:32:42.340
change the value of these little parameters.
00:32:42.340 --> 00:32:44.100
You know, you tweak the numbers inside
00:32:44.260 --> 00:32:46.580
because it's the numbers that encode knowledge.
00:32:46.660 --> 00:32:48.380
Knowledge is stored in the numbers.
00:32:48.380 --> 00:32:49.820
And if you want to change the
00:32:49.820 --> 00:32:51.420
knowledge and the behavior of a model,
00:32:51.420 --> 00:32:53.740
you change those numbers through fine tuning.
00:32:53.740 --> 00:32:55.700
But then you risk that the model is
00:32:55.700 --> 00:32:57.540
going to forget what it knew before, right?
00:32:57.540 --> 00:32:58.980
Because you've changed the numbers.
00:32:59.060 --> 00:33:03.140
Cultural soft prompting is you kind of append a special
00:33:03.140 --> 00:33:05.440
type of a vector in front of, of the prompt
00:33:05.520 --> 00:33:08.760
which in a way sways the behavior of a model
00:33:08.760 --> 00:33:12.280
to a particular like way of say talking.
00:33:12.280 --> 00:33:14.320
So you can kind of prepend this.
00:33:14.320 --> 00:33:16.560
Like there's a preamble, a vector preamble which
00:33:16.560 --> 00:33:19.040
can make the model be more polite or
00:33:19.040 --> 00:33:23.680
more direct, more verbose or like really concise
00:33:23.680 --> 00:33:26.120
or flattering or non flattering.
00:33:26.120 --> 00:33:28.320
And yeah, if that doesn't work, fine
00:33:28.320 --> 00:33:30.240
tuning, partial fine tuning, not full fine
00:33:30.240 --> 00:33:32.080
tuning, full fine tuning is very risky.
00:33:32.080 --> 00:33:34.320
Partial fine tuning, may be the
00:33:34.320 --> 00:33:36.290
solution if that doesn't work.
00:33:36.370 --> 00:33:37.290
Human in the Loop.
00:33:37.290 --> 00:33:38.650
I actually think you always need
00:33:38.650 --> 00:33:39.850
to have human in the loop.
00:33:39.850 --> 00:33:43.330
But human in the loop is not very scalable. Right?
00:33:43.330 --> 00:33:45.090
Human in the loop is not very scalable.
00:33:45.410 --> 00:33:47.330
So there are different levels.
00:33:47.730 --> 00:33:50.450
The best thing would be of course train your own model.
00:33:50.450 --> 00:33:51.890
But this is hardly feasible.
00:33:52.050 --> 00:33:56.610
Yeah, this is too time consuming.
00:33:57.170 --> 00:34:02.210
And yeah, you probably wouldn't have the capabilities and
00:34:02.210 --> 00:34:04.730
the resources to train a large language model.
00:34:04.730 --> 00:34:06.250
It's going to be a small language model.
00:34:06.250 --> 00:34:08.389
And after all the capabilities, capabilities of
00:34:08.389 --> 00:34:11.790
models are directly correlated to their size.
00:34:11.870 --> 00:34:13.790
So the bigger the model, the more capable.
00:34:13.790 --> 00:34:16.190
So that, that is a little bit of a trade off there.
00:34:16.350 --> 00:34:16.909
Yeah.
00:34:17.429 --> 00:34:19.790
And once this is done, like depends on which
00:34:20.110 --> 00:34:22.949
method you have used to actually, get the model
00:34:22.949 --> 00:34:26.870
to produce the culturally fine tuned and appropriate, type
00:34:26.870 --> 00:34:30.469
of text or interaction or images. Can be anything.
00:34:30.469 --> 00:34:32.270
This applies to any modality.
00:34:32.590 --> 00:34:34.480
Right, then monitor all the time.
00:34:34.708 --> 00:34:36.708
Time monitor all the time. Right.
00:34:36.708 --> 00:34:39.708
You need to have constant sampling, monitoring
00:34:39.708 --> 00:34:41.668
to make sure that the model still
00:34:41.668 --> 00:34:44.149
behaves, produces the right output.
00:34:44.149 --> 00:34:46.389
Cultures change, you know, model behavior
00:34:46.389 --> 00:34:47.509
may change for some reason.
00:34:47.909 --> 00:34:51.109
So especially if you are not working with your own model,
00:34:51.109 --> 00:34:54.389
but you are using a model by a third party provider,
00:34:54.900 --> 00:34:57.269
then you have to monitor all the time monitor.
00:34:57.269 --> 00:35:01.829
Okay, we have the little secret recipe here. Yes.
00:35:02.229 --> 00:35:03.429
Try that recipe.
00:35:03.669 --> 00:35:04.869
Tell me what you've cooked.
00:35:08.290 --> 00:35:10.490
Can I ask you, do you have something to add?
00:35:10.490 --> 00:35:12.570
Maybe something like an anecdote, something you would
00:35:12.570 --> 00:35:16.210
like to mention that we haven't covered before?
00:35:16.370 --> 00:35:19.810
Anecdotes, Many of AI models going wrong. Right.
00:35:19.970 --> 00:35:21.250
But I actually want to
00:35:21.250 --> 00:35:22.850
add something more philosophical.
00:35:22.850 --> 00:35:25.410
I've been on a bit of a philosophical wave lately.
00:35:26.550 --> 00:35:28.570
I was thinking a lot about innovation.
00:35:28.570 --> 00:35:30.050
There was an innovation Week,
00:35:30.650 --> 00:35:33.650
last week organized by rws.
00:35:34.590 --> 00:35:37.860
And this is what struck me me one day.
00:35:38.580 --> 00:35:41.620
And I find that like a very interesting observation.
00:35:42.180 --> 00:35:45.620
I don't know, might not be very unique, but here it is.
00:35:46.970 --> 00:35:49.180
Any innovation that we have had so
00:35:49.180 --> 00:35:52.940
far in human history has been fully
00:35:52.940 --> 00:35:56.500
explainable by the creator of the innovation.
00:35:56.900 --> 00:35:59.900
So people invented the wheel, they
00:35:59.900 --> 00:36:01.300
could explain how it works.
00:36:01.850 --> 00:36:03.740
James Watt invented the steam machine.
00:36:03.740 --> 00:36:05.780
He could explain every little detail.
00:36:05.780 --> 00:36:08.760
Like why is it that if you water there and you
00:36:08.760 --> 00:36:11.440
heat it, then steam comes out there and it works? Right.
00:36:11.950 --> 00:36:14.640
People invented Bell, invented the telephone.
00:36:14.640 --> 00:36:17.920
He could explain the telephone in every detail.
00:36:18.160 --> 00:36:20.480
And that was a predictable machine.
00:36:20.560 --> 00:36:22.320
It's a deterministic technology.
00:36:23.040 --> 00:36:26.640
And that has been always true of technological innovations
00:36:26.880 --> 00:36:30.640
until the transformers came out and AI in general.
00:36:31.120 --> 00:36:33.360
Now we are creating Models.
00:36:33.360 --> 00:36:35.440
We are creating technology that we
00:36:35.440 --> 00:36:37.600
have no idea how it works.
00:36:37.680 --> 00:36:39.600
In fact, there is a whole branch of
00:36:39.900 --> 00:36:45.980
science emerging, which is AI neuroscience, researchers, machine
00:36:45.980 --> 00:36:49.500
learning researchers are peeking under the hood of
00:36:49.500 --> 00:36:53.420
AI models, basically using the same methods to
00:36:53.500 --> 00:36:57.020
investigate what is happening there, which neurons get
00:36:57.020 --> 00:36:58.620
activated when I do this?
00:36:58.700 --> 00:37:00.460
How can I tweak the behavior?
00:37:00.460 --> 00:37:02.380
What will happen if I deactivate
00:37:02.380 --> 00:37:04.300
that neuron or activate that one?
00:37:04.460 --> 00:37:07.580
That's exactly what neuroscientists do with our human brains
00:37:07.580 --> 00:37:09.420
when they want to understand how it works.
00:37:09.500 --> 00:37:11.570
What will happen if we kind of send mild
00:37:11.800 --> 00:37:13.960
electricity impulse to that part of the brain?
00:37:14.040 --> 00:37:16.920
Oh, the hand goes up. Interesting. Yeah.
00:37:17.160 --> 00:37:21.320
So we are using the same technology to explore
00:37:21.640 --> 00:37:25.080
the electronic brain as we are exploring the human
00:37:25.080 --> 00:37:27.360
brain, which is one of the big mysteries. Right.
00:37:27.360 --> 00:37:29.040
We still don't understand how that works.
00:37:29.040 --> 00:37:31.120
And it strikes me that we have
00:37:31.120 --> 00:37:35.720
created something that we ourselves then need
00:37:35.720 --> 00:37:37.600
to research to understand how it works.
00:37:37.600 --> 00:37:39.880
We are as if playing God here.
00:37:39.960 --> 00:37:43.200
We have creating something that is beyond us and
00:37:43.200 --> 00:37:45.620
we cannot explain it in the tiniest detail.
00:37:45.700 --> 00:37:47.540
And this is why we are
00:37:47.540 --> 00:37:50.820
having this whole cultural, dimension conversations.
00:37:51.460 --> 00:37:54.660
Because now culture becomes relevant, right?
00:37:54.660 --> 00:37:57.740
When you have created an entity, a type of
00:37:57.740 --> 00:38:02.260
intelligence, which is not just a tool anymore.
00:38:02.260 --> 00:38:03.940
It's not deterministic.
00:38:03.940 --> 00:38:06.140
You cannot predict what it's gonna do.
00:38:06.140 --> 00:38:08.140
Like the way you can predict what a hammer is
00:38:08.140 --> 00:38:11.260
gonna do, or what a telephone, is gonna do, or
00:38:11.260 --> 00:38:14.710
like how electricity, electricity flows through the wires, right?
00:38:15.350 --> 00:38:19.030
That's when you stop interacting with that
00:38:19.030 --> 00:38:21.110
technology as a tool, but you start
00:38:21.110 --> 00:38:24.870
interacting with this technology as a collaborator.
00:38:25.430 --> 00:38:27.750
AI is already a collaborator.
00:38:27.750 --> 00:38:28.710
It's not a tool.
00:38:28.950 --> 00:38:30.190
You know, people use it as
00:38:30.190 --> 00:38:34.390
a brainstorming partner, idea generation partner.
00:38:34.550 --> 00:38:38.230
Some people use it as a psychiatrist and even a friend.
00:38:38.390 --> 00:38:39.030
Okay?
00:38:39.190 --> 00:38:41.590
This becomes a different type of intelligence
00:38:41.590 --> 00:38:43.510
which is now collaborating with us.
00:38:43.970 --> 00:38:45.730
And because of that, we need
00:38:45.730 --> 00:38:48.530
it also culturally attuned to us.
00:38:48.850 --> 00:38:51.130
Until now, you don't really care what the culture
00:38:51.130 --> 00:38:53.850
of your phone is because it's just irrelevant, right?
00:38:53.850 --> 00:38:54.850
That's not even a question.
00:38:55.090 --> 00:38:57.650
But you do want to have the culture.
00:38:57.730 --> 00:38:59.890
You want to have this new type of tool,
00:38:59.969 --> 00:39:04.930
a, collaborative tool to actually share your values, share
00:39:05.010 --> 00:39:10.080
your moral principles, share your ethical principles, because it's
00:39:10.080 --> 00:39:12.520
gone beyond the stage of being just a tool.
00:39:12.840 --> 00:39:15.320
So that's what I've been thinking about lately.
00:39:16.110 --> 00:39:16.760
So interesting.
00:39:16.760 --> 00:39:17.880
Yeah, so interesting.
00:39:18.280 --> 00:39:20.120
Thank you for sharing that with us.
00:39:20.280 --> 00:39:22.560
Thank you, for giving me the opportunity. Maybe.
00:39:22.560 --> 00:39:24.040
I have one last question.
00:39:24.200 --> 00:39:25.760
Very, very quick answer.
00:39:25.760 --> 00:39:27.640
What is cultural intelligence for you?
00:39:27.800 --> 00:39:30.840
So for me Cultural intelligence, both in humans and
00:39:30.840 --> 00:39:35.240
AI, actually means being able to make yourself understood
00:39:35.240 --> 00:39:38.520
and communicate with any culture which is out there.
00:39:38.840 --> 00:39:42.420
Not have any stereotypes, preconditions, conceptions.
00:39:42.890 --> 00:39:47.940
Be open and have the ability of communicating and
00:39:48.020 --> 00:39:52.420
understanding a, human being or an, end user or
00:39:52.420 --> 00:39:55.300
an AI model, whatever it may be, that has
00:39:55.300 --> 00:40:01.460
a different system of values, rules, behavioral patterns.
00:40:01.620 --> 00:40:03.780
That's what cultural intelligence is for me.
00:40:04.020 --> 00:40:05.700
Okay, thank you. That was it.
00:40:05.860 --> 00:40:08.060
Thank you so much for being here with us. Thank you.
00:40:08.060 --> 00:40:09.940
And stay tuned for the next episode.