►
From YouTube: APAC Hybrid Cloud Kopi Hour (E6) | AI in APAC!
Description
Join us for the next session of Kopi hour where we look at some of the exciting developments in AI! We are joined by Red Hat APAC CTO Vincent Caldeira to dive into all things artificial! We’ll learn about what makes AI work, how businesses in APAC are using it today, and how Red Hat’s strategy is all in for AI!
Vincent's Slides: https://speakerdeck.com/redhatlivestreaming/apac-hybrid-cloud-kopi-hour-openshift-ai-20230810
Foundation Models demo: https://github.com/caldeirav/foundation-models-demo
00:00 Stream begins ...
03:47 AI in popular culture
06:25 GenAI and Foundational Models 101
33:45 Live demo!
01:05:34 Red Hat OpenShift AI Strategy
A
A
Is
this
wonderful,
coffee
beverage
you
get
in
Singapore
that
our
current
guest
probably
knows
a
lot
about
so
I
have
my
Kopi
mug
here
full
of
Nescafe,
but
it
is
what
it
is:
I
am
August
seminelli
I'm,
a
technical
marketing
manager
here
at
red
hat
and
as
always,
I
am
joined
by
Dev
Sean
bag,
principal
architect,
Dev,
say
hello.
Today,.
B
Hi
everyone
thanks
August
for
the
warm
welcome
another
exciting
show
of
copy
hour
today.
Unfortunately,
my
coffee
quota
for
today
is
done
so
I'm
staying
with
water.
However,
you
know
very
excited
to
be
here,
and
you
know
listen
to
all
that
our
guest
has
to
say
today
a
true
Trailblazer
in
the
technology
landscape.
B
B
A
C
C
So
generally
is
probably
the
best
Brew
for
us
program.
Also
having
late
night,
you
know
session
trying
to
deal
with
fixing
bugs
or
or
getting
software
out
and
yeah.
You
know
very
happy
to
be
here
and
And
discussing
the
exciting
topic
of
artificial
intelligence
and
in
particular
the
recent
Trend
engineering
TV.
Yes,
absolutely.
A
Well,
and
before
we
dive
into
the
topic,
one
thing
I
have
to
notice
came
across
my
LinkedIn
feed,
so
you
were
named
one
of
the
top
10
Chief
technology
officers
in
APAC
by
technology
magazine
and
I.
I.
Think
it's
really
neat
I
mean
it
really
shows
your
own
abilities
and
skills
and
to
be
part
of
things,
but
I'm
really
proud,
as
a
red
Hatter
to
have
have
you
on
that
list.
It's
really
a
wonderful
thing.
It's
it's
really
cool.
C
Yeah
I
mean
thank
you
I
think
you
know
to
me
is
I
think
what's
important,
is
I'm
very
passionate
about
technology
and
and
evangelizing
more
points
of
solution,
in
particular,
so
I
think
to
me
the
you
know,
the
greatest
recognition
is:
is
that
probably
some
people
out
there
in
the
region?
They
actually
like
to
hear
my
my
message
and,
and,
and
you
know,
as
a
result,
we
have
this
opportunity
to
to
meet
more
people
and
and
share
what
we
do
at
hadat,
which
of
course,
yeah.
C
A
Know
a
lot
right.
Our
company,
Red
Hat,
obviously
is
deeply
involved
and
we're
going
to
learn
more
about
that.
But
I
wanted
to
look
at
like
what
I
actually
kind
of
my
experiences
with
AIS
AI
is
so
I
thought
I'll
do
what
everyone
does
right.
I
went
to
chat,
GPT
and
I
said
you
know
I'm
going
to
get
it
to
do.
My
job
for
me,
write
me
a
fun
but
short
introduction
to
a
podcast
or
live
stream
about
Ai
and
so
chat.
A
Gpt
wrote
this
for
me
and
there
you
go
unravel
the
Mysteries,
correct
the
jokes
and
decode
the
geeky
magic
of
artificial
intelligence
and
I
thought.
That's
interesting.
How
does
it
do
that?
So
then,
I
looked
at
another
one,
I
went
to
Bard
and
I
said
I'm
going
to
ask
it
the
same
question,
and
this
is
how
Bard
responded
to
me:
welcome
to
the
AI
podcast
I'm,
your
host
Bart,
a
language
model
from
AI
Google
AI,
and
then
it
made
a
joke.
It
made
a
joke
that
I
actually
didn't
really
get
so.
A
Know
if
it
was
trying
to
make
the
point
that
I
had
no
clue,
but
it
was
interesting
to
see
how
I
asked
a
robot
the
same
thing
and
got
different
responses
and
then
finally
I
thought.
Well
what
else
do
custody
do?
People
find
Ai
and
they
find
it
in
image
generation,
so
I
went
and
onto
leonardo.ai
and
I
asked
it
to
design
me.
A
Someone
standing
in
the
rain,
looking
mysterious
wearing
a
red
hat,
and
this
is
what
I
got,
which
of
course,
sent
me
down
a
rabbit
hole
that
I
just
wanted
to
show
off
before
I
actually
hear
the
background
of
it.
I
stuck
Elon
Musk
I,
stuck
Ruth
Bader
Ginsburg
in
a
red
hat
that
wasn't
enough,
so
I
decided
to
stick
a
kangaroo
in
a
red
hat.
A
So,
what's
the
point
of
all
that,
it's
not
just
to
show
off
my
amazing
artistic
talent,
but
to
say
what
is
AI
and
and
how
is
it
doing
that
Vincent?
How
is
it
taking
this
type
of
stuff
and
coming
up
with
pictures
of
random
people
or
a
joke
about
synapses
that
I
don't
get?
How?
How
does
this
happen
and
that's
where
I
wanted
to
to
leave
it
with
you
to
educate.
C
Yeah,
absolutely
so
I
mean
spot
on.
What's
really
what
I
wanted
to
focus
on
today,
I'm
by
no
means
a
data
scientist,
so
I'm
already
a
technologist
who's,
trying
to
really
understand
how
AI
thinks
kind
of
come
about
in
our
life
and
I
think
we
will
be
spending
a
bit
of
time
today
to
really
go
through.
I
would
say
initial
understanding
of
how
generative
tvi
comes
about
right
and
and
we'll
go
through
some
simple
demo
as
well.
So
maybe,
let's
get
started.
A
C
Generally
people
are
not
so
clear
in
terms
of
their
understanding,
so
we
start
with
a
little
bit
of
a
primer
on
a
iron
Foundation
model,
and
then
you
know
I
I'll
stop
a
while
with
slides
to
to
just
go
into
a
little
demo
to
explain
some
of
the
concept
so
where
I
wanted
to
start
is
when
you
look
at
generative
AI,
you
can't
really
understand
this
without
the
broader
context
and
and
the
whole
AI
Journey
behind
it
yeah.
So
so,
let's
try
to
define
a
few
terms
before
we
proceed.
C
So
AI
is
really
the
generic
term,
which
is
the
capability
for
computer
system
to
replicate
or
mimic
human
cognitive
function.
Machine
learning
is
really
a
subset
of
AI
which
focuses
on
this,
particularly
an
application
of
AI
in
particular.
So
if
you
look
at
machine
learning
today,
whether
or
not
you
realize
it's
pretty
much
all
around
us
and
very
pervasive
already,
it's
used
in
prediction
engine
a
lot
of
prediction
system.
Very
basically,
look
at
you
know
using
existing
data
compute
the
probability
of
some
event
happening.
C
You
know
like
a
system
failure
and
then
we
pretty
much
generate
predictions
based
on
probability
recommendation
engine
which
we,
you
know
we
basically
use
in
our
daily
lives.
Indirectly.
You
know
you
watch
Netflix,
you,
like
a
show,
it's
going
to
recommend
more
show
for
you,
you
go
and
purchase
some
good
on
Amazon
is
going
to
use
other
people's
purchasing
patterns
and
data
to
to
suggest
as
well.
Some
purchase
that
you
may
like
is
using
financials.
C
So
when
you
go
abroad
and
you
use
your
credit
card
to
to
make
a
purchase
and
basically
the
system
is
not
expecting
you
to
be
spending
money,
invest
particular
country,
it
may
be
stopped
automatically
and
you
may
get
a
message
from
your
bank
and
it's
even
using
very
much
as
a
business
tool.
I
mean
personally
I
I
spent
about
20
years
of
my
life
in
banking.
We
are
now
heavily
using
machine
learning
for
what
we
call
algorithmic
trading,
which
is
really
automating.
C
A
C
Yeah
so
I
mean
essentially
machine
learning
uses
an
algorithm.
You
know,
of
course
the
algorithm
itself
depends
on
the
types
of
machine
learning
models
that
you
use.
So
you
know
it's
a
bit
complex
right,
but
you
know
I
would
say
at
the
what
was
very
important
is
to
understand
at
least
the
the
big
categories
of
of
learning.
C
For
example,
you
have
supervised
learning
where,
typically,
what
we
do
is
we
have
data
that
we
label,
and
so
we
provide
this
data
as
an
input
and
we
have
an
output
that
is
known
and
we
vertically
try
to
find
the
best
algorithm
to
generate
the
output.
So,
typically
how
it's
done
is
you
know
you
build
an
algorithm
based
on
some
data,
and
then
you
have
some
other
data
that
you
use
for
valuation
for
evaluation.
C
Yes,
and
so
you
basically
run
this
algorithm
to
increase
the
the
predictive
ability
of
the
output,
which
is
basically
what
we
call
supervised
learning
now,
if
you
compare
this
with
unsupervised
learning,
it's
again
another
form
of
machine
learning,
but
in
this
case
the
model
is
trained
on
unlabeled
data
and
you
basically
let
the
model
to
discover
the
pattern
on
its
own
and
to
predict
some
kind
of
output.
So
so,
in
that
case,
you
actually
use
pretty
different
technique
and
it's
obviously
used
for
different
types
of
of
needs.
C
Right
so
example
of
supervised
learning,
pretty
simple
one!
You
do
sentimental
analysis
right.
You
have
maybe
some
announcements
on
the
radio
or
attacks
just
telling
you
whether
it's
a
positive
or
negative
sentiment,
they
have
a
supervised
learning,
could
come
in
in
all
daily
lives.
We
use
pan
filtering
for
emails.
You
know
how
does
the
system
detect?
What
is
Spam
is
using
supervised
learning
to
do
this
right,
feeding
a
lot
of
emails
and
then
literally
like
having
enough
data
to
show
what's
what's
a
spam?
C
On
the
other
hand,
unsupervised
learning
is
basically
a
little
bit
more
loose.
An
example
of
this
is,
you
know,
for
example,
marketing
campaign
yeah,
you
you
look
at
patterns
of
a
behavior
or
purchase
of
your
consumer
and
when
you
actually
ask
the
system,
could
you
segment
customer
in
two
different
buying
patterns
for
me,
because
then
the
outcome
will
actually
be
different
depending
on
the
patterns.
You
identify
right.
So
so
that's
that's
pretty
much.
You
know
just
what
machine
learning
can
do
and
the
high
level
categories
of
training.
B
Just
another
question
to
you:
Vincent:
so,
as
you
just
explained
unsupervised
and
supervised,
so
unsupervised
is
more
a
powerful
way
of
you
know
doing
things
and
supervised
days.
You
know
more
similar
to
you
know
you
inputting
something
and
then
getting
some
back.
Is
it
similar
to
you
know
the
weak,
Ai
and
strong
AI,
where
you
know
Rick
AIS.
Probably
everyone
knows
you
know
you
put
that
put
something
in
a
chat
Jeopardy
and
it
throws
things
out,
whereas
strong
AI
is
something.
For
example,
autonomous
cars
am
I
going
the
right
direction.
C
Well,
I
think
you
know
it's
a
little
bit
more
complex
on
this
I
think
you
know,
weak
and
strong.
Ai
refer
more
to
I
would
say
the
the
ability
to
to
kind
of
predict
the
the
result
and
and
to
prove
it.
It's
not.
You
know
it's
it's
not
going
to
have
exactly
the
same
definition
depending
on
the
type
of
model
you
look
at
yeah,
I,
I.
C
Think
to
me
the
way
the
way
I
I
kind
of
look
at
it
is
those
are
just
different
approach
of
of
teaching
a
system
and
learning
they
just
applied
to
different
type
of
problem
yeah.
So
so,
once
you
provide
learning
in
particular
is
really
when
you
are
kind
of
trying
to
make
sense
of
data
and
to
come
with
categories
yeah
and
and
and
so
there
will
never
be
an
actual
right
or
wrong
algorithm
right.
C
It
really
depends
on
your
tolerance
and
how
you
want
these
classifications
to
happen,
and
what
pretty
much
depends
on
the
use
case
yeah,
while
supervised
learning,
you
can
always
kind
of
Benchmark
a
pretty
easy.
The
model
performance
because
you
know
the
expected
output
in
the
case
of
unsupervised
learning
you
actually
don't
reboot
so,
but
but
what's
the
difference
for
me
right
so
so
the
quality
of
the
result
depends
on
actually
your
usage
right.
It's
not
so
easy
to
to
to
kind
of
predict
like
what
is
kind
of
a
good
outcome
or
not
yeah.
B
C
So
that's
machine
learning
and
well,
it
gets
a
bit
more
complex
now,
but
thankfully,
all
these
concepts
are
related
is
weaving
machine
learning.
Now
we
have
the
concept
of
deep
learning.
So
what
what's
learning
deep
learning
does?
Is
it's
trying
to
really
replicate
I
would
say
the
architecture
of
a
human
brain.
So
we
talk
about
neural
network
and
the
idea
here
is:
we
can
actually
connect
a
lot
of
layers
of
different
neurons
that
can,
you
know,
recognize
and
analyze
pattern.
C
So,
instead
of
having
like
a
straight,
you
know
throughput
driver,
oh
yeah,
you
give
inputs,
you
get
output
here,
we
have
potentially
a
high
number
of
steps
that
are
all
helping
us
to
solve
much
more
complex
problems
that
are
not
linear
in
nature
yeah.
So
deep
learning
can
basically
do
this.
Now.
The
benefit
of
deep
learning
is
because
of
this
I
would
say,
architecture
that
mimics
the
human
brain.
We
can
improve
accuracy
and
performance
by
quite
a
lot.
We
reduce
the
need
of
what
we
call
feature
engineering,
which
is
in
traditional
machine
learning.
C
Models
are
very
heavy
type,
so
feature
engineering
is
basically
the
technique
to
refine
the
data
and
the
data
quality
to
produce
new
features
as
input.
So
you
basically
improve
the
accuracy
of
model
in
the
case
of
deep
learning,
a
lot
of
this
work
can
pretty
much
be
delegated
to
the
model
and,
of
course,
it
improves
the
scalability
and
effectivity
for
solving
problem,
and
so
so
when
we
look
at
Deep
learning,
this
is
really
the
base
that
has
enabled,
if
I
may
say,
generative
AI
without
deep
learning,
there's
no
General
TBI.
C
So
now
we
take
deep
learning
and
what
is
really
generative
in
the
context.
Again,
we
are
talking
about
a
subset,
because
generative
AI
is
solely
focused
on
bidding
system
that
can
generate
new
data.
It
can
be
text,
it
can
be
image,
it
can
be
video,
it
can
be
audio
as
well
and
V
Systems.
They
are
not
explicitly
taught
any
language
or
grammar
rules
or,
if
you,
you
know,
if
you
talk
about
image
like
image,
title
definition,
what
we
do
is
we
give
a
massive
amount
of
data,
whether
it's
text
or
image.
C
A
C
I
mean
exactly
so,
but
let's,
let's
put
on
that,
was
also
you
know.
One
of
my
questions
initially
when
I
started
to
look
at
this
domain
is
how
how
recent
is
this
right
was
it?
Was
generic
TV
I
really
born
I
have
a
chat
GPT.
So
so
let
me
go.
You
know
for
a
little
bit
of
history
there
how
how
we
got
well,
so
First
Foundation
models
were
only
defined
in
2021.
C
Oh
and
it's
really
genome
event
supervised
learning
and
it
can
be
adapted
to
accomplish
a
broad
range
of
tasks
right.
So
so
one
element
of
the
foundation
model
is:
they
are
not
specific
in
nature.
I
give
a
lot
of
example,
of
machine
learning
earlier
and
as
you've
seen.
A
lot
of
traditional
machine
learning
is
really
much
towards
a
very,
very
specific
output,
and
if
you
need
to
solve
a
different
problem,
you
very
much
need
to
come
up
with
another
model
Foundation
model
by
Nature.
C
C
The
paper
is
called
attention
is
all
you
need,
and
basically
what
the
Google
researcher
were
trying
to
do
is
to
build
a
neural
network
that
understand
the
context
and
the
meaning
of
relationship
right.
So,
for
example,
if
you
have
an
English
sentence,
yeah,
if
you
as
a
human,
you
read
the
sentence,
you
will
know
that
certain
words
related
to
each
other.
How
can
we
actually
teach
a
model
to
understand
this
collection
ship
yeah
to
track
both
relationship
in
sequential
data
for
just
reading
a
lot
of
similar
data
so
Transformer
model?
C
They
basically
use
mathematics
and
probably
not
good
enough
to
understand
you
know
the
actual
implementation
or
or
this
right,
but
they
use
mathematological
techniques
that
are
called
self-attention.
So
this
technique.
Basically,
they
help
to
to
detect
correlations
between
data
elements
and
how
they
are
related
to
each.
A
C
Exactly
right,
so
so
forcing
many
sentence,
you
start
to
realize
that
hey!
You
know
this
verb
is
related
to
this
subject
and
yeah
in
you
know
the
world
eating
this.
You
know
context.
This
is
what
it
means
right.
It's
essentially
if
it's
building
correlation
in
in
sentence
right,
so
so
pretty
pretty
interesting
concept,
so
so
so,
based
on
this
kind
of
concept,
it
basically
inspired
the
creation
of
large
language
model,
the
first
one
being
birth
in
2018,
which
I
think
is
probably
a
foundational
moment
of
natural
language
processing.
C
And
so
again
it
was
what
is
by
Google
as
an
open
source
software
and
it
really
much
started
the
whole
loud
language
model.
I
would
say
a
field
of
research
and
everything
that
we've
seen
until
today,
with
with
chai
GPT
yeah.
Obviously
Google
did
it
for
a
very,
very
simple
reason,
but
should
be
obvious
to
all
of
us.
He
wanted
to
make
sure
that
in
the
Google
search
engine
you
can
actually
type
a
sentence.
Yeah
and
I'll
just
type
a
keyword,
a
Google
search
engine
will
actually
understand.
C
It
makes
a
lot
of
sense
right
and
I
I
kind
of
look
back
myself
and
I'm
like
hey
yeah.
Actually
there
was
really
this
point
when
suddenly
Google
search
engine
was
able
to
to
return
some
meaningful
data
when
I
put
the
whole
sentence,
as
opposed
to
somehow
select
some
of
the
world
and
return
crap
yeah
yeah,
so
so
I
kind
of
found
out
that
hey.
C
C
So
of
course,
because
it
was
made
highly
available
to
everyone
to
test.
Then
it
started
to
capture
the
the
the
attention.
Mostly
the
I
would
say
what
GPT
3
brought
is
that
this
was
the
first
large
language
model
trained
on
such
a
a
huge
amount
of
data
about
175
billion
parameters.
Yeah
that
had
not
been
done
before
so
I
think
to
me
is
the
the
architecture
was
actually
not
new,
but
the
scale
I
was
and
scale
really
made
a
big
difference.
A
C
So
you
know
it:
it's
definitely
not
cheap,
so
you
kind
of
need
the
podcast
yes
to
do
the
training.
I
think
it
has
been
a
lot
of
kind
of
Phenom
discussion
recently
about
the
the
total
cost
of
trading
Your
Own
Foundation
model
yeah.
It's
not
cheap,
but
also
that's
why
Foundation
models
are
so
exciting
because
you
I
mean
I
would
say
in
most
business
case,
you
don't
need
to
retrain
them.
You
just
need
to
fine-tune
them
or
give
them
additional
data
yeah,
but,
but
essentially
so
that
you
know
the
the
gpt3.
C
Essentially,
the
average
architecture
of
the
previous
model,
which
is
gpt2
gpt2,
was
a
what
we
call
an
auto
regressive
model
based
on
the
Transformer
architecture.
But
what
was
new
compared
to
a
Perth
model
is
basically
the
pre-training
approach
by
GPT
to
use,
instead
of
just
pushing
a
lot
of
data
and
data
set.
Basically,
in
the
training
stage,
the
researchers
started
to
provide
contextual
instruction,
so
they
basically
shared
with
the
model
and
show
to
the
model
what
good
likes.
If
you
want
right,
so
that
was
kind
of
a
new
pre-training
technique.
C
That
was
giving
a
lot
better
result,
and
we
realized
that,
with
this
technique,
the
more
we
actually
scale
with
more
data
and
the
the
the
the
closer
to
a
human
response.
The
model
gets
so
so
they're
pretty
much
started
to
work
just
on
scale
yeah,
so
distributed
training
just
get
huge
access
to
a
lot
of
architecture,
I
mean
we
will
see
later.
Why
this
is
interesting
in
the
concept
of
highlight
is
because
these
need
to
train.
A
C
But
I've
never
kind
of
trained
my
my
own
model,
you
know
end-to-end
on-prem,
we're
doing
mostly
fine
tuning
right
from
the
development
angle.
I
actually
don't
know
how
much
space
and
20
175
billion
parameter
would
take
five,
but
but
I
have
her
training
costs
in
the
range
of
a
million
on
US
dollar,
just
to
run
right
or
to
basically
train
a
foundation
model
of
that
type
of
skill.
So
that
gives
you
an
idea
of
of
how
much
infrastructure
probably
will
be
required
right.
We're
just
talking
about
massive
amount
of
gpud
project
scale.
B
C
Yeah
so
I
mean
you
know,
I
think
that's
interesting.
The
type
of
model
we
see
emerging
I
think
there
will
be
models
of
as
a
service
I
mean,
of
course
charge.
Gpt
is
the
most
famous
one,
but
I
mean,
let's
be
honest
here:
I,
don't
see
a
lot
of
Enterprise,
actually
leveraging
it
in
the
context
of
a
managed
service.
Why?
Because
you
know
essentially
simulator
has
been
built
for
I,
would
say
kind
of
consumer
usage
right.
C
It's
not
really
specific
to
Enterprise
use
case
and
also
you
know,
Enterprise
really
have
problems
with
sharing
their
data,
although
essentially
you
know,
of
course
open.
Ai
is
saying
that
we
don't
take
the
the
data
you
share
in
a
prompt
to
train
their
own
model.
I
mean
this
being
said.
You
are
still
pretty
much
sending
the
data
to
them
through
some
kind
of
API
call.
So
you
have
like
nowhere
to
verify.
C
What's
really
done
with
it
and
when
you
add
italyx,
filtrating
information,
but
maybe
sensitive
and
in
the
process
you
may
be
violating
you
know,
data
residency
laws
or
or
this
sensitivity,
policies
that
you
have
internally
and
so
so
I
think
the
way
I
look
at
it
is
you
know,
and
that's
what
as
well
I
I
want
people
to
really
kind
of
keep
in
their
mind
is
foundation
model.
Is
this
huge
opportunity
for
people
to
leverage
big
investment
in
infrastructure
and
scale
that
the
the
common
individuals
or
organizations
will
not
have
it's
the
point?
C
You
don't
have
to
actually
redo.
Everything
is
opening
fantastic
opportunity
as
well,
for
for
open
source
to
build
not
just
model,
but
also
to
contribute
open
source
data
that
can
be
used
to
to
train
this
model
in
I
would
say
in
a
transparent
way,
yeah.
So
so,
let's
be
reminded
of
one
of
the
key
benefits
of
Open
Source
is
not
just
that
a
is
free
software,
and
so
on
is
that
it
gives
you
visual
transparency
on
the
process.
C
Right,
you
can
see
the
code,
you
can
see
what
data
has
been
used
to
train
the
model
and
then,
if
you
need
to
basically
improve
it
or
fine
tune
it,
you
can
pretty
much
take
it
and
and
start
from
them,
and
so
that's
the
way.
I
look
at
it.
From
my
perspective,
a
lot
of
the
discussions
I
have
with
a
Reddit
customer.
C
C
So,
let's
go
maybe
enough
for
some
implementation,
so
what
what
I
really
want
to
get
there
is
give
a
sense
of
I
would
say
the
steps
and
the
key
terms
involved
as
well,
and
what
you
see
here
is
really
I
would
say,
a
developers
and
quick
point
of
view:
yeah
yeah
most
developers
today.
What
they
do
is
basically
what
we
call
in
context.
C
Learning,
which
is
you
know,
basically
means
that
we
use
a
love
language
model
of
the
Shelf,
and
then
they
control
the
behavior
through
some
capability,
which
we
call
the
prompting,
as
well
as
how
they
actually
ask
the
question,
which
means
the
contextual
data
that
they
actually
pass.
So
there's
kind
of
a
few
sequential
steps
in
this
process
that
are
worth
knowing.
This
is
what
I
try
to
to
demo
very
quickly.
Just
after
this.
C
The
first
stage
is
really
a
data
pre-processing,
a
step
so
in
by
step,
we'll
use
what
we
call
embeddings
to
store
private
data
that
is
used
by
the
model
to
understand
better
your
query.
So
an
embedding.
You
have
to
see
it
as
a
way
to
represent
complex
information
like
text
in
what
we
call
a
vector,
and
so
this
Vector
basically
will
be
stored
in
a
way,
but
the
system
can
actually
quickly
identify
relevant
piece
of
text
and
information
based
on
the
query,
but
by
using
you
know,
obviously,
mathematics
and
some
score.
C
That
indicates
the
hololens
yeah.
So
so
generally,
the
first
step
we
look
at
is:
let's
see
what
data
we
want
to
to
to
give
to
that
model,
so
it
can
provide
us
with
a
better
answer
and
the
way
of
doing
this
is
what
we
call
was
embeddings
now.
The
second
step
is
really
what
we
call
the
pond
construction
or
retrieval.
So
when
you
submit
the
query,
you
can
basically
give
some
instructions
for
the
model
to
return
an
answer.
C
So
you've
probably
heard
about
prompt
engineering,
prompt
engineering
is
really
the
art
of
Designing,
a
text
that
helps
the
model,
get
a
better,
more
compliant
output
with
what
you
expect
yeah.
Ideally,
it's
concise
and
it
increased
the
value
of
of
the
response
by
filtering
the
inappropriate
ones.
Yeah.
C
Yeah
correct
so
yeah,
so
you
know
it's
it's
very
much
still
and
out,
but
you
know
if
you
are
interested
I
was
reading
an
article
recently
on
you
know.
The
future
of
generality
tvi
is
not
prompt
engineering,
which
I
found
quite
interesting,
which
basically
was
explaining
that
you
know
all
the
time
as
we
prompt
them.
The
model
will
will
also
have
the
ability
to
supervise
themselves
because
they
can
learn
from
our
prompting.
C
Yeah
I
mean
exactly
right.
You
could
be
like
like
models
that
are
personal
and
then
we
have
some
kind
of
capability
or
prompting,
but
you
know
some
some
could
be
a
more
aggressive
prompter.
Some
could
be
like
you
know,
maybe
more
dedicated
I,
don't
know
right,
it's
it
it's
kind
of
interesting,
but
it
is
it's
really
a
topic.
Well,
you
know
like
most
of
the
time
it
just
blows
my
mind
like
like
what
is
actually
doable
but
yeah,
so
I
just
wanted
to
share
at
least
that
kind
of
the
high
level
concept.
C
So
you
know
they
are
all
familiar
with
me.
So
maybe
that
let
me
let's
move
on
to
some
demo,
so
we
can
actually,
you
know,
see
the
the
you
know
this
in
context
right,
so
the
demo
will
be
really
three-step
and
you
know,
as
usual,
all
available
on
GitHub.
Those
are
just
simple,
notebooks
python.
Anyone
can
run,
we
will
go
for
model
input
and
prompt
engineering
I.
Do
the
example
of
a
pet
name
generator
then
I'll
show
you
what's.
C
A
number
looks
like
it's
pretty
at
all:
just
full
disclosure,
but
you
know
it's
kind
of
interesting
to
see
how
it
looks
like
and
and
then
we'll
go
through
question
answering
through
embeddings
right.
So
you
just
see
embeddings
in
use.
It
can
do
more
than
just
question
answering,
but
that's
typically
the
most
obvious
use
case,
so
we'll
basically
go
through
that
yeah.
So
let
me
stop
sharing
this
and
then
share
again.
Yeah.
C
C
C
C
So
we're
really
going
to
start
with,
with
inputs
to
to
to
very
known
model
that
we
all
use
for
the
the
Swanky
web
interface.
We
should
charge
GPT
I'm
going
to
use
a
gpt35
turbo,
because
it's
it's
faster,
but
also
it
is
cheaper,
so
bear
in
mind
like
every
API
call,
goes
to
my
credit
card
at
the
back,
but
I
mean
I'm
just
joking,
because
it's
quite
cheap,
so
I
really
like
no,
no,
no
harm
done
here.
C
We
are
actually
using
openshift
data
science
at
the
back,
but
in
particular
just
Jupiter
notebook
because
they
are
pretty
good.
You
know
to
to
basically
do
experiment
and
to
show
these
kind
of
steps,
so
so
essentially
I'm
using
the
python
API
for
open
AI,
which
has
been
pro-installed
because
it
takes
time
and
then
we're
kind
of
going
through
hey.
How
do
you
actually
interact
with
this
model?
So
the
way
it
works?
C
Is
you
specify
the
the
actual
model
that
you
want
to
to
use,
because
openai
gives
you
access
to
a
few
of
them?
And
then
you
have
a
message:
format
where
you
basically
pass
the
prompt
of
three
different
type
of
messenger.
One
is
the
system
which
is
the
actual
prompting
right.
So
it's
kind
of
the
context.
One
is
the
user,
which
is
you
and
assistant,
is
very
much
a
charge,
GPT
itself
replying
what's
interesting,
and
what
I
found
out
is.
C
You
can
actually
pass
an
entire
conversation
with
the
assistant
answer
as
well
as
if
you
want
to
actually
influence
and
give
the
contacts
of
an
entire
discussion
to
the
to
the
the
system
and,
of
course
the
content
is
what
you
pass.
So,
let's
see
a
very
simple
example
of
of
content
right,
so
I
have
a
whole
discussion
here,
I'm
asking
already
open
AI
to
be
a
helpful
assistant
when
I
say,
knock,
knock,
asking
who
is
there
and
when
I
reply,
Orange.
C
So
right
now
I'm
just
kind
of
firing.
This
API
call
to
open
AI.
So
this
is
like
the
whole
message
it
returns.
Right
gives
me
the
model
context.
Interestingly,
for
every
computation,
you
actually
have
the
token
spent,
which
is
basically
an
ID
of
of
how
intensive
the
computation
was
on
the
on
on
the
open
AI
aside,
and
then
it
basically
returns
to
me
the
content
yeah,
which
is
basically
the
next
sentence.
Now,
if
you
just
want
to
extract
the
content
itself,
you
pretty
much
just
need
to
pass
the
the
response.
C
C
A
C
Do
as
you
yeah
as
you
do,
because
I
mean:
why
would
actually
you
kind
of
spice
up
the
Reddit
documentation?
You
know
if,
if
we
were
to
actually
you
know,
use
this
for
a
bit
of
creative
writing
so
essentially
is
generated
a
response
for
me.
It
was
really
I
mean
I
haven't
met
black
belt
personally,
but
that
kind
of
looks
like
you
know
something
that
you
see
in
the
pirate
movies
so
that
that's
all
there
is
right
and
and
I
mean
some
of
those
things.
C
As
you
can
see,
it
will
just
influence
the
type
of
answer
you
you
get
from
the
model
I
mean
if,
if
I
don't
pass,
the
prompt
I'm
still
getting
pretty
much.
You
know
contextual
answer
with
black
belt,
but
maybe
it's
kind
of
influencing
exactly
how
the
model
actually
replies
to
me.
And
then
you
know
what
type
of
welding
right
is
actually
using.
A
C
That's
interesting
right
because
you
know
now
it's
like
a
professor,
but
to
me
it's
like
yeah
and
maybe
Professor,
you
meant
more
of
an
academic,
you
know
style,
and
so
probably
you
will
use
the
prompt
engineering
to
basically
derive
more
of
an
academic
answer
and
not
a
kind
of
context.
Students
all
the
time
in.
A
B
C
You
know
I
I,
think
I
I
strongly
suggest
right.
At
least
you
know
for
for
people
who
are
not
familiar.
I
think
to
me
is
I'm
not
really
afraid
about
hey,
hey
I
is
replacing
or
or
what
not
right,
I.
Think
it's
really
about
at
times
just
efficiency,
I
I,
do
make
a
lot
of
views
right
now
of
open
AI,
which
is
why
I
have
a
subscription
to
at
least
help
me
with
really
structuring
flow
of
information,
because
I
really
have
to
deal
with
a
lot
of
documentation.
C
Information
and
a
lot
of
my
job
is
hey.
You've
got
10
minutes
to
to
explain
a
concept
or
an
ID
to
someone
it.
You
know
it
it's
just
very
good
at
kind
of
summarizing,
giving
you
a
kind
of
way
or
or
an
angle
that
you
could
actually
explain
something
with
like
less
well
or
in
a
simpler
manner
and
yeah.
It's
it's
just
kind
of
helps.
I
mean
the
the
way
that
we
improve,
AI
I,
feel
AI
is
also
improving
us
and
and
what
we
do
if
we
can
use
it
responsibly
and.
A
The
risk
here
is
that
you,
you
don't
know
anything
about
not
you
personally,
but
one
doesn't
know
anything
about
what
they're
doing
and
using
it.
For
that
purpose.
Is
there
a
way
we
can
train
AI
on
our
own?
Our
own
information
could
I
have
it
scan
all
my
Google
Docs
and
and
then
help
me
write
my
own
responses
using
me
as
the
model.
C
Absolutely
so
hang
on.
This
is
what
they're
doing
step
two
with
the
embeddings
yeah,
so
we're
having
it.
Let
me
go
through
this
yeah,
a
novel,
a
fun
example
yeah,
and
then
we
look
at
hey.
How
do
we?
How
do
we
give
data
now
so
so
here
we
are
looking
at
a
very
simple
use
case,
which
is
we
kind
of
trying
to
build
a
a
fancy
pet
name
generator.
C
So
what
I'm
trying
to
do
here
is,
you
know
really
starting
simple:
ask
for
a
name
for
a
horse
and
then
let's
try
again
and
see
well
now,
if
I
ask
for
a
name
for
a
black
horse,
would
it
actually
influence
what
open
AI
response
so
she's,
like
hey
absolutely
yeah
I,
mean
I,
seems
like
a
very
nice
name
for
a
black
horse,
now
I'm
trying
something
a
bit
more
complex,
which
is
suggest
free
names
for
a
horse.
That
is
a
superhero
yeah.
C
So
those
names
are
a
little
bit
more
creative,
which
is
nice,
but
that's
not
exactly
what
I
want
right.
I
mean
like
how
a
superhero
name
typically
is
generated,
as
you
all
know,
like
you
know,
Captain,
America
or
or
thunder,
or
whatever
yeah
so
I'm
going
to
use
contextual
learning.
Now
in
the
next,
you
know
call
to
show
how
this
is
done.
So
if
you
can
see
here,
I'm
actually
providing
open
AI
with
example
of
how
this
kind
of
name
is
actually
generated.
So
I've
used
cats
and
dogs,
as
example,
and
so
I.
C
A
C
Am
yeah
with
three
pretty
cool
horse
names
yeah
now,
interestingly,
if
I
keep
running,
this
you'll
see
that
it
pretty
much
always
generate
very,
very
similar
names.
So
there's
no
real
creativity,
isn't
it
it's
all
probability
is
probability
based,
as
is
finding
the
most
likely
words
to
actually
give
you
an
answer
now.
This
is
where
it
gets
very
interesting.
C
There's
actually
settings
to
the
model
that
can
basically
influence
how
creative
it
is
like.
So
you
are
basically
telling
the
model
hey.
You
may
not
want
to
follow
the
best
probability
every
time.
How
about
you
take
a
bit
of
freedom
and
basically
be
more
creative
yeah,
and
this
setting
is
what
we
call
the
temperature.
A
Yeah,
that's
incredible
and
so
yeah
you
see
how
I
could
I
could
feed
it.
Information
like
information
about
me
or
so
now.
This
makes
sense
to
me
because
one
thing
I
did
and
I
I
I
I
I
swear.
I
didn't
do
this
for
the
kid
but
I
said,
write
me
a
speech
like
a
10
year
old
right
because
she
had
an
assignment
I,
didn't
let
her
see
it
and,
and
you
know
but,
and
it
did
it
right
and
so
I
can
see
how
I
could
set
the
temperature
how
I
could
set.
A
You
know
these
prompts
that
I
didn't
even
know.
I
was
doing
to
get
it
to
to
use
the
information
in
a
certain
way
yeah.
What
I
didn't
try
is
write.
It
like
I,
am
a
50
year
old
and
write
it
like
I'm
100
year
old,
because
it
would
have
come
out
different
and
that
would
have
been
a
wonderful
test,
but
that
was
really
cool.
C
All
right
so
we'll
come
to
the
next
step
now
which
is
well.
You
know
you
may
want
to
have
some
contextual
data
that
you
pass
as
well,
because
there
are
things
that
you
know:
open,
AI
hasn't
really
trained
with
so
in
this
example
and
I'm
not
going
to
run
everything
because
it
generally
takes
a
bit
of
time.
So
I've
run
the
demo
just
before
a
Visa
as
well
to
to
make
sure
everything
processes,
but
essentially
what
I
want
to
show.
C
Is
we
download
a
few
hundred
Wikipedia
articles
about
the
2022
Olympics
and
then
we
basically
chunk
them
and
we
create
what
we
call
those
embeddings
which
I
mentioned
earlier?
We
put
them
into
a
vector
and
then
we
use
this
to
basically
pass
to
the
model
and
help
him
answer
the
question
about
the
2022
Olympics.
Now
you
would
ask
me
why.
Why
do
we
need
to
do
that?
Yeah?
C
Essentially,
because
we
will
see
here
a
letter
when
we
question
answering,
if
you
have
used
in
the
past
the
charge
GPT
open
to
public
which
relies
on
the
on
the
charge,
GPT
3.5
model,
it
actually
does
not
have
any
training
data
after
2021,
so
we
would
actually
not
know
over
the
2022
Olympics
yeah.
So
that's
why
we
may
actually
want
to
to
pass
data
so
in
this
example,
I'm
I'm,
using
a
nice
little
API.
C
That
gives
me
access
to
Wikipedia
data
and
basically
I
I
gather
all
the
2022
winter
Olympics
articles
and
basically
we
have
a
total
of
732
articles
referring
to
Winter
Olympics.
So
from
the
from
there.
Mostly.
What
we
have
to
do
is
it's
really
not
a
very
releasing
task.
If
I
may
say
it's
pretty
boring,
but
we
have
to
spend
time
cleaning
up
the
data
and
basically
spitting
it
into
relevant
paragraph
and
sections,
but
the
hollow
one
by
themselves
and
that
stay
relatively
short
so
that
they
are
usable.
C
So
we
go
through
a
series
of
steps
to
basically
ignore
certain
section
of
the
Articles
stuff,
like
bibliography
sophistic
citation.
They
are
not
interesting
to
us
and
then
we
basically
associate
them
with
their
title
and
we
split
them
into
section
so
so
out
of
732
pages,
I,
basically
created
5742
sections
of
text,
which
I
then
proceed
to
clean
up
by
doing
various
filtering,
removing
blanks
and
so
on
a
lot
a
lot
of
really
not
interesting
code,
but
here
it.
This
is
how
it
looks
like
more
or
less
once
we
have
done
the
cleanup
right.
C
We've
got
basically
some
keywords,
some
short
descriptions
for
each
section,
and
then
we
have
like
a
bit
of
text
that
is
basically
a
single
paragraph
yeah.
So
once
we
have
Osmos
section,
we
want
now
to
create
what
we
call
the
embeddings.
So
for
this
again
we
use
a
model
that
is
provided
by
a
GPT
and
we
we
are
just
going
to
basically
pass
all
these
tracks.
You
know
again
to
to
to
to
create
all
those
different
embeddings
for
us
right.
C
So
so,
essentially
you
know
I'm
I'm,
trying
to
show
you
a
kind
of
readable
section
of
it.
You
know
we
have
something
like
this
yeah
name
of
the
person.
This
is
the
biography,
and
this
is
kind
of
some
information
about
this
person
as
an
example.
So
we
want
to
pass
it
into
a
chunk
that
will
be
usable
in
the
future
for
search
about
this
Michela
mayonnaise.
C
C
If,
if
you
actually
are
really
aware
of
why
they
are
relevant
and
then
we
basically
start
to
to
to
to
run
some
queries,
you
know
patching
them
by
sending
100
elements
at
a
time
to
this
model
and
what
it
really
does
for
us
is.
It
takes
this
chunk
of
text
and
when
it
pretty
much
returns
a
vector
yeah.
So
we
built
a
total
of
about
six
thousand
plus
vectors
and
when
we
say
them
in
a
CSV
file,
not
because
it's
a
good
idea,
it's
actually
a
terrible
idea.
C
C
So,
like
you
see
here
now
we
have
this
rows
of
text
and
then
in
front
of
the
text
we
have
a
vector
which
is
basically
coordinates
multi-dimensional
coordinates,
but
the
system
will
use
to
understand
the
the
the
the
the
likeness
yeah
between
different
parts
of
text,
all
right,
all
good
with
this
step,
not
the
most
yeah
interesting
step
of
all,
but
it's
kind
of
a
preliminary
yeah.
So
back
back
to
your
question
earlier
like
this
is
what
you
would
have
to
do.
C
Of
course,
it's
not
I'm,
not
showing
you
a
state
of
the
after
a
pipeline
to
generate
embedding
I
would
say
it's
a
it's
a
101
explanation
but,
like
you
know,
if,
if
you're
trying
to
teach
a
student
what
it
does-
and
you
are
the
starting
point-
that
kind
of
gives
you
an
idea
of
the
process
that
there's
obviously
a
lot
more
in
terms
of
automating
the
cleaning.
You
know
like
structuring
the
data
for
Optimal
Performance,
but
we
are
not
kind
of
looking
at
or
or
discussing
here.
A
C
Exactly
right
so
so
to
me:
what's
I
mean
you
know,
although
it's
really
a
long
time
ago,
I
I
used
to
be
more
of
you,
know
scientific
and
much
person,
and
but
that
to
me
is
you
know
interesting.
The
fact
that
you
have
this
multi-dimensional
Vector
really
use
and-
and
you
basically
use
a
cosine
a
function
to
to
relate
whether
a
vector
is
close
to
another.
Is
you
know,
I
I
kind
of
never
realized
that
hey?
This
is
actually
what
what
the
Marx
is
helping
us
to
do:
yeah,
yeah
all
right.
C
So
now
we
come
to
the
interesting
parties
with
what
was
embedding.
So
what
are
they
use
for?
So,
let's
start
with
first
some.
You
know
I
I
mentioned
it
earlier.
Why
do
we
even
need
to
do
embedding?
So
the
first
thing
is
GPT
and
large
language
model.
They
typically
can
learn
through
two
ways:
one
is
through
model
weights
and
one
is
from
model
input,
so
model
weight
means
you
go
back
to
the
training
data
set
of
the
model
and
you
fine
tune
it.
This
is
great
for
big
structural
learning.
C
So,
for
example,
you
want
to
teach
a
language
model
to
Learn
Python
programming.
You
have
to
do
this
kind
of
fine
tuning
and
training.
I
start
going
to
run
python
just
on
its
own
yeah,
most
likely
you
can
try,
but
I,
don't
think
the
results
will
be
fantastic
yeah
now,
if
you
just
want
to,
if
you
don't
want
to
to
you,
don't
have
to
build
any
kind
of
structural
capability.
So
in
our
case
we
are
just
still
using
language
and
English
language.
C
It's
actually
much
faster
and
less
costly
to
just
use
what
you
know.
What
is
the
model
input
yeah,
so
we
use
those
embeddings
as
additional
inputs
additional
information.
It's
a
bit
like
short-term
memory
right.
You
are
giving
this
short-term
memory
for
this
person
to
answer
I
mean
I.
Think
everyone
can
understand
this
right.
If
you
ask
me
any
question
about
2022,
Olympics
I
won't
be
able
to
answer.
If
you
give
me
access
to
600
articles
about
the
Olympics.
Well,
you
know
what
I
can
read
the
articles
and
find
the
end
server
yeah.
C
So
that's
pretty
much
what
we
are
doing
yeah
so
so
here
what
we'll
really
go
through
the
four
kind
of
process,
which
here
is
pretty
much?
Let's
try
to
ask
the
model
first
yeah,
so
I
didn't
give
any
additional
context,
and
now
GPT
is
telling
me
well,
you
know,
I,
don't
have
the
real-time
data
at
all
22
2022
Olympics
is
you
know
it's
not
something
I
know
because
I've
been
trained
for
data
before
yeah.
C
For
this
we
are
using
some
maths
again,
which
is
a
cost
in
a
function
which
helped
me
to
to
really
get
a
score
on
how
related
the
articles
are
so
what's
cool
now
is
if
I
pass
a
term
like
curling
gold
medal
is
going
to
start
finding.
You
know
like
going
through
those
embedding
and
finding
relevant
once
and
also
giving
me
a
score
of
collectedness
to
my
ask
yeah.
So
that's
how
we
do
it
essentially
telling
me
hey.
This
stuff
looks
pretty
good
0.879.
This
one
is
slightly
less
good
and
so
on
all
right.
C
C
C
A
C
I
mean
yeah
to
me,
is
you
know,
I
think
what's
important
is
really
to
understand
the
flow
and
and
the
concepts
you
know
the
the
the
the
beauty
is.
You
know
again
what
I
like
in
this
series,
you
know
to
me:
I,
consider
myself
more
as
a
data
engineer
than
a
data
scientist
right,
I've
already
been
in
the
data
field
for
four
years.
This
is
not
new
to
my
to
my
Reddit
skin,
but
you
know
like
these
type
of
capability
has
really
transformed
the
way
I
can
do
data
engineering
at
scale.
C
You
know
they
are
tasks
out
there
that
I
used
to
spend
in
auditory
weeks
to
really
kind
of
improve
data,
quality,
automatically,
classify
and
so
on
versus
literally
changed.
How
I
can?
Actually,
you
know,
do
my
work,
so
you
know
I
I,
just
think
that,
for
anybody
just
to
be
able
to
understand
this
is
to
understand
hey,
this
is
where
it
could
help
me
or
fit.
What
I
do
is
actually
pretty
exciting.
Yeah.
C
Absolutely
I
mean
we're
going
to
try
to
at
least
you
know,
I
I
like
to
spend
hopefully
10-15
minutes.
You
know
at
least
to
to
go
through
the
strategy.
So
let's
just
finish
quickly
with
generally
TBI
in
the
real
world
right,
so
we've
seen
a
lot
of
language
quality.
They
are
probably
the
most
I
would
say
mature
right
now,
because
we've
seen
the
Transformer
model.
C
You
know
it
was
initially
applied
to
text,
but
you
know
we've
seen
through
your
examples
of
yes,
it's
supplied
as
well
for
a
visual
I
mean
I
myself
as
well:
Avid
user
of
Brianna,
Leonardo
or
AI
to
generate
images
and
and
then
it
can
generate
audio
as
well
yeah.
So
one
aspect
that
I
would
like
to
highlight
for
because
to
me
is
also
pretty
fundamental-
is
generative.
Ai
is
a
big
contributor
to
to
improving
AI
as
a
whole.
Why?
C
One
of
the
key
problems
we
have
with
AI
is
having
good
quality
data.
Guess
what
generative
AI
is
also
used
to
generate
synthetic
data,
so
you
can
generate
synthetic
data
with
high
accuracy,
because
it's
now
using
model
and
deep
learning
to
do
that.
And
then
you
can
use
the
synthetic
data
to
train
better,
both
generic
model
and
also
with
machine
learning
algorithm
that
we
discussed
earlier
on
both
fault
detection
and
so
on,
yeah.
C
So
so
that's
pretty
cool.
So
you
know
you
look
at
at
cases
like
Automotive
as
an
example
in
order
to
have
to
retrain
those
models
for
self-driving
car
is
able
to
build.
Novos
3D
was
fake,
3D
World,
but
can
be
used
by
the
car
to
just
automatically
train
out
of
synthetic
data,
so
it
doesn't
have
to
go
out
there
and
kill
a
few
passerby
to
learn
that
this
is
bad
behavior,
because,
generally
theory
is
telling
him,
but
this
is
actually
what
could
happen.
I
mean
like
again:
how
cool
is
that
right?
C
So
a
lot
of
benefits
we've
seen
is
able
to
generate
all
these.
You
know,
algorithm
content
by
itself
is
able
to
help
you
explore
data
much
faster,
it's
able
to
accelerate
tasks
and
processes.
Now
before
we
move
on
to
radar
strategy,
just
a
few
key
aspects
on
the
challenges.
Yeah.
The
key
challenge
is
the
scale
of
the
compute
infrastructure.
C
C
Interestingly,
a
lot
of
people
who
are
good
data
scientists.
We
are
not
necessarily
good
infrastructure
people
and
people
who
are
good
at
infrastructure.
They
don't
necessarily
understand
data
science
yeah.
So
that's
why
this
field
is
seeing
a
huge
demand
in
people.
The
Second
Challenge
is
what
we
call
sampling
speed.
So
this
one
is
really
about
the
fact
that
yeah,
because
this
model
have
so
much
data
and
and
high
quality
data,
is
needed
to
have
accurate
answer.
They
have
really,
you
know
reasonably
slow
speed
to
basically
provide
you
with
an
answer.
C
So
so
from
the
human
perspective.
You
see
an
impact
now,
obviously
in
certain
use
cases
all
right,
because
you
can
just
wait
in
some
other
use
case.
It
may
be
dead
treatment
or,
if
you
know
sometimes
you
may
want
a
fast
answer.
That's
good
enough,
rather
than
a
long
answer
that
is
very
accurate,
yeah.
C
So
so
there
may
be
some
limitations
there,
which
is
why
you
see
a
lot
of
models
that
are
trade
with
different
amount
of
parameters
which
basically
give
you
the
choice
in
terms
of
amount
of
compute
and
speed
that
you
get
in
terms
of
getting
an
answer
yeah,
so
you
will
see
a
lot
of
model.
They
have
some
name.
C
You
know
like
your
c7b
or
170b
b
stands
for
billion
and
billion
is,
is
pretty
much
the
number
of
parameters
that
are
being
passed
yeah
so
that
that
should
kind
of
give
you
a
you
know
an
idea
of
what
it
is
and
why
it
makes
sense.
Yeah
over
issue
is,
of
course,
data
quality.
You
know
to
generate
data
well
with
data,
so
I
mentioned
it
earlier.
The
3D
model,
the
biggest
issue
is,
is,
is
actually
data.
C
There
are
not
a
lot
of
3D
models
of
the
world
today
available
yeah,
so
you
need
more
to
train
the
generative
AI,
so
it
can
actually
do
this
yeah.
It
cannot
just
miraculously
create
model
to
to
to
kind
of
train
off
yeah.
So,
let's,
let's
be
aware
of
it
and
the
last
one
is
you
focus
on
since
some
articles
recently
is
I
would
say
not
just
data
but
model
licenses
yeah.
When
people
say
this
model
are
open
source,
not
exactly
in
the
true
sense
of
the
term.
C
Yeah
I
think
a
lot
of
people
are
kind
of
now
rising
to
hey.
You
know:
why
are
they
calling
them
themselves
open,
source
and
yeah?
The
most
of
these
models
are
not
open
source
from
the
the
the
you
know,
the
true
sense
of
the
term
as
it
is
defined
by
the
institution
right,
but
my
view
is
this
is
expected
because
you
know
I
think
it
will
take
a
few
years,
for
you
know
AI,
which
is
a
new
field
to
find
kind
of
the
right
way
and
terms
and
Licensing
model
to
basically
operate
right.
C
We
we
can't
be
so
impatient,
I
mean
again.
This
thing
did
not
exist.
Much
like
just
a
few
years
back
so
I
would
tell
all
the
open
source
Fanatics
I
mean
I'm,
one
myself,
you
know
we
we've
got
to
work
with
the
openness
of
community
to
help
people
understand
the
fundamentals
of
Open
Source.
What
we're
trying
to
achieve,
and
hopefully
we
will
see
new
permissive
licenses
models
being
born
so
that
we
can
actually
spread
the
capabilities.
Yeah.
A
And
we'd
be
confident
of
that
I
mean
open
source.
Is
it's
not
new,
and
so
that's
very
clear
and
I
think
that
that
puts
us
in
a
great
spot,
but
yeah
patience
is
an
important
message
there
that
you
know
it's
taking
a
lot
of
money
and
a
lot
of
folks
to
do
this.
So
let's
get
there
and
how
could
Red
Hat
help
that
kind
of
stuff.
C
Exactly
so
perfect,
perfect
introduction
to
the
to
the
final
section,
which
is
what
is
what
I'm
doing
about
this
right?
So
so
the
way
I
look
at
it
is,
you
know
again:
I
tend
to
have
the
point
of
view
of
the
developer.
At
the
end
of
the
day,
what
I
like
to
do
is
develop.
Software
generative,
AI
or
AI
in
general,
is
an
additional
tool,
but
I
now
get
into
my
my
toolbox.
C
So
the
way
I
look
at
it
is
that
if
you
look
at
the
Software
System
engineering
cycle,
you
know
we
talk
a
lot
about
devops,
CI,
CD
and
so
on.
Now
we've
got
this
additional
Loop,
which
is
the
AI
modeling
cycle,
and
so,
if
you
build
an
an
application
today
chance
is
you
want
it
to
be
an
intelligent
application,
not
just
the
dumb
application,
with
a
lot
of
even
else
and
and
hard-coded
choices,
and
you
want
to
use
AI.
So
you
have
to
see
this
AI
capability
as
something
that
needs
to
integrate
with
the
software.
C
Now
why
this
is
interesting
from
that
point
of
view
is
because,
of
course,
button
of
what
we
do
is
helping
people
to
to
build
their
devops
capability
work
with
infrastructure
platform
to
really
make
sure
that
the
data
scientists,
the
data
engineer,
the
developer,
the
Ops
Team-
they
all
work
together
to
happily
integrate
all
those
capabilities
in
a
way
that
is
really
agile
yeah.
So
that's
that's
the
way
we
look
at
it.
C
So
we
really
focus
on
I,
would
say
the
life
cycle
and
making
the
life
cycle
less
painful
as
much
as
possible,
which
is
how
do
we
help
people
to
develop
the
model?
How
do
we
help
people
to
serve
and
monitor
the
model
once
it's
ready,
how
we
automate
the
whole
lifecycle
management
and
how
we
help
the
collaboration
with
the
developers
yeah
now?
C
So
this
is
an
important,
a
part
of
our
messaging,
because
it
really
defines
I
would
say
what
is
the
the
the
the
DNA?
In
my
view
of
the
red
acceleration
is.
Why
would
you
like
to
to
work
with
a
lot
in
this
space?
There's
a
lot
happening,
there's
a
lot
of
great
AI
company
out
there.
So
why
read
that
so?
For
me,
I
see
really
four
main
aspects
of
what
we
do
and
and
why
customer
would
actually
want
to
partner
with
us.
The
first
one
is.
We
are
really
focused
on
collaboration.
C
So
the
way
we
look
at
it
is
we
don't
want
to
provide
you
a
platform
that
is
great
for
the
data
scientists,
but
then
the
developer
is
totally
isolated,
and
so
we
are
really
looking
at
building
a
unified
platform
for
the
developer
and
the
data
scientist
to
build
intelligent
application.
The
second
aspect
is
foreign.
C
I
work
with
personally
public
sector
Healthcare,
you
know
financial
sector
as
well
the
the
kind
of
hybrid
model,
the
the
ability
to
to
to
to
train
a
model
consistently
on-prem
and
then
deploy
it
on
the
cloud
and
it
just
works
is
very
important
at
this
stage.
The
next
reason
is
ease
of
access.
Generally.
All
Focus
for
us
is
not
to
build
the
actual
model.
You
know
that
focus
is
not
the
actual
AI.
Expertise
is
really
to
make
access
to
this
complex
resource,
like
GPU,
very,
very
easy
right.
C
So
a
lot
of
my
customers
when
I
asked
them
today.
How
long
does
it
take
for
data
scientist
to
get
a
new
cluster
to
start
a
project?
We
are
talking
about
weeks
of
tax
month.
The
way
we
look
at
it
is
we
want
to
build
service
capability.
You
go
in
a
in
some
kind
of
a
web
interface
Define
how
much
space?
How
much
you
know
how
much
power
you
need
press
the
button,
no
it
guy
behind
the
scene.
You
know
kind
of
deploying
a
cluster
or,
like
you
know
like
like.
Let
me
show
you
up.
C
The
proper
size
is
the
right
type
of
GPU
and
so
on
and
all
about
Automation
and
then
last
but
not
least,
of
course,
all
approach
is
totally
based
on
open
source
yeah.
So
so
you
you've
seen
in
this
discussion
so
far
open
source
I,
think
in
AI,
just
like
it
has
been
in
the
Health
Sciences,
actually
has
been
really
The
Innovation
mechanism
for
for
AI
energy,
Network.
C
What
is
by
Google
to
the
first
few
model,
to
what
we
see
now
happening
with
you
know,
with
provider
like
a
ging
phase,
providing
all
these
capability
in
their
in
their
open
source,
offering
it's
all.
This
Innovation
is
driven
by
open
source.
So
what
we
know
is
it's
going
to
keep
changing
as
you
actually
adopt
the
technology
right.
C
Let's
say
you
spend
one
year
of
your
time
implementing
this
fancy
new
AI
platform,
which
has
the
smaller
ABC,
and
when
you
find
out
that,
one
year
later,
those
models
are
probably
obsolete,
always
like
better
version
of
it.
The
only
way
to
stay
all
about
is
to
be
able
to
quickly
take
a
piece
of
open
source
of
technology
and
then
deploy
it,
and
so
that's
pretty
much
what
we
focus
on
essential
and
I'm
not
going
to
to
Deep
dive
here,
but
we
I
like
to
describe
that.
C
We
are
really
a
provider
of
mL
of
solution,
we're
running
trying
to
piece
the
entire
process
to
build
model,
deploy
them
and
integrate
them
to
be
repeatable
and
automated
yeah,
and
so
this
is
how
complex
a
pipeline
can
be.
So,
just
to
give
you
an
idea
of
hey
I,
you
know
actually
it's
kind
of
nice
to
start
with
some
kind
of
a
model
and
just
kind
of
tweak
it
to
your
to
your
nets.
Yeah.
A
C
A
C
That
nicely
keep
going
exactly
yeah.
So
so,
but
that's
like
a
great
point
because
to
me
what
you
know,
I
think
some.
Some
people
try
to
compare
this
with
some
machine
learning
product,
for
example,
which
have
like
automl
capability
right.
Those
are
those
products
like
hey.
You
know
you
pass
them
some
data,
they
help
you
to
tell
you
hey.
Maybe
this
model
would
work
well,
I
mean
we
are
not
this
kind
of
product.
Typically,
we
work
with
this
kind
of
Provider.
C
We
are
more
addressing
the
more
fundamental
stage,
which
is
the
infrastructure,
the
automation
behind
it.
Now
let
me
go
to
you
know
to
the
transition
right,
because
what
we've
done
so
far,
which
is
openshift
data
science,
which
is
part
of
open
as
a
value
offering,
is
really
very
much
the
the
wires
right
now
we
are
going
in
the
foundation
model
era,
and
so
what
what's
next
for
for
openshift
AI
is.
C
Well,
the
the
the
part
of
the
problem
that
you're
trying
to
focus
on
is
really
what
will
be
going
through
in
in
this
a
show
today.
Foundation
model
require
this
huge
amount
of
infrastructure
and
data
to
actually
be
useful,
so
a
formal
point
of
view
as
a
platform
provider,
a
lot
of
the
contribution
and
work
we've
been
doing
internally,
we
will
engineer,
is
really
how
do
we
make
it
easier
for
these
people,
who
are
data
scientists
to
have
access
to
this
complex
infrastructure
and
resources?
How
do
they
queue
workload?
C
How
do
we
manage
the
scaling
automatically
for
them
and
how
do
we
pretty
much
run
the
training
and
how
do
we
also
avoid
the
situation
where
those
people
who
are
the
engineer,
the
cluster
administrator?
They
are
not
constantly
kind
of
having
to
do
this
sort
of
data
scientists
like
more
or
less
manually
yeah.
So
that's
really
what
we're
trying
to
to
do,
and
so
I
really
finished
with
this.
But
so
that's
the
exciting
part
for
me,
because
we
are
entering
into
this
this
next
stage.
C
Now
of
openshift
AI
roadmap
we've
actually
published
a
foundation
model
architecture
jointly
with
IBM.
So
it's
been
built
with
with
IBM
research
and
collaboration.
We've
read
that,
and
this
architecture
has
an
open
source
stack
called
code
flare.
You
can
find
the
repository
of
that
on
GitHub,
again,
obviously
open
source
and
this
code
flare.
Stacks
is
all
the
number
of
challenges
faced
by
the
data
scientists
and
the
administrator.
C
Let's
look
quickly
at
the
stack,
rather
you
see
on
the
left,
so
it
provides
you
an
SDK,
and
this
SDK
is
basically
a
way
like
you
know,
YouTube
to
you
as
a
data
scientist
in
your
python
code,
you
can
basically
start
running
job.
You
can
specify
them.
You
can
contribute
architect,
you
know
infrastructure
remotely
to
distribute
Computing
jobs,
so
you
don't
even
need
to
understand
how
many
servers
how
distributed
they
are
where
they
are
located.
C
Yeah,
you
don't
need
to
understand
anything
around
kubernetes
and
and
how
kubernetes
scale
we
give
you
a
simple
SDK,
which
is
a
programmatic
way
to
access
this
infrastructure
for
heavy
parallel
compute,
and
for
this
SDK
you
use
these
two
components,
which
is
the
multi-cluster
application
dispatcher,
which
handle
the
queuing
the
resource
quota,
the
management
of
past
job,
and
this
component
called
instascale,
which
is
doing
the
the
on-demand
resource
scaling
on
the
openshift
platform.
And
so
that's
really
what
what
it's
about.
C
So,
thanks
to
this,
we
basically
are
able
to
provide
easy
access
to
data
scientists
to
infrastructure
from
something
that
they
are
familiar
with,
which
is
a
pythonic
interface
important
to
mention
they
don't
need
to
rewrite
any
of
their
model
training
code.
So
if
you
are
using
a
framework
like
pie,
torch
tensorflow,
you
pretty
much
think
pytorch
code
and
you
just
distribute
the
training
for
this.
We
are
using
a
technology
which
is
called
red.io.
This
is
a
very,
very
cool,
open
source
project.
I'll,
look
it
up!
C
If
you,
if
you
don't
know
about
it,
the
it
could
be
a
session
on
its
own,
but
you
know
that's!
Basically,
what
we've
been,
what
we've
been
leveraging
and
contributing
to
the
back
now?
That
means
as
well
for
the
the
it
administrator
you
guys
out
there,
who
you
know
who,
like
me,
don't
like
to
or
have
to
look
up
tickets
and
collision
infrastructure
manually.
C
We've
got
a
solution
to
pretty
much,
have
a
stack
Diplomat
model
that
runs
consistently
on
any
infrastructure,
Cloud
on-prem,
but
pretty
much
abstract
the
management
of
resources
for
you.
So
you
can
focus
on
finops,
optimization,
around
aiops
and
so
on,
and
so
that's
what
I
wanted
to
share
today.
Wow
we've
overrun
by
about
15
minutes
but
hope
was.
It
was
actually
useful
for.
C
Yeah
I
mean
you
know,
I
think,
but
but
the
thing
that
is
always
you
know
important
I,
think
you
know
you
know,
interactions
with
people
right,
I
I
spend
a
lot
of
time
with
customers
for
some
of
them.
It's
good
to
understand
that
we
have
this
capability
now
and
you
know,
probably,
is
going
to
be
stable
more
like
next
year,
but
then
for
an
early
adopter
or
you
can
pretty
much
work
right
now.
We're
ready
to
to
kind
of
try
it
out
to
to
to
to
teach
the
technology
to
your
internal
people.
C
So
you
you
kind
of
get
to
choose,
based
on
your
on
your
appetite
you
know
and
and
like
I
expect.
I
like
is,
if
you
want
to
be
an
early
adopter.
You
have
this
this
fantastic
opportunity
to
to
channel
some
of
your
inputs
via
people
like
myself
and
the
the
most
exciting
part
of
my
job
is
to
take
all
the
inputs
from
from
people
who
use
the
technology
and
then
go
back
to
engineer
and
and
share
how
we
can
improve
the
product.
Yeah
yeah.
A
B
I
was
just
all
the
things
that
we
learned
today,
amazing
amazing
demo
that
you
shared
with
us.
You
know
took
us
through
the
inner
workings
really
glad,
and
you
know
just
like
we
gravitated
towards
you
know
all
that
you
know
the
passion
really
comes
out.
Vincent
I'll,
be
honest
here,
I'm
sure
people
would
like
to
connect
to
you
and
they've.
You
know
got
your
details,
but
thanks
for
doing
this,
we
really
appreciate
it.
C
A
So,
head
out
there
you
can
see
Vincent
at
those
events,
live,
there'll,
be
coverage
of
a
lot
of
AI
stuff
and
a
lot
more.
What
a
great
show
again
we'll
be
back
again
in
a
couple
weeks
with
actually
I
think
our
next
show
is
on
automation,
which
is
super
exciting
and
we'll
touch
again
on
AI,
because
it
is
really
everywhere
right
now
and
it's
important
to
the
strategy
and
to
the
way
things
are
going
in
technology,
so
that
is
super
exciting
to
have
that
option
as
well.