Up next in 10
"As AI adoption increases rapidly and organizations clamor to take advantage of OpenAI, Copilot, and other powerful tools, many mistakes and oversights are made. These mistakes manifest themselves in security breaches, invalid results, and even offensive content. The immediate results are lost time, money, and embarrassment. What do these problems have in common? Bad data!
Quality AI solutions require clean, documented, and validated data. This session dives into data quality with a strong focus on how data is maintained as it moves from transactional to analytic workloads.
Some topics include:
• How transactional data be maintained effectively without compromising performance.
• The importance of documentation in ensuring data is used correctly.
• How to validate data as it is moved, transformed, and crunched.
• Security implications of handing off data to AI applications.
• Ensuring that a source-of-truth exists for each data source.
This is a fast-paced session that promises both helpful best practices and also some fun along the way."
🔗 Conference Website: https://softwarearchitecture.live
📺 CSharp TV - Dev Streaming Destination http://csharp.tv
🌎 C# Corner - Community of Software and Data Developers
Show More Show Less View Video Transcript
0:03
all right well thank you very much for
0:04
all right well thank you very much for
0:04
all right well thank you very much for having me this is exciting this is a
0:06
having me this is exciting this is a
0:06
having me this is exciting this is a topic that is both old and new uh data
0:09
topic that is both old and new uh data
0:09
topic that is both old and new uh data quality is nothing new we've been
0:10
quality is nothing new we've been
0:10
quality is nothing new we've been dealing with it for years and years and
0:11
dealing with it for years and years and
0:11
dealing with it for years and years and years even before databases existed
0:13
years even before databases existed
0:14
years even before databases existed people had filing cabinets full of paper
0:16
people had filing cabinets full of paper
0:16
people had filing cabinets full of paper and there were sure to be mistakes
0:17
and there were sure to be mistakes
0:17
and there were sure to be mistakes inside of them and so data quality is an
0:19
inside of them and so data quality is an
0:19
inside of them and so data quality is an eternal issue we've been dealing with
0:20
eternal issue we've been dealing with
0:20
eternal issue we've been dealing with forever it doesn't stop being an issue
0:22
forever it doesn't stop being an issue
0:22
forever it doesn't stop being an issue but the more Downstream things we do the
0:25
but the more Downstream things we do the
0:25
but the more Downstream things we do the more complex our usage that data gets
0:27
more complex our usage that data gets
0:27
more complex our usage that data gets the more ways we have to muck it up and
0:29
the more ways we have to muck it up and
0:29
the more ways we have to muck it up and make our lives very difficult so our
0:30
make our lives very difficult so our
0:30
make our lives very difficult so our goal here is to talk about data quality
0:33
goal here is to talk about data quality
0:33
goal here is to talk about data quality through the lens of AI that is we're
0:35
through the lens of AI that is we're
0:35
through the lens of AI that is we're going to throw complicated algorithms at
0:37
going to throw complicated algorithms at
0:37
going to throw complicated algorithms at our data what does that mean for data
0:39
our data what does that mean for data
0:39
our data what does that mean for data quality I'm G to move kind of quickly my
0:41
quality I'm G to move kind of quickly my
0:41
quality I'm G to move kind of quickly my goal is to save some time at the end for
0:43
goal is to save some time at the end for
0:43
goal is to save some time at the end for questions those questions are great um
0:45
questions those questions are great um
0:45
questions those questions are great um but if there aren't any we can also just
0:46
but if there aren't any we can also just
0:46
but if there aren't any we can also just chat and have fun as well so I'm just
0:48
chat and have fun as well so I'm just
0:48
chat and have fun as well so I'm just going to Dive Right on in very quickly
0:50
going to Dive Right on in very quickly
0:50
going to Dive Right on in very quickly I've already gotten the intro um this is
0:52
I've already gotten the intro um this is
0:52
I've already gotten the intro um this is just more stuff I will skip it because
0:54
just more stuff I will skip it because
0:54
just more stuff I will skip it because we already got a cool intro the agenda
0:56
we already got a cool intro the agenda
0:57
we already got a cool intro the agenda is straightforward uh really we're going
0:59
is straightforward uh really we're going
0:59
is straightforward uh really we're going to talk about data quality
1:00
to talk about data quality
1:00
to talk about data quality the classic problem how to solve it why
1:03
the classic problem how to solve it why
1:03
the classic problem how to solve it why it matters and then what are all the
1:05
it matters and then what are all the
1:05
it matters and then what are all the things we do in AI that make our lives
1:07
things we do in AI that make our lives
1:07
things we do in AI that make our lives more difficult because of this we'll
1:10
more difficult because of this we'll
1:10
more difficult because of this we'll bring it all together at the end and
1:12
bring it all together at the end and
1:12
bring it all together at the end and save a little bit of time for
1:15
save a little bit of time for
1:15
save a little bit of time for questions so this is an example of a
1:18
questions so this is an example of a
1:18
questions so this is an example of a software development life cycle there
1:20
software development life cycle there
1:20
software development life cycle there are many out there different
1:21
are many out there different
1:21
are many out there different organizations different sizes have
1:23
organizations different sizes have
1:23
organizations different sizes have different variations of this if you're a
1:25
different variations of this if you're a
1:25
different variations of this if you're a small organization you have a little
1:26
small organization you have a little
1:26
small organization you have a little less if you're big you may have more but
1:28
less if you're big you may have more but
1:28
less if you're big you may have more but the basic idea is there's design there's
1:30
the basic idea is there's design there's
1:30
the basic idea is there's design there's architecture you build some software you
1:33
architecture you build some software you
1:33
architecture you build some software you test a whole bunch iterate and
1:34
test a whole bunch iterate and
1:34
test a whole bunch iterate and eventually deploy maintain so on and so
1:37
eventually deploy maintain so on and so
1:37
eventually deploy maintain so on and so forth and there's a lot of arrows a lot
1:39
forth and there's a lot of arrows a lot
1:39
forth and there's a lot of arrows a lot of boxes and these are all here
1:41
of boxes and these are all here
1:42
of boxes and these are all here primarily so that we make good decisions
1:44
primarily so that we make good decisions
1:44
primarily so that we make good decisions we add quality new features make good
1:47
we add quality new features make good
1:47
we add quality new features make good decisions make few mistakes and then
1:49
decisions make few mistakes and then
1:49
decisions make few mistakes and then release something that users love right
1:51
release something that users love right
1:51
release something that users love right that's the whole point of software
1:53
that's the whole point of software
1:53
that's the whole point of software development but where that is kind of
1:56
development but where that is kind of
1:56
development but where that is kind of where a mature product would behave in
1:58
where a mature product would behave in
1:58
where a mature product would behave in terms of how we go from an aidea to code
2:01
terms of how we go from an aidea to code
2:01
terms of how we go from an aidea to code in a production environment what very
2:03
in a production environment what very
2:03
in a production environment what very often happens in AI is a little bit like
2:06
often happens in AI is a little bit like
2:06
often happens in AI is a little bit like this not exactly I'm exaggerating but
2:09
this not exactly I'm exaggerating but
2:09
this not exactly I'm exaggerating but it's funny right somebody comes and says
2:11
it's funny right somebody comes and says
2:11
it's funny right somebody comes and says we want the bright shiny new thing now
2:13
we want the bright shiny new thing now
2:13
we want the bright shiny new thing now and you say okay when do you want to buy
2:15
and you say okay when do you want to buy
2:15
and you say okay when do you want to buy and they say next week and you work
2:17
and they say next week and you work
2:17
and they say next week and you work really fast and a new feature goes out
2:18
really fast and a new feature goes out
2:18
really fast and a new feature goes out and it's cool because the new shiny
2:20
and it's cool because the new shiny
2:20
and it's cool because the new shiny thing's worth a lot right and so very
2:22
thing's worth a lot right and so very
2:22
thing's worth a lot right and so very often we skip steps we hurry we skip we
2:25
often we skip steps we hurry we skip we
2:26
often we skip steps we hurry we skip we do everything we can to get things out
2:28
do everything we can to get things out
2:28
do everything we can to get things out and that's dangerous right we're being a
2:30
and that's dangerous right we're being a
2:30
and that's dangerous right we're being a software design life cycle so part of
2:32
software design life cycle so part of
2:32
software design life cycle so part of this entire conversation is that we
2:34
this entire conversation is that we
2:34
this entire conversation is that we should treat anything we develop in AI
2:37
should treat anything we develop in AI
2:37
should treat anything we develop in AI to be very much like any other software
2:40
to be very much like any other software
2:40
to be very much like any other software development right it's just features in
2:42
development right it's just features in
2:42
development right it's just features in the same way that database development
2:44
the same way that database development
2:44
the same way that database development is just development software development
2:46
is just development software development
2:46
is just development software development is development AI development machine
2:48
is development AI development machine
2:48
is development AI development machine learning algorithms analytics
2:49
learning algorithms analytics
2:49
learning algorithms analytics dashboarding all those things that we do
2:51
dashboarding all those things that we do
2:51
dashboarding all those things that we do with our data later it's just
2:53
with our data later it's just
2:53
with our data later it's just development so treat it like development
2:55
development so treat it like development
2:55
development so treat it like development uh if you do what I show on the screen
2:56
uh if you do what I show on the screen
2:56
uh if you do what I show on the screen right now things may not work the way we
2:59
right now things may not work the way we
2:59
right now things may not work the way we want them to you'll get some surprises
3:01
want them to you'll get some surprises
3:01
want them to you'll get some surprises that you know surprise cake for your
3:02
that you know surprise cake for your
3:03
that you know surprise cake for your birthday good surprise bugs not so good
3:06
birthday good surprise bugs not so good
3:06
birthday good surprise bugs not so good right there's more challenges though
3:09
right there's more challenges though
3:09
right there's more challenges though what are the challenges we face with
3:10
what are the challenges we face with
3:10
what are the challenges we face with data and Ai and these are common
3:13
data and Ai and these are common
3:13
data and Ai and these are common challenges some of them for example data
3:15
challenges some of them for example data
3:15
challenges some of them for example data Grows Right data gets bigger it always
3:17
Grows Right data gets bigger it always
3:17
Grows Right data gets bigger it always gets bigger we have more databases more
3:19
gets bigger we have more databases more
3:19
gets bigger we have more databases more data May more data sources and it can be
3:22
data May more data sources and it can be
3:22
data May more data sources and it can be from anywhere right data begins
3:23
from anywhere right data begins
3:23
from anywhere right data begins somewhere it begins in a transactional
3:25
somewhere it begins in a transactional
3:25
somewhere it begins in a transactional database in Edge devices in some data
3:30
database in Edge devices in some data
3:30
database in Edge devices in some data Source element place somewhere where
3:32
Source element place somewhere where
3:32
Source element place somewhere where it's being created and then it exists
3:34
it's being created and then it exists
3:34
it's being created and then it exists it's used right it's used by
3:36
it's used right it's used by
3:36
it's used right it's used by applications used in uis used by users
3:39
applications used in uis used by users
3:39
applications used in uis used by users used by people but then more things
3:41
used by people but then more things
3:41
used by people but then more things happen to that data we begin getting
3:43
happen to that data we begin getting
3:43
happen to that data we begin getting bigger and asking more questions about
3:45
bigger and asking more questions about
3:45
bigger and asking more questions about our data and so we begin copying it we
3:47
our data and so we begin copying it we
3:47
our data and so we begin copying it we make replicas of it maybe we'll do some
3:49
make replicas of it maybe we'll do some
3:49
make replicas of it maybe we'll do some transformations of that data to make it
3:52
transformations of that data to make it
3:52
transformations of that data to make it easier to report on we want to provide
3:54
easier to report on we want to provide
3:54
easier to report on we want to provide dashboards and reports to users or to
3:56
dashboards and reports to users or to
3:56
dashboards and reports to users or to internal to an organization or to
3:59
internal to an organization or to
3:59
internal to an organization or to executive sort who knows who else
4:00
executive sort who knows who else
4:00
executive sort who knows who else shareholders there's a million places
4:02
shareholders there's a million places
4:02
shareholders there's a million places that data can be used and so after
4:04
that data can be used and so after
4:04
that data can be used and so after transforming it and playing with it a
4:06
transforming it and playing with it a
4:06
transforming it and playing with it a bit we may eventually put into other
4:08
bit we may eventually put into other
4:08
bit we may eventually put into other places uh into red shift or snowflake a
4:12
places uh into red shift or snowflake a
4:12
places uh into red shift or snowflake a data Lake a warehouse a lake house
4:13
data Lake a warehouse a lake house
4:13
data Lake a warehouse a lake house whatever data goes somewhere else and we
4:16
whatever data goes somewhere else and we
4:16
whatever data goes somewhere else and we work with it
4:17
work with it
4:17
work with it Downstream uh and so because data is
4:19
Downstream uh and so because data is
4:19
Downstream uh and so because data is growing and evolving over time and
4:21
growing and evolving over time and
4:21
growing and evolving over time and moving a lot we have a way of making
4:24
moving a lot we have a way of making
4:24
moving a lot we have a way of making mistakes along the way and those
4:26
mistakes along the way and those
4:26
mistakes along the way and those mistakes get copied well as soon as a
4:27
mistakes get copied well as soon as a
4:27
mistakes get copied well as soon as a mistake is made with data you have a
4:29
mistake is made with data you have a
4:29
mistake is made with data you have a data quality issue it just passes
4:31
data quality issue it just passes
4:31
data quality issue it just passes through this process whatever it looks
4:33
through this process whatever it looks
4:33
through this process whatever it looks like and continues its way to the end
4:35
like and continues its way to the end
4:35
like and continues its way to the end which isn't very good and you know
4:38
which isn't very good and you know
4:38
which isn't very good and you know downside to AI processes which is an
4:41
downside to AI processes which is an
4:41
downside to AI processes which is an additional is they like returning
4:43
additional is they like returning
4:43
additional is they like returning answers they don't give results they
4:45
answers they don't give results they
4:45
answers they don't give results they give answers we want our AI to tell
4:47
give answers we want our AI to tell
4:48
give answers we want our AI to tell people what they want to know like I
4:50
people what they want to know like I
4:50
people what they want to know like I want to know what a plane ticket costs
4:51
want to know what a plane ticket costs
4:51
want to know what a plane ticket costs give me the price I want to know what
4:54
give me the price I want to know what
4:54
give me the price I want to know what car you'd recommend for my lifestyle go
4:56
car you'd recommend for my lifestyle go
4:56
car you'd recommend for my lifestyle go and give me the right answer so they're
4:57
and give me the right answer so they're
4:57
and give me the right answer so they're very authoritative in nature so if you
4:59
very authoritative in nature so if you
5:00
very authoritative in nature so if you have bad data then you're going to
5:01
have bad data then you're going to
5:02
have bad data then you're going to authoritatively give bad answers which
5:04
authoritatively give bad answers which
5:04
authoritatively give bad answers which is a challenge so you want to be able
5:06
is a challenge so you want to be able
5:06
is a challenge so you want to be able authoritatively give good answers and
5:08
authoritatively give good answers and
5:08
authoritatively give good answers and data quality is the Lynch pin to all of
5:11
data quality is the Lynch pin to all of
5:11
data quality is the Lynch pin to all of this so many of this is going to be a
5:14
this so many of this is going to be a
5:14
this so many of this is going to be a classic answer how do we prevent bad
5:16
classic answer how do we prevent bad
5:16
classic answer how do we prevent bad data this is going to begin in the land
5:18
data this is going to begin in the land
5:18
data this is going to begin in the land of classic data problems and end in the
5:20
of classic data problems and end in the
5:20
of classic data problems and end in the land of AI we're going to start with
5:22
land of AI we're going to start with
5:22
land of AI we're going to start with things you may have heard before or
5:24
things you may have heard before or
5:24
things you may have heard before or maybe things you do already maybe things
5:26
maybe things you do already maybe things
5:26
maybe things you do already maybe things you don't do
5:27
you don't do
5:27
you don't do already the simplest way to prevent bad
5:30
already the simplest way to prevent bad
5:30
already the simplest way to prevent bad data is to prevent it at the source so
5:33
data is to prevent it at the source so
5:33
data is to prevent it at the source so you're at the beginning of the data life
5:34
you're at the beginning of the data life
5:34
you're at the beginning of the data life cycle you're in an application you're
5:36
cycle you're in an application you're
5:37
cycle you're in an application you're pulling in data from Ed devices people
5:39
pulling in data from Ed devices people
5:39
pulling in data from Ed devices people are processing orders or buying things
5:42
are processing orders or buying things
5:42
are processing orders or buying things or Bank transactions or whatever you
5:45
or Bank transactions or whatever you
5:45
or Bank transactions or whatever you could be anywhere data is being created
5:47
could be anywhere data is being created
5:47
could be anywhere data is being created for the first time it's being stored
5:49
for the first time it's being stored
5:49
for the first time it's being stored somewhere and we want to make sure it's
5:51
somewhere and we want to make sure it's
5:51
somewhere and we want to make sure it's correct this is where data begins so how
5:53
correct this is where data begins so how
5:53
correct this is where data begins so how do we do that and there's different ways
5:55
do we do that and there's different ways
5:55
do we do that and there's different ways to do it you can be strict about it and
5:58
to do it you can be strict about it and
5:58
to do it you can be strict about it and have things like check constraints data
6:00
have things like check constraints data
6:00
have things like check constraints data constraints foreign Keys uh database
6:02
constraints foreign Keys uh database
6:02
constraints foreign Keys uh database constructs that ensure that if you try
6:05
constructs that ensure that if you try
6:05
constructs that ensure that if you try to put bad data in there it will fail
6:07
to put bad data in there it will fail
6:07
to put bad data in there it will fail and those are good things to have if
6:09
and those are good things to have if
6:09
and those are good things to have if they're available to you because they
6:10
they're available to you because they
6:10
they're available to you because they ensure you can't have bad data if your
6:12
ensure you can't have bad data if your
6:12
ensure you can't have bad data if your database says you can't do this then you
6:15
database says you can't do this then you
6:15
database says you can't do this then you can't do this problem solved right we
6:18
can't do this problem solved right we
6:18
can't do this problem solved right we can't do that in every system but when
6:19
can't do that in every system but when
6:20
can't do that in every system but when we can it's nice uh some systems put
6:22
we can it's nice uh some systems put
6:22
we can it's nice uh some systems put data in out of order and do other things
6:24
data in out of order and do other things
6:24
data in out of order and do other things where you just can't do it or you're in
6:25
where you just can't do it or you're in
6:25
where you just can't do it or you're in a platform where those sorts of
6:27
a platform where those sorts of
6:27
a platform where those sorts of constraints don't make sense you can
6:29
constraints don't make sense you can
6:29
constraints don't make sense you can still validate data though the
6:31
still validate data though the
6:31
still validate data though the application can validate data the
6:33
application can validate data the
6:33
application can validate data the database could have a process to
6:34
database could have a process to
6:34
database could have a process to validate data or you can simply run
6:36
validate data or you can simply run
6:36
validate data or you can simply run reports after the fact or validation
6:38
reports after the fact or validation
6:38
reports after the fact or validation processes after the fact that check data
6:40
processes after the fact that check data
6:40
processes after the fact that check data and say is it good is it bad and then
6:43
and say is it good is it bad and then
6:43
and say is it good is it bad and then deal with it in whatever the appropriate
6:45
deal with it in whatever the appropriate
6:45
deal with it in whatever the appropriate way is it's important to note that any
6:48
way is it's important to note that any
6:48
way is it's important to note that any bad data created here lasts forever
6:51
bad data created here lasts forever
6:51
bad data created here lasts forever anything else you do with it from this
6:52
anything else you do with it from this
6:52
anything else you do with it from this point on will have mistakes in it and
6:55
point on will have mistakes in it and
6:55
point on will have mistakes in it and it's going to stay there so if we could
6:57
it's going to stay there so if we could
6:57
it's going to stay there so if we could fix it at the source that's a very good
7:00
fix it at the source that's a very good
7:00
fix it at the source that's a very good thing and note that I mentioned
7:02
thing and note that I mentioned
7:02
thing and note that I mentioned validating data at multiple levels
7:04
validating data at multiple levels
7:04
validating data at multiple levels there's value in that sometimes we like
7:05
there's value in that sometimes we like
7:05
there's value in that sometimes we like to think to ourselves a if the app is
7:07
to think to ourselves a if the app is
7:07
to think to ourselves a if the app is checking our data for us we're good
7:08
checking our data for us we're good
7:09
checking our data for us we're good right we don't have to worry about
7:09
right we don't have to worry about
7:09
right we don't have to worry about anything else the database can be kept
7:11
anything else the database can be kept
7:11
anything else the database can be kept simple straightforward small and compact
7:13
simple straightforward small and compact
7:13
simple straightforward small and compact but we're assuming the app is perfect
7:16
but we're assuming the app is perfect
7:16
but we're assuming the app is perfect and if you've built an app where there's
7:17
and if you've built an app where there's
7:17
and if you've built an app where there's no bugs and that's the way it's been for
7:19
no bugs and that's the way it's been for
7:19
no bugs and that's the way it's been for its life cycle that's awesome uh but
7:22
its life cycle that's awesome uh but
7:22
its life cycle that's awesome uh but most apps don't behave that way things
7:24
most apps don't behave that way things
7:24
most apps don't behave that way things go wrong we have bugs we fix bugs bad
7:26
go wrong we have bugs we fix bugs bad
7:26
go wrong we have bugs we fix bugs bad data is created so consider when you're
7:28
data is created so consider when you're
7:28
data is created so consider when you're validating data and in the its creation
7:30
validating data and in the its creation
7:30
validating data and in the its creation State having it checked in multiple
7:32
State having it checked in multiple
7:32
State having it checked in multiple places it's a great way in say we you
7:35
places it's a great way in say we you
7:35
places it's a great way in say we you layer security layer data validation and
7:37
layer security layer data validation and
7:37
layer security layer data validation and and and data Integrity checking because
7:40
and and data Integrity checking because
7:40
and and data Integrity checking because this is the beginning of data it's bad
7:41
this is the beginning of data it's bad
7:41
this is the beginning of data it's bad here it's bad
7:43
here it's bad
7:43
here it's bad forever from here where do data go we
7:46
forever from here where do data go we
7:46
forever from here where do data go we start with transactional data we start
7:47
start with transactional data we start
7:47
start with transactional data we start with application data and it begins
7:49
with application data and it begins
7:49
with application data and it begins moving along it may it copied may end up
7:52
moving along it may it copied may end up
7:52
moving along it may it copied may end up in analytic environment because we want
7:53
in analytic environment because we want
7:53
in analytic environment because we want to run reports whether they're inward
7:55
to run reports whether they're inward
7:55
to run reports whether they're inward facing outward facing if there's any
7:57
facing outward facing if there's any
7:57
facing outward facing if there's any sort of analytics going on at all
8:00
sort of analytics going on at all
8:00
sort of analytics going on at all decisions are being made based on it and
8:01
decisions are being made based on it and
8:01
decisions are being made based on it and so we want those decisions to be made
8:03
so we want those decisions to be made
8:03
so we want those decisions to be made correctly so how do we do that
8:06
correctly so how do we do that
8:06
correctly so how do we do that validating data is not hard honestly you
8:08
validating data is not hard honestly you
8:08
validating data is not hard honestly you don't have to check every single value
8:09
don't have to check every single value
8:09
don't have to check every single value you have to make sure everything is
8:10
you have to make sure everything is
8:10
you have to make sure everything is correct very often simply checking the
8:13
correct very often simply checking the
8:13
correct very often simply checking the general data size and data shape after
8:15
general data size and data shape after
8:15
general data size and data shape after data gets moved or copied somewhere or
8:17
data gets moved or copied somewhere or
8:17
data gets moved or copied somewhere or transformed is good enough if the amount
8:20
transformed is good enough if the amount
8:20
transformed is good enough if the amount of data that I have changes by more than
8:22
of data that I have changes by more than
8:22
of data that I have changes by more than x percent day over day or like over like
8:25
x percent day over day or like over like
8:25
x percent day over day or like over like we want to know if the data return is
8:27
we want to know if the data return is
8:27
we want to know if the data return is zero for a day if the number of data
8:28
zero for a day if the number of data
8:28
zero for a day if the number of data sources goes up or down by a lot if the
8:30
sources goes up or down by a lot if the
8:30
sources goes up or down by a lot if the bite count goes up or down by a lot
8:32
bite count goes up or down by a lot
8:32
bite count goes up or down by a lot things like that very very easy to check
8:34
things like that very very easy to check
8:34
things like that very very easy to check and very easy to tell and something's
8:36
and very easy to tell and something's
8:36
and very easy to tell and something's wrong if we get a thousand rows from
8:38
wrong if we get a thousand rows from
8:38
wrong if we get a thousand rows from somewhere every single day and suddenly
8:40
somewhere every single day and suddenly
8:40
somewhere every single day and suddenly it's a million that's unusual and
8:42
it's a million that's unusual and
8:42
it's a million that's unusual and there's a reason right similarly we can
8:44
there's a reason right similarly we can
8:44
there's a reason right similarly we can look at values as well uniqueness of
8:46
look at values as well uniqueness of
8:46
look at values as well uniqueness of values there certain values that should
8:48
values there certain values that should
8:48
values there certain values that should always be unique are nulls and blanks
8:50
always be unique are nulls and blanks
8:50
always be unique are nulls and blanks allowed or not are there values that are
8:53
allowed or not are there values that are
8:53
allowed or not are there values that are invalid you know very often we'll say
8:55
invalid you know very often we'll say
8:55
invalid you know very often we'll say take an integer store value but we know
8:57
take an integer store value but we know
8:57
take an integer store value but we know it can't be negative but storing as in
9:00
it can't be negative but storing as in
9:00
it can't be negative but storing as in anyway you can check and see are they
9:01
anyway you can check and see are they
9:01
anyway you can check and see are they negatives and you may say like that
9:03
negatives and you may say like that
9:03
negatives and you may say like that can't possibly happen but we know how
9:05
can't possibly happen but we know how
9:05
can't possibly happen but we know how these things work if you allow a
9:06
these things work if you allow a
9:06
these things work if you allow a negative somebody will put a negative in
9:08
negative somebody will put a negative in
9:08
negative somebody will put a negative in if you allow it to happen it's just what
9:10
if you allow it to happen it's just what
9:10
if you allow it to happen it's just what QA does right it's how we do things
9:12
QA does right it's how we do things
9:12
QA does right it's how we do things negative a million just as valid as
9:14
negative a million just as valid as
9:14
negative a million just as valid as positive a million if you allow it right
9:16
positive a million if you allow it right
9:16
positive a million if you allow it right missing data is important as is
9:17
missing data is important as is
9:17
missing data is important as is duplicate data these are key validations
9:20
duplicate data these are key validations
9:20
duplicate data these are key validations that are not hard to check for if you
9:21
that are not hard to check for if you
9:22
that are not hard to check for if you have some key value that should be in
9:23
have some key value that should be in
9:23
have some key value that should be in unique make sure it only appears once
9:25
unique make sure it only appears once
9:25
unique make sure it only appears once per value again in the analytic world we
9:27
per value again in the analytic world we
9:27
per value again in the analytic world we very often we have less constraints have
9:29
very often we have less constraints have
9:29
very often we have less constraints have less validation that's built into data
9:31
less validation that's built into data
9:31
less validation that's built into data structures but we can check it right you
9:33
structures but we can check it right you
9:33
structures but we can check it right you can load data check it you can copy data
9:35
can load data check it you can copy data
9:35
can load data check it you can copy data check it you can transform it check it
9:37
check it you can transform it check it
9:37
check it you can transform it check it it's very fast it's very easy and once
9:39
it's very fast it's very easy and once
9:39
it's very fast it's very easy and once you've checked any data you've loaded
9:40
you've checked any data you've loaded
9:40
you've checked any data you've loaded you can put that then onto your giant
9:41
you can put that then onto your giant
9:42
you can put that then onto your giant pile mountain of of existing data and
9:45
pile mountain of of existing data and
9:45
pile mountain of of existing data and not worry about it again also consider
9:47
not worry about it again also consider
9:47
not worry about it again also consider edge cases uh edge cases happen a lot in
9:50
edge cases uh edge cases happen a lot in
9:50
edge cases uh edge cases happen a lot in data are they valid or not valid edge
9:53
data are they valid or not valid edge
9:53
data are they valid or not valid edge cases do they mean something and we
9:55
cases do they mean something and we
9:55
cases do they mean something and we should have them or are they crazy if I
9:57
should have them or are they crazy if I
9:57
should have them or are they crazy if I see in my order processing system that
9:59
see in my order processing system that
9:59
see in my order processing system that somebody placed an order for 10 trillion
10:01
somebody placed an order for 10 trillion
10:01
somebody placed an order for 10 trillion doll I would assume that education is
10:03
doll I would assume that education is
10:03
doll I would assume that education is probably invalid nobody is making an
10:06
probably invalid nobody is making an
10:06
probably invalid nobody is making an order that large right on the other hand
10:09
order that large right on the other hand
10:09
order that large right on the other hand if I saw an order for a million dollars
10:11
if I saw an order for a million dollars
10:11
if I saw an order for a million dollars I would say all right that number is
10:12
I would say all right that number is
10:12
I would say all right that number is really big probably somebody should look
10:14
really big probably somebody should look
10:14
really big probably somebody should look at it um but it could be valid right
10:16
at it um but it could be valid right
10:16
at it um but it could be valid right somebody could spend a million dollars
10:18
somebody could spend a million dollars
10:18
somebody could spend a million dollars on something as improbable as it may be
10:20
on something as improbable as it may be
10:20
on something as improbable as it may be that's at least feasible uh negative
10:22
that's at least feasible uh negative
10:22
that's at least feasible uh negative million dollars not possible negative
10:24
million dollars not possible negative
10:24
million dollars not possible negative numbers you can't spend a negative
10:25
numbers you can't spend a negative
10:25
numbers you can't spend a negative number of dollars right so we can check
10:27
number of dollars right so we can check
10:27
number of dollars right so we can check for things like that and understand end
10:29
for things like that and understand end
10:29
for things like that and understand end is it correct or not and importantly
10:32
is it correct or not and importantly
10:32
is it correct or not and importantly here this validation should all occur
10:34
here this validation should all occur
10:34
here this validation should all occur before the data is moved again before
10:35
before the data is moved again before
10:35
before the data is moved again before it's used for anything else this is your
10:37
it's used for anything else this is your
10:37
it's used for anything else this is your second chance so you check it at its
10:39
second chance so you check it at its
10:39
second chance so you check it at its source you check it after it's been
10:41
source you check it after it's been
10:41
source you check it after it's been moved and copied and each time it gets
10:43
moved and copied and each time it gets
10:43
moved and copied and each time it gets moved and copied there needs to be
10:45
moved and copied there needs to be
10:45
moved and copied there needs to be validation and it amazes me how many
10:46
validation and it amazes me how many
10:47
validation and it amazes me how many systems I've worked with over my
10:48
systems I've worked with over my
10:48
systems I've worked with over my lifetime and seen and heard about where
10:50
lifetime and seen and heard about where
10:50
lifetime and seen and heard about where there's no validation data simply moves
10:51
there's no validation data simply moves
10:51
there's no validation data simply moves all over the place and people just use
10:53
all over the place and people just use
10:53
all over the place and people just use it and that's it uh keep in mind that if
10:57
it and that's it uh keep in mind that if
10:57
it and that's it uh keep in mind that if you do that you're riding on Hope and
10:59
you do that you're riding on Hope and
10:59
you do that you're riding on Hope and dreams but hopes and dreams aren't very
11:01
dreams but hopes and dreams aren't very
11:01
dreams but hopes and dreams aren't very scientific uh nor do they really help
11:03
scientific uh nor do they really help
11:04
scientific uh nor do they really help you get good data I've hoped and dreamed
11:06
you get good data I've hoped and dreamed
11:06
you get good data I've hoped and dreamed a lot and my data quality is never gone
11:07
a lot and my data quality is never gone
11:07
a lot and my data quality is never gone up as a result so I highly recommend
11:09
up as a result so I highly recommend
11:09
up as a result so I highly recommend validation as opposed to hopes uh they
11:12
validation as opposed to hopes uh they
11:12
validation as opposed to hopes uh they don't really help as
11:13
don't really help as
11:13
don't really help as much and of course one big piece that
11:16
much and of course one big piece that
11:16
much and of course one big piece that doesn't always get uh considered is what
11:19
doesn't always get uh considered is what
11:19
doesn't always get uh considered is what happens when things change uh software
11:21
happens when things change uh software
11:21
happens when things change uh software releases happen all the time in
11:22
releases happen all the time in
11:23
releases happen all the time in applications in the analytic world in
11:24
applications in the analytic world in
11:25
applications in the analytic world in the AI World things change right you
11:26
the AI World things change right you
11:26
the AI World things change right you decide to go from you know chat GP to a
11:29
decide to go from you know chat GP to a
11:29
decide to go from you know chat GP to a new version of chap GPT we're going to
11:31
new version of chap GPT we're going to
11:31
new version of chap GPT we're going to upgrade and try a new version we're
11:32
upgrade and try a new version we're
11:32
upgrade and try a new version we're going to move everything to that but
11:33
going to move everything to that but
11:33
going to move everything to that but that changes things right or maybe a
11:35
that changes things right or maybe a
11:35
that changes things right or maybe a software release goes into a
11:36
software release goes into a
11:36
software release goes into a transactional application well check and
11:39
transactional application well check and
11:39
transactional application well check and make sure that any data impacted by the
11:41
make sure that any data impacted by the
11:41
make sure that any data impacted by the release is checked validated looks good
11:44
release is checked validated looks good
11:44
release is checked validated looks good also ensure that new data being created
11:46
also ensure that new data being created
11:46
also ensure that new data being created is also checked validated good uh and
11:49
is also checked validated good uh and
11:49
is also checked validated good uh and this is really what QA is for right
11:51
this is really what QA is for right
11:51
this is really what QA is for right quality assurance is there to ensure
11:52
quality assurance is there to ensure
11:52
quality assurance is there to ensure that changes happen well that's
11:55
that changes happen well that's
11:55
that changes happen well that's important uh but if we don't do that
11:57
important uh but if we don't do that
11:57
important uh but if we don't do that though there's always a chance that
11:58
though there's always a chance that
11:58
though there's always a chance that things will go wrong going forward and
12:00
things will go wrong going forward and
12:00
things will go wrong going forward and now you're in a world where some of your
12:01
now you're in a world where some of your
12:01
now you're in a world where some of your data is good and some of your data is
12:02
data is good and some of your data is
12:02
data is good and some of your data is bad we have to come back and deal with
12:04
bad we have to come back and deal with
12:04
bad we have to come back and deal with that later which I suppose is better
12:06
that later which I suppose is better
12:06
that later which I suppose is better than all bad data but just keep in mind
12:09
than all bad data but just keep in mind
12:09
than all bad data but just keep in mind that uh changes to any tier of your
12:11
that uh changes to any tier of your
12:12
that uh changes to any tier of your application whether it's a transactional
12:14
application whether it's a transactional
12:15
application whether it's a transactional app whether it's a production database
12:17
app whether it's a production database
12:17
app whether it's a production database whether it's ETL elt data movement data
12:21
whether it's ETL elt data movement data
12:21
whether it's ETL elt data movement data copying analytics AI machine learning
12:22
copying analytics AI machine learning
12:22
copying analytics AI machine learning whatever those are all changes you
12:24
whatever those are all changes you
12:24
whatever those are all changes you should QA tested data validated it's
12:27
should QA tested data validated it's
12:27
should QA tested data validated it's important
12:30
important
12:30
important something else too that's kind of a a
12:31
something else too that's kind of a a
12:31
something else too that's kind of a a fundamental in database design that gets
12:33
fundamental in database design that gets
12:33
fundamental in database design that gets often thrown away when we start talking
12:35
often thrown away when we start talking
12:35
often thrown away when we start talking about things so far Downstream are what
12:38
about things so far Downstream are what
12:38
about things so far Downstream are what do our names of things look like what do
12:41
do our names of things look like what do
12:41
do our names of things look like what do our data types look like these are
12:43
our data types look like these are
12:43
our data types look like these are classic problems we've dealt with that
12:45
classic problems we've dealt with that
12:45
classic problems we've dealt with that are are very often seen as beginning of
12:47
are are very often seen as beginning of
12:47
are are very often seen as beginning of the lifeline um for our data but they
12:50
the lifeline um for our data but they
12:50
the lifeline um for our data but they matter more now than ever because we're
12:51
matter more now than ever because we're
12:51
matter more now than ever because we're gonna take data and throw it into an AI
12:54
gonna take data and throw it into an AI
12:54
gonna take data and throw it into an AI algorithm and say hey look at my data
12:57
algorithm and say hey look at my data
12:57
algorithm and say hey look at my data process it understand it then answer my
12:59
process it understand it then answer my
12:59
process it understand it then answer my questions about it if you have M if you
13:01
questions about it if you have M if you
13:01
questions about it if you have M if you have metadata problems not just data
13:03
have metadata problems not just data
13:03
have metadata problems not just data problems but you have metadata problems
13:06
problems but you have metadata problems
13:06
problems but you have metadata problems then there is a chance the AI will make
13:08
then there is a chance the AI will make
13:08
then there is a chance the AI will make a mistake based on it this is really
13:09
a mistake based on it this is really
13:09
a mistake based on it this is really common we've all probably run into
13:13
common we've all probably run into
13:13
common we've all probably run into metadata problems like this and if you
13:14
metadata problems like this and if you
13:14
metadata problems like this and if you haven't you will but you're lucky so far
13:17
haven't you will but you're lucky so far
13:17
haven't you will but you're lucky so far I suppose for example I have a couple
13:18
I suppose for example I have a couple
13:18
I suppose for example I have a couple fun examples here just for you know heck
13:21
fun examples here just for you know heck
13:21
fun examples here just for you know heck to throw some ideas out here you know
13:22
to throw some ideas out here you know
13:22
to throw some ideas out here you know what if you saw a data element that was
13:25
what if you saw a data element that was
13:25
what if you saw a data element that was integer named invoice what would that
13:28
integer named invoice what would that
13:28
integer named invoice what would that what would that be if you're AI
13:29
what would that be if you're AI
13:29
what would that be if you're AI algorithm consuming this what do you
13:30
algorithm consuming this what do you
13:30
algorithm consuming this what do you think it would be does it mean there is
13:32
think it would be does it mean there is
13:32
think it would be does it mean there is an invoice yes or no is it the invoice
13:35
an invoice yes or no is it the invoice
13:35
an invoice yes or no is it the invoice amount is it an invoice number I don't
13:38
amount is it an invoice number I don't
13:38
amount is it an invoice number I don't really know what it is I can guess I can
13:40
really know what it is I can guess I can
13:40
really know what it is I can guess I can look at the data and guess uh but
13:42
look at the data and guess uh but
13:42
look at the data and guess uh but there's no guarantee I'll be right I'll
13:44
there's no guarantee I'll be right I'll
13:44
there's no guarantee I'll be right I'll be right sometimes and I might be wrong
13:46
be right sometimes and I might be wrong
13:46
be right sometimes and I might be wrong sometimes so look at any data you have
13:49
sometimes so look at any data you have
13:49
sometimes so look at any data you have and look at the names the data types the
13:51
and look at the names the data types the
13:51
and look at the names the data types the sizes and validate is this correct or
13:54
sizes and validate is this correct or
13:54
sizes and validate is this correct or not for example if I had a a date time
13:56
not for example if I had a a date time
13:56
not for example if I had a a date time named entry time is it entry time or is
14:00
named entry time is it entry time or is
14:00
named entry time is it entry time or is it really a date time sometimes we store
14:02
it really a date time sometimes we store
14:02
it really a date time sometimes we store dates or times as date time or even as a
14:05
dates or times as date time or even as a
14:05
dates or times as date time or even as a string strings are dangerous because
14:07
string strings are dangerous because
14:07
string strings are dangerous because they could have bad values right
14:08
they could have bad values right
14:08
they could have bad values right February 31st is not possible in uh a
14:12
February 31st is not possible in uh a
14:12
February 31st is not possible in uh a date time that's typed correctly uh but
14:14
date time that's typed correctly uh but
14:14
date time that's typed correctly uh but if it was a string February 31st didn't
14:17
if it was a string February 31st didn't
14:17
if it was a string February 31st didn't happen and or you just mix up European
14:19
happen and or you just mix up European
14:19
happen and or you just mix up European and American times for example you know
14:21
and American times for example you know
14:21
and American times for example you know some of us will write two 29 for
14:24
some of us will write two 29 for
14:24
some of us will write two 29 for February 29th but what about 292 or 29th
14:27
February 29th but what about 292 or 29th
14:27
February 29th but what about 292 or 29th of February or year month day Year Day
14:30
of February or year month day Year Day
14:30
of February or year month day Year Day month so on and so forth those things
14:32
month so on and so forth those things
14:32
month so on and so forth those things matter here AI is going to consume this
14:34
matter here AI is going to consume this
14:34
matter here AI is going to consume this look at it make decisions based on it it
14:36
look at it make decisions based on it it
14:36
look at it make decisions based on it it needs to be clear similarly if I had a
14:39
needs to be clear similarly if I had a
14:39
needs to be clear similarly if I had a call and this is kind of a fun one uh
14:41
call and this is kind of a fun one uh
14:41
call and this is kind of a fun one uh very often we soft delete data because
14:43
very often we soft delete data because
14:43
very often we soft delete data because we want to keep it around for posterity
14:45
we want to keep it around for posterity
14:45
we want to keep it around for posterity for reference purposes for compliance
14:47
for reference purposes for compliance
14:47
for reference purposes for compliance purposes and so we'll soft delete data
14:49
purposes and so we'll soft delete data
14:49
purposes and so we'll soft delete data and we'll have some column or element
14:51
and we'll have some column or element
14:51
and we'll have some column or element like is deleted or something and then is
14:53
like is deleted or something and then is
14:53
like is deleted or something and then is archived whatever and we'll have that
14:55
archived whatever and we'll have that
14:55
archived whatever and we'll have that out there and then at some point AI
14:58
out there and then at some point AI
14:58
out there and then at some point AI comes in consumes this and the algorithm
15:00
comes in consumes this and the algorithm
15:00
comes in consumes this and the algorithm will then maybe say well hey uh it's
15:03
will then maybe say well hey uh it's
15:03
will then maybe say well hey uh it's deleted okay does that mean I should use
15:05
deleted okay does that mean I should use
15:05
deleted okay does that mean I should use this data as I Crunch and return
15:08
this data as I Crunch and return
15:08
this data as I Crunch and return responses or should I skip the ones that
15:10
responses or should I skip the ones that
15:10
responses or should I skip the ones that are deleted because they're deleted
15:11
are deleted because they're deleted
15:11
are deleted because they're deleted realistically it doesn't really know
15:13
realistically it doesn't really know
15:13
realistically it doesn't really know what the right answer is uh you and I
15:14
what the right answer is uh you and I
15:14
what the right answer is uh you and I may um but it won't so consider that as
15:17
may um but it won't so consider that as
15:17
may um but it won't so consider that as well that if there's any kind of
15:19
well that if there's any kind of
15:19
well that if there's any kind of ambiguous metadata or metadata that may
15:21
ambiguous metadata or metadata that may
15:22
ambiguous metadata or metadata that may influence decisions you may have to
15:23
influence decisions you may have to
15:23
influence decisions you may have to train it or give it additional
15:24
train it or give it additional
15:24
train it or give it additional information to understand what it means
15:26
information to understand what it means
15:26
information to understand what it means or just remove it all together if a
15:28
or just remove it all together if a
15:28
or just remove it all together if a column is something you can simply
15:30
column is something you can simply
15:30
column is something you can simply resolve and say you know what I just
15:32
resolve and say you know what I just
15:32
resolve and say you know what I just don't want deleted data in there at all
15:34
don't want deleted data in there at all
15:34
don't want deleted data in there at all then remove all the rows is deleted one
15:36
then remove all the rows is deleted one
15:36
then remove all the rows is deleted one call it a day never train with it don't
15:38
call it a day never train with it don't
15:38
call it a day never train with it don't bring it in as rag data just leave it
15:40
bring it in as rag data just leave it
15:40
bring it in as rag data just leave it out don't even give it to the algorithm
15:42
out don't even give it to the algorithm
15:42
out don't even give it to the algorithm don't let it think about it it's just
15:43
don't let it think about it it's just
15:43
don't let it think about it it's just noise at that point and providing noise
15:45
noise at that point and providing noise
15:46
noise at that point and providing noise to an AI algorithm is only gonna create
15:48
to an AI algorithm is only gonna create
15:48
to an AI algorithm is only gonna create bad responses or or just you know things
15:50
bad responses or or just you know things
15:50
bad responses or or just you know things people don't expect the last thing you
15:52
people don't expect the last thing you
15:52
people don't expect the last thing you want to do is get information on an
15:54
want to do is get information on an
15:54
want to do is get information on an order you placed last year that you
15:55
order you placed last year that you
15:55
order you placed last year that you deleted and never actually placed right
15:57
deleted and never actually placed right
15:57
deleted and never actually placed right I filled my card up I went to place the
15:59
I filled my card up I went to place the
15:59
I filled my card up I went to place the order I said oh I don't want that stuff
16:02
order I said oh I don't want that stuff
16:02
order I said oh I don't want that stuff now just start over and then it's giv me
16:05
now just start over and then it's giv me
16:05
now just start over and then it's giv me answers back about it later that's not
16:07
answers back about it later that's not
16:07
answers back about it later that's not so
16:09
so
16:09
so good one note is we have different kinds
16:12
good one note is we have different kinds
16:12
good one note is we have different kinds of data uh I'll these are two examples
16:14
of data uh I'll these are two examples
16:14
of data uh I'll these are two examples you may have other data that goes into
16:16
you may have other data that goes into
16:16
you may have other data that goes into your algorithms goes in your machine
16:17
your algorithms goes in your machine
16:17
your algorithms goes in your machine learning keep in mind that each of these
16:20
learning keep in mind that each of these
16:20
learning keep in mind that each of these are different data sets so you have
16:22
are different data sets so you have
16:22
are different data sets so you have training data uh the purpose of training
16:24
training data uh the purpose of training
16:24
training data uh the purpose of training data is to ultimately tell your AI how
16:26
data is to ultimately tell your AI how
16:26
data is to ultimately tell your AI how to behave what is its purpose what is is
16:28
to behave what is its purpose what is is
16:29
to behave what is its purpose what is is it going to do how is it going to
16:31
it going to do how is it going to
16:31
it going to do how is it going to respond what is it trying to and if it's
16:34
respond what is it trying to and if it's
16:34
respond what is it trying to and if it's bad what happens well if you have bad
16:35
bad what happens well if you have bad
16:35
bad what happens well if you have bad training data your model will simply
16:38
training data your model will simply
16:38
training data your model will simply return not behave the way it should it
16:40
return not behave the way it should it
16:40
return not behave the way it should it may return the right answers potentially
16:42
may return the right answers potentially
16:42
may return the right answers potentially but it may not do it in the right way or
16:44
but it may not do it in the right way or
16:44
but it may not do it in the right way or it may give irrelevant answers that are
16:46
it may give irrelevant answers that are
16:46
it may give irrelevant answers that are still correct you ask me what the color
16:48
still correct you ask me what the color
16:48
still correct you ask me what the color of the sky is and I say 42 that's not a
16:51
of the sky is and I say 42 that's not a
16:51
of the sky is and I say 42 that's not a very helpful response maybe I was
16:52
very helpful response maybe I was
16:52
very helpful response maybe I was thinking of a wavelength instead or
16:54
thinking of a wavelength instead or
16:54
thinking of a wavelength instead or something else but that isn't relevant
16:56
something else but that isn't relevant
16:56
something else but that isn't relevant or correct so I can give the right
16:58
or correct so I can give the right
16:58
or correct so I can give the right answer to the wrong question that's
17:00
answer to the wrong question that's
17:00
answer to the wrong question that's often what happens here uh on the other
17:02
often what happens here uh on the other
17:02
often what happens here uh on the other hand you have retrieval augmented
17:04
hand you have retrieval augmented
17:04
hand you have retrieval augmented Generation Um that is rag data that's
17:06
Generation Um that is rag data that's
17:06
Generation Um that is rag data that's the data you bring so you train a model
17:09
the data you bring so you train a model
17:09
the data you bring so you train a model it's behaving the way you want now you
17:11
it's behaving the way you want now you
17:11
it's behaving the way you want now you swap in your data the actual data you
17:13
swap in your data the actual data you
17:13
swap in your data the actual data you care about you've trained it to be a you
17:17
care about you've trained it to be a you
17:17
care about you've trained it to be a you know a bot that will answer questions
17:19
know a bot that will answer questions
17:20
know a bot that will answer questions about an airline for example now you're
17:22
about an airline for example now you're
17:22
about an airline for example now you're gonna bring your airline data in so it
17:24
gonna bring your airline data in so it
17:24
gonna bring your airline data in so it can use current information to do what
17:25
can use current information to do what
17:25
can use current information to do what it has to do if your rag data is
17:28
it has to do if your rag data is
17:28
it has to do if your rag data is incorrect what do you get you have
17:30
incorrect what do you get you have
17:30
incorrect what do you get you have invalid responses bad answers so it's
17:32
invalid responses bad answers so it's
17:32
invalid responses bad answers so it's important to know you have different
17:33
important to know you have different
17:33
important to know you have different sets of data and bad data in any of
17:36
sets of data and bad data in any of
17:36
sets of data and bad data in any of these areas will result in different
17:38
these areas will result in different
17:38
these areas will result in different results at the end of the line so you
17:40
results at the end of the line so you
17:40
results at the end of the line so you can have bad training data which will
17:42
can have bad training data which will
17:42
can have bad training data which will provide one set of problems and
17:43
provide one set of problems and
17:43
provide one set of problems and headaches and bad your data frag data
17:46
headaches and bad your data frag data
17:46
headaches and bad your data frag data which will result in other headaches
17:49
which will result in other headaches
17:49
which will result in other headaches different headaches and if you have bad
17:50
different headaches and if you have bad
17:50
different headaches and if you have bad data in both you may have a hard time
17:52
data in both you may have a hard time
17:52
data in both you may have a hard time diagnosing where the problem
17:57
is all right so now we're going to talk
17:59
is all right so now we're going to talk
17:59
is all right so now we're going to talk about AI specific how do we cheat how do
18:03
about AI specific how do we cheat how do
18:03
about AI specific how do we cheat how do we do things we're not supposed to do to
18:06
we do things we're not supposed to do to
18:06
we do things we're not supposed to do to try to fix the problem because we have
18:08
try to fix the problem because we have
18:08
try to fix the problem because we have an app we're close to releasing some new
18:10
an app we're close to releasing some new
18:10
an app we're close to releasing some new stuff it's not behaving correctly we've
18:13
stuff it's not behaving correctly we've
18:13
stuff it's not behaving correctly we've got to fix it what are the things we do
18:16
got to fix it what are the things we do
18:16
got to fix it what are the things we do and then don't look back on that
18:18
and then don't look back on that
18:18
and then don't look back on that probably we shouldn't do all right this
18:22
probably we shouldn't do all right this
18:22
probably we shouldn't do all right this sounds fun right this is the most common
18:24
sounds fun right this is the most common
18:24
sounds fun right this is the most common one I see it's very often that we're
18:26
one I see it's very often that we're
18:26
one I see it's very often that we're close we have a a interface that works
18:29
close we have a a interface that works
18:29
close we have a a interface that works nicely it's almost correct all the time
18:32
nicely it's almost correct all the time
18:32
nicely it's almost correct all the time sometimes you get bad responses
18:34
sometimes you get bad responses
18:34
sometimes you get bad responses sometimes we get answers that aren't
18:35
sometimes we get answers that aren't
18:35
sometimes we get answers that aren't correct so what do we do we take our
18:37
correct so what do we do we take our
18:37
correct so what do we do we take our prompt and try to engineer it further
18:39
prompt and try to engineer it further
18:39
prompt and try to engineer it further and we say hey you know what we're
18:40
and we say hey you know what we're
18:40
and we say hey you know what we're really close we'll just adjust it a
18:43
really close we'll just adjust it a
18:43
really close we'll just adjust it a little bit so that it doesn't give the
18:44
little bit so that it doesn't give the
18:44
little bit so that it doesn't give the wrong responses anymore and this goes
18:47
wrong responses anymore and this goes
18:47
wrong responses anymore and this goes down a rabbit hole of kind of cause and
18:50
down a rabbit hole of kind of cause and
18:50
down a rabbit hole of kind of cause and effect cause and effect the purpose of
18:51
effect cause and effect the purpose of
18:51
effect cause and effect the purpose of prompt engineering is to provide purpose
18:54
prompt engineering is to provide purpose
18:54
prompt engineering is to provide purpose to an algorithm so that it knows what is
18:57
to an algorithm so that it knows what is
18:57
to an algorithm so that it knows what is it what is its role what is it supposed
18:58
it what is its role what is it supposed
18:58
it what is its role what is it supposed to be doing how does it answer things
19:01
to be doing how does it answer things
19:01
to be doing how does it answer things like that that's what we're trying to do
19:03
like that that's what we're trying to do
19:03
like that that's what we're trying to do it keeps things relevant it makes sure
19:05
it keeps things relevant it makes sure
19:05
it keeps things relevant it makes sure that it's doing what you want it to do
19:07
that it's doing what you want it to do
19:07
that it's doing what you want it to do if you have bad data coming in from
19:08
if you have bad data coming in from
19:09
if you have bad data coming in from anywhere in your process and it consumes
19:12
anywhere in your process and it consumes
19:12
anywhere in your process and it consumes that bad data you really can't prompt
19:14
that bad data you really can't prompt
19:14
that bad data you really can't prompt your way out of it you can try you can
19:16
your way out of it you can try you can
19:16
your way out of it you can try you can include details in your prompt to try to
19:18
include details in your prompt to try to
19:18
include details in your prompt to try to get around it but the problem is you're
19:20
get around it but the problem is you're
19:20
get around it but the problem is you're changing its purpose now and so you're
19:22
changing its purpose now and so you're
19:22
changing its purpose now and so you're trying to prompt way of bad data but as
19:24
trying to prompt way of bad data but as
19:24
trying to prompt way of bad data but as you do so you're going to make it behave
19:28
you do so you're going to make it behave
19:28
you do so you're going to make it behave differently in an effort to avoid bad
19:30
differently in an effort to avoid bad
19:30
differently in an effort to avoid bad data and if you think of it more from a
19:32
data and if you think of it more from a
19:32
data and if you think of it more from a human perspective like if you were going
19:34
human perspective like if you were going
19:34
human perspective like if you were going to ask somebody to answer questions for
19:36
to ask somebody to answer questions for
19:36
to ask somebody to answer questions for you on the phone and you went to them
19:38
you on the phone and you went to them
19:38
you on the phone and you went to them and said listen this is really important
19:40
and said listen this is really important
19:40
and said listen this is really important when you're answering questions about
19:41
when you're answering questions about
19:41
when you're answering questions about the airline don't respond with that
19:44
the airline don't respond with that
19:44
the airline don't respond with that flight going to Bangkok tomorrow just
19:45
flight going to Bangkok tomorrow just
19:45
flight going to Bangkok tomorrow just don't mention it at all and if they if
19:47
don't mention it at all and if they if
19:47
don't mention it at all and if they if they ask about it don't say anything
19:49
they ask about it don't say anything
19:49
they ask about it don't say anything it's not a$ thousand dollar like you
19:50
it's not a$ thousand dollar like you
19:50
it's not a$ thousand dollar like you begin giving these weird pointers and a
19:52
begin giving these weird pointers and a
19:52
begin giving these weird pointers and a human being would say what what are you
19:55
human being would say what what are you
19:55
human being would say what what are you talking about that doesn't make any
19:56
talking about that doesn't make any
19:56
talking about that doesn't make any sense um but the algorithms have to
19:58
sense um but the algorithms have to
19:58
sense um but the algorithms have to consume what whatever you give it and so
19:59
consume what whatever you give it and so
19:59
consume what whatever you give it and so you prompt engineer you put details in
20:02
you prompt engineer you put details in
20:02
you prompt engineer you put details in and you might start getting correct
20:03
and you might start getting correct
20:03
and you might start getting correct answers to the problem you found but I
20:06
answers to the problem you found but I
20:06
answers to the problem you found but I guarantee you will introduce
20:08
guarantee you will introduce
20:08
guarantee you will introduce complexities and wrong answers elsewhere
20:10
complexities and wrong answers elsewhere
20:10
complexities and wrong answers elsewhere so you can't prompt your way out of bad
20:12
so you can't prompt your way out of bad
20:12
so you can't prompt your way out of bad data and if you try you will prompt
20:13
data and if you try you will prompt
20:13
data and if you try you will prompt yourself into other bad responses not a
20:17
yourself into other bad responses not a
20:17
yourself into other bad responses not a good way to go similarly we have the
20:20
good way to go similarly we have the
20:20
good way to go similarly we have the world of the data we bring retrieval me
20:21
world of the data we bring retrieval me
20:22
world of the data we bring retrieval me to generation rag data is data we bring
20:24
to generation rag data is data we bring
20:24
to generation rag data is data we bring so you train your model on one data set
20:26
so you train your model on one data set
20:26
so you train your model on one data set and then when you're happy with what you
20:28
and then when you're happy with what you
20:28
and then when you're happy with what you have you bring your data and continue
20:30
have you bring your data and continue
20:30
have you bring your data and continue working with it um you can't fix bad
20:32
working with it um you can't fix bad
20:32
working with it um you can't fix bad training with this though the purpose of
20:34
training with this though the purpose of
20:34
training with this though the purpose of rag data is to be your data for the
20:37
rag data is to be your data for the
20:37
rag data is to be your data for the flight you know for the airline example
20:39
flight you know for the airline example
20:39
flight you know for the airline example I will have flight numbers and prices
20:41
I will have flight numbers and prices
20:41
I will have flight numbers and prices and layovers and T durations and first
20:44
and layovers and T durations and first
20:44
and layovers and T durations and first class economy whatever I have all that
20:47
class economy whatever I have all that
20:47
class economy whatever I have all that data coming in for upcoming flights
20:49
data coming in for upcoming flights
20:49
data coming in for upcoming flights people will ask questions they'll get
20:50
people will ask questions they'll get
20:50
people will ask questions they'll get answers the nice thing about this though
20:52
answers the nice thing about this though
20:53
answers the nice thing about this though is if you do have bad data in the rag
20:54
is if you do have bad data in the rag
20:54
is if you do have bad data in the rag world that's gonna mean bad responses
20:56
world that's gonna mean bad responses
20:56
world that's gonna mean bad responses just inaccurate answers that's all it's
20:59
just inaccurate answers that's all it's
20:59
just inaccurate answers that's all it's also easy to fix though this is the data
21:01
also easy to fix though this is the data
21:01
also easy to fix though this is the data set you're bringing you simply identify
21:03
set you're bringing you simply identify
21:03
set you're bringing you simply identify where's the bad data you fix it and then
21:05
where's the bad data you fix it and then
21:05
where's the bad data you fix it and then after you fix it you can go back and
21:06
after you fix it you can go back and
21:06
after you fix it you can go back and find the source of the bad data and fix
21:08
find the source of the bad data and fix
21:08
find the source of the bad data and fix the source as well and then we're good
21:11
the source as well and then we're good
21:11
the source as well and then we're good and you move on with life don't make it
21:13
and you move on with life don't make it
21:13
and you move on with life don't make it more complicated than it has to be don't
21:15
more complicated than it has to be don't
21:15
more complicated than it has to be don't try to work around it don't try to train
21:16
try to work around it don't try to train
21:17
try to work around it don't try to train around it don't try to fine tune around
21:18
around it don't try to fine tune around
21:18
around it don't try to fine tune around it don't try to unlearn stuff anything
21:20
it don't try to unlearn stuff anything
21:20
it don't try to unlearn stuff anything else you do to try to fix your bad rag
21:22
else you do to try to fix your bad rag
21:22
else you do to try to fix your bad rag data is just going to make things more
21:24
data is just going to make things more
21:24
data is just going to make things more complicated and there really is a beauty
21:26
complicated and there really is a beauty
21:26
complicated and there really is a beauty and simplicity in this world the more
21:28
and simplicity in this world the more
21:28
and simplicity in this world the more complic cated your code gets for
21:30
complic cated your code gets for
21:30
complic cated your code gets for algorithms the harder it is to
21:31
algorithms the harder it is to
21:31
algorithms the harder it is to troubleshoot and to get back what you
21:33
troubleshoot and to get back what you
21:34
troubleshoot and to get back what you really
21:36
want semantic search is fun because
21:38
want semantic search is fun because
21:38
want semantic search is fun because there's a lot of math involved here I
21:40
there's a lot of math involved here I
21:40
there's a lot of math involved here I like personally I love math math is fun
21:41
like personally I love math math is fun
21:41
like personally I love math math is fun to me not everyone believes that but I
21:43
to me not everyone believes that but I
21:43
to me not everyone believes that but I enjoy it purpose of semantic search is
21:45
enjoy it purpose of semantic search is
21:45
enjoy it purpose of semantic search is take and create associations in data
21:48
take and create associations in data
21:48
take and create associations in data this helps us create more qualitative
21:50
this helps us create more qualitative
21:50
this helps us create more qualitative analysis and so you have qu quantitative
21:52
analysis and so you have qu quantitative
21:52
analysis and so you have qu quantitative mathematical analysis that results in
21:54
mathematical analysis that results in
21:54
mathematical analysis that results in our ability to make things associate to
21:57
our ability to make things associate to
21:57
our ability to make things associate to each other answer questions so if I ask
21:59
each other answer questions so if I ask
22:00
each other answer questions so if I ask a question of an algorithm and it
22:01
a question of an algorithm and it
22:01
a question of an algorithm and it doesn't recognize all the words or
22:03
doesn't recognize all the words or
22:03
doesn't recognize all the words or sentences it can figure out what I'm
22:05
sentences it can figure out what I'm
22:05
sentences it can figure out what I'm talking about based on meaning and and
22:07
talking about based on meaning and and
22:07
talking about based on meaning and and sort it out synonyms things like that
22:09
sort it out synonyms things like that
22:09
sort it out synonyms things like that and so under the covers it's going to
22:11
and so under the covers it's going to
22:11
and so under the covers it's going to vectorize data it's going to break it
22:13
vectorize data it's going to break it
22:13
vectorize data it's going to break it into chunks it's going to create
22:14
into chunks it's going to create
22:14
into chunks it's going to create associations assign numbers of
22:16
associations assign numbers of
22:16
associations assign numbers of similarity or dissimilarity and then go
22:18
similarity or dissimilarity and then go
22:18
similarity or dissimilarity and then go from there and this is really cool but
22:22
from there and this is really cool but
22:22
from there and this is really cool but if you have bad data in here you'll get
22:23
if you have bad data in here you'll get
22:23
if you have bad data in here you'll get bad associations the good news is this
22:26
bad associations the good news is this
22:26
bad associations the good news is this is kind of like a very basic example I
22:28
is kind of like a very basic example I
22:28
is kind of like a very basic example I like to use to show what vectorization
22:30
like to use to show what vectorization
22:30
like to use to show what vectorization looks like so you're basically taking
22:31
looks like so you're basically taking
22:31
looks like so you're basically taking this many dimensional model of
22:34
this many dimensional model of
22:34
this many dimensional model of relationships and obviously this is
22:35
relationships and obviously this is
22:35
relationships and obviously this is smaller than any data you have right you
22:37
smaller than any data you have right you
22:37
smaller than any data you have right you have more than eight things and seven
22:40
have more than eight things and seven
22:40
have more than eight things and seven things you associate them with you'll
22:42
things you associate them with you'll
22:42
things you associate them with you'll probably have hundreds thousands
22:43
probably have hundreds thousands
22:43
probably have hundreds thousands millions whatever but simplify it you
22:45
millions whatever but simplify it you
22:45
millions whatever but simplify it you look at this here and you see positive
22:47
look at this here and you see positive
22:47
look at this here and you see positive and negative numbers indicate similarity
22:49
and negative numbers indicate similarity
22:49
and negative numbers indicate similarity or dissimilarity if I were to be talking
22:52
or dissimilarity if I were to be talking
22:52
or dissimilarity if I were to be talking to an algorithm chapot whatever and
22:54
to an algorithm chapot whatever and
22:54
to an algorithm chapot whatever and getting back wrong answers because of
22:58
getting back wrong answers because of
22:58
getting back wrong answers because of this I can trace it back so tracing back
23:01
this I can trace it back so tracing back
23:01
this I can trace it back so tracing back back data from semantic search is
23:02
back data from semantic search is
23:02
back data from semantic search is actually not that bad because if I see
23:05
actually not that bad because if I see
23:05
actually not that bad because if I see that cat is associated into K9 and
23:07
that cat is associated into K9 and
23:07
that cat is associated into K9 and there's a high relevance factor for it
23:10
there's a high relevance factor for it
23:10
there's a high relevance factor for it I'm gonna say all right there's
23:12
I'm gonna say all right there's
23:12
I'm gonna say all right there's something wrong in my data there's no
23:14
something wrong in my data there's no
23:14
something wrong in my data there's no way that would happen otherwise let me
23:15
way that would happen otherwise let me
23:15
way that would happen otherwise let me go dig into it and figure out where that
23:17
go dig into it and figure out where that
23:17
go dig into it and figure out where that came from and solve it so the good news
23:19
came from and solve it so the good news
23:19
came from and solve it so the good news here is that if you see bad data here
23:21
here is that if you see bad data here
23:21
here is that if you see bad data here you can reverse engineer it solve it and
23:23
you can reverse engineer it solve it and
23:24
you can reverse engineer it solve it and be done U but you can't work around it
23:26
be done U but you can't work around it
23:26
be done U but you can't work around it though if there's bad data here it will
23:28
though if there's bad data here it will
23:28
though if there's bad data here it will continue associate badly in the future
23:30
continue associate badly in the future
23:30
continue associate badly in the future and K9 doesn't just mean dog can mean
23:32
and K9 doesn't just mean dog can mean
23:32
and K9 doesn't just mean dog can mean wolf it mean something else and so
23:33
wolf it mean something else and so
23:33
wolf it mean something else and so there's potential for more bad
23:35
there's potential for more bad
23:35
there's potential for more bad associations that I simply haven't hit
23:37
associations that I simply haven't hit
23:37
associations that I simply haven't hit on yet it's because I found one does not
23:39
on yet it's because I found one does not
23:40
on yet it's because I found one does not mean that more aren't there
23:43
mean that more aren't there
23:43
mean that more aren't there somewhere a few more fun ones here um
23:46
somewhere a few more fun ones here um
23:46
somewhere a few more fun ones here um before I leave some time here for
23:47
before I leave some time here for
23:47
before I leave some time here for questions because questions are fun fine
23:50
questions because questions are fun fine
23:50
questions because questions are fun fine tuning is something that we often misuse
23:53
tuning is something that we often misuse
23:53
tuning is something that we often misuse fine-tuning lets you take an algorithm
23:55
fine-tuning lets you take an algorithm
23:55
fine-tuning lets you take an algorithm and make it specific to a certain use
23:57
and make it specific to a certain use
23:58
and make it specific to a certain use case like for example let's say that I
24:00
case like for example let's say that I
24:00
case like for example let's say that I have my airline model and I like it but
24:02
have my airline model and I like it but
24:02
have my airline model and I like it but I want to make a couple of specific use
24:04
I want to make a couple of specific use
24:04
I want to make a couple of specific use cases out of that one for cargo and one
24:07
cases out of that one for cargo and one
24:07
cases out of that one for cargo and one for first class because those just tend
24:08
for first class because those just tend
24:08
for first class because those just tend to be very different from everything
24:10
to be very different from everything
24:10
to be very different from everything else I want them to kind of be tailored
24:12
else I want them to kind of be tailored
24:12
else I want them to kind of be tailored specifically to the crowds that will be
24:14
specifically to the crowds that will be
24:14
specifically to the crowds that will be interested in paying money for those
24:16
interested in paying money for those
24:16
interested in paying money for those services to really do this correctly
24:19
services to really do this correctly
24:19
services to really do this correctly though you have to spend a little bit of
24:20
though you have to spend a little bit of
24:20
though you have to spend a little bit of time understanding the use cases the
24:22
time understanding the use cases the
24:22
time understanding the use cases the business model and what you're trying to
24:24
business model and what you're trying to
24:24
business model and what you're trying to do an important key is that you can't
24:27
do an important key is that you can't
24:27
do an important key is that you can't find like you can't r your way out of
24:28
find like you can't r your way out of
24:28
find like you can't r your way out of bad data you can't uh unlearn it untrain
24:31
bad data you can't uh unlearn it untrain
24:31
bad data you can't uh unlearn it untrain it you can't prompt engineer your way
24:33
it you can't prompt engineer your way
24:33
it you can't prompt engineer your way you can't fine-tune your way out of bad
24:35
you can't fine-tune your way out of bad
24:35
you can't fine-tune your way out of bad data if you have a data problem fine
24:37
data if you have a data problem fine
24:38
data if you have a data problem fine tuning is not going to fix it the
24:39
tuning is not going to fix it the
24:39
tuning is not going to fix it the purpose of fine-tuning is to make a
24:41
purpose of fine-tuning is to make a
24:41
purpose of fine-tuning is to make a model more specific that's all it's
24:43
model more specific that's all it's
24:43
model more specific that's all it's supposed to allow it to handle a subset
24:46
supposed to allow it to handle a subset
24:46
supposed to allow it to handle a subset of use cases that are important to you
24:48
of use cases that are important to you
24:48
of use cases that are important to you under certain circumstances and fine
24:51
under certain circumstances and fine
24:51
under certain circumstances and fine tuning for that purpose is easy uh but
24:53
tuning for that purpose is easy uh but
24:53
tuning for that purpose is easy uh but it takes time and effort do it correctly
24:56
it takes time and effort do it correctly
24:56
it takes time and effort do it correctly it's not going to solve bad data and if
24:57
it's not going to solve bad data and if
24:57
it's not going to solve bad data and if you try to solve bad data here again
25:00
you try to solve bad data here again
25:00
you try to solve bad data here again you're saying hey model I'm giving you a
25:02
you're saying hey model I'm giving you a
25:02
you're saying hey model I'm giving you a purpose but by the way as part of that
25:04
purpose but by the way as part of that
25:04
purpose but by the way as part of that purpose avoid this and don't do this and
25:07
purpose avoid this and don't do this and
25:07
purpose avoid this and don't do this and and this should really be this that's
25:09
and this should really be this that's
25:09
and this should really be this that's not a purpose that isn't providing a use
25:12
not a purpose that isn't providing a use
25:12
not a purpose that isn't providing a use case that's confusing uh whenever you're
25:15
case that's confusing uh whenever you're
25:15
case that's confusing uh whenever you're not really sure if something you're
25:16
not really sure if something you're
25:16
not really sure if something you're doing is correct think about talking to
25:17
doing is correct think about talking to
25:17
doing is correct think about talking to a human being to ask them to do these
25:19
a human being to ask them to do these
25:19
a human being to ask them to do these things would it make sense or not and if
25:21
things would it make sense or not and if
25:22
things would it make sense or not and if it's total nonsense then it's probably
25:24
it's total nonsense then it's probably
25:24
it's total nonsense then it's probably nonsense to the algorithm as well so
25:26
nonsense to the algorithm as well so
25:26
nonsense to the algorithm as well so fine-tuning is to create specific use
25:29
fine-tuning is to create specific use
25:29
fine-tuning is to create specific use cases is not meant to solve that
25:33
cases is not meant to solve that
25:33
cases is not meant to solve that data here's a new one that's been coming
25:35
data here's a new one that's been coming
25:35
data here's a new one that's been coming up more and more and this has been
25:37
up more and more and this has been
25:37
up more and more and this has been proven through recent studies that are
25:39
proven through recent studies that are
25:39
proven through recent studies that are really interesting um many models have
25:42
really interesting um many models have
25:42
really interesting um many models have way many apps have ways to unlearn data
25:45
way many apps have ways to unlearn data
25:45
way many apps have ways to unlearn data basically say all right you know what I
25:47
basically say all right you know what I
25:47
basically say all right you know what I have an algorithm is doing great but I
25:48
have an algorithm is doing great but I
25:48
have an algorithm is doing great but I want to forget certain things there's
25:51
want to forget certain things there's
25:51
want to forget certain things there's bad data with pii there's copyrighted
25:53
bad data with pii there's copyrighted
25:53
bad data with pii there's copyrighted material there's there's something I
25:54
material there's there's something I
25:54
material there's there's something I just don't want it to ever talk about so
25:57
just don't want it to ever talk about so
25:57
just don't want it to ever talk about so remove it and the reality of this right
26:00
remove it and the reality of this right
26:00
remove it and the reality of this right now is that all the unlearning methods
26:01
now is that all the unlearning methods
26:01
now is that all the unlearning methods that are out there right now are pretty
26:02
that are out there right now are pretty
26:02
that are out there right now are pretty rudimentary um they might kind of work
26:05
rudimentary um they might kind of work
26:05
rudimentary um they might kind of work but there's more of a chance they're
26:06
but there's more of a chance they're
26:06
but there's more of a chance they're gonna harm your model than help it it's
26:08
gonna harm your model than help it it's
26:08
gonna harm your model than help it it's very much like me saying one day like
26:10
very much like me saying one day like
26:10
very much like me saying one day like you know what last week was horrible I
26:12
you know what last week was horrible I
26:12
you know what last week was horrible I had a terrible week everything went
26:13
had a terrible week everything went
26:13
had a terrible week everything went wrong I want to go in my brain and pull
26:15
wrong I want to go in my brain and pull
26:15
wrong I want to go in my brain and pull out all the neurons from last week and
26:17
out all the neurons from last week and
26:17
out all the neurons from last week and make all those memories go away um can
26:19
make all those memories go away um can
26:19
make all those memories go away um can you try to do that I'm sure there's some
26:21
you try to do that I'm sure there's some
26:21
you try to do that I'm sure there's some way to try to do that in science right
26:24
way to try to do that in science right
26:24
way to try to do that in science right now I'm sure it'll have negative
26:26
now I'm sure it'll have negative
26:26
now I'm sure it'll have negative repercussions that I really really
26:27
repercussions that I really really
26:27
repercussions that I really really wouldn't like
26:28
wouldn't like
26:28
wouldn't like so it's this is just technology that's
26:31
so it's this is just technology that's
26:31
so it's this is just technology that's getting there it's not there yet if you
26:34
getting there it's not there yet if you
26:34
getting there it's not there yet if you try to do unlearning be very cautious
26:37
try to do unlearning be very cautious
26:37
try to do unlearning be very cautious check save back up everything you're
26:39
check save back up everything you're
26:39
check save back up everything you're doing before you make anything permanent
26:41
doing before you make anything permanent
26:41
doing before you make anything permanent change to it and be cautious unlearning
26:44
change to it and be cautious unlearning
26:44
change to it and be cautious unlearning especially you unlearn more data will
26:46
especially you unlearn more data will
26:47
especially you unlearn more data will very often result in bad things
26:49
very often result in bad things
26:49
very often result in bad things happening if you need to get rid of data
26:51
happening if you need to get rid of data
26:51
happening if you need to get rid of data remove it from the data source if your
26:53
remove it from the data source if your
26:53
remove it from the data source if your rag data has stuff in it you don't want
26:55
rag data has stuff in it you don't want
26:55
rag data has stuff in it you don't want to bring anymore um remove the data
26:58
to bring anymore um remove the data
26:58
to bring anymore um remove the data present a new set of data and go from
27:01
present a new set of data and go from
27:01
present a new set of data and go from there that's far better than trying to
27:02
there that's far better than trying to
27:02
there that's far better than trying to tell it after the fact exclude data
27:04
tell it after the fact exclude data
27:04
tell it after the fact exclude data unlearn data make it go away for
27:06
unlearn data make it go away for
27:06
unlearn data make it go away for training data this is even more
27:08
training data this is even more
27:08
training data this is even more important uh you've built this model you
27:11
important uh you've built this model you
27:11
important uh you've built this model you built the way it behaves you've worked
27:13
built the way it behaves you've worked
27:13
built the way it behaves you've worked with it you like it you you've adjusted
27:15
with it you like it you you've adjusted
27:15
with it you like it you you've adjusted it now to go and begin ripping stuff out
27:17
it now to go and begin ripping stuff out
27:17
it now to go and begin ripping stuff out after the fact it may have unintended
27:20
after the fact it may have unintended
27:20
after the fact it may have unintended consequences so I highly
27:22
consequences so I highly
27:22
consequences so I highly recommend caution about unlearning data
27:25
recommend caution about unlearning data
27:25
recommend caution about unlearning data because you don't really know what the
27:28
because you don't really know what the
27:28
because you don't really know what the results are going to be like and in
27:30
results are going to be like and in
27:30
results are going to be like and in current methods current models a lot of
27:32
current methods current models a lot of
27:32
current methods current models a lot of harm can be done and so I really just
27:34
harm can be done and so I really just
27:34
harm can be done and so I really just caution caution here um before doing it
27:38
caution caution here um before doing it
27:38
caution caution here um before doing it these will get better with time I'm sure
27:40
these will get better with time I'm sure
27:40
these will get better with time I'm sure they will but there's always a danger in
27:42
they will but there's always a danger in
27:42
they will but there's always a danger in telling a model to forget stuff because
27:44
telling a model to forget stuff because
27:44
telling a model to forget stuff because it may forget more than you want it to
27:47
it may forget more than you want it to
27:47
it may forget more than you want it to or it may forget less or it may forget
27:50
or it may forget less or it may forget
27:50
or it may forget less or it may forget who knows what you may try to get rid of
27:52
who knows what you may try to get rid of
27:52
who knows what you may try to get rid of all the pii and accidentally get rid of
27:53
all the pii and accidentally get rid of
27:53
all the pii and accidentally get rid of addal information you need uh or you may
27:56
addal information you need uh or you may
27:56
addal information you need uh or you may not get rid of all of it and then you'll
27:57
not get rid of all of it and then you'll
27:57
not get rid of all of it and then you'll still give to people inadvertently but
28:00
still give to people inadvertently but
28:00
still give to people inadvertently but confidently after the
28:02
confidently after the
28:02
confidently after the fact so I want to wrap up I have a few
28:04
fact so I want to wrap up I have a few
28:04
fact so I want to wrap up I have a few minutes or questions at the end um this
28:07
minutes or questions at the end um this
28:07
minutes or questions at the end um this this presentation really had a lot of
28:09
this presentation really had a lot of
28:09
this presentation really had a lot of information in it that was pre-existing
28:11
information in it that was pre-existing
28:11
information in it that was pre-existing that existed long before AI was talked
28:13
that existed long before AI was talked
28:13
that existed long before AI was talked about the public it's been around for
28:15
about the public it's been around for
28:15
about the public it's been around for years and a lot of newer things as well
28:17
years and a lot of newer things as well
28:17
years and a lot of newer things as well and it's all the same it all relates
28:19
and it's all the same it all relates
28:19
and it's all the same it all relates we've been using data for analysis for a
28:21
we've been using data for analysis for a
28:21
we've been using data for analysis for a long time it's not going to change uh
28:24
long time it's not going to change uh
28:24
long time it's not going to change uh all the new algorithms out there are
28:25
all the new algorithms out there are
28:25
all the new algorithms out there are really just an extension of things we've
28:27
really just an extension of things we've
28:27
really just an extension of things we've already had
28:28
already had
28:28
already had and this will keep happening and keep
28:30
and this will keep happening and keep
28:30
and this will keep happening and keep going on and keep going on and so just
28:32
going on and keep going on and so just
28:32
going on and keep going on and so just keep in mind the best place to solve bad
28:35
keep in mind the best place to solve bad
28:35
keep in mind the best place to solve bad data is to do it at the source solve it
28:38
data is to do it at the source solve it
28:39
data is to do it at the source solve it early cut it off before it can go
28:41
early cut it off before it can go
28:41
early cut it off before it can go Downstream uh AI you know messing with
28:43
Downstream uh AI you know messing with
28:43
Downstream uh AI you know messing with your AI models can do great things for
28:45
your AI models can do great things for
28:45
your AI models can do great things for them but it won't be a substitute for
28:47
them but it won't be a substitute for
28:47
them but it won't be a substitute for good data it can't solve good bad data
28:49
good data it can't solve good bad data
28:50
good data it can't solve good bad data it can't make your data better you can
28:51
it can't make your data better you can
28:51
it can't make your data better you can try but you're far better better off
28:54
try but you're far better better off
28:54
try but you're far better better off solving it earlier in your processes
28:56
solving it earlier in your processes
28:56
solving it earlier in your processes once you're into a model once you're
28:58
once you're into a model once you're
28:58
once you're into a model once you're testing it once you're working with it
28:59
testing it once you're working with it
28:59
testing it once you're working with it once you're Qing it once you're letting
29:00
once you're Qing it once you're letting
29:00
once you're Qing it once you're letting people use it for real keep testing
29:03
people use it for real keep testing
29:03
people use it for real keep testing carefully and check for responses
29:04
carefully and check for responses
29:05
carefully and check for responses whenever bad responses happen find the
29:07
whenever bad responses happen find the
29:07
whenever bad responses happen find the source of it was it bad training was it
29:09
source of it was it bad training was it
29:09
source of it was it bad training was it bad rag data was it some changes you
29:11
bad rag data was it some changes you
29:11
bad rag data was it some changes you made that backfired figure out what the
29:13
made that backfired figure out what the
29:13
made that backfired figure out what the source is and resolve the source this is
29:15
source is and resolve the source this is
29:15
source is and resolve the source this is the best way of solving problems in your
29:17
the best way of solving problems in your
29:17
the best way of solving problems in your models that really really will make
29:20
models that really really will make
29:20
models that really really will make things better so I want to stop here I'm
29:23
things better so I want to stop here I'm
29:23
things better so I want to stop here I'm going to provide a little bit of
29:24
going to provide a little bit of
29:24
going to provide a little bit of information about me this will be in the
29:26
information about me this will be in the
29:26
information about me this will be in the slides later that get shared I'll share
29:28
slides later that get shared I'll share
29:28
slides later that get shared I'll share them afterwards so you have them are
29:30
them afterwards so you have them are
29:30
them afterwards so you have them are there any questions
29:35
[Music]