AI Architecture: What Is Possible Today and Future Trends | Software Architecture Conference

Name: AI Architecture: What Is Possible Today and Future Trends | Software Architecture Conference
Uploaded: 2025-08-06T12:06:14+00:00
Duration: 35 min 35 s

0:03
all right today I'm going to talk about
0:05
all right today I'm going to talk about
0:05
all right today I'm going to talk about AI architecture Trends and components
0:07
AI architecture Trends and components
0:07
AI architecture Trends and components and share with you my personal points of
0:09
and share with you my personal points of
0:09
and share with you my personal points of view as a frequent speaker on II topics
0:12
view as a frequent speaker on II topics
0:12
view as a frequent speaker on II topics and Linkedin author with some courses on
0:15
and Linkedin author with some courses on
0:15
and Linkedin author with some courses on generative VII I have one in semantic
0:17
generative VII I have one in semantic
0:17
generative VII I have one in semantic colel another in N that should be
0:20
colel another in N that should be
0:20
colel another in N that should be published end of this month and soon
0:24
published end of this month and soon
0:24
published end of this month and soon starting one in aogen gen Ki I hope this
0:27
starting one in aogen gen Ki I hope this
0:27
starting one in aogen gen Ki I hope this gives me some credibility I will not
0:30
gives me some credibility I will not
0:30
gives me some credibility I will not call myself an expert as you know this
0:33
call myself an expert as you know this
0:33
call myself an expert as you know this AI W is crazy it's super fast evolving
0:36
AI W is crazy it's super fast evolving
0:36
AI W is crazy it's super fast evolving and this makes impossible for anybody to
0:40
and this makes impossible for anybody to
0:40
and this makes impossible for anybody to be an expert but I would say know one or
0:42
be an expert but I would say know one or
0:43
be an expert but I would say know one or two things at
0:45
two things at
0:45
two things at least so yeah changes I don't need to
0:49
least so yeah changes I don't need to
0:49
least so yeah changes I don't need to ask you if you are familiar with ch gbt
0:52
ask you if you are familiar with ch gbt
0:52
ask you if you are familiar with ch gbt or generative Ai and probably the terms
0:54
or generative Ai and probably the terms
0:54
or generative Ai and probably the terms AI orchestrator or even L chain maybe
0:57
AI orchestrator or even L chain maybe
0:57
AI orchestrator or even L chain maybe semantic kernel ring a bell if not many
1:01
semantic kernel ring a bell if not many
1:01
semantic kernel ring a bell if not many by now um I'm pretty sure that you have
1:05
by now um I'm pretty sure that you have
1:05
by now um I'm pretty sure that you have been playing with these Technologies uh
1:08
been playing with these Technologies uh
1:08
been playing with these Technologies uh and I can tell you that these are the
1:10
and I can tell you that these are the
1:10
and I can tell you that these are the core Technologies of this new generative
1:13
core Technologies of this new generative
1:13
core Technologies of this new generative AI World
1:15
AI World
1:15
AI World [Music]
1:16
[Music]
1:17
[Music] um so welcome to it if you did not
1:20
um so welcome to it if you did not
1:20
um so welcome to it if you did not realize last year every at least in my
1:23
realize last year every at least in my
1:23
realize last year every at least in my work I hear openi I hear Chachi hia like
1:28
work I hear openi I hear Chachi hia like
1:28
work I hear openi I hear Chachi hia like every now and then it's it's incredible
1:31
every now and then it's it's incredible
1:31
every now and then it's it's incredible amount of times I hear these
1:34
amount of times I hear these
1:34
amount of times I hear these words so one thing before get into it
1:38
words so one thing before get into it
1:38
words so one thing before get into it that I love about AI is uh the the
1:43
that I love about AI is uh the the
1:43
that I love about AI is uh the the feeling when you're doing it that it can
1:46
feeling when you're doing it that it can
1:46
feeling when you're doing it that it can do surprising things and it makes you
1:48
do surprising things and it makes you
1:48
do surprising things and it makes you feel like a magician like Orchestra
1:50
feel like a magician like Orchestra
1:50
feel like a magician like Orchestra director which you are playing with
1:52
director which you are playing with
1:52
director which you are playing with literally with digital brains right and
1:56
literally with digital brains right and
1:56
literally with digital brains right and uh this is what amazed me as well the
1:59
uh this is what amazed me as well the
1:59
uh this is what amazed me as well the way you have to change your mind when
2:01
way you have to change your mind when
2:01
way you have to change your mind when interaction interacting with these
2:03
interaction interacting with these
2:03
interaction interacting with these components that can think on their own
2:06
components that can think on their own
2:06
components that can think on their own and listen to these words think on their
2:08
and listen to these words think on their
2:08
and listen to these words think on their own and they are non deterministic right
2:12
own and they are non deterministic right
2:12
own and they are non deterministic right for a developer that something really or
2:14
for a developer that something really or
2:14
for a developer that something really or an AR that getes something really
2:16
an AR that getes something really
2:16
an AR that getes something really strange at the beginning it it gets you
2:19
strange at the beginning it it gets you
2:19
strange at the beginning it it gets you this aha moment when AI delivers an
2:22
this aha moment when AI delivers an
2:22
this aha moment when AI delivers an unexpected yet perfect
2:25
unexpected yet perfect
2:25
unexpected yet perfect result that is truly priceless
2:33
all right you will agree on me that
2:37
all right you will agree on me that
2:37
all right you will agree on me that these technology AI has a incredible
2:40
these technology AI has a incredible
2:40
these technology AI has a incredible disruptive
2:42
disruptive
2:42
disruptive potential this writing and translation
2:45
potential this writing and translation
2:45
potential this writing and translation in industry is in shock at the moment it
2:48
in industry is in shock at the moment it
2:48
in industry is in shock at the moment it was already last year and some call
2:50
was already last year and some call
2:50
was already last year and some call setup professionals lately have already
2:52
setup professionals lately have already
2:52
setup professionals lately have already started career conversions due to the
2:54
started career conversions due to the
2:54
started career conversions due to the voice chat AI voice and voice
2:57
voice chat AI voice and voice
2:57
voice chat AI voice and voice recognition interfaces which in in some
3:00
recognition interfaces which in in some
3:00
recognition interfaces which in in some products already in the market they are
3:02
products already in the market they are
3:02
products already in the market they are performing better than 40 and 60% of
3:05
performing better than 40 and 60% of
3:05
performing better than 40 and 60% of some of colleagues or people that work
3:07
some of colleagues or people that work
3:07
some of colleagues or people that work on it I got myself a friend that worked
3:11
on it I got myself a friend that worked
3:11
on it I got myself a friend that worked in Translation
3:12
in Translation
3:12
in Translation well their business has not grown the
3:15
well their business has not grown the
3:15
well their business has not grown the last two years and and I'm falling quite
3:17
last two years and and I'm falling quite
3:17
last two years and and I'm falling quite short but enough talk let's get started
3:20
short but enough talk let's get started
3:20
short but enough talk let's get started let's take a look at the current state
3:23
let's take a look at the current state
3:23
let's take a look at the current state of
3:26
a oh yeah I got some here uh
3:30
a oh yeah I got some here uh
3:30
a oh yeah I got some here uh disruptions will AI replace the writers
3:32
disruptions will AI replace the writers
3:32
disruptions will AI replace the writers you search for it AI assistant novels
3:35
you search for it AI assistant novels
3:35
you search for it AI assistant novels are winning more AI is disrupting the
3:38
are winning more AI is disrupting the
3:38
are winning more AI is disrupting the billion translation industry well yeah
3:41
billion translation industry well yeah
3:41
billion translation industry well yeah it's quite disrupting everything that we
3:43
it's quite disrupting everything that we
3:43
it's quite disrupting everything that we know as complete Industries now going to
3:47
know as complete Industries now going to
3:47
know as complete Industries now going to call centers so let's on the current
3:50
call centers so let's on the current
3:50
call centers so let's on the current state of
3:52
state of
3:52
state of AI I love this slide let's start by
3:56
AI I love this slide let's start by
3:56
AI I love this slide let's start by understanding the limitations of current
3:59
understanding the limitations of current
3:59
understanding the limitations of current AI
4:00
AI
4:00
AI systems first rack retrieval and
4:05
systems first rack retrieval and
4:05
systems first rack retrieval and generation do they always when you do
4:08
generation do they always when you do
4:08
generation do they always when you do rack which rack is okay I do some take
4:12
rack which rack is okay I do some take
4:13
rack which rack is okay I do some take some text a book or some information
4:15
some text a book or some information
4:15
some text a book or some information from your company I put in a vector
4:17
from your company I put in a vector
4:17
from your company I put in a vector database and through Vector similarity I
4:19
database and through Vector similarity I
4:19
database and through Vector similarity I try to retrieve some information through
4:22
try to retrieve some information through
4:22
try to retrieve some information through keywords so this is often very
4:26
keywords so this is often very
4:26
keywords so this is often very inaccurate because how much cont text
4:30
inaccurate because how much cont text
4:30
inaccurate because how much cont text can we put in there in in in this Vector
4:33
can we put in there in in in this Vector
4:33
can we put in there in in in this Vector database usually you put that you have
4:36
database usually you put that you have
4:36
database usually you put that you have to split in chunks of text which
4:38
to split in chunks of text which
4:38
to split in chunks of text which sometimes you are dividing their
4:39
sometimes you are dividing their
4:40
sometimes you are dividing their knowledge in different pieces and when
4:41
knowledge in different pieces and when
4:41
knowledge in different pieces and when you are retrieving you don't get all of
4:44
you are retrieving you don't get all of
4:44
you are retrieving you don't get all of them and sometimes you retrieve the
4:45
them and sometimes you retrieve the
4:45
them and sometimes you retrieve the wrong one which happens quite a lot so
4:48
wrong one which happens quite a lot so
4:48
wrong one which happens quite a lot so rack is really really hard it's getting
4:50
rack is really really hard it's getting
4:50
rack is really really hard it's getting better we will see that in a
4:53
better we will see that in a
4:53
better we will see that in a moment and finally um well next the AI
4:57
moment and finally um well next the AI
4:57
moment and finally um well next the AI is hallucinating a lot so that's a
5:00
is hallucinating a lot so that's a
5:00
is hallucinating a lot so that's a significant issue where models generate
5:01
significant issue where models generate
5:01
significant issue where models generate information that seems okay but it's
5:04
information that seems okay but it's
5:04
information that seems okay but it's absolutely not it's absolutely made up
5:07
absolutely not it's absolutely made up
5:07
absolutely not it's absolutely made up right and finally maintaining the
5:09
right and finally maintaining the
5:09
right and finally maintaining the context over long conversations so we
5:12
context over long conversations so we
5:12
context over long conversations so we went more from AI now we have models
5:14
went more from AI now we have models
5:14
went more from AI now we have models that have one and half million tokens
5:16
that have one and half million tokens
5:16
that have one and half million tokens and this trend is growing but they keep
5:20
and this trend is growing but they keep
5:20
and this trend is growing but they keep forgetting what we told is like uh we
5:22
forgetting what we told is like uh we
5:22
forgetting what we told is like uh we are talking with somebody and suddenly
5:24
are talking with somebody and suddenly
5:24
are talking with somebody and suddenly hey what did you tell at the beginning I
5:26
hey what did you tell at the beginning I
5:26
hey what did you tell at the beginning I don't remember anything of it so the
5:27
don't remember anything of it so the
5:27
don't remember anything of it so the same happens with ai ai is different so
5:30
same happens with ai ai is different so
5:30
same happens with ai ai is different so they have the tendence to remember what
5:32
they have the tendence to remember what
5:32
they have the tendence to remember what was told at the beginning and at the end
5:34
was told at the beginning and at the end
5:34
was told at the beginning and at the end but in the middle sometimes it forgets
5:38
but in the middle sometimes it forgets
5:38
but in the middle sometimes it forgets right so these are some of the
5:40
right so these are some of the
5:40
right so these are some of the limitations in here but despite these
5:44
limitations in here but despite these
5:44
limitations in here but despite these limitations AI continues to devop
5:46
limitations AI continues to devop
5:47
limitations AI continues to devop various Fields with its potential and
5:49
various Fields with its potential and
5:50
various Fields with its potential and and it's really really good at uh couple
5:53
and it's really really good at uh couple
5:53
and it's really really good at uh couple three or four things one is NLP natural
5:56
three or four things one is NLP natural
5:56
three or four things one is NLP natural language processing this has
5:58
language processing this has
5:58
language processing this has revolutionized revolutionize tasks like
6:01
revolutionized revolutionize tasks like
6:01
revolutionized revolutionize tasks like text summarization translation question
6:04
text summarization translation question
6:04
text summarization translation question answering sentiment analysis has also
6:07
answering sentiment analysis has also
6:07
answering sentiment analysis has also improve tremendously so I would say now
6:10
improve tremendously so I would say now
6:10
improve tremendously so I would say now it's mostly perfect and we have ai
6:14
it's mostly perfect and we have ai
6:14
it's mostly perfect and we have ai systems that can understand perfectly
6:16
systems that can understand perfectly
6:16
systems that can understand perfectly human emotions and intentions this is
6:19
human emotions and intentions this is
6:19
human emotions and intentions this is known as theory of mind and lately very
6:24
known as theory of mind and lately very
6:24
known as theory of mind and lately very recently it has been determined that it
6:26
recently it has been determined that it
6:26
recently it has been determined that it can also simulate and predict them not
6:29
can also simulate and predict them not
6:29
can also simulate and predict them not yet at the level of Harry S Seldon
6:32
yet at the level of Harry S Seldon
6:32
yet at the level of Harry S Seldon psycho history if any body gets and
6:36
psycho history if any body gets and
6:36
psycho history if any body gets and knows what I'm talking about please
6:37
knows what I'm talking about please
6:37
knows what I'm talking about please raise the hand that's very sneaky if you
6:40
raise the hand that's very sneaky if you
6:40
raise the hand that's very sneaky if you like science fiction like I do and yeah
6:44
like science fiction like I do and yeah
6:44
like science fiction like I do and yeah but those skills are already at human
6:46
but those skills are already at human
6:46
but those skills are already at human level if not more and finally prompting
6:49
level if not more and finally prompting
6:49
level if not more and finally prompting techniques including meta prompting are
6:51
techniques including meta prompting are
6:51
techniques including meta prompting are enhancing AI responses making
6:53
enhancing AI responses making
6:53
enhancing AI responses making interactions more relevant and
6:56
interactions more relevant and
6:56
interactions more relevant and precise next I want to show you the
6:59
precise next I want to show you the
6:59
precise next I want to show you the Garner AI High which you may recognize
7:09
yeah
7:11
yeah
7:11
yeah so here um you can see that it provides
7:15
so here um you can see that it provides
7:15
so here um you can see that it provides a snapshot for various AI Technologies
7:17
a snapshot for various AI Technologies
7:17
a snapshot for various AI Technologies stand in terms of maturity and adoption
7:19
stand in terms of maturity and adoption
7:19
stand in terms of maturity and adoption on the different uh phases I recommend
7:22
on the different uh phases I recommend
7:22
on the different uh phases I recommend you to search for Garner AI have I will
7:24
you to search for Garner AI have I will
7:24
you to search for Garner AI have I will not disclose or explain it in detail but
7:27
not disclose or explain it in detail but
7:27
not disclose or explain it in detail but basically you can see that composite AI
7:31
basically you can see that composite AI
7:31
basically you can see that composite AI uh and hii these white ones here are
7:35
uh and hii these white ones here are
7:35
uh and hii these white ones here are already in under two years which is
7:37
already in under two years which is
7:37
already in under two years which is plateau of productivity IND indicating
7:40
plateau of productivity IND indicating
7:40
plateau of productivity IND indicating their growing maturity in real world
7:43
their growing maturity in real world
7:43
their growing maturity in real world applicability uh whatever their uh
7:45
applicability uh whatever their uh
7:45
applicability uh whatever their uh Innovation Trigger or pick of
7:48
Innovation Trigger or pick of
7:48
Innovation Trigger or pick of expectations or it's this
7:50
expectations or it's this
7:50
expectations or it's this disillusionment like more disappointment
7:53
disillusionment like more disappointment
7:53
disillusionment like more disappointment as well but whatever composite AI is the
7:55
as well but whatever composite AI is the
7:55
as well but whatever composite AI is the combined application or Fusion of
7:56
combined application or Fusion of
7:56
combined application or Fusion of different tactics to improve the AI
7:59
different tactics to improve the AI
7:59
different tactics to improve the AI solution and it's at the center of geni
8:01
solution and it's at the center of geni
8:01
solution and it's at the center of geni as we know it at the moment and also
8:04
as we know it at the moment and also
8:04
as we know it at the moment and also agent based modeling is the next wave of
8:08
agent based modeling is the next wave of
8:08
agent based modeling is the next wave of composite AI which is important I will
8:10
composite AI which is important I will
8:10
composite AI which is important I will talk a bit about that multi agent system
8:13
talk a bit about that multi agent system
8:13
talk a bit about that multi agent system are on the five to I think should be to
8:16
are on the five to I think should be to
8:16
are on the five to I think should be to 10 years so it's on is uh black and it's
8:22
10 years so it's on is uh black and it's
8:22
10 years so it's on is uh black and it's quite
8:24
quite
8:24
quite okay I cannot find it I S I saw it
8:30
okay I cannot find it I S I saw it
8:30
okay I cannot find it I S I saw it uh
8:32
uh
8:32
uh yeah all right so it's uh yeah um multi
8:38
yeah all right so it's uh yeah um multi
8:38
yeah all right so it's uh yeah um multi agent system is on the five 10 year
8:40
agent system is on the five 10 year
8:40
agent system is on the five 10 year Orizon even I cannot find it uh yeah
8:43
Orizon even I cannot find it uh yeah
8:43
Orizon even I cannot find it uh yeah when we must need it but it's important
8:47
when we must need it but it's important
8:47
when we must need it but it's important to see that it's uh two to five years uh
8:52
to see that it's uh two to five years uh
8:52
to see that it's uh two to five years uh a lot of the AI technologies that you
8:54
a lot of the AI technologies that you
8:54
a lot of the AI technologies that you see here it it tells a lot that in two
8:56
see here it it tells a lot that in two
8:56
see here it it tells a lot that in two five years all the Technologies that we
8:59
five years all the Technologies that we
9:00
five years all the Technologies that we know as AI will be measure so that means
9:03
know as AI will be measure so that means
9:03
know as AI will be measure so that means that yep uh we should get already there
9:06
that yep uh we should get already there
9:06
that yep uh we should get already there into learning if we are not learning
9:08
into learning if we are not learning
9:08
into learning if we are not learning them and get our hands on it if we want
9:10
them and get our hands on it if we want
9:10
them and get our hands on it if we want to retain our skills as an architect or
9:13
to retain our skills as an architect or
9:13
to retain our skills as an architect or as a developer as simple as
9:15
as a developer as simple as
9:15
as a developer as simple as that as well uh special mention to
9:18
that as well uh special mention to
9:18
that as well uh special mention to intelligent applications and knowledge
9:20
intelligent applications and knowledge
9:20
intelligent applications and knowledge grabs and Cloud AI Services I'm curious
9:23
grabs and Cloud AI Services I'm curious
9:23
grabs and Cloud AI Services I'm curious as why AI engineering and geni are at
9:25
as why AI engineering and geni are at
9:25
as why AI engineering and geni are at the top of the slope and about to go
9:27
the top of the slope and about to go
9:27
the top of the slope and about to go down the expectation hype
9:30
down the expectation hype
9:30
down the expectation hype here as that's not my opinion but that's
9:34
here as that's not my opinion but that's
9:34
here as that's not my opinion but that's Garner so we should trust Garner um but
9:38
Garner so we should trust Garner um but
9:38
Garner so we should trust Garner um but yeah to my to me personally I would put
9:40
yeah to my to me personally I would put
9:40
yeah to my to me personally I would put some AI engineering on the plateau of
9:42
some AI engineering on the plateau of
9:42
some AI engineering on the plateau of productivity already as partially but I
9:44
productivity already as partially but I
9:45
productivity already as partially but I mentioned partially take it with a gr of
9:46
mentioned partially take it with a gr of
9:46
mentioned partially take it with a gr of sale to me is business ready that is the
9:49
sale to me is business ready that is the
9:49
sale to me is business ready that is the case for me in a very personal opinion
9:51
case for me in a very personal opinion
9:51
case for me in a very personal opinion with semantic
9:52
with semantic
9:52
with semantic kernel but take it as a with a grain of
9:55
kernel but take it as a with a grain of
9:55
kernel but take it as a with a grain of salt not everything is already there but
9:58
salt not everything is already there but
9:58
salt not everything is already there but they are in a great Direction
10:00
they are in a great Direction
10:00
they are in a great Direction but what is in a great Direction and
10:02
but what is in a great Direction and
10:02
but what is in a great Direction and it's working and it's out of
10:04
it's working and it's out of
10:04
it's working and it's out of experimental it's working and you can
10:06
experimental it's working and you can
10:06
experimental it's working and you can use it in business applications as of
10:10
use it in business applications as of
10:10
use it in business applications as of today here I put the Garner AI hype link
10:15
today here I put the Garner AI hype link
10:15
today here I put the Garner AI hype link I will share the links later in my
10:18
I will share the links later in my
10:18
I will share the links later in my Discord and I will put yeah semantic
10:21
Discord and I will put yeah semantic
10:21
Discord and I will put yeah semantic kernel in in this part to my opinion but
10:25
kernel in in this part to my opinion but
10:25
kernel in in this part to my opinion but that's completely my
10:27
that's completely my
10:27
that's completely my opinion already
10:30
opinion already
10:30
opinion already aome we can see here um
10:34
aome we can see here um
10:34
aome we can see here um this is kind of a positioning of
10:37
this is kind of a positioning of
10:37
this is kind of a positioning of different language large language models
10:39
different language large language models
10:40
different language large language models as of June 2024 this is taken from life
10:43
as of June 2024 this is taken from life
10:43
as of June 2024 this is taken from life architect. and you can see in the last
10:46
architect. and you can see in the last
10:46
architect. and you can see in the last uh models of oppus and Sonet how good
10:49
uh models of oppus and Sonet how good
10:49
uh models of oppus and Sonet how good they score in GP
10:51
they score in GP
10:52
they score in GP QA yeah this year artificial
10:55
QA yeah this year artificial
10:55
QA yeah this year artificial intelligence is far smarter than humans
10:58
intelligence is far smarter than humans
10:58
intelligence is far smarter than humans across many method
11:00
across many method
11:00
across many method I'm not saying SP or more intelligent I
11:03
I'm not saying SP or more intelligent I
11:03
I'm not saying SP or more intelligent I mean they are scoring
11:05
mean they are scoring
11:05
mean they are scoring better so it's they doing better and
11:08
better so it's they doing better and
11:09
better so it's they doing better and playing smarter that we may think this
11:11
playing smarter that we may think this
11:11
playing smarter that we may think this is a GP QA this is Google proof qna
11:15
is a GP QA this is Google proof qna
11:15
is a GP QA this is Google proof qna Benchmark is one of the hardest
11:18
Benchmark is one of the hardest
11:18
Benchmark is one of the hardest benchmarks that determines the according
11:21
benchmarks that determines the according
11:21
benchmarks that determines the according to the score the IQ so at the moment if
11:25
to the score the IQ so at the moment if
11:25
to the score the IQ so at the moment if we look at the IQ this should be at the
11:28
we look at the IQ this should be at the
11:28
we look at the IQ this should be at the level of 100 50 which is quite
11:33
level of 100 50 which is quite
11:33
level of 100 50 which is quite impressive oppus and Lama are setting
11:36
impressive oppus and Lama are setting
11:36
impressive oppus and Lama are setting new benchmarks there are more uh I think
11:39
new benchmarks there are more uh I think
11:39
new benchmarks there are more uh I think a few days ago we have the new gro from
11:42
a few days ago we have the new gro from
11:42
a few days ago we have the new gro from Elon Musk and it's about
11:44
Elon Musk and it's about
11:44
Elon Musk and it's about 506 so it's not yet updated this graph
11:49
506 so it's not yet updated this graph
11:49
506 so it's not yet updated this graph but basically it's evolving and this
11:52
but basically it's evolving and this
11:52
but basically it's evolving and this does not seem to stop this evolution of
11:55
does not seem to stop this evolution of
11:55
does not seem to stop this evolution of large language models multimodal AI
11:58
large language models multimodal AI
11:58
large language models multimodal AI models like gp4 or are integrating text
12:01
models like gp4 or are integrating text
12:01
models like gp4 or are integrating text uh image audio and even video data
12:03
uh image audio and even video data
12:03
uh image audio and even video data opening up new possibilities for
12:05
opening up new possibilities for
12:05
opening up new possibilities for holistic and context aware
12:07
holistic and context aware
12:07
holistic and context aware application including application in
12:10
application including application in
12:10
application including application in robotics as we are seeing with um open
12:14
robotics as we are seeing with um open
12:14
robotics as we are seeing with um open Ai and other companies that are jumping
12:16
Ai and other companies that are jumping
12:16
Ai and other companies that are jumping in this robotic uh one wagon as AI
12:20
in this robotic uh one wagon as AI
12:20
in this robotic uh one wagon as AI becomes more P per pervasive uh
12:23
becomes more P per pervasive uh
12:23
becomes more P per pervasive uh addressing security and ethical
12:25
addressing security and ethical
12:25
addressing security and ethical considerations as data protection and
12:26
considerations as data protection and
12:26
considerations as data protection and fairness is crucial and finally meta
12:30
fairness is crucial and finally meta
12:30
fairness is crucial and finally meta prompting is emerging as a valuable
12:32
prompting is emerging as a valuable
12:32
prompting is emerging as a valuable technique for guiding AI behavior and
12:34
technique for guiding AI behavior and
12:34
technique for guiding AI behavior and controlling it with guard
12:36
controlling it with guard
12:36
controlling it with guard rails and this is kind of big uh
12:40
rails and this is kind of big uh
12:40
rails and this is kind of big uh overview of the current state of
12:47
AI so they are increasing their
12:49
AI so they are increasing their
12:49
AI so they are increasing their reasoning skills they are increasing
12:52
reasoning skills they are increasing
12:52
reasoning skills they are increasing their multimodel skills yeah and also
12:55
their multimodel skills yeah and also
12:55
their multimodel skills yeah and also there is this uh PR engineering
12:57
there is this uh PR engineering
12:57
there is this uh PR engineering Evolution towards prompting which is
12:59
Evolution towards prompting which is
13:00
Evolution towards prompting which is nothing more to put a proper system
13:02
nothing more to put a proper system
13:02
nothing more to put a proper system prompt that kind of guides uh and
13:05
prompt that kind of guides uh and
13:05
prompt that kind of guides uh and protects the model that it does not do
13:08
protects the model that it does not do
13:08
protects the model that it does not do something that you don't want
13:11
something that you don't want
13:11
something that you don't want to rapers with code data set GP QA is
13:15
to rapers with code data set GP QA is
13:15
to rapers with code data set GP QA is some more information on the
13:17
some more information on the
13:17
some more information on the gpq is a very common frame
13:21
gpq is a very common frame
13:21
gpq is a very common frame Benchmark and let's say uh we have seen
13:25
Benchmark and let's say uh we have seen
13:25
Benchmark and let's say uh we have seen a bit the current trends we are going to
13:28
a bit the current trends we are going to
13:28
a bit the current trends we are going to pick at the future future Trends and
13:30
pick at the future future Trends and
13:30
pick at the future future Trends and some of the current
13:33
ones a journey begins with the
13:36
ones a journey begins with the
13:36
ones a journey begins with the significant improvements in a models
13:39
significant improvements in a models
13:39
significant improvements in a models that we just
13:40
that we just
13:40
that we just saw as probably you will this will
13:43
saw as probably you will this will
13:43
saw as probably you will this will remember uh song they are getting better
13:47
remember uh song they are getting better
13:47
remember uh song they are getting better faster stronger and also
13:49
faster stronger and also
13:49
faster stronger and also multimodel they are not only faster they
13:52
multimodel they are not only faster they
13:52
multimodel they are not only faster they are also more efficient so they are uh a
13:56
are also more efficient so they are uh a
13:56
are also more efficient so they are uh a lot cheaper than they were some years
13:58
lot cheaper than they were some years
13:58
lot cheaper than they were some years ago multimodal capabilities they allow
14:01
ago multimodal capabilities they allow
14:01
ago multimodal capabilities they allow them to seamlessly integrate text image
14:04
them to seamlessly integrate text image
14:04
them to seamlessly integrate text image video audio and other kind of data
14:07
video audio and other kind of data
14:07
video audio and other kind of data leading to more comprehensive outputs in
14:10
leading to more comprehensive outputs in
14:10
leading to more comprehensive outputs in near future I expect they will combine
14:13
near future I expect they will combine
14:13
near future I expect they will combine more abilities to generate code 3D
14:15
more abilities to generate code 3D
14:15
more abilities to generate code 3D objects or understand uh 3D Vision a vis
14:20
objects or understand uh 3D Vision a vis
14:20
objects or understand uh 3D Vision a vis video and more and they will continue to
14:23
video and more and they will continue to
14:23
video and more and they will continue to evolve that's not stopping it's kind of
14:27
evolve that's not stopping it's kind of
14:27
evolve that's not stopping it's kind of a commercial race at the moment
14:30
a commercial race at the moment
14:30
a commercial race at the moment Additionally the Adent of natural voice
14:32
Additionally the Adent of natural voice
14:32
Additionally the Adent of natural voice interfaces like the one the GPT 40 voice
14:36
interfaces like the one the GPT 40 voice
14:36
interfaces like the one the GPT 40 voice that everybody's edly waiting for it to
14:39
that everybody's edly waiting for it to
14:39
that everybody's edly waiting for it to be mainstream as an app and API will
14:42
be mainstream as an app and API will
14:42
be mainstream as an app and API will make AI interactions incredibly
14:45
make AI interactions incredibly
14:45
make AI interactions incredibly humanlike enhancing user experience
14:47
humanlike enhancing user experience
14:48
humanlike enhancing user experience through various
14:49
through various
14:49
through various applications and we come to one of my
14:52
applications and we come to one of my
14:52
applications and we come to one of my favorites agents what are agents I mean
14:56
favorites agents what are agents I mean
14:56
favorites agents what are agents I mean AI agents I see them as an entity which
14:59
AI agents I see them as an entity which
14:59
AI agents I see them as an entity which brings together some reasoning abilities
15:02
brings together some reasoning abilities
15:02
brings together some reasoning abilities determined by the model A Persona which
15:05
determined by the model A Persona which
15:05
determined by the model A Persona which is a personality that determines how it
15:08
is a personality that determines how it
15:08
is a personality that determines how it reacts or he or she reacts to situation
15:11
reacts or he or she reacts to situation
15:11
reacts or he or she reacts to situation or tries to perform a task right I told
15:14
or tries to perform a task right I told
15:14
or tries to perform a task right I told him hey you are Thor and you're giving a
15:18
him hey you are Thor and you're giving a
15:18
him hey you are Thor and you're giving a a tool he will try a hammer he will try
15:21
a tool he will try a hammer he will try
15:21
a tool he will try a hammer he will try to beat and break everything and solve
15:23
to beat and break everything and solve
15:23
to beat and break everything and solve everything with the hammer I give it a
15:25
everything with the hammer I give it a
15:25
everything with the hammer I give it a banana it will try to solve everything
15:26
banana it will try to solve everything
15:26
banana it will try to solve everything with a banana right or I give it humor I
15:29
with a banana right or I give it humor I
15:29
with a banana right or I give it humor I thought hey your personality is humorous
15:31
thought hey your personality is humorous
15:31
thought hey your personality is humorous try to solve everything with humor you
15:34
try to solve everything with humor you
15:34
try to solve everything with humor you guess where I'm going as well these
15:38
guess where I'm going as well these
15:38
guess where I'm going as well these agents also have tools and the ability
15:41
agents also have tools and the ability
15:41
agents also have tools and the ability to decide which tool to use and
15:44
to decide which tool to use and
15:44
to decide which tool to use and how the personality can range from being
15:47
how the personality can range from being
15:47
how the personality can range from being friendly and conversational to
15:49
friendly and conversational to
15:49
friendly and conversational to analytical and precise or pragmatic if
15:52
analytical and precise or pragmatic if
15:52
analytical and precise or pragmatic if you want an agent action driven they can
15:55
you want an agent action driven they can
15:55
you want an agent action driven they can interact and respond to a user or to
15:57
interact and respond to a user or to
15:57
interact and respond to a user or to other agents or interact and react to
16:00
other agents or interact and react to
16:00
other agents or interact and react to the environment as
16:02
the environment as
16:02
the environment as well that's an
16:06
agent yeah this tool it should be the
16:08
agent yeah this tool it should be the
16:08
agent yeah this tool it should be the personality reasoning
16:09
personality reasoning
16:09
personality reasoning abilities yeah then we come to agentic
16:12
abilities yeah then we come to agentic
16:12
abilities yeah then we come to agentic AI which is when we have more than one
16:14
AI which is when we have more than one
16:15
AI which is when we have more than one agent working together this is a system
16:19
agent working together this is a system
16:19
agent working together this is a system when multiple intelligent agents
16:21
when multiple intelligent agents
16:21
when multiple intelligent agents collaborate to perform complex task or
16:24
collaborate to perform complex task or
16:24
collaborate to perform complex task or one task which is complex these agents
16:27
one task which is complex these agents
16:27
one task which is complex these agents are usually designed with concrete
16:29
are usually designed with concrete
16:29
are usually designed with concrete capabilities so it's focus focus focus
16:33
capabilities so it's focus focus focus
16:33
capabilities so it's focus focus focus this enables them to do a single task
16:36
this enables them to do a single task
16:36
this enables them to do a single task becoming expert on it and that's highly
16:38
becoming expert on it and that's highly
16:38
becoming expert on it and that's highly important because models are kind of
16:40
important because models are kind of
16:40
important because models are kind of limited and they looks context very
16:42
limited and they looks context very
16:42
limited and they looks context very quickly and hallucinate so the more
16:44
quickly and hallucinate so the more
16:44
quickly and hallucinate so the more Focus you give them the better they will
16:47
Focus you give them the better they will
16:47
Focus you give them the better they will perform and on a conversation of
16:49
perform and on a conversation of
16:49
perform and on a conversation of workflow that usually design around them
16:52
workflow that usually design around them
16:52
workflow that usually design around them so they can interact in a very optimized
16:54
so they can interact in a very optimized
16:54
so they can interact in a very optimized way usually you have to streamline to
16:57
way usually you have to streamline to
16:57
way usually you have to streamline to resolve a task or set of tasks together
17:00
resolve a task or set of tasks together
17:00
resolve a task or set of tasks together as a human team would some Advanced
17:03
as a human team would some Advanced
17:03
as a human team would some Advanced agents can perform self-critic and
17:05
agents can perform self-critic and
17:05
agents can perform self-critic and improve over time
17:07
improve over time
17:07
improve over time learning which is the case for for
17:10
learning which is the case for for
17:10
learning which is the case for for example autogen right this collaborative
17:14
example autogen right this collaborative
17:14
example autogen right this collaborative approach opens up a lot of new
17:16
approach opens up a lot of new
17:16
approach opens up a lot of new possibilities for automation efficiency
17:18
possibilities for automation efficiency
17:18
possibilities for automation efficiency and scalability in various domains and
17:20
and scalability in various domains and
17:20
and scalability in various domains and if you think about that which kind of
17:23
if you think about that which kind of
17:23
if you think about that which kind of very similar workflow that you do with
17:25
very similar workflow that you do with
17:25
very similar workflow that you do with your chpt just that instead of you
17:27
your chpt just that instead of you
17:27
your chpt just that instead of you having to type you have agent which will
17:30
having to type you have agent which will
17:30
having to type you have agent which will type for you for example so it saves a
17:33
type for you for example so it saves a
17:33
type for you for example so it saves a lot of
17:37
time also it's not just
17:40
time also it's not just
17:40
time also it's not just me so the CEO of Nvidia yeah a are going
17:45
me so the CEO of Nvidia yeah a are going
17:45
me so the CEO of Nvidia yeah a are going to be like employees in your
17:48
companies also Andrew and G yeah if
17:52
companies also Andrew and G yeah if
17:52
companies also Andrew and G yeah if anything to keep an eye on its AI agents
17:55
anything to keep an eye on its AI agents
17:55
anything to keep an eye on its AI agents also and dng uh from Deep learning
17:58
also and dng uh from Deep learning
17:58
also and dng uh from Deep learning incredible person AI gentic workflows
18:01
incredible person AI gentic workflows
18:01
incredible person AI gentic workflows will drive massive AI progresses year
18:04
will drive massive AI progresses year
18:04
will drive massive AI progresses year perhaps even more than the next
18:06
perhaps even more than the next
18:06
perhaps even more than the next generation of foundational
18:08
generation of foundational
18:08
generation of foundational models and so on and ACC Center and
18:12
models and so on and ACC Center and
18:12
models and so on and ACC Center and McKinzie and also Bill Gates is saying
18:15
McKinzie and also Bill Gates is saying
18:15
McKinzie and also Bill Gates is saying that oh and I could feel three or four
18:19
that oh and I could feel three or four
18:19
that oh and I could feel three or four more pages which quotes from people that
18:23
more pages which quotes from people that
18:23
more pages which quotes from people that apparently know what they are talking
18:25
apparently know what they are talking
18:25
apparently know what they are talking about not like me I know two or three
18:28
about not like me I know two or three
18:28
about not like me I know two or three things
18:29
things
18:29
things but I want to show you this picture atic
18:33
but I want to show you this picture atic
18:33
but I want to show you this picture atic workflows are already providing huge
18:36
workflows are already providing huge
18:36
workflows are already providing huge benefits um on some T for example this
18:40
benefits um on some T for example this
18:40
benefits um on some T for example this is uh it's a comparison between GPT 3.5
18:44
is uh it's a comparison between GPT 3.5
18:44
is uh it's a comparison between GPT 3.5 and GPT 4 doing zero shot and using
18:47
and GPT 4 doing zero shot and using
18:47
and GPT 4 doing zero shot and using agent workflows doing this human evil
18:50
agent workflows doing this human evil
18:50
agent workflows doing this human evil human evil is coding Benchmark GPT
18:54
human evil is coding Benchmark GPT
18:54
human evil is coding Benchmark GPT 3.5 has 48% accuracy zero shot GPT for
18:59
3.5 has 48% accuracy zero shot GPT for
18:59
3.5 has 48% accuracy zero shot GPT for it has quite a lot more
19:01
it has quite a lot more
19:01
it has quite a lot more 67% but if we apply a gentic AI a simple
19:08
67% but if we apply a gentic AI a simple
19:08
67% but if we apply a gentic AI a simple critic workflow it gets up to
19:12
critic workflow it gets up to
19:12
critic workflow it gets up to 96% of course we do that with gp4 it
19:15
96% of course we do that with gp4 it
19:15
96% of course we do that with gp4 it gets a bit more like 97 98% more or less
19:20
gets a bit more like 97 98% more or less
19:20
gets a bit more like 97 98% more or less so you can see the amount of improvement
19:22
so you can see the amount of improvement
19:22
so you can see the amount of improvement that this is
19:23
that this is
19:23
that this is bringing is pretty
19:27
bringing is pretty
19:27
bringing is pretty impressive so and dng and I took this
19:30
impressive so and dng and I took this
19:30
impressive so and dng and I took this slide from
19:32
slide from
19:32
slide from him
19:34
him
19:34
him already other thing that I believe is
19:38
already other thing that I believe is
19:38
already other thing that I believe is current Trend and it start to appear it
19:41
current Trend and it start to appear it
19:41
current Trend and it start to appear it was just launched uh I think one or two
19:44
was just launched uh I think one or two
19:44
was just launched uh I think one or two weeks ago from Lang WP which launch I
19:46
weeks ago from Lang WP which launch I
19:46
weeks ago from Lang WP which launch I think um it's graph driven agent so
19:50
think um it's graph driven agent so
19:50
think um it's graph driven agent so representing agent and agent workflows
19:52
representing agent and agent workflows
19:52
representing agent and agent workflows like a graph you know a node and a
19:54
like a graph you know a node and a
19:54
like a graph you know a node and a vector that determines a direction and
19:57
vector that determines a direction and
19:57
vector that determines a direction and uh this is the way AR tectonically to
19:59
uh this is the way AR tectonically to
19:59
uh this is the way AR tectonically to Define properly um agentic
20:03
Define properly um agentic
20:03
Define properly um agentic workflow which involves uh two or more
20:06
workflow which involves uh two or more
20:06
workflow which involves uh two or more agents using graphic structures with
20:08
agents using graphic structures with
20:08
agents using graphic structures with nodes and edges the interactions between
20:10
nodes and edges the interactions between
20:10
nodes and edges the interactions between agents can be perfectly ma uh and pro
20:13
agents can be perfectly ma uh and pro
20:13
agents can be perfectly ma uh and pro providing some benefits like cycles
20:15
providing some benefits like cycles
20:15
providing some benefits like cycles controllability and persistence which
20:17
controllability and persistence which
20:17
controllability and persistence which enable Loops conditionals and also keep
20:20
enable Loops conditionals and also keep
20:20
enable Loops conditionals and also keep a state after each step which is good if
20:22
a state after each step which is good if
20:22
a state after each step which is good if we want to know what happened and why
20:24
we want to know what happened and why
20:24
we want to know what happened and why something happened right uh for business
20:27
something happened right uh for business
20:27
something happened right uh for business perspective we will like to see why this
20:29
perspective we will like to see why this
20:29
perspective we will like to see why this fail and why did it succeed so we can
20:32
fail and why did it succeed so we can
20:32
fail and why did it succeed so we can back do backtracking and control how
20:34
back do backtracking and control how
20:35
back do backtracking and control how things are
20:36
things are
20:36
things are working which is pretty
20:39
working which is pretty
20:39
working which is pretty cool also we have graph rack to solve
20:43
cool also we have graph rack to solve
20:43
cool also we have graph rack to solve these
20:44
these
20:44
these hallucination process of retrieving the
20:46
hallucination process of retrieving the
20:46
hallucination process of retrieving the wrong chunks um a graph approach to do
20:49
wrong chunks um a graph approach to do
20:49
wrong chunks um a graph approach to do rack retrial aent generation is an
20:51
rack retrial aent generation is an
20:51
rack retrial aent generation is an Innovative way to combin the strength of
20:53
Innovative way to combin the strength of
20:53
Innovative way to combin the strength of graph structures uh creating something
20:56
graph structures uh creating something
20:56
graph structures uh creating something called knowledge graphs um basically it
21:01
called knowledge graphs um basically it
21:01
called knowledge graphs um basically it has um the the
21:04
has um the the
21:04
has um the the previous implementation of vector rack
21:07
previous implementation of vector rack
21:07
previous implementation of vector rack databases uh has limited contextual
21:09
databases uh has limited contextual
21:10
databases uh has limited contextual understanding lacks scalability and is
21:11
understanding lacks scalability and is
21:12
understanding lacks scalability and is very limiting in real scenarios cack
21:16
very limiting in real scenarios cack
21:16
very limiting in real scenarios cack does its magic by creating relations
21:18
does its magic by creating relations
21:18
does its magic by creating relations between the data chunks creating a
21:20
between the data chunks creating a
21:20
between the data chunks creating a knowledge graph that interconnects them
21:23
knowledge graph that interconnects them
21:23
knowledge graph that interconnects them uh indexes them and organizes the data
21:26
uh indexes them and organizes the data
21:26
uh indexes them and organizes the data hierarchically in semantic clusters and
21:28
hierarchically in semantic clusters and
21:28
hierarchically in semantic clusters and searchable and connecting basically
21:31
searchable and connecting basically
21:31
searchable and connecting basically everything of course this is
21:33
everything of course this is
21:33
everything of course this is computationally very expensive but also
21:35
computationally very expensive but also
21:35
computationally very expensive but also very successful like day and night in ra
21:38
very successful like day and night in ra
21:38
very successful like day and night in ra and seems to be having a lot of
21:44
success another one is distributed
21:47
success another one is distributed
21:47
success another one is distributed agentic AI so taking AI agents and how
21:51
agentic AI so taking AI agents and how
21:52
agentic AI so taking AI agents and how thinking how will they run at
21:54
thinking how will they run at
21:54
thinking how will they run at scale um how can we enable efficient
21:58
scale um how can we enable efficient
21:58
scale um how can we enable efficient collaboration collation across multiple
22:00
collaboration collation across multiple
22:00
collaboration collation across multiple locations maybe in the cloud maybe in
22:02
locations maybe in the cloud maybe in
22:02
locations maybe in the cloud maybe in local and how can we achieve this uh
22:06
local and how can we achieve this uh
22:06
local and how can we achieve this uh desire having F tolerance and revness uh
22:09
desire having F tolerance and revness uh
22:09
desire having F tolerance and revness uh sounds familiar right I'm talking like a
22:11
sounds familiar right I'm talking like a
22:11
sounds familiar right I'm talking like a distributed application in the web so
22:13
distributed application in the web so
22:13
distributed application in the web so the same is for agentic AI this setup
22:17
the same is for agentic AI this setup
22:17
the same is for agentic AI this setup allows for real-time collaboration as
22:19
allows for real-time collaboration as
22:19
allows for real-time collaboration as well making ideal for applications that
22:22
well making ideal for applications that
22:22
well making ideal for applications that require continuous distributed and
22:24
require continuous distributed and
22:24
require continuous distributed and parallelized processing such as smart
22:26
parallelized processing such as smart
22:26
parallelized processing such as smart grids and Global monitoring system just
22:29
grids and Global monitoring system just
22:29
grids and Global monitoring system just to put an example the site attributes
22:31
to put an example the site attributes
22:31
to put an example the site attributes for this approach include being
22:33
for this approach include being
22:33
for this approach include being distributed scalable and both cloud and
22:35
distributed scalable and both cloud and
22:35
distributed scalable and both cloud and local radio even the the first
22:38
local radio even the the first
22:38
local radio even the the first attributes mostly appear in the cloud
22:40
attributes mostly appear in the cloud
22:40
attributes mostly appear in the cloud key technologies that I see that will
22:42
key technologies that I see that will
22:42
key technologies that I see that will support this idea and these are
22:44
support this idea and these are
22:44
support this idea and these are important is actor pattern kind of for
22:47
important is actor pattern kind of for
22:47
important is actor pattern kind of for example I would say Orleans I mean I
22:49
example I would say Orleans I mean I
22:49
example I would say Orleans I mean I know a lot of bit Asher and I know it's
22:53
know a lot of bit Asher and I know it's
22:53
know a lot of bit Asher and I know it's well used it's one of the best actor
22:55
well used it's one of the best actor
22:55
well used it's one of the best actor patterns I would say this is
22:59
patterns I would say this is
22:59
patterns I would say this is want to properly consider and then
23:02
want to properly consider and then
23:02
want to properly consider and then approach for graph and microservice
23:05
approach for graph and microservice
23:05
approach for graph and microservice architecture but you have to see how
23:07
architecture but you have to see how
23:07
architecture but you have to see how this evolves and maybe explore a bit L
23:13
graph so scalable efficient
23:15
graph so scalable efficient
23:15
graph so scalable efficient collaboration cloud ready resilient
23:17
collaboration cloud ready resilient
23:17
collaboration cloud ready resilient actor pattern based and kind of akin to
23:21
actor pattern based and kind of akin to
23:21
actor pattern based and kind of akin to microservices that's what I would like
23:23
microservices that's what I would like
23:23
microservices that's what I would like and of course more important for a
23:26
and of course more important for a
23:26
and of course more important for a company resilient
23:31
and that's it we are going to kind of uh
23:35
and that's it we are going to kind of uh
23:35
and that's it we are going to kind of uh in almost last part which is AI
23:39
in almost last part which is AI
23:39
in almost last part which is AI architecture how can we with all we have
23:42
architecture how can we with all we have
23:42
architecture how can we with all we have seen the current AI
23:44
seen the current AI
23:44
seen the current AI components and the future ones how can
23:46
components and the future ones how can
23:46
components and the future ones how can we map them into AI
23:52
architecture so first I want to start by
23:56
architecture so first I want to start by
23:56
architecture so first I want to start by understanding the role of AI in modern
23:59
understanding the role of AI in modern
23:59
understanding the role of AI in modern architectures my point of view is that
24:01
architectures my point of view is that
24:01
architectures my point of view is that just as databases and API systems were
24:04
just as databases and API systems were
24:04
just as databases and API systems were revolutionary in their time AI is
24:06
revolutionary in their time AI is
24:06
revolutionary in their time AI is emerging as a crucial component in today
24:09
emerging as a crucial component in today
24:09
emerging as a crucial component in today uh in today's technological landscape
24:12
uh in today's technological landscape
24:12
uh in today's technological landscape and these AI components such as models
24:15
and these AI components such as models
24:15
and these AI components such as models as a service and AI orchestrators are
24:18
as a service and AI orchestrators are
24:18
as a service and AI orchestrators are seamlessly integrating into existing
24:20
seamlessly integrating into existing
24:20
seamlessly integrating into existing architectures enhancing functionality
24:22
architectures enhancing functionality
24:22
architectures enhancing functionality and enabling new possibilities like
24:24
and enabling new possibilities like
24:25
and enabling new possibilities like literally connecting a brain that does a
24:27
literally connecting a brain that does a
24:27
literally connecting a brain that does a concrete thing to your application
24:29
concrete thing to your application
24:29
concrete thing to your application something you can do since last
24:32
something you can do since last
24:32
something you can do since last year I done that so so can you I put
24:36
year I done that so so can you I put
24:36
year I done that so so can you I put here some links if you curious
24:39
here some links if you curious
24:39
here some links if you curious that's not it so one of the key
24:42
that's not it so one of the key
24:42
that's not it so one of the key components in mod AI architecture is
24:45
components in mod AI architecture is
24:45
components in mod AI architecture is model as a service and to my opinion
24:49
model as a service and to my opinion
24:49
model as a service and to my opinion this allows AI model to be access as a
24:51
this allows AI model to be access as a
24:51
this allows AI model to be access as a service when you connect to ashure AI
24:53
service when you connect to ashure AI
24:53
service when you connect to ashure AI service you connect to Google AWS or
24:56
service you connect to Google AWS or
24:56
service you connect to Google AWS or directly to open AI you are connecting
24:59
directly to open AI you are connecting
24:59
directly to open AI you are connecting to a service in the cloud right this
25:02
to a service in the cloud right this
25:02
to a service in the cloud right this approach to me is not only enhan
25:04
approach to me is not only enhan
25:04
approach to me is not only enhan scalability makes it easier to update
25:06
scalability makes it easier to update
25:06
scalability makes it easier to update and manage um in a in in in a way that
25:10
and manage um in a in in in a way that
25:10
and manage um in a in in in a way that they remain effective and relevant and I
25:13
they remain effective and relevant and I
25:13
they remain effective and relevant and I see that in a c4 syntax which I love and
25:16
see that in a c4 syntax which I love and
25:16
see that in a c4 syntax which I love and I hope you are familiar with it as a
25:19
I hope you are familiar with it as a
25:19
I hope you are familiar with it as a another simple component right to me
25:22
another simple component right to me
25:22
another simple component right to me this is nothing more than external
25:24
this is nothing more than external
25:24
this is nothing more than external Software System if we are seeing here U
25:27
Software System if we are seeing here U
25:27
Software System if we are seeing here U one uh system
25:36
diagram then we have uh the component I
25:40
diagram then we have uh the component I
25:40
diagram then we have uh the component I call it AI orchestrator for me that
25:42
call it AI orchestrator for me that
25:42
call it AI orchestrator for me that would be either L chain either sematic
25:44
would be either L chain either sematic
25:44
would be either L chain either sematic kernel or anything that is managing an
25:47
kernel or anything that is managing an
25:47
kernel or anything that is managing an AI to get something done right and you
25:50
AI to get something done right and you
25:50
AI to get something done right and you put that as a component somewhere in
25:52
put that as a component somewhere in
25:52
put that as a component somewhere in your architecture to me that's an A
25:55
your architecture to me that's an A
25:55
your architecture to me that's an A orchestrator and this plays uh Cal role
25:58
orchestrator and this plays uh Cal role
25:58
orchestrator and this plays uh Cal role in managing and coordinating the
26:00
in managing and coordinating the
26:00
in managing and coordinating the different AI components within an
26:02
different AI components within an
26:02
different AI components within an architecture so to me is if you see a
26:05
architecture so to me is if you see a
26:05
architecture so to me is if you see a component diagram um I kind of put it
26:09
component diagram um I kind of put it
26:09
component diagram um I kind of put it here in this case component AI
26:11
here in this case component AI
26:11
here in this case component AI orchestrator so that's kind of my
26:13
orchestrator so that's kind of my
26:13
orchestrator so that's kind of my Approach when I do that in C4 syntax
26:16
Approach when I do that in C4 syntax
26:16
Approach when I do that in C4 syntax which I particularly love and they
26:18
which I particularly love and they
26:18
which I particularly love and they ensure that the different Services work
26:21
ensure that the different Services work
26:21
ensure that the different Services work together efficiently optimizing
26:23
together efficiently optimizing
26:23
together efficiently optimizing performance and integration and act as
26:25
performance and integration and act as
26:25
performance and integration and act as the central nervous system of an AA
26:27
the central nervous system of an AA
26:27
the central nervous system of an AA driven architecture managing the
26:28
driven architecture managing the
26:28
driven architecture managing the workflows uh delegating the work to in
26:31
workflows uh delegating the work to in
26:31
workflows uh delegating the work to in this case to the AI service which can be
26:34
this case to the AI service which can be
26:34
this case to the AI service which can be in the cloud or can be deployed in a
26:37
in the cloud or can be deployed in a
26:37
in the cloud or can be deployed in a kubernetes cluster or wherever you want
26:39
kubernetes cluster or wherever you want
26:39
kubernetes cluster or wherever you want it to be depending on your data
26:43
it to be depending on your data
26:43
it to be depending on your data policies here we can observe this
26:46
policies here we can observe this
26:46
policies here we can observe this internal component also the component
26:48
internal component also the component
26:48
internal component also the component that we saw
26:49
that we saw
26:49
that we saw before very simple right and I don't
26:52
before very simple right and I don't
26:52
before very simple right and I don't think it takes more and then uh there
26:55
think it takes more and then uh there
26:55
think it takes more and then uh there are two questions
26:57
are two questions
26:57
are two questions that I got from a good friend how do you
27:00
that I got from a good friend how do you
27:00
that I got from a good friend how do you test a non-deterministic system such as
27:02
test a non-deterministic system such as
27:02
test a non-deterministic system such as AI hm well to me it's clear testing AI
27:06
AI hm well to me it's clear testing AI
27:06
AI hm well to me it's clear testing AI system is a critical aspect of ensuring
27:08
system is a critical aspect of ensuring
27:09
system is a critical aspect of ensuring the reliability so if you do AI you
27:10
the reliability so if you do AI you
27:10
the reliability so if you do AI you should test it right um and but how
27:13
should test it right um and but how
27:13
should test it right um and but how given their non deterministic nure
27:15
given their non deterministic nure
27:15
given their non deterministic nure traditional testing will fall short it
27:19
traditional testing will fall short it
27:19
traditional testing will fall short it requires specialized testing Frameworks
27:22
requires specialized testing Frameworks
27:22
requires specialized testing Frameworks and this must account for the
27:23
and this must account for the
27:23
and this must account for the variability in the AA outputs so
27:26
variability in the AA outputs so
27:26
variability in the AA outputs so ensuring that the systems perform
27:28
ensuring that the systems perform
27:28
ensuring that the systems perform consistently and accurately under
27:30
consistently and accurately under
27:30
consistently and accurately under different conditions essentially simply
27:33
different conditions essentially simply
27:33
different conditions essentially simply say you test AI with AI to assess these
27:36
say you test AI with AI to assess these
27:36
say you test AI with AI to assess these variability and understand the outcome
27:38
variability and understand the outcome
27:38
variability and understand the outcome and if it matches expected one within a
27:41
and if it matches expected one within a
27:41
and if it matches expected one within a range of course so nothing strange
27:45
range of course so nothing strange
27:45
range of course so nothing strange here and then
27:47
here and then
27:47
here and then security which is quite key security is
27:51
security which is quite key security is
27:51
security which is quite key security is Paramount to AI system to ensure
27:53
Paramount to AI system to ensure
27:53
Paramount to AI system to ensure alignment and proper behavior techniques
27:56
alignment and proper behavior techniques
27:56
alignment and proper behavior techniques like meta prompting which I mentioned
27:59
like meta prompting which I mentioned
27:59
like meta prompting which I mentioned this is a prompt engineering technique
28:01
this is a prompt engineering technique
28:01
this is a prompt engineering technique to structure system prompts to help
28:04
to structure system prompts to help
28:04
to structure system prompts to help guide AI responses reducing the risk of
28:07
guide AI responses reducing the risk of
28:07
guide AI responses reducing the risk of inappropriate or misleading outputs and
28:09
inappropriate or misleading outputs and
28:09
inappropriate or misleading outputs and of course avoiding potential
28:11
of course avoiding potential
28:11
of course avoiding potential manipulations kind of kind of a
28:13
manipulations kind of kind of a
28:14
manipulations kind of kind of a jailbreak for example additionally
28:16
jailbreak for example additionally
28:16
jailbreak for example additionally specialized security solution are
28:18
specialized security solution are
28:18
specialized security solution are essential to protect against issues kind
28:21
essential to protect against issues kind
28:21
essential to protect against issues kind of hallucinations and jailbreaks I I
28:24
of hallucinations and jailbreaks I I
28:24
of hallucinations and jailbreaks I I think a few months ago
28:26
think a few months ago
28:26
think a few months ago Airline um went uh um had to give a
28:32
Airline um went uh um had to give a
28:32
Airline um went uh um had to give a incredible offer that their chatbot
28:35
incredible offer that their chatbot
28:35
incredible offer that their chatbot proposed to a customer
28:37
proposed to a customer
28:37
proposed to a customer because the the B seem to have legal uh
28:42
because the the B seem to have legal uh
28:42
because the the B seem to have legal uh uh commitment so he say yeah I give you
28:44
uh commitment so he say yeah I give you
28:44
uh commitment so he say yeah I give you that for $5 and he got like kind of a
28:48
that for $5 and he got like kind of a
28:48
that for $5 and he got like kind of a incredible flight to somewhere and they
28:50
incredible flight to somewhere and they
28:50
incredible flight to somewhere and they had to comply so I would be careful
28:54
had to comply so I would be careful
28:54
had to comply so I would be careful because if somebody drives your chat bot
28:58
because if somebody drives your chat bot
28:58
because if somebody drives your chat bot into selling one of your cars if you're
29:00
into selling one of your cars if you're
29:00
into selling one of your cars if you're selling cars for
29:02
selling cars for
29:02
selling cars for $10 maybe it's a good idea to put a meta
29:05
$10 maybe it's a good idea to put a meta
29:05
$10 maybe it's a good idea to put a meta prom there right and secure that and put
29:08
prom there right and secure that and put
29:08
prom there right and secure that and put some G rails and avoid some
29:09
some G rails and avoid some
29:10
some G rails and avoid some manipulations of evil customers
29:13
manipulations of evil customers
29:13
manipulations of evil customers customers that want to have fun this
29:15
customers that want to have fun this
29:15
customers that want to have fun this another approach for meta prompting
29:17
another approach for meta prompting
29:17
another approach for meta prompting aside from keep putting a strong um meta
29:22
aside from keep putting a strong um meta
29:22
aside from keep putting a strong um meta prom which is a c prom with uh
29:25
prom which is a c prom with uh
29:25
prom which is a c prom with uh additional
29:26
additional
29:26
additional layers is to put um a modifiable agents
29:31
layers is to put um a modifiable agents
29:31
layers is to put um a modifiable agents that do a supervision on the input and
29:34
that do a supervision on the input and
29:34
that do a supervision on the input and on the output so these agents cannot be
29:36
on the output so these agents cannot be
29:36
on the output so these agents cannot be manipulated cannot be modified and only
29:39
manipulated cannot be modified and only
29:39
manipulated cannot be modified and only have to access this uh input and the
29:43
have to access this uh input and the
29:43
have to access this uh input and the output so to ensure that they are safe
29:47
output so to ensure that they are safe
29:47
output so to ensure that they are safe and this input and output verification
29:49
and this input and output verification
29:49
and this input and output verification adds an extra layer security and of
29:51
adds an extra layer security and of
29:51
adds an extra layer security and of course complexity and processing but I
29:54
course complexity and processing but I
29:54
course complexity and processing but I think it's better that you're safer
29:56
think it's better that you're safer
29:56
think it's better that you're safer right I think I
30:00
right I think I
30:00
right I think I put an example here and think I put a
30:04
put an example here and think I put a
30:04
put an example here and think I put a link
30:05
link
30:05
link no uh if you look for Microsoft meta
30:09
no uh if you look for Microsoft meta
30:09
no uh if you look for Microsoft meta prompt template you will you will find
30:12
prompt template you will you will find
30:12
prompt template you will you will find it definitely
30:15
it definitely
30:15
it definitely already and
30:17
already and
30:17
already and then to finish it a bit let's take a
30:21
then to finish it a bit let's take a
30:21
then to finish it a bit let's take a look uh a bit at a couple of more things
30:25
look uh a bit at a couple of more things
30:25
look uh a bit at a couple of more things uh I would like to share a bit my b on
30:29
uh I would like to share a bit my b on
30:29
uh I would like to share a bit my b on AI agents on this
30:31
AI agents on this
30:31
AI agents on this case we will soon be able to instantiate
30:34
case we will soon be able to instantiate
30:34
case we will soon be able to instantiate teams of AI agents to perform advanced
30:36
teams of AI agents to perform advanced
30:36
teams of AI agents to perform advanced task at the scale of course technology
30:39
task at the scale of course technology
30:39
task at the scale of course technology has to evolve so it's already doing so
30:42
has to evolve so it's already doing so
30:42
has to evolve so it's already doing so with langra probably it will soon evolve
30:46
with langra probably it will soon evolve
30:46
with langra probably it will soon evolve in
30:48
in
30:48
in um semantic Kel so yesterday in in the
30:51
um semantic Kel so yesterday in in the
30:51
um semantic Kel so yesterday in in the office hours it's if you're curious
30:53
office hours it's if you're curious
30:53
office hours it's if you're curious probably it's already available the team
30:56
probably it's already available the team
30:56
probably it's already available the team announced a road map for agent and
30:58
announced a road map for agent and
30:58
announced a road map for agent and eating bols grabs as well and also they
31:01
eating bols grabs as well and also they
31:02
eating bols grabs as well and also they are adopting the cre AI way which is
31:05
are adopting the cre AI way which is
31:05
are adopting the cre AI way which is Task driven I think it's uh not that
31:10
Task driven I think it's uh not that
31:10
Task driven I think it's uh not that nice but very efficient for simple tasks
31:12
nice but very efficient for simple tasks
31:12
nice but very efficient for simple tasks but if they manage to combine both of
31:14
but if they manage to combine both of
31:14
but if they manage to combine both of them it will be good but I think it's
31:16
them it will be good but I think it's
31:16
them it will be good but I think it's pretty pretty pretty interesting to to
31:19
pretty pretty pretty interesting to to
31:19
pretty pretty pretty interesting to to check that out but it's going in a nice
31:23
check that out but it's going in a nice
31:23
check that out but it's going in a nice Direction and there are a lot of changes
31:25
Direction and there are a lot of changes
31:25
Direction and there are a lot of changes they aiming for the 9 of night timeline
31:29
they aiming for the 9 of night timeline
31:29
they aiming for the 9 of night timeline which is
31:30
which is
31:30
which is amazing but what I'm talking about is
31:33
amazing but what I'm talking about is
31:33
amazing but what I'm talking about is not two or three agents or five or six
31:35
not two or three agents or five or six
31:35
not two or three agents or five or six or 10 which we can have now but having
31:38
or 10 which we can have now but having
31:38
or 10 which we can have now but having hundreds collaborating seamlessly and
31:40
hundreds collaborating seamlessly and
31:40
hundreds collaborating seamlessly and this capability will revolutionize uh
31:44
this capability will revolutionize uh
31:44
this capability will revolutionize uh how we approach complex problems and
31:46
how we approach complex problems and
31:46
how we approach complex problems and will be a game changer a disruptor for
31:49
will be a game changer a disruptor for
31:49
will be a game changer a disruptor for in many Industries if not
31:56
all here is my call to action to embrace
31:59
all here is my call to action to embrace
31:59
all here is my call to action to embrace and Implement
32:01
and Implement
32:01
and Implement AI a call to you is to embrace this
32:03
AI a call to you is to embrace this
32:03
AI a call to you is to embrace this feature large language models will
32:05
feature large language models will
32:05
feature large language models will continue to improve significantly in
32:07
continue to improve significantly in
32:08
continue to improve significantly in collaboration with agentic AI which you
32:11
collaboration with agentic AI which you
32:11
collaboration with agentic AI which you seen what's the improvement over uh what
32:13
seen what's the improvement over uh what
32:13
seen what's the improvement over uh what a model can do right from 48 to
32:18
a model can do right from 48 to
32:18
a model can do right from 48 to 94% in accuracy in one of the hardest
32:22
94% in accuracy in one of the hardest
32:22
94% in accuracy in one of the hardest coding uh benchmarks human evil that we
32:25
coding uh benchmarks human evil that we
32:25
coding uh benchmarks human evil that we saw at the beginning um how about
32:28
saw at the beginning um how about
32:28
saw at the beginning um how about ever um this means that in collaboration
32:32
ever um this means that in collaboration
32:32
ever um this means that in collaboration with atic AI they will be able to
32:33
with atic AI they will be able to
32:34
with atic AI they will be able to achieve amazing things autonomously so
32:37
achieve amazing things autonomously so
32:37
achieve amazing things autonomously so chances are that we will put a PBI and
32:40
chances are that we will put a PBI and
32:40
chances are that we will put a PBI and it will be done by itself and it will
32:43
it will be done by itself and it will
32:43
it will be done by itself and it will require very small or not too much
32:47
require very small or not too much
32:47
require very small or not too much supervision but however someone will
32:50
supervision but however someone will
32:50
supervision but however someone will need to design and Implement these
32:53
need to design and Implement these
32:54
need to design and Implement these agents along with the car rails and
32:56
agents along with the car rails and
32:56
agents along with the car rails and quality controls and fine tune them
32:58
quality controls and fine tune them
32:58
quality controls and fine tune them regularly and to ensure they do what
33:01
regularly and to ensure they do what
33:01
regularly and to ensure they do what they are supposed to do by nothing evil
33:04
they are supposed to do by nothing evil
33:04
they are supposed to do by nothing evil and do not devate from that this is
33:06
and do not devate from that this is
33:06
and do not devate from that this is where you as a generative AI expert or
33:09
where you as a generative AI expert or
33:09
where you as a generative AI expert or developer or generative AI architect um
33:13
developer or generative AI architect um
33:13
developer or generative AI architect um may come
33:15
may come
33:15
may come in we need to
33:18
in we need to
33:18
in we need to learn how to set up and manage those
33:21
learn how to set up and manage those
33:21
learn how to set up and manage those teams of AI agents and all the relevant
33:24
teams of AI agents and all the relevant
33:24
teams of AI agents and all the relevant components and skills you need to learn
33:28
components and skills you need to learn
33:28
components and skills you need to learn AI how to manage and orchestrate AI
33:30
AI how to manage and orchestrate AI
33:30
AI how to manage and orchestrate AI agents and teams of agents and do so as
33:34
agents and teams of agents and do so as
33:34
agents and teams of agents and do so as a good team Le does it's not just
33:36
a good team Le does it's not just
33:36
a good team Le does it's not just development it's just you have to think
33:38
development it's just you have to think
33:38
development it's just you have to think as a manager when you do that it's kind
33:40
as a manager when you do that it's kind
33:40
as a manager when you do that it's kind of a uh it blows your mind things you
33:44
of a uh it blows your mind things you
33:44
of a uh it blows your mind things you have to have in in at the same time but
33:48
have to have in in at the same time but
33:48
have to have in in at the same time but it's also very exciting when you see
33:50
it's also very exciting when you see
33:50
it's also very exciting when you see that they you do what they are expected
33:52
that they you do what they are expected
33:52
that they you do what they are expected you have to become common thing like a
33:54
you have to become common thing like a
33:54
you have to become common thing like a manager plan as an architect and
33:56
manager plan as an architect and
33:56
manager plan as an architect and understand each component in detail
33:58
understand each component in detail
33:58
understand each component in detail and more fun its personality as well
34:01
and more fun its personality as well
34:01
and more fun its personality as well people management skills alone are not
34:03
people management skills alone are not
34:03
people management skills alone are not enough you need Cloud knowledge AI
34:05
enough you need Cloud knowledge AI
34:05
enough you need Cloud knowledge AI component knowledge architectural skills
34:07
component knowledge architectural skills
34:07
component knowledge architectural skills from engineering agentic Ai and coding
34:10
from engineering agentic Ai and coding
34:10
from engineering agentic Ai and coding knowledge if you fully want to
34:13
knowledge if you fully want to
34:13
knowledge if you fully want to understand what's going on maybe it's a
34:14
understand what's going on maybe it's a
34:14
understand what's going on maybe it's a bit overwhelming but uh you just have to
34:17
bit overwhelming but uh you just have to
34:17
bit overwhelming but uh you just have to start and everything will come to you
34:20
start and everything will come to you
34:20
start and everything will come to you now is the moment as fast as this field
34:23
now is the moment as fast as this field
34:23
now is the moment as fast as this field is evolving if you haven't started
34:25
is evolving if you haven't started
34:25
is evolving if you haven't started already it may already bit too late so
34:29
already it may already bit too late so
34:29
already it may already bit too late so think
34:30
think
34:30
think seriously about starting this journey as
34:33
seriously about starting this journey as
34:33
seriously about starting this journey as soon as possible chanes are that in two
34:36
soon as possible chanes are that in two
34:36
soon as possible chanes are that in two three years somebody with the skills I
34:38
three years somebody with the skills I
34:39
three years somebody with the skills I just mentioned will be able to do what
34:41
just mentioned will be able to do what
34:41
just mentioned will be able to do what your complete team is doing if not more
34:44
your complete team is doing if not more
34:44
your complete team is doing if not more and
34:45
and
34:45
and that's kind of a disruptor
34:49
that's kind of a disruptor
34:49
that's kind of a disruptor right and that's say it I want to thank
34:53
right and that's say it I want to thank
34:53
right and that's say it I want to thank you for listening to me I hope this
34:55
you for listening to me I hope this
34:55
you for listening to me I hope this presentation has resonated with you and
34:57
presentation has resonated with you and
34:58
presentation has resonated with you and Spark some new ideas thank you so
35:03
Spark some new ideas thank you so
35:03
Spark some new ideas thank you so much if you find my thoughts and ideas
35:05
much if you find my thoughts and ideas
35:05
much if you find my thoughts and ideas interesting you can follow me on Twitter
35:07
interesting you can follow me on Twitter
35:07
interesting you can follow me on Twitter LinkedIn or even check my LinkedIn
35:10
LinkedIn or even check my LinkedIn
35:10
LinkedIn or even check my LinkedIn courses which I have this uh one uh on
35:16
courses which I have this uh one uh on
35:16
courses which I have this uh one uh on semantic kernel one about TV publish on
35:18
semantic kernel one about TV publish on
35:18
semantic kernel one about TV publish on Asher Ai and working on a new one on um
35:22
Asher Ai and working on a new one on um
35:22
Asher Ai and working on a new one on um AO chain you can also meet me at the
35:23
AO chain you can also meet me at the
35:24
AO chain you can also meet me at the events I organized such as donet surc
35:25
events I organized such as donet surc
35:25
events I organized such as donet surc and Global AI surc sometimes H in
35:28
and Global AI surc sometimes H in
35:28
and Global AI surc sometimes H in personal online and again it has been an
35:31
personal online and again it has been an
35:31
personal online and again it has been an absolute pleasure to be here with you
35:34
absolute pleasure to be here with you
35:34
absolute pleasure to be here with you thank you

AI Architecture: What Is Possible Today and Future Trends | Software Architecture Conference

CSharpCorner

Build an MCP Server from Your API Fast with Zuplo + Supabase

What is New in GPT 5

Top 10 Vibe Coding Tools to Try Now

What Developers Should Know about Observability | Software Architecture Conference

Keynote: AI and GenAI for the Busy Software Architect | Software Architecture Conference

Code Quality/Performance & Application Architecture - AMA

Taking code to next level with Open Source (Git & Github) || Code Quality & Performance Conference

Rockin' The Code World with dotNetDave ft. Mahesh Chand Ep 15

Navigating the Role of Software Architecture in the AI Era Amid Economic Volatility

Crafting Resilient Integration Solutions | Software Architecture Conference

KISS - Keep It Simple Stupid | Software Architecture Conference

Unleashing the power of Blazor: Future of web development

Up next in 10

AI Architecture: What Is Possible Today and Future Trends | Software Architecture Conference

CSharpCorner