Up next in 10
The Cloud Show with Magnus Mårtensson ft. Rene Schulte- Ep: 68
0 views
Jun 20, 2025
Join this live session with Magnus Mårtensson ft. Tore Nestenius for the next episode of The Cloud Show with Magnus Mårtensson on April 23, at 01:05 PM (EST). The show is about cloud leadership and all the important questions relating to cloud projects. Certainly, many matters when a company is going to and wants to be successful in the cloud, are about technology. However, there are many additional matters, adjacent to technology, that we also need to tend to regarding business strategy, human resources, organizational change, planning for a technical cloud approach, and many more questions. These conversations are critical for a healthy cloud and for a swift and accurate cloud approach. GUEST SPEAKER Tore Nestenius is a passionate software development expert who enjoys independently sharing his knowledge with others. He provides training, consulting, and coaching services to businesses and aspiring developers, helping them build innovative, secure, and scalable systems. 📺 CSharp TV - Dev Streaming Destination http://csharp.tv 🌎 C# Corner - Community of Software and Data Developers https://www.c-sharpcorner.com #CSharpTV #csharpcorner #TheCloudShow #CSharp #interview 🎙️ New to streaming or looking to level up? Check out StreamYard and get $10 discount! 😍 https://streamyard.com/pal/d/5348850450038784
View Video Transcript
0:02
hello hello everyone and welcome to yet
0:04
another episode of the cloud show of
0:07
course we have a new guest today and
0:09
it's going to be an interesting
0:10
conversation about a very hot topic
0:12
which is complicated for many and um
0:15
people feel like oh my god there's so
0:17
much I need to learn fortunately we have
0:19
a star of the show who knows all about
0:22
AI and the trends sort of shaping the
0:25
future of AI rather that's what we're
0:27
going to talk about three different
0:29
trends today and uh I'm going to do that
0:32
together with a with a good friend of
0:34
mine uh Mr renee Schult
0:42
[Music]
0:52
hello my friend welcome to the cloud
0:54
show
0:55
hey Magnus thanks for having me
0:57
absolutely my pleasure um it's great to
1:00
to talk to you we saw each other just a
1:02
couple of weeks ago yeah it feels like
1:04
yesterday but it's been like four weeks
1:07
three weeks i know i know time flies it
1:09
does but it was it was good to see you
1:11
in in uh the same physical space but
1:14
right now you're not in this physical
1:16
space in my home office you are in what
1:18
your home office yeah it's my home
1:21
office in in Dson in Germany in Germany
1:23
all right so for the audience uh tell us
1:26
a little bit about yourself and what you
1:28
do sure sure um so my name is Renee i
1:31
work for a company called Reply uh we're
1:33
a system integrator um 16,000 employees
1:36
worldwide and you know quite a bit of
1:39
revenue um not a lot of folks know us
1:41
that's why I talk about it um but look
1:44
mainly headquartered in Italy but also
1:46
big presence in in Germany and actually
1:48
also in the US we have a thousand people
1:50
or so what I do is I work for our CTO
1:53
and I manage multiple groups which we
1:56
call communities of practices and they
1:59
work on innovative topics and you know
2:01
my teams are digital human synthetic
2:03
data general AI
2:06
um spatial computing and quantum and
2:08
future of computing all the fun great
2:10
topics um and I enjoy this because you
2:14
know working with a lot of smart young
2:15
folks and you know doing some fun stuff
2:18
in the end but also some some good
2:20
bleeding edge stuff there i'm also a
2:22
content creator making bunch of you know
2:25
stuff like we all do on social media um
2:28
but also have my own podcast if you so
2:30
shameless pluck your digital dialogues
2:32
if you check
2:34
out family yeah we're we're right now
2:38
talking with lots of experts in humanoid
2:40
robotics so if that is your thing you
2:42
might want to check this out and yeah
2:44
what else um MVP and Microsoft regional
2:47
director just like you and all these
2:49
other things we do right it's like all
2:52
these things we do all right we keep
2:53
hanging out together that's that's for
2:55
sure and I I like that a lot so um All
2:58
right cool so now we know a little bit
3:00
about you so you just you just picked
3:02
the the easiest teams to lead i mean on
3:04
just a simple topics no complicated
3:06
anything at all in in the stuff that you
3:08
do in business right like quantum or you
3:11
know easy stuff easy easy stuff where
3:14
you always you know I keep on saying
3:15
it's the bleeding edge and sometimes we
3:17
really bleed on the edge I guess where
3:20
it comes from that's exactly you know
3:23
where that comes from
3:25
and you know what's the challenge also
3:27
is like these days when when we talk
3:29
about AI and you know bipe coding and
3:31
all the good stuff um there's not a lot
3:34
of data for these topics available which
3:37
these models could have been trained on
3:39
so you know it's still untouched field
3:42
so sometimes early in some areas yeah
3:44
for sure the bleeding edge the front
3:46
line where things happen and so so one
3:50
of those things that has has been
3:52
happening for quite a while now is is AI
3:55
and and I'm sure a lot of people are
3:57
confused about what's going on and like
4:00
where is this going and what is that
4:02
front line the edge and where is the
4:04
trend in in this space right the the
4:08
future of AI you gave me this wonderful
4:10
topic to talk about today said "Hey
4:12
let's talk about the the three key
4:14
trends shaping the future of AI." I was
4:16
like "That sounds really impressive
4:18
let's talk about that."
4:21
Right great let's let's uh let's do this
4:25
um
4:26
you know if we of course everyone might
4:29
have a different view on this but like
4:31
from the work I'm doing like these three
4:34
trends clearly emerge but let me give
4:36
you an intro to this first i think like
4:38
when we look at AI and as in particular
4:41
the applied usage of AI I think the two
4:44
biggest transformative applied use cases
4:47
will be one computer use agents so
4:51
basically AI agents that control your
4:54
computer in the end so think about you
4:56
can give it an old uh let's say
4:58
windforms application or some
5:00
complicated ERP system no one really
5:02
understands how to use and so you can
5:05
apply a computer use agent and will
5:07
automatically you know learn how to do
5:09
this and basically fulfill a task on
5:11
behalf of you and that will be a big
5:13
transformative automation in office work
5:16
for sure um it will be massive right if
5:18
you think about business process
5:20
outsourcing and and similar kind of
5:22
things like that might actually be you
5:24
know changing from like a nearshoring or
5:27
shoring into a silicon shoring right
5:29
where you leverage agents and you
5:31
control these agentic systems so that's
5:33
one part computer use agents right and
5:35
and the second is likely humanoid
5:37
robotics and humanoid robots we might
5:39
see them first in factories but like in
5:42
the next decade I'm pretty optimistic
5:43
that we will see them you know around us
5:46
pretty much um I keep on joking right in
5:48
Germany we might still use fax machines
5:50
um by 2030 when the rest of the world
5:53
employs humanoid robotics but maybe the
5:56
final grand touring test is when we get
5:58
a humanoid robot to actually operate a
6:00
fax machine i don't know maybe that's
6:03
We'll see we'll see yeah yeah know so
6:05
it's that when when we see ro robots
6:07
doing things in our environment
6:10
autonomously
6:11
uh is that going to be a really bad day
6:13
for everyone who drives deliveries or
6:16
what's going to happen you think um you
6:19
know that the that's a common theme and
6:21
I think also the uh there's a
6:24
misconception um coming from especially
6:27
science fiction movies they ruined the
6:29
perception of a lot of people right when
6:31
you hear about robots you immediatically
6:33
see the Terminator intro scene in your
6:35
head right and you think I robot or
6:37
something like that exactly or I robot
6:39
and all these crazy crazy movies right
6:42
where it's always the bad and so that is
6:44
the perception a lot of folks are having
6:46
and that is not going to happen and also
6:48
they will not replace humans but we are
6:51
facing a challenge if you look at the
6:53
market and I was just talking with a
6:56
professor from from Italy that is very
6:58
in-depth with humanite robotics and he
7:00
was basically saying it's not just the
7:02
western world we see decline of birth
7:04
rate now happening in a lot of places in
7:06
the world and that actually means we
7:08
don't have enough workforce so who's
7:10
going to do the delivery driving uh you
7:12
know they're already like I don't know
7:14
how it is in in Sweden or the rest of
7:15
the world but for sure here in Germany
7:18
it's like truck drivers like they're
7:20
looking for truck drivers all over the
7:22
place they cannot even find enough of
7:24
them right so it's just going to um you
7:28
know fulfill the gap basically and it's
7:31
going to allow us to scale so
7:34
yeah
7:35
know it might mean that some things will
7:38
change but but likely the opportunities
7:42
for doing something is not going to go
7:44
away you're not going to be replaced by
7:46
a a silicon chip it's not that's not how
7:48
it works it might take over some some
7:50
work task that you used to do maybe um
7:53
but you will do something else it's not
7:55
like it's a threat uh you know that's
7:58
that's exactly the the right way to look
8:00
at it and in the end these are waves of
8:02
automation right we have all we also had
8:05
in the past right there's multiple waves
8:07
of automation the difference now is it
8:09
goes really fast and that's that's the
8:11
challenge we're facing right people need
8:14
to adapt to this new modern work world
8:17
and that is a challenge and so I think
8:20
the biggest challenge is actually the
8:21
cultural shift indeed that you mentioned
8:23
right but we we got to support everyone
8:27
and you know take everyone with us like
8:29
my recommendation always when I do talks
8:31
and so on is like since I don't know a
8:33
couple of years always keep on saying to
8:34
folks like try out these services try
8:37
out these tools experiment with those
8:40
because one thing is for sure like if
8:43
you have no clue about AI tools if
8:46
you're not if you're simply not
8:48
interested in that and say "Well this is
8:50
just going to go away." Then you're in
8:52
the same position like the folks in the
8:53
90s when they said "Oh the internet is
8:55
just a thing that's going to go away."
8:56
Right that's not like this is here to
8:59
stay and if you want to be relevant in
9:01
the market you got to use you got to
9:04
upskill yourself in that field that's
9:06
for sure yeah yeah no for sure no I
9:09
totally agree and and uh that's why you
9:12
know change is is scary for some people
9:15
and and we already know that so let's
9:17
let's try to um kind of look forward on
9:20
on these uh you were saying three things
9:23
right yeah yeah yes yes um so let's
9:25
let's dive into each of those right we
9:27
have multimodel models and this is the
9:31
first big trend we're seeing right now
9:33
um everyone knows generative AI right
9:35
it's a subcategory of neural networks in
9:38
the end that generate content and they
9:41
generate you know language models like
9:43
the chat GPTs of the world they generate
9:45
text um and of course you have other
9:47
models that generate images and so on
9:50
and now we also have models that can not
9:52
only generate images but they can also
9:55
have vision they have vision
9:56
capabilities so they can also kind of
9:58
understand these um images that you're
10:01
passing in and send them an image and
10:04
say create create use this image and
10:06
create you know me flying in an airplane
10:09
or something exactly and that's what we
10:12
have seen with the latest GPT4 image
10:14
generation which is pretty interesting
10:16
in itself from a technical standpoint
10:18
because they're using a different
10:20
approach than all these other image
10:21
generation models like you know stable
10:24
diffusion midjourney deli and so on they
10:26
use what is called diffusion method and
10:30
LLMs they are auto reggressive models
10:33
transformer architectures be behind this
10:36
and if you use the GPT4 image generation
10:38
it actually does not use a diffusion
10:40
model to generate the image But it also
10:41
does it auto reggressively so instead of
10:44
text token by token it's pixels by
10:46
pixels and that's why you get much
10:48
better consistency in the GPD4 image
10:50
generation also when you say hey take
10:52
this image and to put me into a plane
10:55
you really get reassembled right try to
10:57
do this with diffusion model much much
10:59
harder and that's the beauty of this
11:01
auto reggression but it also takes
11:02
longer then right so it's always kind of
11:04
trade-off there um but what I wanted to
11:06
say is these are multimodel models like
11:09
true multimodel models can interpret and
11:11
generate content across different
11:13
multiple modalities so image text audio
11:17
video and so on and the really good ones
11:20
they are actually have so-called joint
11:21
embeddings what that means is when you
11:24
train that model with these text tokens
11:26
you for example if you also want to do a
11:28
real-time audio model you basically give
11:30
the same audio snippet with it right so
11:32
internally that model learns then text
11:35
snippet you know combined to the audio
11:37
snippet and that's how these real time
11:38
audio models work you know like if
11:40
you're using Germany from Google or GPT
11:44
uh from OpenAI if you have the mobile
11:46
app you have this like real-time audio
11:47
conversation that's only possible with
11:49
these true multimodel models right so
11:52
different data types in different data
11:53
types out that's multimodality and we're
11:55
really just at the beginning um there's
11:57
a research from um Switzerland um some
12:01
university where they have 24 different
12:03
modalities like I could not even come up
12:05
with what are 24 different modalities
12:07
right but apparently there are so many
12:09
like depth data you know and all these
12:12
spectral image data and what have you
12:14
right there's all these different
12:17
more than I know yeah I was like what 24
12:20
like what could it be um I could
12:22
probably come up with 10 or so but hey
12:25
what do I know and that is multimodality
12:28
so different input types
12:31
in and also you can generate different
12:33
output types um that's very important
12:36
for these models to then also understand
12:38
the world and what we're also seeing is
12:42
of course you don't just use a single
12:44
model but you use actually multiple
12:46
instances even even it could be the same
12:49
model or could also be different models
12:51
because some of these language models
12:53
have you know certain specifics what
12:56
they're trained for like some are very
12:57
good at math or coding models are very
12:59
good at at copy you know text copy
13:01
writing that kind of a stuff and when
13:04
you combine those you end up with a
13:07
multi- aent system and Right what an
13:10
agent like you have multiple levels
13:12
there with agents right the first the
13:13
first experience is basically you're
13:15
using a chat interface and you have an
13:17
assistant kind of a thing the next level
13:20
is then the next level is then a single
13:22
agent where you have one one model but
13:24
you give it a specific role if you will
13:27
right and you give it also data
13:28
contextual data and so for example we
13:32
created an agent for RFPs right who
13:34
likes to answer RFPs of course we have
13:36
to do it um but no one likes that so we
13:38
created an agent that is very good and
13:40
has contextual data for RFPs like
13:43
request for proposals right like if
13:45
you're not familiar what what RFP is and
13:48
um that basically helps a lot and it's
13:50
it's very
13:51
specifically configured for that task so
13:54
that's a single agent and if we take
13:56
then a single agent and actually make
13:58
multiple versions of it or like other
14:00
agents then we have a multi- aent system
14:02
and that's where it really becomes
14:03
interesting so one example is take a
14:07
software project right like a software
14:08
development project you have all these
14:10
different roles you have software
14:11
developer you have QA analyst you have
14:14
project manager uh you have business
14:16
analyst whatever right depends always on
14:18
the project and the clients but you have
14:20
all these different roles now imagine
14:22
like each of these roles becomes an
14:24
agent and you as a developer actually
14:27
you are orchestrating uh a team of
14:30
agents and that's what the shift a lot
14:32
of folks need to make mentally Right
14:34
like developers will still be needed
14:36
that's for sure but instead of writing
14:39
lots of lines of code you will be an
14:41
orchestrator you will be rather a
14:43
technical program manager of an agentic
14:45
team and the crazy part is also um you
14:50
can apply this to whole organizations
14:52
right like your your whole company
14:54
organizational structure you can think
14:55
about like how can I implement agentic
14:58
teams in different roles wow and that's
15:00
that's going to be a massive shift
15:02
actually oh yeah all right so we're
15:06
going for that and uh there was another
15:08
trend about synthetic
15:11
um data yes yes synthetic data and that
15:14
that is becoming interesting for a
15:16
couple of points so first of all um if
15:19
you listen to one of the latest OpenAI
15:21
podcasts um with Sam Elkman and and some
15:23
of his colleagues they actually
15:25
mentioned a quite interesting thing that
15:28
the development of the novel language
15:30
models is not so much limited by the
15:33
compute power anymore right there they
15:35
say there there's enough compute power
15:37
available for them what they're rather
15:39
limited is which is surprising what
15:41
they're rather limited on is data and
15:44
most of the models are data bound
15:46
because they use so much data to train
15:47
these models and they're kind of
15:49
reaching a peak you know with the
15:51
available data they can use and so what
15:54
what some companies are doing and
15:55
Microsoft is very good there with the
15:57
fee four model or as our US colleagues
16:00
would say FI for um but you know the
16:03
Greek letter fee and basically that
16:06
model but also Google with Gemma what
16:08
they are leveraging also a lot of
16:09
synthetic data to actually train the
16:11
models so instead of using data from the
16:14
real world they're actually taking some
16:16
little data samples from the real world
16:18
and then augment it with synthetic data
16:20
so with you know artificially generated
16:22
data and that could be textual data but
16:25
it could also be visual data and that's
16:27
that's where it's also becoming
16:29
interesting when we talk about the other
16:30
trend about embodied intelligence or
16:32
physical AI thinking about robotics um
16:36
like synthetic data let me give you a
16:38
bunch of examples um before we dive into
16:41
the robotics aspect but one case where
16:43
we leveraged synthetic data was with
16:45
healthcare data and as you know it's of
16:48
course very sensitive data right patient
16:50
data you don't want to actually leverage
16:52
the real data too much or or just use a
16:55
small set of that
16:57
and in of course what you lose if you
17:00
then train a model on that um you lose a
17:03
lot of generalization because if you
17:04
don't have enough data you know the
17:06
model will not be good so what you
17:08
leverage then is is certain algorithms
17:11
where you can take a very small data set
17:14
and augment and expand this into you
17:17
know with a higher diversity and you get
17:19
a higher variance in the end from the
17:21
statistical standpoint which means your
17:23
model will generalize much better and
17:25
that's what we actually did with
17:27
synthetic medical document generation
17:29
and we got like a what is it I have the
17:32
slide here it's like three times higher
17:34
variance right so three times better
17:36
generalization
17:37
by just using existing data and
17:39
augmenting this like you know adding
17:42
more stuff so that's pretty interesting
17:44
I think from the classic tabular textual
17:47
machine learning data but even even more
17:50
interesting I think is it when we think
17:52
about the um application of synthetic
17:55
data for computer vision and also for
17:57
robotics scenarios so Nvidia for example
18:00
is is pretty prominent there in the
18:02
space with Nvidia Omniverse uh where
18:04
they provide you kind of simulation
18:06
environment and we're also working with
18:08
clients for example we're replicating
18:11
warehouses uh or simulating certain um
18:15
stations certain packaging stations in
18:17
the warehouse and so on and we do this
18:19
in order to identify
18:22
let me make it more practical um so
18:25
there's a case with a client they need
18:27
to load a pallet you They produce goods
18:29
and they need to put boxes on a pallet
18:32
on a pallet ship it out so usual use
18:35
case and of course the pallet loading is
18:36
done by robotic arm right there's no
18:38
human on the pallet and so it's on the
18:41
factory floor well there's still a lot
18:43
of humans there for sure but you know
18:45
like who wants to I mean these are heavy
18:47
goods right so it's good that point it's
18:50
like they don't need to take the heavy
18:51
boxes and put it on the pallet the issue
18:53
only becomes is sometimes the robot
18:55
makes mistakes and puts the boxes on the
18:58
wrong order or something which means
19:00
then if if then someone comes in with a
19:02
forklift and picks it up it might fall
19:03
over and that is expensive with the
19:06
goods right easily 100k broken stuff or
19:08
so and it only happens every few weeks
19:10
or months and that's a challenge because
19:13
how would you detect this you don't have
19:14
the real world data available to train
19:16
your model so what we're doing is we're
19:18
simulating the whole thing in a 3D
19:20
environment and then we can simulate all
19:22
the different variations you can imagine
19:24
we can render a million different
19:26
permutations out of this with different
19:28
light with different shading different
19:30
backgrounds and all of this and this is
19:32
the data we then use to train a computer
19:34
vision model which will then detect the
19:37
issue before it's too late and that is
19:40
one case another one is also if you
19:42
think about like dangerous uh locations
19:45
think about a mine right like a coal
19:46
mine or whatever right where miners are
19:48
inside imagine if there's a fire
19:50
happening right and you want to send in
19:52
a robot and the robot needs to navigate
19:54
and needs to know the way and so on um
19:57
it's very hard to simulate to have this
20:00
train trending data on the real world
20:01
right when there's a fire in the mind
20:02
like when when would you get training
20:04
data you don't have a million fires to
20:06
to to with real data you have to right
20:10
that would be bad if you had a million
20:11
fires it would be really bad especially
20:15
but you would have enough data to train
20:16
your your your AI on it that would be
20:19
good but it would be terrible to have a
20:21
million fires yeah exactly so so don't
20:24
do this um what role do is um leverage
20:28
also Unreal Engine right or other like
20:30
3D rendering you know methodologies
20:33
these days if you look at some of the
20:35
stuff you know if look at some of the
20:37
games you probably have seen it's
20:39
incredible like the quality 3D rendering
20:41
so what we can do is we can send a robot
20:44
once into the mine with a laser scanner
20:47
and it does a 3D scan of the environment
20:49
then we take this 3D scan put it in
20:51
Unreal or whatever or Omniverse doesn't
20:54
matter and then we add artificial smoke
20:56
or artificial fire and all of this and
20:58
all different variations you can imagine
21:00
so it's all fake but the AI is is
21:03
trained on the the this generated
21:06
synthetic data exactly and then we have
21:09
we have the trained model and then we
21:12
take this and this is called sim tore so
21:14
we do a simulation first and then we
21:16
deploy it on the real robotic system and
21:19
then of course that's the final real
21:20
test if it actually works if it was
21:22
trained well um but that's that's a lot
21:25
of the methodologies we see these days
21:27
it's also called robotics gyms just if
21:29
you go to the gym to train right you can
21:31
have these robotic systems actually
21:33
train in these omniverse or other
21:34
environments cool and and that's when we
21:37
when we come into embodied intelligence
21:39
right we take That's right it's like
21:41
that's like the next uh the next thing
21:43
or like the five well out of the three
21:45
we're talking about today so let's
21:46
finish off strong on embodied
21:48
intelligence then what are we talking
21:50
about now yeah exactly so that's the
21:52
point where we take this um synthetic
21:55
data for example and train a robotic
21:56
system and that's is where we have
21:58
embodied intelligence or physical AI
22:00
right you have these these novel AI
22:03
algorithms these novel AI services now
22:06
applied into an embodiment and that
22:08
could be a quadrupled a four-legged
22:10
robot like the Boston dynamic spot or a
22:13
humanoid robot we also have one of those
22:15
like a bipedal robot that walks like a
22:18
human and so we leverage synthetic data
22:20
also to train these ones and then to
22:22
apply it so that's one part of the
22:24
physical AI uh sorry one part of the
22:26
embodied intelligence is the physical
22:28
embodiment the other one is a digital
22:30
embodiment where I have an avatar a
22:32
digital human you know like a like I
22:34
give a face to a chatbot if you will um
22:37
but I can also attach to this a
22:39
effective computing which we're also
22:41
working on think about you giving it a
22:43
personality um it's actually a fun
22:45
anecdote um you probably know the Inside
22:48
Out movie right this this famous
22:50
animated movie and like the the girl
22:52
there she has eight emotions in her head
22:54
and actually these eight emotions they
22:56
did not just come up with that that's
22:57
actually based on a psychological model
22:59
from um an expert called Paul Eggman and
23:02
so that's the Paul Eggman model psych
23:04
we're actually leveraging the same model
23:05
for our digital human so eight different
23:07
emotions are firing each of them is an
23:10
agent multi- aent system so let's say
23:12
the fear or the surprise agent they're
23:15
they're getting an input from the user
23:17
and then they're firing right and then
23:18
we have an orchestration agent which
23:20
then assembles the final answer and then
23:22
we have a system that reacts more
23:23
empathetically and has more you know
23:26
sentiment and emotion where you can as a
23:28
human uh you know talk talk to and also
23:31
feel more connected i would say wow
23:34
that's that's incredibly cool you know
23:37
what unfortunately we are running out of
23:40
time for this show we could have talked
23:43
about this for the longest time but
23:45
we'll stay true to form it's a 20-minut
23:47
show so for the audience we're This is
23:50
where we cut Renee off this time thank
23:52
you so much for being on the cloud show
23:54
today i appreciate it thanks for having
23:56
me and take care my friends and don't
23:58
worry just make sure you look into AI
24:00
services and stay up to date that's
24:02
right and I'll see you guys next time on
24:04
the Cloud Show
24:06
bye-bye
24:11
[Music]