►
From YouTube: Tech.ml - new Clojure Machine Learning and ETL platform - Chris Nuernberger - Scicloj meeting 1
Description
Chris Nuernberger discussed Tech.ml, the new Machine Learning and ETL platform that he developed. Generateme gave a short presentation of his new plotting library, Cljplot. Followed by some broader discussion of the Clojure data science landscape.
Transcript by joinr: https://github.com/joinr/sciclojminutes/blob/master/meeting.org
This was the first meeting of the Scicloj community, on April 15th, 2019, 3pm UTC. This is an open online gathering of the Clojure data science community.
More details here:
https://twitter.com/scicloj/status/1114978949380083712
Talk slides:
https://drive.google.com/file/d/1PCGdt0TxWvuHiyli8sV9Ml4ZfIgkTuV1/view?usp=sharing
Meeting's text chat:
https://docs.google.com/document/d/1oNc7RTPryiyFzcEo6_RYSgPYGHoI0IbefsB7SKgVTuI/edit
A
So
hello,
everybody
I
guess
some
of
us
are
already
here
and
maybe
a
little
bit
more
several
minutes
so
that
few
more
people
can
join
and
I.
Guess
this
time
it's
like
a
good
opportunity
for
people
to
look
into
the
chair
and
see
that
you,
you
understand
the
environment.
If
you
haven't
had
an
opportunity
to
look
into
soon
and
and
I
guess
there
are
many
people,
or
at
least
in
potentially
maybe
40
people
to
join.
A
A
So
this
is
psych
launched
a
group
of
people
that
started
creating
some
things
together
in
the
last
few
months
and
what
we
are
trying
to
do
is
build
the
closure
data
science,
community
and
I.
Think
this
meeting
is
was
originally
an
idea
of
Alan
Murat,
see
that
some
of
you
know-
and
he
is
not
here
today-
but
I-
just
use
from
the
way
he
phrased
it.
A
They
will
talk
for
maybe
after
an
hour
or
if
there
are
more
questions,
then
more
and
then
we
have
some
discussions
discuss
and
what
could
be
nice
is
that
if
anybody
is
beginning
to
talk,
then
that
somebody
could
say
some
things
about
themselves
like
I'm.
My
name
is
that
and
I'm
from
there
and
I'm
doing
and
I
mean
setting
that
you
can
take
like
a
minute
to
talk
about
yourself
would
be
nice.
Does
it
make
sense
yeah
any
question.
A
And
so
one
more
thing,
all
these
the
process
of
AIT
meetings
and
other
kinds
of
communication
that
they're
working
on
it
in
the
last
few
weeks
is
a
joint
work
of
a
group
of
people
and
Chris
known
belt
and
Thomas
relay
and
Allen
Marazzi
and
join
up
that
you
know
by
this
name
or
part
of
it,
and
everybody
are
welcome
to
become
part
of
what
we're
building
and
to
influence
the
process
of
building
a
community.
And
so
let
us
talk
about
it.
A
C
Can
you
hurry?
Can
everybody
hear
me
yeah
sure
I
would
like
to
begin.
I
have
a
set
of
slides
that
I'm
gonna
attempt
to
share,
but
when
I'm
sharing
our
slides
I
won't
be
able
to
see
the
screen
or
the
chat.
So
if
Daniel,
if
you
could
stay
unmuted
and
let
me
know,
something's,
not
working
I'd
appreciate.
D
C
So
let's
talk
about
that,
let's
start
a
conversation
on
machine
learning
and
numerix
enclosure
and
I
I
think
there
have
been
existing
things
in
this
space.
I
do
not
think
that
those
things
represent
the
extent
of
what's
possible,
with
machine
learning,
our
numerix
and
so
I
think
there's
a
lot
of
space
for
conversation.
Yet
so
real,
quick,
I'm,
not
gonna,
talk
that
much
about
tech
assent,
but
we're
three
people.
C
We
have
a
broad
spectrum
of
contracts,
lots
of
industry
experience,
doing
lots
of
different
things
and
we
all
three
of
us
are
senior
engineers
and
senior
developers
and
seeing
somewhat
senior
c-level
executives.
So
we've
managed
people,
we've
managed
projects,
I've,
seen
languages
come
and
go
I've
seen
some
really
good
ideas
get
forgotten
in
the
history
of
time,
and
so
I
think
that
you
know
that's
some
of
some
where
I'm
coming
from
this.
C
C
One
branch
is
for
clawed
wrists
and
it's
really
for
closures,
who
don't
know
much
about
machine
learning
and
who,
for
instance,
really
haven't
worked
in
MATLAB
once
maybe
ever
never
used
pandas
and
what
you
do
is
you
know
you,
you
program
a
bit
of
closure
or
day
in
and
day
out,
and
you
have
access
to
day,
Tomic
or
an
SQL
system
or
an
Excel
spreadsheet.
So
you
have
access
to
a
rectangle
of
numbers
and
a
lot
of
times
in
Clojure.
That
rectangle
is
going
to
be
a
sequence
of
maps
and
then
how?
C
C
So
I
can't
see
the
chat
while
I'm
doing
this,
so
if
somebody
could
notify
me
if
there's
an
interesting
question
popping
up
or
maybe
we'll
address,
questions
at
the
end,
probably
the
best
thing
is
will
address
questions
at
the
end
of
the
top.
Oh,
so
you
are
kind
of
the
things
I
was
shooting
for
when
I
built,
DML
system
and
I,
don't
actually
think
I
hit
all
of
them
and
I
will
talk
about
that
in
a
second,
but
basically.
C
C
But
columnstore
mathematic
data
is
a
different
animal
than
a
sequence
of
maths,
and
so
the
more
you
get
to
the
actual
base.
Numerix
of
data
I.
Think,
though,
the
more
awkward
any
base
language
gets,
and
you
start
moving
into
what
you
really
need.
A
numerix
language
to
do
it
in
a
way
that
is
gonna
feel
really
good.
Then
model
search
and
I
want
to
talk
about
this.
This
is
really
important
for
a
lot
of
reasons,
and
the
main
reason
is
that
machine
learning
is
really
hard
and
there
are
a
lot
of
details.
C
C
Not
not
a
human
brain
is
not
a
machine
to
do
that
type
of
processing.
So
we
just
want
a
way
to
run
a
range
of
models
across
the
problem
and
see
which
one's
did
dusts
and
sometimes
a
lot
of
times.
You're
done
you
don't
need
even
worry
about
why
that
model
did
best.
You've
got
the
accuracy
you
needed
to
move
on
and
you
can
stop,
and
sometimes
you
can't
stop
and
you
need
to
understand
like
why
did
a
model
do
better
or
not?
But
that's
the
second
step.
The
first
shotgun
step
is
I'm.
C
A
rectangle
data
I
want
to
predict
one
of
the
columns
of
the
data
which
model
does
best
and
once
you
figure
out
which
model
is
best,
you
can
look
at
the
hyper
Crammers
used
to
train
up
all
the
BLA,
but
basically
a
searching
mechanism
across
this
massive
space
of
different
things
that
might
possibly
solve
your
problem.
I
think
for
me
is
really
important,
because
there's
no
other
way,
I
figured
out
to
do
information
hiding
where
you
don't
have
to
be
an
expert
in
each
model
and
still
get
decent
results,
performance,
reasonable
performance.
C
If
you
want
to
put
the
time,
on
the
other
hand,
I
think
nobody
will
probably
want
to
put
the
time
in
or
will
need
to
unless
you
have
a
very
specific
problem,
and
these
are
out
of
order,
but
that's
okay,
I'm
gonna
flip.
The
order
in
my
head
and
I
will
reorder
them
so
play
well
with
C
and
that's
key
for
me,
and
maybe
that's
because
I
have
such
a
strong
C
background.
But
that's
important
for
me
and
playing
well
with
the
JVM
means.
C
Obviously,
that's
important
closures,
ability
to
handle
just
basic
data
processing
and
transformations
if
your
problem
looks
like
a
transformation
from
a
data
structure
to
data
structure,
closures
dominate,
and
it's
amazing
and
it's
beautiful,
and
it's
probably
a
lot
of
the
reason
why
I'm
listener
here,
but
there's
got
to
be
good
ways
to
go
from
the
ways
that
make
closure
really
efficient
and
effective.
To
these
other
ways
of
representing
the
problems
that
closure
doesn't
play
is
well
with
in
its
native
form,
playing
well
with
Python
and
R.
C
C
You
take
an
efficiency
hit
in
a
lot
of
different
ways
and
in
terms
of
programming
time,
but
if
you're,
for
instance,
training
images
is
sorry
if
you're
doing
the
image
rec,
your
actual
efficiency
is
irrelevant
because
all
the
time
is
spent
in
the
neural
net
in
the
training
process.
It's
not
spent
in
closure
even.
C
These
things
need
to
be
totally
ok,
so
I'm
not
trying
to
build
something
that
I'm
gonna
stand
on
a
platform
and
say:
look
if
you're
not
using
what
I
built
you're
just
being
irresponsible
or
you're
being
ignorant
or
I'm,
not
trying
to
build
anything
like
that
and
I
have
felt
that
energy
before
building
things
in
closure.
I,
just
have
to
be
honest,
so
I
want
to
distance
myself
and
say
mainly
the
most
effective
use
of
Tekken,
generating
conversations
and
generating
ideas
that
potentially
don't
use
it
or
go
different
directions.
C
Then
it
does
and
that's
fine
with
me
and
that's
a
key
concept
behind
what
I
build.
That's
one
reason
why
I
don't
do
certain
things
with
this
tack?
So
openness
in
that
sense,
in
terms
of
furthering
conversations
in
constructive
ways,
is
really
important
to
me
and
the
foundation
behind
tech,
ascent
and
lots
of
different
things,
though.
C
I
can
build
nice
wrappers
around
things
which
shorten
the
pathway
to
doing
X,
Y
or
Z,
and
we
can
all
talk
about
certain
things,
work
or
not
work
or
feel
awkward
or
whatever,
but
for
any
branch
a
machine
learning,
I
can
think
of
and
so
I
just
don't
see
a
big
hole
anywhere
walking
through
this
list.
You
know
we
have
literally
20
matrix
libraries
selling
sports
bars,
sell
them.
Don't
some
extend
to
matrices
that
don't
fit
memory,
some
of
them,
don't
whatever
we
got
a
lot
of
those
a
lot
of
those.
C
C
Data
set,
Protestant
you
know
table
saw
is
not
a
new
project.
We
just
I
just
heard
of
it
three
months
ago,
but
it's
not
any
project.
It's
it's
mature.
It's
been
around
I,
wrapped
it
with
ml
dataset
and
there's
spork,
which
I
hope
you
guys
can
see.
Also
and
spork
is
a
column-wise
data
store,
that's
built
on
persistent
data
structures,
so
in
that
sense
it's
more
cordial
oriented
it's
easier
to
use
from
closure
and
integrates
deeper
with
closer
than
the
stuff
that
I
built.
On
the
other
hand,
we
found
that
table
saw
is
very
ni.
C
Table
saw,
is
very
efficient
for
data
sets
for
its
stores
data
much
more
efficiently
and
that's
potentially
something
that
solvable
in
sport
using
different
sorts
of
persistent
vector
type
things.
But
who
knows
classic
in
law?
Algorithms?
Oh
man
there's
a
lot
so
both
me
and
Tomas.
If
I
said
your
name
right,
I
apologize
if
I
didn't
wrap
smile
and
smile
has
more
than
what
I
did
and
fast
math
does
considerably
more
of
a
percentage
of
smile
than
what
I
did.
C
But
smile
has
more
than
that
Shogun,
which
I
think
is
the
way
to
go
moving
forward,
just
based
off
what
I've
heard
of
it
and
a
little
tiny
bit
of
research
and
extra
boost
and
there's
ten
others
than
this
I'm
not
kidding.
So
Intel
has
one
there's
a
lot
of
tree
gradient
boosted,
trees
and
this
kind
of
thing.
So
in
terms
of
the
classic,
you
know
logarithms
I,
just
don't
believe
you
need
to
write
any
more
right
now.
C
I
think
that
there's
plenty
of
ways
to
get
at
different
things
in
ml
and
that
Anna
based
approaches,
I
can't
say
enough
about
MX
net.
A
lot
of
the
recent
stuff
have
has
been
really
impressive
and
I
really
would
strongly
advise
anybody
going
that
direction
to
try
and
X
at
first
and
really
try
it.
On
the
other
hand,
I
think
the
classic
Camela
algorithms
apply
to
exponentially
more
problems
than
the
neural
net,
based
approaches
and
I
have
to
say
that
so.
But
that
being
said,
I'm
stopping
in
X
net
is
brilliant.
C
Karen's
work
in
that
is
really
really
good
and
then
there's
another
one.
Tensorflow
Scala,
since
the
flow
is
a
different
design
than
a
max
net.
So
tensorflow
is
more
like
you
have
an
abject
syntax
tree
of
math
and
I.
Do
the
things
on
it,
and
that's
also
interesting
to
me
for
a
lot
of
reasons,
but
you
you've
to
invasive
birches.
C
C
Unless,
potentially,
you
really
know
one
of
those
algorithms
and
you
can
think
of
an
edge
that
the
other
libraries
don't
have
and
then,
if
everything
I
said,
sounds
like
nonsense
to
you,
my
strong
strong
advice.
It's
just
just
use
XD
boost
just
use
extremists,
maybe
via
tech
ml.
Maybe
just
raw
I,
don't
I,
don't
really
care,
but
if
everything
sounds
like
nonsense,
just
use
excuse.
C
So
what
are
the
requirements
to
if
we're
gonna
ask
somebody
to
say:
hey,
invest
time
in
our
machine
learning
platform
like
we
think
it's
really
good.
We
think
it's
worth
it
for
you
to
invest
a
lot
of
time
in
to
to
bet
your
career
on
what
are
the
requirements
that
we
have
to
support
in
order
to
validate
that
question
or
in
order
to
feel
good
about
asking
somebody
to
do
this.
C
You
know
cuz
I
feel
like
machine
learning,
it's
an
investment
and
if
you
invest
all
your
time
in
a
platform
that
never
gets
used,
you
get
to
invest
some
of
that
time.
Just
relearning
another
platform
that
does
get
used
or
if
you
invest
time
in
it
or
worse.
If
you
invest
time
in
a
platform
that
doesn't
scale
or
doesn't
actually
solve
the
problem,
then
that
potentially
you
walk
away
from
that
entire
investment.
C
So
the
first
thing
I
think
is
important
to
me-
is
proving
that
the
platform
does
actually
work
and
right
now,
I
think
it's
actually
easy
in
machine
learning
as
compared
to
have,
for
instance,
computer-graphics,
because
you
have
a
set
of
these
fairly
difficult
machine
learning
problems.
You
have
them
solved
in
an
open
way
with
a
bunch
of
tools,
and
we
can
take
pick
off
the
problems
just
one
by
one
and
make
sure
we're
getting
good
results
for
those
problems,
and
that
to
me
is
really
really
important.
C
I,
don't
believe
that
people
will
spend
a
lot
of
time
learning
how
to
do
machine
learning
and
closure
unless
there's
some
subset
of
problems
that
we
can
prove.
We
get
world-class
results
for,
and
that's
just
my
opinion,
but
that
that
is
really
important
for
me
and
I.
Think.
As
far
as
that
goes,
there
are
simple
machine
learning
problems
in
cattle
that
don't
have
closure
analogues
grouped
out
for
them,
and
we
haven't
Allen,
set
up
a
Trello
board,
I.
C
Think
Allen
did
it
I'm,
not
sure
who
did
but
Allen
somebody
set
up
a
Trello
board
where
we
listed
kind
of
like
here's,
a
bunch
for
NLP.
Here's
a
bunch
for
this
type
of
thing:
here's
a
bunch
for
that
type
of
thing
and
having
regardless
of
the
toolkit
you
choose
to
use
having
those
demonstrated
in
clear
ways.
I
think
goes
a
really
long
way
because,
as
I
said,
I
feel
like-
and
this
has
been
true
for
a
while
this
isn't
just
in
the
last
month.
I
feel
like.
C
We
have
the
tools
enclosure
to
do
all
this,
but
because
of
the
complexity
of
the
domain
of
machine
learning,
without
having
clear
and
simple
examples
of
using
those
tools
to
get
world-class
results.
You're
still
asking
a
lot
from
someone
to
try
to
do
a
lot
of
their
machine
learning
in
closure
and
Plus.
Every
one
of
these
things
that
I
do
I,
learn
a
lot
myself
and
I
learned
about
deficiencies
and
the
things
that
I'm
working
with.
C
C
Changing
changing
here's
a
little
bit
I
just
want
to
talk
about
my
experience
with
cloture
a
little
bit
and
talk
about
kind
of
what
I
hear
a
lot
from
people
and
trying
to
sell
consultants,
services
in
closure
and
trying
to
talk
to
people
about
why
I
love
closure
and
this
I
feel,
like
you
know,
popularity
contests
are
not
something
I
respect
in
any
great
amount.
I've
never
won
one,
and
never
probably
will
I,
don't
think
cloture
will
either
I
accept
salary
that
came
out
recently.
That's
a
powerful
one.
C
If
you're
gonna
win
one
it
probably
salary
is
a
good
one
to
win,
but
for
a
lot
of
reasons,
I
feel.
Like
closures,
not
gonna
win
the
popularity
contest
and
that's
fine.
It's
not
actually
a
problem,
that's
meaningful,
and
so
how
relevant
is
this
to
what
I'm
talking
about
and
so
I
feel
like
this
informs
the
direction
I
take
when
I'm
building
something
and
so
closes
the
underdog
and
everybody
loves
to
root
for
the
underdog.
C
So
that's
an
advantage,
but
because
it's
an
underdog
most
of
the
software,
that's
really
great
in
the
swirl,
is
gonna
be
built
in
other
languages
and
so
making
it
really
easy
and
building
effective
low
code
integrations
with
powerful
software
is
the
fastest
way
to
scale
the
language,
in
my
opinion,
and
that's
absolutely
an
opinion
but
I
think
there's
just
so
much
really
good
software
written
in
other
languages,
especially
C
and
C++,
and
now
rust
that
closure.
It's
just
such
a
natural
companion
to
closure.
You
know
for
any
problem.
C
You
come
up
with
a
really
any
realistic
system.
You
come
up
with
in
computer
science
at
all.
I
can
tell
you
the
best
way
to
solve
that.
That
problem
is
going
to
be
a
mixture
of
a
garbage-collected
language,
garbage-collected
language
right
there.
It's
just
obvious
to
me-
and
that
includes
you
know,
a
computer
game
that
includes
most
of
the
big
server
systems
or
microservices
systems
that
you're
talking
about
any
complex
thing.
You're
talking
about
it's
going
to
be
a
mix
of
garbage
collected
and
native
and,
unfortunately,
I
feel
like
mixing.
C
Those
two
is
still
awkward
and
that's
funny
to
me,
because
I
feel,
like
Microsoft,
made
a
storm
push
to
mix
them
years
ago
and
it
failed,
but
that's
just
my
opinion.
It's
just
natural
to
move
between
garbage
collected
and
on
garbage
collected
worlds
for
for
any
given
problem
in
the
given
complexity.
C
So
that
being
said,
I
feel
strongly
that
see.
Family
integration
is
part
of
what
makes
everything
tick
so
I've
researched
this
a
lot.
One
important
thing
is:
how
do
I
take
a
native
resource
and
bind
it
to
the
garbage
collector
in
a
meaningful
way
or
bind
it
to
my
program,
stack
frames
and
we
basically
I've.
After
doing
the
research
I
write
a
little
library
and.
C
That's
about
where
the
research
stops,
so
we
have
a
set
of
libraries
that
tech
assent
that
are
built
around
see
family
integration.
If
you
want
to
integrate
with
a
see
thing,
I
think
I've
built
things
that'll
make
it
easier,
even
if
you
choose
not
to
use
one
or
two
of
them,
some
of
them
were
probably
no
help
and
as
far
as
that
goes,
some
of
this
is
also
asking
nicely
for
our
brethren
programming,
C++
and
rust
to
produce
api's
binary
interfaces
that
are
in
a
c
subset.
C
So
that
means
structs,
plain
old
data
types,
data
types
that
don't
have
meaningful,
constructors
or
destructors,
and
that
are
just
data
where
mem
copy
is
fundamentally
means
min
copy,
equal
semantics
work
and
these
functions.
These
function
pointers
and
they
use
contiguous
arrays
of
plain
old
data
types.
C
That's
it
anything
more
than
that
and
you're
drastically
reducing
the
number
languages
that
can
you
can
work
with,
and
so
some
of
this
is
is
trying
to
work
with
the
experts
in
C++
for
producing
these
amazing
things
and
say:
I
know
that
it's
like
cutting
off
your
arm,
but
a
capi
would
help
me
use
this,
and
the
truth
is
once
there's
a
C
API.
You
can
use
it
from
C
Python.
You
can
use
it
from
rust.
You
can
use
it
from
Julia.
C
You
can
probably
use
it
from
our,
although
I
don't
know,
and
you
can
definitely
use
it
from
Java
using
J&A
with
no
further
work
on
their
part.
That
is
substantially
less
work
than
than
building
their
own
J&I
wrappers.
Then
then
building
their
own
Python
using
boost
Python.
Then
then,
building
who
knows
what
I
mean
it's
just
obvious,
put
a
nice
C
interface
on
your
shed
and
then
stop
walk
away.
C
So
the
JVM
is
another
area
and
because
of
really
good
design
decisions
really
early
on
it's
not
that
hard
I'm,
not
gonna
talk
about
it
much
one
of
the
base,
libraries
that
we
built
everything
on
is
meant
to
work
with
contiguous
buffers
of
primitive
data
types
and
that's
because
it's
so
core
to
integrating,
with
the
JVM
in
so
many
different
ways
and
with
C
systems
in
so
many
different
ways,
and
that's
where
it
comes
back
to
the
nail
buffer
point
I
made
earlier,
but
it's
more
than
that.
You'll
see
more
about
this
later.
C
C
That
being
said,
all
the
time,
almost
every
Java
library,
I
use
I
end
up
using
reflection,
Devorah
pathway
through
they
almost
always
make
something
private
that
isn't
really
important
for
somebody
else
to
get
to
especially
for
performance
reasons,
and
it's
just
the
nature
of
the
game.
And,
quite
honestly,
that's
always
been
my
experience
with
languages
that
have
strong
language
based
protections
in
them.
So
that's
true
of
C++,
where
I've
seen
people,
for
instance,
for
the
people
who
program
C++
in
this
room.
C
Real,
quick,
consulting
with
closures,
an
interesting
game
and
it
could
be
we're
still
learning
how
to
do
it.
We
could
be
doing
it
wrong,
but
it's
the
first
few
years
of
it
have
been
a
lot
of
learning,
so
the
the
the
only
costs
so
far
to
tech,
ascent
of
Claude
ruse,
and
this
stash
has
been
that
I
talk
about
a
lot.
So
we
go
to
d3
meetings
and
I.
C
Go
to
coffee
clubs
in
the
morning
and
I
go
to
these
things
and
I
like
to
give
presentations
on
things,
and
so,
if
I
can
find
a
group
of
people
who
are
interested
in
this
thing,
I
might
build.
It
include
her
and
then
talk
about
it
and
that
way,
it's
a
nice
way
to
say
hey.
This
is
this
is
one
of
the
strengths
of
culture
here
or
this
is
not
potentially
a
lot
of
people
who
want
to
talk
to
us
are
worried
about
what
happens
when
the
engagement
ends.
C
So
they
are
concerned
about
their
ability
to
hire
or
train
Clojure
engineers,
and
that's
just
is
what
it
is
they're.
Always
it's
always
a
big
system
we're
working
with.
We
always
have
to
interplay
with
lots
of
different
things
and
where
that's
concerned
embedding
clojurescript
and
larger
JavaScript
applications
is
unwieldy.
C
So
if
they
have
a
large,
for
instance,
view
j/s
app
and
we'd
like
to
write
some
portion
of
that
app
enclosure
script
that
gets
pretty
hairy
pretty
fast,
so
I
think
that
that's
just
a
thing
that
being
said
once
the
contract
sign,
people
are
really
happy
with
the
outcomes.
Closure
is
such
high
leverage
that
the
three
of
us
can
do
things
that
teams
of
people
have
problems
with,
and
that
is
an
outcome
of
the
language
being
hoho
iconic.
C
That
is
an
outcome
of
us
having
four
years
of
experience
doing
these
things,
and
that
is
an
outcome
of
our
very
diverse
backgrounds.
The
three
of
us
have
extremely
different
backgrounds,
and
so,
if
you
combine
all
those
things
together,
then
man
you,
you
have
a
lot
of
potential
for
speed,
so
the
last
thing
I
want
to
say-
and
some
of
you
probably
heard
this
before
and
I
actually
I
saw
lately
joined
the
call
and
I
want
to
credit
I
have
to
credit
Ben
camp
house
with
this.
C
A
Yeah
thanks
guys
that
was
beautiful
and
eye-opening.
Can
you
feel
me?
Okay,
I
can
yeah
yeah,
wonderful,
so
I
guess
that
it
would
be
a
good
time
for
questions
and
discussion,
and
and
can
you
see
the
chat
now
many
people
are
saying
thanks
and
I
think
it's
beautiful
for
everybody,
I
guess,
and
so
it
would
be
a
good
time
for
you
to
open
it
for
discussion
for
questions
absolutely.
A
Wonderful
and
I
guess,
because
we
are
not
like
40
people
as
I
thought
we
would
be,
then
you
can
just
open
your
audio
and
speak
and
maybe,
if
you
feel
comfortable,
open
your
video
to
everybody
so
that
we
can
see
each
other
and
get
to
know
a
little
bit
of
who
we
are.
It
wouldn't
be
okay.
If
anybody
wants
to
ask
any
and
also
as
I
mentioned
when
you
speak
for
the
first
time.
Please
talk
about
yourself,
so
maybe
I'll
ask
something:
I'll
talk
about
myself
I'm,
my
name
is
Danielle
and
I'm.
A
C
So
thank
you
that
that
actually
is
a
big
piece
so
that
you
were
there
at
the
conception
of
that
idea.
We
were
talking
about
it
and
I.
Basically,
Alan
said
you
know
what
a
lot
of
people
are
building
these
complex
training
pathways
and
then
they
go
to
productize,
and
then
it's
tough
and
if
you
have
a
big
data
set
I
think
it's
really
easy
to
see
why
it's
tough,
because
you're
going
to
do
things
like
replace
the
missing
values
with
the
me
know
so
part
of
your
data
set
and
then
during
production.
C
You
have
to
remember
that
mean.
So
if
you
get
the
same
missing
value,
you
can
use
the
same
mean,
and
so
there's
these
bindings
of
data
that
you
generate
during
training
beyond
the
models
that
you
trained
and
they
need
to
be
represented
in
in
the
in
the
actual
production
process,
so
that
combined
with
a
complex
regression
demonstration
on
Kaggle
that
I
found
basically
I,
cut
and
pasted
the
pandas
and
translate
it
into
Clojure
data
structures
and
then
I
implemented
a
system
that
would
run
those
data
structures
and
produce
the
same
result.
C
C
C
C
So
I'm
going
to
show
us
some
subsections
from
it.
That
I
think
are
interesting.
I
when
I
said
this
is
actually
one
of
the
things.
I
got
a
lot
of
feedback
right
when
I
made
this
and
I
agree
with
all
of
it,
but
my
I
like
to
let
feedback
sit
for
a
while
and
let
people
try
to
use
it
in
various
different
ways
before
I
change
anything.
But
you
can
see
that
here
you
know
the
pipeline.
C
C
But
it's
a
really
error-prone
process
to
carry
this
sort
of
data
set
engineering
and
re-engineer
it,
as
you
can
see
like.
Sometimes
we
do
pretty
in-depth
replacing
of
values
for
various
reasons,
and
you
get
to
re-engineer
all
that
in
a
different,
slightly
different
contexts.
So
if
it's
a
fixed
value
of
replacing,
you
know,
that's
not
very
hard,
but
when
it
does
come
down
to,
for
instance,
well
yeah
replacing
things
with
means,
or
some
specific
thing.
You
really
like
a
lot
of
data.
The
end
result
of
your
training
process
isn't
just
your
models.
C
It
should
also
be
a
pipeline,
and
the
truth
of
the
matter
is
is
really
the
pipeline
should
be
a
graph
and
running
the
pipeline
is
a
translation
from
one
graph
to
another
graph,
and
the
second
graph
has
you're
not
models
and
all
your
animal
UDI
data
that
was
too
aggressive,
I
couldn't
I
didn't
want
to
do
that.
As
my
first
try
at
this
I
wanted
to
get
one
clear
kaggle
example
working
and
I
did
and
what
was
really
nice.
C
Let's
see
if
I
can
find
the
correlation
table
that
I
produced
matches,
two
I
think
the
fourth
does
third
decimal
place
with
the
correlation
table
that
pandas
produced
and
that
there's
a
lot
of
processing
before
we
ever
get
to
this
correlation
table.
So
that
was
really
that
felt
really
good
when
that
happened.
That
was
the
result
of
a
lot
of
work
that
didn't
happen
by
chance
so
anyway.
This
notebook
example
is
literally
line
for
line
almost
exactly
the
same
as
the
Capitol
example
and
I.
C
Don't
always
want
to
hold
to
that
sort
of
exact
matching,
because
there's
things
that
are
more
efficient
to
do
in
one
system.
If
you
structure
them
slightly
different
but
for
this
example
I
felt
like
it
was
important
and
I
think
that
was
you
know.
That
was
the
impetus
behind
a
lot
of
the
pipeline.
I've
gotten
a
lot
of
feedback
on
it.
C
One
of
the
most
important
points
I
made
earlier
was
that
you
know
the
whole
system
needs
to
be
simple
without
drama
and
I
feel
like
the
pipeline
misses
the
simple
aspect,
because
it
has
so
it's
so
data
oriented.
You
can't
really
use
the
pipeline
without
quoting
things
and
without
doing
a
lot
of
other
stuff,
and
that
is
your
second
step.
C
Your
first
step
is
going
to
be
a
bunch
of
data
set
engineering
and
getting
a
model
that
trains
well
and
once
you
get
them
all
that
trains
well,
your
second
stuff's
gonna
be
productizing,
it
so
I
think
I'm,
a
the
privatizing
portion
relatively
too
important
and
the
actual
data
manipulation
portion
while
is
very
efficient
in
this
format.
It
still
awkward
to
use
from
the
repple
and
in
an
exploratory
fashion.
So
that
is
something
I'm
going
to
address
and
version
two
will
have
a
set
of
fixes
around
all
this
kind
of
stuff.
A
F
G
G
So
closure
is
definitely
helping,
but
again
it's
I
totally
agree
on
the
outreach
outreach
like
it's
just
it's
very
difficult
to
get
both
newbies
and
experts
being
interested
in
something
like
this.
The
only
way
that
I
got
people
interested
was
to
show
them
the
salary
which
just
came
out
like
you
know
like
two
weeks
ago
and
I
thought
that
was
that
was
really
interesting
so
that
the
Stack
Overflow
results
actually
helped
a
lot
in
terms
of
you
know:
adding
credibility
to
closure
so,
for
instance,
before
it'll
be
like
yeah.
G
You
know,
we've
built
this
really
cool
thing.
You
know,
and
people
will
be
like
what
language
you're
using
and
it
will
be
all
closure.
What
is
this
language
you
know?
And
now
and
now
there's
as
a
you
know,
I
just
show
them
the
celery
and
people
go.
Oh
okay
and
there's
an
argument
around
that
anymore.
So
that's
that's
been
pretty
cool.
G
But
again
you
know.
I
I
know
that
experienced
closure
developers
take
like
yeah
like
experienced.
Closure
developers
are
very
rare,
and
people
that
want
to
get
experienced
in
this
language
is
also
very
rare,
so
I
think
I
think
I've
I
really
enjoyed
the
talk
and
and
you've
said
a
lot
of
things
that
really
resonated
with
me.
Sorry
thank.
E
H
A
G
Yeah
I
thought
it
was
great.
The
one
thing
you
talked
about,
which
I
think
is
super
critical,
is
the
see
linkage
so
I
mean
you
made
a
point
at
the
end,
which
was
now
let's
build
bridges
and
leverage
other
things
that
people
other
people
are
making,
and
it's
that's
sort
of
the
point
of
closure
in
the
first
place
in
the
original
design.
Being
able
to
take
advantage
of
Java
libraries
was
was
a
fundamental
idea
right,
you
know,
a
new
language
is
going
to
be
small.
You
need
a
base
of
libraries.
G
It
seems
clear,
though,
that
the
libraries
that
need
to
support
are
and
Python
do
so
via
C,
and
so
a
closure,
in
my
opinion,
is
missing.
Most
is
the
C
linkage
the
ability
to
sort
of
automatically
tap
into
libraries
written
for
Python
by
a
C,
linkage
and
stuff
you
short
around.
That
was
particularly
interesting.
My
question
is
how
how
much
of
the
C
interface
stuff
is
actually
that
C
subset
as
opposed
to
C++
and
I'm
more
challenging?
G
What's
how
common
is
the
C
binary
model,
because
I
think
that
that's
the
that
is
the
first
problem
we
have
here
that
if
we
could
transparently
tap
into
sort
of
anything,
that's
been
written
for
Python
we're
in
a
great
place
we're
in
the
same
place.
We
are
relative
to
Java,
and
so
that's
my
question.
You
know
how
how
come
and
health
the
stuff
you've
built
that
does
that
bridging?
How
useful
is
it?
G
How
often
can
we
apply
it,
because
what
I
see
happening
too
often
I
think
is
we
build
things
enclosure
on
top
of
Java
or
Scala
who've,
already
sort
of
encapsulated
the
good
bits
of
interacting
with
C,
and
we
don't
get
the
power
that
they
have,
for
instance,
to
do
model
definition.
We
have
to
use
their
model
definition
stuff,
a
step
removed
so
is
C.
Is
that
C
subset
common?
You
know
50%
of
the
time
80%
20%.
C
It
really
depends
on
how
good
the
team
is
that
built
the
library.
So
let's
talk
about
really
good
teams,
put
the
C
interface
on
it
themselves,
so
TVM,
for
instance,
which
is
the
tensor
VM
that
I'm
really
impressed
with
one
of
the
languages
that
has
world-class
numerix
performance.
Undeniably,
it
is
built
by
DM
LC,
and
really
it's
headed
by
this
man
named
chunky
Chen.
He
puts
a
C
interface
on
everything
that
he
does
before
it
puts
anything
else
on
it.
C
C
Not
so
really
good
teams
are
doing
that
at
times,
but
it's
very
rare
in
my
opinion
that
you
know
somebody
who
wrote
a
lot
of
C++
telling
them
they
have
to
define
an
interface
with
C,
really
is
very
painful
to
them,
because
now
it's
like,
oh
man,
that's
so
much
extra
code
and
all
my
template
parameters.
I
now
have
to
code
up
explicitly
and
did
it
today,
so
they're,
not
fans
than
that
rust.
I.
C
Think
people
are
more
fans
of
that
and
then
there
is
an
option
that
is
not
based
on
a
C
interface
and
that's
Java.
C
VB
and
Java
CPP
uses
a
parser
to
parse
a
header
to
parse
the
actual
C++
headers
and
can
export
raw
j'ni
bindings.
It's
in
auto
generates
the
j'ni
bindings
union
and
you
compile
them
and
you
build
your
own
library.
This
is
so
far
less
ideal
in
lots
of
ways
than
I
would
like
number
one.
C
They
they
ease
their
own
super
floss
parser
and
they
didn't
I'm
hearing
a
little
bit
of
something
from
you
right,
but
they
use
their
own
super
puzzles.
Parser
they
shouldn't
use
cars,
so
clang
would
correct
it.
Although
in
C++,
but
instead
they
wrote
my
house
telling
that
being
said,
when
I
have
wanted
to
bind
to
a
c
plus
source
library,
that
java
c
vp
supported
poorly
I
wrote
a
C++
header,
only
wrapper
that
would
get
pileup
compiled
in
this
library.
C
Its
Duggars,
it's
just
awkward
and
Java,
is
modest,
well
design
this
Dana,
so
you
have
two
bubbles.
You
can
be
two
key
plus
plus
it's
just
not
as
good
as
when
they're
seen.
How
was
the
percentage
of
running
into
more
modern
things,
I
think
are
done
with
everybody
wants
our
in
Python
to
work.
They
don't
want,
have
to
write
their
own
thing,
so
I
think
at
least
in
machine
learning,
potentially
the
world's
coming
around.
G
C
I
I
At
that
yeah
so
like
cause,
our
is
like
an
alternative
implementation
of
our
poor,
growl
and
yeah.
They
do
have
fighting
as
well
drew
be
even
and
yeah
there
you
get
zero
interrupts
between
all
those
languages
and,
like
you
could
even
like,
theoretically
or
yeah.
They
do
have
an
LLVM
compiler
as
well
and
interpret
a
little
bit
code
as
well.
F
H
Why
I
looked
at
the
Python
stuff
was
kind
of
like
POC,
they
wanted
to
get
it
working
and
they
knew
that
getting
all
the
the
see
stuff
and
all
the
fiddly
bits
working
with,
and
they
were
just
beginning
to
work
on
that.
So
I
have
no
idea
exactly
where
it
is
it
was.
It
was
just
kinda
like
in
the
hello
world
stage.
The
last
time
I
looked
yeah.
I
A
A
Anywhere
again,
yeah,
so
so
maybe
I
guess
we
still
have
time
and
before
we
move
to
the
more
broader
discussion
of
what
we
are
about
in
the
community
questions
so
Chris.
What
do
you
think?
Maybe
you
could
show
a
little
bit
of
your
source
code
so
that
people
could
like
get
to
know
a
little
bit
of
what
you
create
and
the
values
and
tech
libraries?
Is
it
a
good
idea.
J
Yeah
Cheers,
so
my
name
is
Joe
I'm,
a
full
time:
Python
slash
Scala
person
at
a
a
massive
fortune,
100
company
and
I
dream
of
working
in
closure
and
try
to
find
clever
ways
to
convince
my
boss.
It's
the
go-to
route,
so
I
could
get
those
giant
salaries
like
you
guys,
so,
just
from
my
perspective,
I
think.
A
lot
of
the
reason
closure
is
is
so
interesting
to
folks,
like
me,
is
because
its
core
principles
around
it's
its
core
data
structures,
its
leaning
towards
the
mutability,
all
that
good
stuff.
J
When
I
see
the
data
science
that
we
do
at
my
company,
it
is
notoriously
undisciplined.
It
is
a
mess.
So
I
think
just
one
of
the
things
that
comes
to
mind
for
me
is
is
I
understand
growing
the
community
is,
is
definitely
important,
but
I
guess
I'm
I'm
worried
a
bit
about
growing
the
community
at
the
expense
of
compromising
kind
of
those
core
priorities
of
closure.
That
I
think
ultimately
made
it
so
successful.
So
yeah
I
guess
that's
more
than
a
comment
than
a
question,
but
I
don't
know.
C
Makes
some
sense
and
I
I
definitely
feel
like
the
libraries
I
built
are
not
as
axiomatic
as
a
lot
of
the
original
language
design.
Of
course,
I
didn't
spend
nearly
the
time
on
it
and,
among
other
things,
I
just
need
more
input
and
more
good
ideas
to
make
it
that
way.
I
think
that,
for
me
you
know
it's
it.
Some
of
this
goes
back
to
a
simple
made.
Easy
we're
back.
Somatic
stuff
is
orthogonal
and
infinitely
composable,
but
my
target
audience
is
at
least
for
the
tech.
J
Think
I
read
that
and
someone's
log,
and
who
knows
somebody
be
on
this
call,
but
the
repple
might
be
the
thing
that
beats
out
the
notebooks
for
me
someday
if
there's
a
way
to
put
that
graph
or
that
scatter
plot
on
the
repple
and
then
kind
of
get
our
data
scientists
to
start
thinking
about
this.
This
is
a
function
rather
than
a
mishmash
of
all
this
other
garbage,
but
yeah.
That
sounds
good
well,.
C
C
So
there
are
that's
we're
having
a
real
numerix
library
and
having
a
strong
lyrics
background
where
you
can
have
the
option
in
the
dewclaw
immutable,
but
you
have
the
option
for
mutable
when
you
get
in
a
hole,
because
you
have
a
lot
of
data.
That's
you
know.
That
was
also,
in
my
opinion,
one
of
the
core
designs
behind
closure.
That
was
really
smart.
Was
it
wasn't
functional
or
bust?
C
It
was
functional
first,
but
pretty
mutable
if
you
want
second-
and
that
is
fair,
I
think
and
that
allows
you
to
get
the
performance
which
I've
at
least
my
experience
talking
to
people
who
have
tried
to
use
closure
at
times
for
machine
learning
has
been
when
the
performance
isn't
promised,
and
then
it's
not
there.
That
leaves
a
very
strong
bad
taste
in
someone's
mouth
yeah.
D
C
C
So
as
far
as
that
question
goes
I
don't
at
this
point,
if
I'm
the
one
doing
the
machine
learning,
then
I
will
do
it
in
closure,
but
I
wouldn't
go
to
somebody
out
of
the
blue
and
say:
hey:
do
you,
for
instance,
they're
getting
pitched
by
a
Python
shop?
I'm,
not
gonna,
go
with
them
and
be
like
dude
enclosure.
I
wouldn't
assume
I
could
win
that
battle
and
it
doesn't
matter
how
right
you
are
at
the
end
of
the
day
in
that
situation.
So
you
know
I
I,
like
the
language
design
behind
closure.
C
A
lot
more
than
Python
I
did
a
lot
of
Python
I
done.
A
lot
of
I
did
a
whole
lot
of
code
generation
in
Python
back
in
my
C++
in
game
development
days.
So
all
the
reasons
that
make
closure
great
aren't
things
that
you
explain
to
a
business
exec
or
at
least
they're,
not
things
that
I
would
feel
comfortable
trying
to
pitch
on
a
lot
of
times
so
I.
If,
if
you're
gonna
do
a
small
web,
app
or
or
an
application
where
it
can,
I
can
really
know,
I
can
hit
closure.
C
Strengths
then
sure
I'll
pitch
that
against
anybody.
What
machine
learning
is
not
there
yet?
So
the
answer
is
I.
Wouldn't
right
now
attempt
that
battle,
and
hopefully
soon
I
would
but
before
I
can
attempt
that
battle.
Lots
of
people
have
to
have
had
success
with
closure
in
machine
learning.
There
have
to
be
lots
of
notebooks
of
people
doing
it
successfully.
There
has
to
be
at
least
a
couple
Python
people
who
are
like
man
I
did
this
enclosure
and
it
worked
great
for
me
and
there
have
to
be
a
few.
A
Yeah,
so
now
we
a
little
bit
into
the
second
hour
of
the
meeting.
So
maybe
let
us
try
to
conclude
this
path
soon
and
I
thought.
If
there
isn't
any
other
question
to
Chris,
then
I
thought
it
could
be
nice
to
see
a
little
bit
of
the
code
structure
of
what
you
have
been
so
that
people
can
get
to
know
well-to-do
if
they
wish
to
contribute
or
whatever,
and
is
it
a
good
day?
You
show
something.
C
E
C
C
A
F
Just
under
the
line
quickly
and
so
I'm
Henry
gana
I
wrote
the
data
science
I
am
dropping
data
scientists
I
try
to
use
closure
as
much
as
possible,
but
equally
I
can't
write
a
lot.
The
libraries
that
I
need-
and
it's
easier
sometimes
to
reach
for,
are,
and
so,
although
I
don't
have
any
particular
proficiency
as
an
art,
programmer
I'm
sure
I
write
terrible,
are
code.
I
write
what
I
need
in
order
to
get
the
job
done
and
I
have
a
sort
of
site
project.
Was
a
series
of
cyber
attacks
really
kicks?
He
stats.
F
Is
that
main
library
that
I
I'm
known
for
uses
transducers
to
do
simple
statistics?
I
love
the
way
that
library
is
written
by
the
way
I
think
it's
beautiful?
Well,
it's
very
kind
of
you.
Thank
you
very
much.
I
would
I'd
love
to
write
more
I
have
a
seven
month
old
daughter
at
the
moment.
I
can
hear
her
downstairs
from
time
to
time,
so
my
output
has
definitely
decreased
recently.
The
key
thing
on
my
wish
list
and
I
think
Jo
kind
of
touched
on
this.
F
I've
never
gone
on
with
them,
with
the
notebook
style
of
working
I
mean
I'm
sure
it's
just
me,
obviously
they're
incredibly
popular
I'm,
most
at
home
in
the
repple.
Mr.
omen,
Emacs
I
love
to
do
as
much
as
possible
there,
and
so
the
thing
that
I
miss
most
is
exploratory
data
science.
You
know
I
know
that
I
can
use
closure
to
build.
F
You
know
anything
if
I,
if
I
give
it
enough
time,
but
when
I'm
confronted
with
a
new
data
set
and
all
I
want
to
do,
is
its
discover
kind
of
you
know
it's
its
shape
in
that
kind
of
abstract
sense.
I
want
the
quickest
path
to
getting
some
visualizations
and
at
the
moment
accomplishing
that
in
the
repple.
F
F
Think
that
the
substance
of
your
your
talk,
which
I
thought
was
was
fantastic,
but
my
my
particular
headspace
is
quite
different
at
the
moment,
and
it's
it's
in
how
to
how
to
scratch
my
own
itch
I,
don't
think
it
would
make
necessarily
closure
or
more
popular
language
for
data
science,
because
you're,
maybe
not
going
to
reach
that
audience
of
people
who
don't
even
know
why
the
repple
is
important
to
them.
They
want
to
do
their
job
in.
C
Vega
was
ported
to
one
of
the
graphing
libraries
in
Java,
so
that
I
could
use
Vega
without
having
to
bounce
to
clojurescript
and
back
because,
when
you're
in
the
repple,
obviously
bouncing
actually
I
do
a
big
markdown
file
or
something
like
that
and
then
I
I
just
view
the
whole
thing
at
once
and
just
refresh
it
all
the
time.
So
it's
clumsy
I
think
actually
he's
on
the
call
Tomas
has
a
good
way:
yeah
Tomas.
Could
you
speak
a
little
bit
about
how
you
do
these
things?
B
E
B
Perfect
and
it's
a
fast
math,
it's
a
library
which
Rob's
a
patent
on
mascara
and
turned
and
smiled
library
written
in
Scala
and
Java,
and
currently
I'm
producing
the
chat
charting
library
which
is
written
directly
into
in
on
JVM.
So
it's
it's
in.
In
the
contrary
of
current
current
approaches,
which
is
web
based
approach
and
you
have
a
lot
of
libraries
based
on
Vega
and
JavaScript,
so
I
decided
to
because
of
you
mutation
mentioned
partly
by
Henry.
B
I
have
I
need
that
I
needed
the
library
which
is
easy
to
use
from
the
from
the
rail
rattle
and
it's
able
to
save
and
display
I'm
going
to
generate
hundreds
of
charts
charts
at
once
without
needing
running
web
browser
and
savings
from
the
screen.
So
I'm
producing,
plotting
libraries,
enclosure,
plot,
CL
g,
j
plot
I
will
I
will
share
a
link
soon.
It's
still
in
a
very,
very
early
stage
stage.
It's
a
work
in
progress,
so
what
I'm
doing
now
is
producing
I
working
on
Emacs
using
him
ax
and
read
inside
there.
B
So
I'm
go
through
the
things.
That's
book,
and
probably
you
know
this.
This
is
a
very,
very
basic
book
for
which
explains
basic
statistics,
but
I
use
it
for,
for
my
test
test
force
and
such
an
examples,
I
use
dissing
examples
to
test
to
test
my
library
and
test.
Does
they
approach
you?
If
you
click
on
link
here,
it's
a
markdown
file,
markdown
file
produced
directly
from
the
closure.
Oh,
you
still
show
your
skin.
B
B
B
So
that's
that's.
Does
the
idea
so
I
don't
need
to
any
notebook
librarian
in
the
club,
Jupiter
cetera,
I
I,
just
write
in
the
Emacs
and
generate
the
markdown
I
will
show
you
the
scripts.
This
is
the
script
which
which
generates
the
the
markdown
file.
It's
very
speaking,
23
lines,
long,
it's
using
marginalia
margin,
area,
a
parser
and
just
generates
not
nothing
special
and
for
me,
it's
easier,
easier
and
very,
very
easy
as
a
very
easy
to
work
with
just
pure
editor
and
later
before,
publishing
content
generate
market
on
self.
That's
all.
A
Okay,
yeah,
that
is
really
exciting,
and
if
I
may
ask
about
it,
then
yeah
I,
guess
what
Henry
said
is
not
only
about
needing
visualization
library,
but
also
about
being
able
to
have
a
decent
record
session
so
exploring
right
now
so
could
you
speak
a
little
bit
about
how
your
interactive
process
for
working
with
data
looks
like
because
you
wouldn't
want
to
render
everything
all
the
time
I
would
like
to
look
into
little
bits
and
pieces
of
it.
Yeah.
B
Okay,
it's
yes,
I
hear
it.
It's
problematic,
still
problematic.
I,
really
wait
for
forward
creases,
a
tecmo
ml
dataset
library
for
fur
for
using
it.
I
asked
Chris
already
to
take
a
look
of
this
tomb.
I
think
starts
examples
and
to
think
how
this
direct
direct
table
saw
calls
could
be,
it
could
be
changed
to
the
technical
data
set
they
currently
currently
I
use.
Let
me
show
you
the
it
out.
Currently
I
have
some
small
wrapper
around
table,
so
it
just
wraps.
Could
you
zoom
in
a
little
bit
yeah.
B
B
Accessing
the
roles,
operating
columns,
etc,
etc.
It's
really
really
simple,
and
here
we
have
missing
values.
Successor,
select
values,
create
account,
they
would
creators
some
simple
protocols
about
the
number
value,
values
or
missing
values
and
something
which
brings
stable
and
F
map
map,
which
maps
maps,
data
and
stays
with
the
type
of
table
so
type
the
table
so
column.
B
Here's
the
example
and
I,
and
we
have
kind
of
exploratory
analysis
of
the
data
set
so
I'm,
taking
the
shape,
prints,
some
rows
and
columns,
column
names,
etc,
etc
and,
like
liking,
a
Henry's
book,
and
then
it's
a
chapter
and
when
you
do
exploratory
analysis,
so
we
can
see
the
size
access.
Some
some
values,
for
example,
value
counts,
is
the
same
as
and
take
the
frame
mated
method
frequencies.
B
C
B
I,
like
it
too
much
yes
as
an
reset,
and
we
discussed
before
data
frame
library.
Yes,
it
would
be
very,
very
helpful
or
I'm
happy
to
that.
That
table
saw
exists.
I'm
also
happy
to
operate
with
a
sec,
you
of
sequence
of
maps,
but
yeah
we
need.
We
need
to
think
how
to
produce
a
provide
common
command
set
of
functions
of
common
set
of
protocols.
Part
of
them
are
done
on
car
matrix,
but
not
for
exploratory
analysis.
I
think.
B
G
I'm
I'm
using
this
is
Chris,
saying
I'm
using
Thomas's
library.
Clj
plot
is
really
good.
The
lattice
feature
is
really
good,
sir.
It's
like
when
you
have
like
three
plots
and
four
plots
like
yeah.
You
put
12
plots
all
in
a
row
at
the
same
time
and
he's
got
a
really
good
language
for
adding
plots
to
the
bottom
to
the
top
to
the
sides
like
it's
it's
it's
very
cool.
G
B
B
B
Replicate
Vega
and
replicate
lattice
our
library
Gigi
plots
so
so
and
add
up
at
something
more
so
you
can
see
here.
You
see
this
is
the
screen
yeah?
Okay,
so
we
can
see
here
exact
the
same
examples
of
as
on
Vega
side.
So
we
can
see
every
plot
every
possible
power
plot
here
with
Stavros
bar
charts
line,
charts
area,
charts,
scatter
plots
bubble
plots,
and
this
is
a
chart
with.
B
B
L
B
B
D
D
This
is
really
awesome,
stuff,
you're
doing
and
I've
been
following
it
on
the
Koreans
discussion
so
on.
Nor
can
there
enough
it
looks.
Awesome.
I
was
just
wondering
you
have
your
exploratory
session
there.
You
have
kind
of
a
nice
DSL,
maybe
for
for
you
know,
I,
guess
in
the
data
and
shaping
and
so
on.
Do
you
have
a
plan?
Do
you
mention
that
you
were
thinking
about
porting
that
to
the
tech
ml
star?
B
B
Concepts
behind
like
like
the
table
so
like
just
pure
second
seconds
of
maps.
Oh
so
this
is
question
mark
to
crease.
If
it's
well
how
to
use
it
as
exploratory
his
dataset
librarian,
how
to
use
it
as
exploratory
library
and
what
we.
What
should
be
there
to
have
such
such
a
pleasure
pleasure
to
work
with
like
like
an
error
or
pandas
or
whatever
yeah.
C
C
So
in
that
sense,
if
you
the
difference
so
far,
what
I've
seen
the
difference
between
the
two
is
not
much.
So
if
you,
for
instance,
I
I,
will
take
a
look
at
these
and
look
at
how
hard
to
do
this
exploratory
stuff
was
will
be
to
do
in
in
the
tech
ml
system,
but
I,
don't
it's
I
would
say
that
if
you
chose
to,
for
instance,
just
go
with
what
Tomas
wrote
and
then
later
on,
you
wanted
to
use
tech
ml.
C
C
Yeah,
that's
that's
kind
of
all
I
said
and
Isis
when
I
built
the
the
data
set
abstraction.
Specifically,
it
was
more
geared
towards
taking
a
concept
in
training
and
making
it
production
because
I
felt
like
that
was
one
thing
that
I
could
add
to
the
conversation
that
pandas
didn't
have,
but
the
truth
of
the
matter
is
and
I
do
think
that
was
true.
That
was
accurate,
but
the
truth
matter
is
most
of
your
work
is
exploration,
so
it
I
kind
of
geared
it
towards
adding
to
a
conversation
and
not
it
towards
the
most
practical
thing.
C
I
could
have
done
mmm-hmm
and
so
version.
Two
will
be
geared
much
more
towards
exploration.
I
heard
that
loud
and
clear
from
like
several
people
and
the
people
who've
done
exploration
with
the
toolset
that
I
built
successfully
so
far
have
been
basically
building
big
pipeline
generators.
So
they've
been
doing
big
grid
searches
across
the
space
of
pipelines,
they
could
possibly
do
and
that's
a
different
form
of
exploration,
but
that's
much
higher
level.
Then
I
got
a
new
data
set
and
I
have
no
idea
what
I'm
looking
on?
How
do
I?
J
J
I
have
had
a
lot
of
success,
even
insofar
as
kind
of
calling
models
built
by
other
statisticians
midstream
and
things
along
those
lines.
So
I
don't
think
it's
in
the
context
of
this
discussion,
but
maybe
it's
useful
for
folks
to
kind
of
I'm
sure
everyone's
thinking
that
that
sort
of
thing
would
certainly
be
possible
with
a
lot
of
the
tools
and
techniques
that
we're
talking
about.
J
J
A
K
Sorry,
my
VDC
assistant
crap
has
tangled
up
my
camera
and
I've
tried
to
kill
all
and
it
hasn't
freed
it
up
for
zoom,
so
I'd
have
to
zoom
on
the
camera
on
otherwise
I.
So
I
was
going
to
talk
about
that
a
little
bit
so
one
of
the
kind
of
closure
data
science
success
stories-
that's
worth
looking
at
on
the
streaming
side,
is
what
the
HCA
guys
are
doing
and
they've.
K
Given
a
couple
conch
talks,
they
have
this
kind
of
Capet
caf-co
data
streaming,
workflow
I,
think
one
of
them
is
called
black
closure
versus
cancer
or
no
sorry
for
versus
sepsis.
That's
the
one
and
there's
another
one.
The
other
thing
is
the
notebook
thing
like
this
is
something
that
happens
a
lot
and
why
I'm
personally
kind
of
cautious
about
how
to
reshape
the
work
of
like
the
notebook
workflow
in
doing
things
on
the
closure
side.
K
Notebooks
are
a
good
tool
for
sharing,
and
the
main
thing
they
accomplish
is
that
they
inline
visualization
and
programming.
So
it's
less
of
the
literate
programming
side
and
moreso
of
the
being
able
to
do
visualization
without
context
shifting
to
looking
to
the
even
to
another,
monitor
or
you
are
you
know,
website
or
something
or
a
folder
directory,
where
your
visualizations
are
getting
done.
I
think
that's
the
main
time-saving
thing,
as
I
kind
of
you
know
commented.
K
You
know
you
know
a
side
way
there.
The
problem
with
that
is
people
get
in
this
conditioning
environment,
where
their
only
means
of
getting
okay
I
did
something
I'm
gonna
do
something
else
like
that:
I
don't
have
an
abstraction
where
I'm
forced
to
use
or
led
towards
using
functions
and
parameters
and
doing
stuff.
K
So
that's
what
most
of
that
stack
is
around
and
I
know:
they've
got
some
open
source
platform
there.
The
other
thing
is:
I
am
doing
some
client
work
right
now.
It's
a
fairly
open
client,
but
just
you
know,
in
terms
of
being
a
little
conservative
want
to
talk
too
much
about
it,
but
I'm
curious
if
other
people
are
doing
much
integration
of
diatomic
anywhere
and
the
data
science
stack
or
how
much
you're
integrating
with
Python
or
our
teams.
It
seems
like
this
is
a
common
pain
point.
E
K
J
Not
yet
for
us,
I
was
I
have
gone
on
my
to-do
list
to
put
some
kind
of
proof
of
concept
work
together
with
day
Tomic
as
compared
to
well.
It's
actually
a
cosmos
DB
collection,
which
is
Microsoft's
slant
on
I,
guess
and
that
stuff
that
we
really
don't
need
and
immutability
would
be
super
helpful.
So
I
need
to
turn
that
into
kind
of
my
executives,
language
and
then
weigh
the
pros
and
cons
and
see
if
it's
a
good
fit
for
our
system,
but
not
yet
at
our
shop
anyway.
Cool.
K
The
reason
I'm
mentioning
this
is
it's
already
on
our
plate
to
open-source.
We
just
don't
know
what
the
time
table
is
for
these
things
and
we're
not
sure
with
the
ETL
library.
If
that's
going
to
be
open
source
is
a
library
if
we
end
up
leaving
it
kind
of
as
is,
and
it
just
comes
out
as
an
example,
then,
where.
K
Apache
arrow
is
something
that
or
any
of
these
other
you
know
wrapping
or
similar
JVM
efforts.
They
they
seem
like
something
that's
right
on
the
side,
that's
proximal
to
the
model
or
to
the
person
doing
numerical
analysis.
So
so,
for
example,
if
you
were
doing
this
entirely
enclosure-
and
you
didn't
have
people
doing
things
in
our,
it
would
be
on
the
other
side.
So
you
would
have
queries
that
would
come
back.
That
would
get
parsed
automatically
into
your
pandas
like
table.
K
So
then,
all
the
follow
on
that
work
that
you
do
against
that,
which
is
what
they're
doing
right
now
effectively.
Is
they
issue
these
queries
and
they
get
back?
You
know
the
tuples
in
the
data
log
query,
get
parsed
into
tables
on
the
our
side
or
our
data
frames,
and
they
do
all
that
work
against
our
data
frames
with
visualization
and
whatnot.
Sometimes
if
the
queries
are
for
really
large
objects
or
blobby
things,
it's
there's
another
step
to
read
that
out.
K
The
detail
side
is
the
interesting
one,
because
what
it
is
doing
as
closure
becomes
a
vehicle
by
which
they
they're
kind
of
harness
or
harnessing
or
tangling
the
mess
of
all
the
data
science
practices
of.
Oh,
we
have
in
different
schemas,
because
every
time
somebody
comes
up
with
an
idea
about
how
to
look
at
data
in
an
are
that
in
our
that,
basically
entails
a
schema
for
the
data
and
and
rather
than
becoming
a
one-off
view
or
something
like
that
or
somebody's
on.
You
know,
idiosyncratic
working
thing.
K
Everything
that
gets
preserved
gets
preserved
into
a
table
of
that
one-off
form.
So
if
you
need
to
sit
down
as
a
researcher
to
combine
insights
from
multiple
data
sets,
the
problem
is
trying
to
make
sense
of
in
different
schemas,
where
every
single
data
science
workflow
produces
data.
That
of
an
arbitrary
you
know,
scheme
of
you
have
to
figure
arbitrary
in
the
in
the
opposite
sense
of
what
it
means
right,
a
schema
that
could
be
anything
that
you
have
to
figure
out
immediately
when
we
sit
down
with
the
data
set.
K
Don't
even
remember
all
the
different
like
kind
of
low-level,
optimization
things
or
whatever
over
the
winter
and
I
realized,
I'd,
probably
bitten
off
more
than
I
could
chew,
and
I
was
delighted
to
see
that
you
guys
said
all
independently
jumped
into
that
on
the
tech
ascent
side.
So
I'm
super
grateful
to
see
all
that
TV
and
work
pay
off
and
there
being
a
way
to
reach
directly
into
that,
rather
rather
than
just
the
MX
net
abstraction
at
the
top,
which
is
also
super
powerful
on
another
important
tool
chain
there.
K
So
thanks
on
that
side,
the
other
thing
I
wanted
to
say
is
in
terms
of
how
to
frame
it
is
I
think
it
makes
sense
to
frame
the
goals
for
closure
data
science.
The
same
way
that
closures
goals
are
largely
framed
right,
which
is
closure,
should
produce
a
clear
advantage.
If
you
work
in
closure,
you
want
it
to
produce
a
clear
advantage.
You
don't
want
to
go,
convert
every
R
or
Python
person
that
you
know
to
closure.
K
It's
already
good
at
addressing
some
of
those
things
and
it
can
go
in
and
do
that
without
necessarily
having
to
displace
the
software,
that's
being
used
to
produce
models
or
to
do
the
low-level
numerix
here
in
there
and
so
I
think.
That's
a
reasonable
strategy
to
take
when
you're
trying
to
come
up
with
a
compelling
case
for
closure
for
the
business,
where
it's
not
being
used,
of
course,
to
get
made.
K
A
scientist
writing
closure,
primarily
I,
think,
is
a
much
longer
struggle
and
I
think
it
has
to
be
kind
of
an
iterative
path
towards
figuring
out
what
the
other
advantages
you
can
pull
in
here
and
there.
The
other
thing,
I,
wasn't
sure,
is
how
much
people
were
looking
at.
The
julia
community
are
paying
attention
to
the
things
they're
doing,
but
I'm
yeah,
because
that's
really
much.
C
So
I
pay
real
close
attention
to
julia
and
so
far
what
I've
seen
is
they
have
a
more
thorough
numeric
stack.
Numeric
stock
goes
up
through
complex
numbers,
which
is
similar
to
APL
by
the
way
which
had
that
long
time
ago.
So
the
same
kind
of
stuff
you
can
do
include
her,
where
your
numerix
can
go
from
a
byte
to
a
double
depending
on
your
operation
and
real
fractions.
Julia
has
this,
but
one
interesting
thing
that
they
have
is:
they
have
auto
compilation,
a
thing
like
TVM
right.
C
You
can
just
write,
straight-up
julia
code
and
it
has
auto
compiler
that
can
put
it
on
your
GPU
for
some
subset
of
problems
or
it
can
distribute
across
a
thousand
GPUs
for
some
subset
of
problems
which
is
I,
think
yeah.
You
know
it's
hard
to
do
that
with
closure
and
meta
programming.
I've
thought
about
this
a
lot
and
some
of
it
comes
down
to
the
language
that
you're
using
to
do
it.
C
C
K
Yes,
so
the
thing
that's
interesting
about
zygote
is
I
go
so
if
you
guys
aren't
familiar
with
zygote,
it
does
auto
differentiation
as
a
compiler
extension
rather
than
by
making
you
use
a
tensor
library,
like
you
know,
pi
torch,
tensor
flow,
etc.
Whatever
the
way
they
work
is,
if
you
use
their
library
extraction,
that
will
provide
some
means
in
the
way
that
the
operators
are
overloaded
for
doing
matrix,
math
to
track
everything
done,
whether
it's
teint
based
auto
diff
or
whatnot,
and
use
that
as
the
foundation
for
auto
differentiation
in
those
frameworks.
K
What
zygote
does
is
it
says
I'm
actually
going
to
do
Auto
differentiation
on
a
more
primitive
representation
in
the
compiler
and
that
that
is
based
on
the
fact
that
julia
is
built
by
Lisp
weenies
and
in
the
middle
of
it
there's
finto
Lisp
and
it's
using
s,
expression,
expansions
and
con
cells
as
the
primitive
as
the
basis
for
doing
that,
auto
differentiation
dude,
the
only
thing
the
reason
I
bring.
That
up
is
because
this
comes
out
of
them.
K
K
I
can
just
use
normal
operators
and
matrix
operators
etc
and
I
get
that
functionality
for
free,
because
the
language
does
a
whole
lot
of
heavy
lifting
for
me
and
in
general,
I
think
this
is
the
weakness
that
Python
and
I
are
going
to
have
over
time,
as
the
language
does
not
do
any
heavy
lifting
for
you
languages.
The
work
is
always
done
behind
an
API
call
and
it's
done
explicitly
in
C++,
and
if
there
are
things
lessons
learned
or
time
saved
at
the
library
level,
those
things
are
always
either.
C
Interesting
I've
thought
about
that
a
bit
with
all
this,
because
I
do
so
much
bridging
between
writing
code
on
the
JVM,
that's
literally
stored
in
like
native
JVM
stuff,
that's
confusing
to
say
that
stored
in
a
java
array
and
I
write
a
bunch
of
code.
That's
then
using
a
native
buffers,
so
it's
allocated
with
malloc
and
I.
C
You
know
one
because
pythons
so
primitive
in
a
lot
of
ways
and
the
interpreters
not
very
powerful.
They
had
to
go
to
native
things
really
early,
which
gave
them
a
lot
of
performance
once
they
got
over
that
hump
of
having
to
go
to
native
things
and
some
of
what
I'm
talking
about
like
building
bridges
to
other
worlds.
C
This
isn't
really
a
closure
problem
as
much
as
it's
at
the
JVM
problem
is
that
jaebeum
is
dangerously
powerful
with
in
and
of
itself
not
having
to
bridge
to
another
world,
and
it
makes
those
bridges
much
more
rare,
so
they're
coming
much
later
in
the
process.
So
a
lot
of
the
ML
stuff
were
writing
now
I'm.
C
You
know
like
I'm,
finding
out
how
to
build
a
bridge
to
C++
or
C
and
in
Python
they
had
to
do
that
literally
15
years
ago
to
get
anything
working
whatsoever,
whereas
in
Java
we
got
a
lot
of
like
pair
local
and
things
like
this
that
work
on
the
JVM,
but
they're
never
going
to
be
optimizable
by
T
VM
or
if
they
are
it's
going
to
be
much
weaker,
because
hotspot
has
so
much
worse,
Sindhi
performance
than
native
C++.
Those
are
a
lot
of
other
things.
C
E
A
It
doesn't
make
sense,
sure
yeah,
so
so,
I
guess
and
what
what
we
have
had
here
is
like
one
talk
and
actually
several
other
things
that
are
worth
the
talk
it
may
be
in
the
one
in
the
upcoming
meetings
and
what
we
intend
to
do
is
meet
every
couple
of
weeks,
two
or
three
or
four
and
I
guess
anybody
here.
If
you
wish
to
talk
about
anything,
then
let
us
talk
about
it
and
set
a
meeting,
because
we
just
saw
that
it
wasn't
too
difficult
and
what
we
are
about
in
the
like.
A
The
next
few
weeks
is
doing
other
other
parts
of
building
of
a
community
which
means
having
a
website
which
is
a
have
all
kind
of
information
of
tutorials
and
libraries
and
places
where
people
can
discuss.
And
what
we
are
also
have
been
doing
is
trying
to
build
little
work
groups
where
people,
small
books
or
people
can
have
focused
discussions
of
topics
they
wish
to
think
about
together.
The
one
group
is
doing
it
discussing
ml
start.
A
If
anybody
has
to
leave
now,
then
thank
you
so
much
for
joining
this
meeting
and
and
let
us
meet
again
soon
and
now.
I
guess.
Let
us
just
continue
as
much
as
you're
willing
to
keep
chatting,
because
it
was
fascinating
in
the
last
few
minutes
yeah.
So
so
it
was
stopping
you
from
all
right.
Then
I
guess
I
just
stopped
you
because
you
were
going
to
say
something
really
fascinating.
G
K
You
know
quasi,
declarative
or
whatever
and
then
would
hit.
You
know
fast
targets
and
then
I
knew
had
already
been
battle
tested
and
would
work
for
a
machine
learning
framework,
since
it
was
working
with
him
x
net,
but
that
was
also
partially
kind
of
just
a
learning
experience
for
me.
I
didn't
I,
think
Chris
came
at
it
with
way
more
clear
objectives
and
with
a
lot
more
familiarity
with
the
domain
in
which
it
was
operating.
So
that's
kind
of
an
aside,
that's
probably
not
worth
digging
into
in
detail.
C
One
thing
I
will
say
and
I've
said
this
before
TBM
is
important,
not
necessarily
just
because
tedium
does
amazing
things,
but
because
it
teaches
you
a
different
way
to
think
about
performance,
optimization
and
any
tool
that
teaches
you
a
different
way
to
think.
It's
probably
worth
it,
in
my
opinion,
to
get
to
use
a
tool
enough
to
warp
your
mind
to
its
way
of
thinking,
because
similar
to
how
functional
programming
changes
the
way
you
think
the
TVM
system
will
do
the
same.
A
Yeah
and
Ben
I
guess
one
thing
of
the
many
things
that
you
told
I
was
really
interesting
to
me
and
I
hope
you
could
tell
about
it
more,
and
that
was
that
the
exact
part
of
the
process.
Well,
closure
was
useful
to
our
people
in
what
they
do
and
and
the
way
you
created
communication
between
these
two
parts
and
that
in
a
sense,
I
guess
that
seems
very
similar
to
what
we
are
seeing
here
at
my
office.
A
K
I'd
like
to
talk
more
about
I,
have
to
jump
off
for
another
meeting
right
now,
unfortunately,
with
with
those
people,
but
I
will
say
that
we
enough
of
the
project
has
been
considered
the
success
that
we
are
looking
at.
The
next
steps
for
presentation,
evangels
and
whatever
both
for
the
project
itself
and
in
general,
the
the
tools
and
possible
open
sourcing
of
those
tools
are
that
planned,
open
sourcing
of
those
tools
so
I'll.
Let
people
know
as
soon
as
I
know
more
on
that
front,
where
we
are
currently
having
that
conversation.
A
J
So
I
just
kind
of
again
a
comment,
I
guess
so
I
think
we
all
probably
have
seen
the
blogs
and
are
aware
of
the
kind
of
next
superpower.
From
my
perspective,
anyway,
with
closure,
dot
speck
I've
been
using
that
for
kind
of
data,
profiling
and
I'm
kind
of
excited
to
try
to
plug
that
into
some
of
Thomas's
repple
driven
pitcher
workflow.
There
I
think
that'd
be
really
cool.
If
I
had
you
know
directory
full
of
specs
for
different
types
of
regular
expressions.
J
For
you
know
a
hundred
thousand
a
hundred
thousand
maps
in
a
seek
that
I
could
profile
and
then
get
into
pitchers
and
things
like
that
and
then
I
started
thinking.
I
wonder
if
spec
to
be
used
for
feature
engineering
at
all
or
things
along
those
lines
and
how
messy
that
would
get,
but
that's
pretty
much
yeah.
Maybe
that's
something
I'll
play
with
that's
all.
J
C
Much
I
have
struggled
with
speck
it.
It
I
am
a
very
much
no-holds-barred
know.
Stipe
programmer.
Okay
and
I
know
that
that
causes
problems
and
I'm
not
saying
there
is
no
place
for
speck.
But
for
me
strongly
typing.
My
data
structures
is
rarely
something
that
I
wanted
to
do.
That
being
said,
what
you
said
reminds
me
of
very
strong,
like
a
general-purpose,
strong
NLP
that
can
pull
structure
out
of
pools
of
text
or
streams
of
text.
C
So
all
the
you
know
all
the
all
the
what
are
called
q10
reports
from
the
mark,
all
the
companies
in
the
stock
market
and
being
able
to
pull
that
out
into
some
structured
way
and
do
something
interesting
with
it
to
me.
It
feels
like
there's
some
parallel
there
between
how
you
see
using
spec
and
kind
of
being
able
to
pull
structured
data
out
of
streams
of
text.
J
A
L
L
You
know
initiating
this
one
and
you
if
you,
if
you
guys,
have
an
idea
about
setting
of
the
website
or
that
kind
of
stuff
why
I'd
be
happy
to
help
and
the
other
one
I
was
I
was
a
bit
late
joining
the
meeting,
because
I
was
somehow
thought.
I
thought
this
was
yesterday,
but
anyway,
so
the
other
one.
That
I
was
wondering.
Is
you
guys
anybody?
L
Anybody
has
looked
into
something
called
ml
flow,
mph
law
org
by
the
guys
who
are
building
the
data
breaks,
that's
basically
to
to
have
like
a
complete
platform
for
from
experiment
to
production
kind
of
idea.
It's
it's
written
in
Python,
and
this
is
this
is
the
kind
of
tool
that
not
a
tool
but
a
kind
of
a
big
framework,
II
sort
of
thing
that
helps
a
lot
in
putting
data
science
projects
into
production.
L
For
me,
at
least
so
that's
what
I
was
following
on,
so
maybe
this
could
be
an
interesting
thing
to
look
into
if
you
haven't
already
seen
it
so
basically,
two
things
one.
If
you
have
a
small
working
group
that
you
are
talking
about
Daniel
to
to
set
up
the
website
and
then
compile
all
the
closure
ml
related
things,
I
would
love
to
contribute
and
the
other
one
is.
L
Technically
it
can
be
well,
they
said
you
can
it.
You
can
use
it
for
anything.
You
would
like
that's
the
idea,
so
the
the
point,
because
I
work
with
a
big
bank
in
the
Netherlands
to
set
up
a
machine
learning
project
for
them,
Indian
machine
learning
pipeline
using
spark.
So
one
of
the
challenges
is
okay.
How
do
we
create
multiple
experiments?
L
I
know
how
do
we
pick
the
right
model
and
then
push
it
into
production
and
deploy
it
onto
cloud
or
on-premise
or
something
so
ml
flow
is
basically
like
the
whole
package,
or
at
least
that's
the
aim
to
coordinate
or
not
coordinate,
but
to
design
the
entire
flow
of
your
ml
projects
from
designing
experiments
to
all
the
way
up
to
deploying
them
using
sage
maker
on
AWS
or
a
zoo,
or
something
like
that.
So
I'll
paste
a
link.
A
L
Sage
maker
is
a
double
use:
machine
learning,
environment.
You
can
put
your
own
docker
containers
on
to
Sage
maker
of
its
like
it.
That's
that's
how
we
deployed
our
machine
learning
programs,
because
our
data
scientists
they
use
Python
and
then
they
built
Sai
kit
based
models
and
sage
maker
gives
them
an
ability
to
put
everything
into
a
docker
container.
So
we
can
put
an
API
in
front
of
scoring
or
prediction
API
so
and
it
can
scale
automatically
as
well.
So
this
is
the
is
probably
some
different
world
that
I'm
coming
from
enterprise
Z.
L
But
the
thing
is
that
I
understand
there
is
a
lot
of
usage
of
Python
because
of
the
Sai
kit
and
you
know
correlation
of
libraries
being
available,
but
at
the
same
time
there
is
no
tool
that
can
or
no
agreed
way
of
coding
machine
learning
projects
into
production
as
quickly
as
possible.
That
is
the
biggest
pain
that
we
are
facing
right
now,
so
the
rate
assigned
is
build
all
that
stuff
in
notebooks
and
then
that's
super
nice.
But
then
I
need
to
convert
that
productionize.
C
Well,
it's
pretty
hard
because
you
generate
a
whole
lot
of
intermediate
states
and
data
and
know
who
knows
what
else
and
so
I
think
that
moving
forward
and
and
and
how
much
introspection
can
a
third
entity
who
didn't
build
this
giant
training
pipeline?
How
much
information
can
they
get
out
of
this
training
pipeline
like
what
tables
doesn't
need
to
join
with?
Does
it
you
know,
doesn't
need
to
filter
them,
doesn't
need
to
store
them
or
whatever
else.
What
feeds
into
this
whole
data
pipeline
I.
C
L
That's
that's
the
biggest
pain
point
that
I
see
in
the
machine
learning
things
because
mostly
the
data
scientists.
They
show
me
some
nice
graphs
and
the
notebook
is
fine
and
everything
is
cool
and
then
now
I
need
to
put
into
production
and
then
I
need
reproducibility,
and
then
they
severally
come
up
with.
Oh,
you
know
we
have
a
different
model
now
and
how
can
I
m
data
into
into
that
model?
And
then,
when
can
you
put
this
model
production
for
us?
L
They
of
course
the
technical
evaluation
with
all
the
recall,
etcetera,
not
the
standard
machine
learning.
Statistics
are
okay
but
at
the
same
time
I
need
to
think
like
a
product
owner
where
I
need
to
a
be
test,
different
models
in
production
mmm-hmm.
These
are
the
challenges
that
that
you
know
there
are
not
solved
by
any
of
the
current
tools.
So
maybe
you
know
we
I
know.
Closure
community
has
like
amazing
ideas,
and-
and
maybe
this
could
be
one
of
the
challenges
that
we
can
crack
and
then
I
know
there
are
lots
of
layers.
L
F
A
A
They
enjoy
it
like
they're,
like
the
joy
of
working
with
data
in
closure,
without
having
it
as
a
main
language,
and
actually
what
we
are
discovering
is
that
our
and
closure
so
similar
being
functional
in
a
sense
and
being
least
in
a
sense.
And
so
what
is
the
coming
here
is
just
beautiful.
The
way
of
communication,
which
does
allow
at
least
on
the
outside.
It
allows
to
enjoy
closure
more
and
more
and
hopefully
enjoy
it
closure
in
its
fullest
flexibility,
and
so
it
is
still
a
process.
A
C
A
And
so
anybody
else
willing
to
tell
the
story
Oh
what
they
about
nowadays
and
if
not,
then,
if
a
you
everybody
will
hear
now,
maybe
we
can
talk
a
little
bit
about
where
we're
going
so
so
there
is
a
little
group
of
people
who
are
communicating
almost
every
day
and
we
use
the
zuly.
Do
everybody
feel
wrong
zuly,
it's
like
a
check
yeah
and
for
for
us
at
least
for
several
parties.
A
There
is
a
coop
talking
about
you,
tow
amps
and
a
group
talking
about
building
a
website
and
another
group
which
is
both
a
group
open
for
open
discussions
of
data
science,
and
there
is
also
a
group
which
discusses
how
to
be
the
community
and
you
are
invited
to
join
these
groups
and
see
if
you
wish
to
influence
them
so
which
I.
Let
us
talk
about
it
later.
A
If
you
wish
how
we
wish
you
to
be
involved
in
that
and
I
guess,
maybe
you
can
ask
ourselves
what
is
missing
in
this
thing,
what
is
missing
for
the
community
to
be
productive
if
anything
is
missing
at
all
all
right,
so
part
of
what
we
found
important
is
actually
to
be
able
to
me
to
have
this
kind
of
video
meeting
and
maybe
face-to-face
meeting
one
day
and
and
what
else?
What
else
could
be
helpful
for
us
to
be
more
creative
and
more
maybe
coordinated,
I.
L
L
So
probably
if
there
is
a
public
log
or
something
like
that,
where
you
can
say
okay,
this
is
what
we
need
to
maybe
a
get
a
project
with
their
issues
or
whatnot.
That
would
probably
help
and
and
of
course,
I
mean
having
the
you
know
like
the
like
the
rust
guys.
That
is,
are
we.
Where
are
we
ml
at?
You
know
some
something
like
that
know
that
they
show
all
the
components.
Are
there
or
not
kind
of
it?
We
can
also
say
all
we
are.
L
We,
you
know
ml
yet
non
closure,
for
example,
and
then,
if
we
can
for
me
that
the
most
interesting
part
is
if
I
want
to
talk
to
somebody
from
Python,
then
I
can
say
for
Sai
kit.
We
have
this,
you
know,
for
you
know,
for
chaos.
We
have
this.
You
know
deeply
linked
for
J,
for
example,
so
that
that
kind
of
a
central
place
where
we
can
point
people
to
for
graphing
this
kind
of
libraries.
L
You
know
for
notebook,
gorilla,
Apple
or
whatever
different
things
that
people
have
built
over
the
period,
so
that
I
think
that
could
be
very
helpful
because
then
it
would
guide
the
people
in
the
right
direction.
In
my
opinion,
so
we
can
collect
the
ideas
and,
as
I
said,
I
would
be
happy
to
work
on
the
web
site
because
that's
what
I'm
most
familiar
with
I'm,
not
a
data
scientist,
so
I
can't
help
you
in
that
sense,
and
also
on
the
engineering
side.
A
A
Popular
is
important,
probably
more
than
for
usual
engineers,
well-liked,
more
ready
to
face
the
technical
problems,
because
data
scientists
have
to
focus
on
the
problem
at
hand.
The
technical
story
is
only
like
a
bipod,
a
byproduct
of
walking
and
and
I
guess
in
this
sense
feeling
that
closure
has
a
bias
towards
high
salaries
and
more
experienced.
A
People
is
not
good
news,
because
what
it
means
it
is
is
there
it
is
not
so
friendly
to
they're
more
newcomers
and
people
are
less
experienced
and
what
we
feel
be
locally
at
our
community
in
Tel
Aviv
is
that
we
need
to
wait
we'd
some
to
wait
for
some
of
the
new
advances
like
what
to
mash
and
quiz
were
talking
about.
We
need
for
things
to
cool
them
a
little
bit
before
we
can
try
to
teach
newcomers
and
because
before
we
can
try
to
sell
it
to
other
communities,
that's
the
way
we
see
here.
A
So
maybe
it
is
a
long
time
to
conclude
and
I
guess
we
will
meet
again,
maybe
in
three
or
four
weeks
we
need
to
see
when
the
next
speaker
can
can
set
it,
and
what
were
you
trying
to
do
is
meet
on
different
parts
of
the
week
so
that
different
people
of
the
community
can
can
attend
and
I
know
for
some
of
you.
It
is
late
night,
especially
in
Greece,
then
thank
you
so
much
for
being
able
to.