►
From YouTube: A backstage pass with GitHub Copilot - Universe 2022
Description
Remember when AI was the stuff of science fiction? Or, AI helping with everyday work seemed an aspiration for the distant future? Today, GitHub Copilot is defining a new paradigm for software development. Hundreds of thousands of developers use GitHub Copilot every day as their AI pair programmer, and get to experience efficiency and delight. That deserves a closer look! Join us for a deep dive on how GitHub Copilot works, how research and feedback have shaped its evolution, and what’s coming next.
As always, feel free to leave us a comment below and don't forget to subscribe: http://bit.ly/subgithub
Thanks!
Connect with us.
Facebook: http://fb.com/github
Twitter: http://twitter.com/github
LinkedIn: http://linkedin.com/company/github
About GitHub
GitHub is the best place to share code with friends, co-workers, classmates, and complete strangers. Millions of people use GitHub to build amazing things together. For more info, go to http://github.com
A
Nerds
Engineers
scientists
analysts
people
who
love
to
build
software.
My
name
is
Ryan
J.
Seven,
we
heard
so
many
great
things
about
copilot
and
code
spaces
and
actions
and
projects
and
everything
this
morning.
I
know
that
not
everyone
here
has
had
an
opportunity
to
play
with
copilot
yourselves
and
so
there's
really
no
better
way
to
get
to
know
what
product
than
than
to
dive
right
into
it.
So
I'm
not
going
to
waste
any
time
with
slides
I'm
going
to
get
directly
to
a
demo.
Should
we
do
it
yeah
all
right?
Let's
do
it
all
right.
A
All
right,
all
right,
so
we're
gonna
start
here
by
creating
a
little
method
in
Python
that
is
going
to
just
create
find
the
smallest
of
three
elements
in
a
list.
All
right.
We're
going
to
do
that
by
providing
a
couple,
little
comments
here
and
there's
you
know
you
can
use
sort
which
is
fine
is
good,
but
you
know
what
let's
actually
not
use
sort,
let's
not
use
the
language
features
here,
let's
actually
make
copilot,
do
it
itself.
In
fact,
you
know
what
I
know.
Some
of
those
aren't
very
helpful.
A
Some
are
longer,
let's
make
it
a
little
bit
harder
for
ourselves
and
let's
actually
make
it
the
N,
smallest.
Okay,
all
right
good!
This
looks
pretty
reasonable,
so
I'm
gonna.
Now,
maybe
that
last
line
is
a
little
bit
difficult
and
so
I'm
going
to
say
you
know
what
tell
me
how
exactly
this
thing
works.
A
A
Okay,
what
if
we
want
it
to
tell
us
why
it
has
that
time,
complexity,
all
right!
That's
a
good
thing
too,
a
little
bit
long,
but
now
we
know
a
little
bit
more
than
we
did
before.
Okay,
this
is
solid.
Okay.
What
about
now?
Let's
do
some
unit
tests
make
it
a
random
number
and
copilot
works
best
when
you
give
it
kind
of
good,
Specific,
Instructions,
and
so
we're
going
to
ask
it
to
use
like
you
know,
dynamically
generated
random
list
of
numbers
or
integers
more
like
it
all
right.
A
And
you
know
what
here,
let's
just
kind
of
go
with
that
we'll
try.
Oh
this
looks
pretty
good.
This
looks
pretty
good.
Let's
see
here,
random
up
to
100
different
tasks,
but
you
know
what
I
don't
think
that
I
want
this
to
actually
have
any
duplicates
in
it.
So
I'm
going
to
modify
it
here
to
say:
do
not
allow
duplicates
in
the
list.
Okay
a
little
bit
different,
but
this
actually
looks
about
right.
A
So
what
was
actually
happening
in
this
well
co-pilot
itself:
Works,
maybe
a
little
bit
like
a
kind
of
like
a
language
translation
tool
right
when
you
go
to
Google
Translate
and
you
say:
translate
hello
world
to
Dutch
or
Zulu
or
polish
right.
It
translates
it
there.
But
if
you
want
to
go
from
English
to
Java,
let's
say
well,
it
looks
an
awful
lot
like
public
glass,
so
on
and
so
forth,
and
that's
exactly
what
co-pilot
is
doing.
It's
essentially
learning
language
patterns,
nouns
verbs,
direct
objects,
syntax
to
better
understand.
A
We,
if
you
have
a
sentence
or
an
intent
that
you
wish
to
convey
to
translate
it.
It's
not
referring
back
to
specific
I,
don't
know,
they've,
never
seen
that
data
before
it's
not
referring
back
to
specific
instances.
No,
it's
synthesizing
that,
based
upon
its
understanding
of
those
language
patterns,
now
there's
a
lot
that
we
do
there
to
kind
of
get
from
the
initial
context.
That's
in
your
code,
editor
back
over
to
the
actual
model
and
back
to
a
suggestion.
A
Your
machine,
one
of
the
first
things
that
we
do
is
we
mess
around
with
what
we
call
Prompt
crafting
right.
What
prompt
crafting
is
essentially
doing
is
it's
playing
lots
of
little
tricks
in
the
code
Editor
to
send
context,
maybe
a
little
before
a
little
bit
after
your
cursor
to
the
model,
but
we
want
to
send
good
juicy
context,
so
we
get
the
most
robust
suggestions
back.
One
of
the
things
that
we
do
is
we
play
with
neighboring
tabs
here
right.
A
A
What
we'll
do
is
we'll
find
in
a
sliding
window
on
that
unopened,
tab
or
rather
neighboring
tab
on
a
sliding
window
will
perform
a
comparison
deduplicated
for
all
the
symbols,
so
we're
only
showing
unique
symbols
and
we
perform
what's
called
a
Jacquard
similarity
comparison
between
the
context.
That's
in
your
active
Tab
and
the
actual
content.
That's
in
your
neighboring
tab.
What
this
does
is.
It
draws
a
denture
additional
variable
names,
class
names,
method,
names,
logic
into
the
prompt,
so
that
when
we
send
it
back
over
to
the
model,
we
can
actually
get
the
best.
A
All
of
that
results
in
aggregate
across
all
languages
and
about
a
30
percent
acceptance
rate.
We
talk
about
acceptance
rate.
That
means
that
if
you
as
a
developer,
see
10
suggestions,
you
might
accept
three
of
those
suggestions
in
a
minute
or
in
a
session
right
and,
as
you
saw
before,
in
the
keynote
up
to
40
percent
of
code
being
written
by
copilot
and
our
best
languages
for
languages
like
python
now
I
talked
about
prompt
crafting.
A
It
then
passes
through
a
proxy
server
before
it
goes
on
to
the
model,
to
return
back
kind
of
the
suggestion,
I
want
to
focus
a
little
bit
right
now
on
that
middle
part
that
that
proxy
server,
because
a
lot
of
interesting
things
happen
in
that
middle
part.
You
know
we
talk
about
co-pilot
as
being
an
AI
pair
programmer
right
and
many
of
you
I'm
sure
have
probably
worked
with
a
pair
programmer
before.
Maybe
you
were
debugging
a
live
site
incident,
maybe
you
were
learning
a
new
part
of
the
code
base.
A
A
You
stay
focused
on
the
code
that
you're
trying
to
write
and
not
introducing
weird
stuff
so
I'm
going
to
show
you
a
little
kind
of
video
here
taken
from
one
of
our
kind
of
test,
harnesses
I'm
gonna
go
ahead
and
tell
you
right
now
up
front.
There's
going
to
be
a
phrase
on
the
next
one
that
introduces
bias.
It
declares
bias,
get
up.
We
do
not
tolerate
bias
of
any
kind.
This
is
merely
for
demonstration
purposes
so
that
you
can
understand
how
the
neural
networks
work
right.
A
So
here
we
go
we're
going
to
run
through
this
and
we're
in
our
test
harness
here,
and
we
have
a
phrase
that
implies
bias
of
one
group
against
another,
but
none
of
the
individual
words
are
naughty
words
themselves
right,
I
are
smart.
None
of
those
are
bad
words
on
their
own.
You
have
to
understand
what
the
intent
of
that
is
in
order
to
understand
whether
or
not
it
is
a
biased
statement.
A
So
we're
going
to
run
it
here
up
against
a
bunch
of
different
checks,
and
what
we'll
find
here
is
that
for
hate
speech
it
evaluates
is
true,
because
the
neural
network
is
identified
as
being
likely,
but
the
pattern
that
is
to
say,
were
any
of
the
words
naughty
words
evaluates
to
false.
So
now,
let's
use
another
example
here:
let's
use
the
example
of
unicorn
Candy
now
unicorn
Candy
is
delicious
and
wonderful
and
I'm
sure
everyone
here
would
love
some,
but
it
will
not
evaluate
as
true
here
neural
network,
no
hate
speech.
A
No,
if
we
go
to
co-pilot-
and
we
may
be
attempt
to
prompt
unicorn
Candy
in
here
using
what
are
we
in
now
are
we
in
Python
I
believe
again
right,
Define,
a
new
method
for
unicorn
Candy
copilot
will
only
be
too
happy
to
provide
a
suggestion
back,
allowing
us
to
kind
of
stay
focused
here:
I'm,
a
unicorn
baby
right
now
we'll
go
back
and
now
we're
going
to
add
unicorn
Candy
to
our
list
here.
This
is
going
to
teach
the
neural
network
that
hey.
A
You
need
to
start
recognizing
things
like
this,
as
possibly
being
biased,
possibly
being
a
naughty
word
right
here.
We're
going
to
compare
it
up
against
a
large
kind
of
Corpus
of
essentially
language,
that's
being
used,
whether
with
copilot
or
elsewhere.
Unicorn
Candy
doesn't
evaluate
true,
because
no
one
says
bad
things
about
unicorn
Candy
and
then
we're
going
to
insert
into
our
database
where
we'll
run
that
orchestrator
again
here,
you'll
notice
that
it
evaluates
to
true
as
soon
as
we've
added
it
into
it.
Now
language
is
an
evolving
thing
right
and
so
don't
worry.
A
We
can
realize
our
mistake
realize
that
unicorn
Candy
is
good
and
we
can
remove
it
back
from
our
index
so
that
everyone
can
now
go
and
play
around
and
co-pilot
using
unicorn,
Candy
feeling
safe
and
knowing
that
you'll
actually
get
some
back.
But
when
we
attempt
to
write
out
unicorn
baby
good
to
go,
we
tend
to
write
out
unicorn
Candy,
maybe
not
so
much
right,
and
this
is
just
one
of
many
kind
of
sort
of
adjustments
that
we
make
on
that
proxy
server.
A
A
So
we've
heard
a
little
bit
about
how
copilot
Works,
under
the
hood,
but
there's
so
much
more,
that
happens
between
the
developer
and
their
keyboard
and
the
productivity
gains
and
the
happiness
and
the
flow
that
they
realize
there.
For
that.
We
really
need
a
human
researcher,
and
so
now
I
want
to
invite
our
lead
human
researcher
Arena
to
Stage
Irene.
Please
come
on
out.
B
B
Oh
sorry,
let's
go
back
a
slide.
We
have
done
plenty
of
research
on
co-pilot
and
we
use
it
both
to
inform
how
we
build
the
product
and
also
to
understand
what
kind
of
effect
it
has
so
today,
I
wanted
us
to
take
a
look
at
what
we
have
learned
from
developers
is
their
experience
with
copilot
and
how
is
it
that
this
tool
helps
them
and
I'm
going
to
take
you
all
in
a
little
bit
of
a
detour,
because
the
process
to
understand
how
is
it
that
copilot
helps
with
developer
productivity
and
happiness
for
us?
B
B
The
the
first
piece
of
research
that
we
did
was
called
The,
Good
Day
project,
and
here
we
were
trying
to
understand
what
is
it
that
makes
developers
days
good.
So
we
run
a
internal
experiment.
At
GitHub
we
had
40
GitHub
Engineers
participate
in
kind
of
a
a
proof
of
concept
and
for
two
weeks
every
day
they
filled
out
a
two-minute
survey,
and
so
let's
imagine
this
it's
the
end
of
your
day.
Whatever
time
you
told
us
that
is
for
you,
you
get
a
pink
from
us.
We
ask
you
how
your
work
day
was.
B
We
give
you
these
options,
and
we
also
ask
you
some
follow-up
questions
about
that
day.
How
many
meetings
you
had
how
many
breaks
you
took
how
much
you
worked
with
others
that
day,
that
sort
of
thing
every
day
the
questions
are
the
same,
and
what
that
allows
us
to
do
in
the
background
is
to
start
associating
how
good
of
a
day
you're
having
with
all
these
other
things
that
are
happening
in
your
day.
B
Research
was
that
for
developers
the
state
of
flow
is
precious
anything
we
can
do
to
help
them
stay
in
the
state
of
flow
really
goes
a
long
way
to
help
them
and
that
interruptions
of
any
form
are
really
disruptions
for
developers.
They
really
cause
them
to
have
a
bad
day.
So
let's
remember
that
the
second
research
piece
that
we
did-
and
that
is
relevant
here-
was
from
last
year's
state
of
the
octopus
report
and
I
will
do
a
small
parenthesis
here
to
say.
There's
a
brand
new
state
of
the
octopus
report
for
2022.
B
definitely
make
sure
to
check
it
out,
but
what
I'm
saying
here
is
from
last
year's
investigation.
Last
year
we
saw
that
teams
that
introduced
Automation
and
how
they're
doing
their
work
not
only
reported
43
percent,
better
performance.
They
also
reported
30
percent
higher
fulfillment
with
their
work,
and
why
is
that?
B
Let's
go
back
to
our
main
path
enough
with
the
detour.
How
does
all
of
this
relate
to
Copilot?
We
believe-
and
we
see
it
in
people's
feedback,
that
it's
these
same
Concepts.
The
importance
of
flow,
the
benefits
of
automation
for
doing
satisfying
work,
these
same
Concepts
apply
to
co-pilot
and
they
resonate
with
developers
as
they're
using
the
tool
we
surveyed
over
2
000
developers
who
used
copilot,
because
we
wanted
to
understand
their
experience
and
what
it
is
that
they're
getting
out
of
it-
and
here
are
some
of
the
things
that
we
saw.
B
We
saw
that
59
reported
that
they
feel
less
frustrated
as
their
coding
when
they
use
copyright.
77
reported
that
they
spent
less
time
searching
and
73
percent
told
us
that
using
copilot
helps
them
stain
the
flow.
You
saw
how
copilot
Works
in
raya's
demo
as
developers
are
in
the
editor
and
they're
typing.
Copilot
delivers
them
suggestions
that
include
a
lot
of
the
answers
and
the
examples
that
developers
otherwise
would
have
to
go
somewhere
else
to
to
look
for,
and
that
would
have
created
a
form
of
interruptions,
and
we
already
know
that.
B
B
We
also
heard
in
the
same
feedback
very
similar
benefits
to
what
we
have
heard
when
teams
introduce
other
forms
of
automation,
we're
at
87
percent
reported
that
they
spend
less
mental
effort
on
repetitive
tasks,
60
reported
that
they
feel
more
fulfilled
with
their
job
when
they're
using
co-pilot
and
74
reported
that
they
are
able
to
focus
on
more
satisfying
work.
This
is
excellent
ground
for
that
positive
effect
on
fulfillment
and
retention
that
I
mentioned
earlier.
B
So
this
is
all
sounds
really
great.
You
know
for
developer
happiness
for
developer
satisfaction
for
productivity
developers
seem
to
be
able
to
stay
in
the
flow
when
they're
using
copilot.
They
are
able
to
focus
on
on
satisfying
work.
There
was
one
more
thing
that
we
noticed
in
the
same
survey:
88
of
the
people
who
participated
reported
that
when
they're
using
copilot,
they
complete
tasks
faster
and
96
percent,
which
is
almost
everyone
said
the
same
thing
specifically
for
repetitive
tasks.
B
This
has
been
one
of
the
earliest
and
most
common
hypotheses
around
co-pilot
is
that
when
developers
use
it
they're
able
to
complete
task
faster
and
from
the
survey,
we
certainly
know
that
developers
feel
faster
as
they're
using
copilot
and
they're
coding
with
it,
and
so
for
our
next
step.
We
wanted
to
measure
to
find
out
how
much
faster
we're
talking
about
more
research,
so
we
run
a
controlled
experiment.
We
recruited
95
professional
Developers.
B
One
group
used
GitHub
co-pilot
to
complete
the
task
and
the
other
one
didn't,
and
we
would
like
to
give
you
a
little
bit
of
a
taste
of
what
that
research,
what
the
task
were
actually
like,
and
since
we
had
Ryan
here,
maybe
we
can
recruit
him
hello,
Ryan,
hey
thanks
for
having
me
back.
Would
you
like
to
be
our
96th
participant
for
this
study?
Can.
A
B
A
A
Implement
me
that
looks
like
the
right
place
to
go
so
we'll
do
just
that
all
right,
so
we're
going
to
start
by
giving
it
some
instructions
to
just
parse
the
string
response,
and
we
will
give
it
a
kind
of
request
that
we
need
to
create
so
good
to
go
there
all
right,
we're
doing
kind
of
tab,
easy
one-liners,
which
this
feels
good.
This
is
easy.
A
A
A
A
A
And
the
HTTP
one
that
have
not
passed,
though
so
I'm
going
to
do
a
little
bit
of
defensive
coding
back
here,
I'm
going
to
debug
the
401
and
I'll
add
in
a
console.log
and
with
that
I'll
have
a
little
bit
more
visibility
into.
What's
going
on
plus
I'll.
Do
some
checks
here
make
sure
that,
like
all
of
the
headers,
actually
show
up
inside
of
it
and
I'm
guessing
here
that,
if
I
also
check
to
make
sure
that
HTTP
is
in
there.
B
A
A
B
B
B
For
being
here
on
stage
is
enough:
yes,
thank
you.
Yes,
thank
you
so
much
for
for
going
along
with
us,
but
let's
take
a
look
at
what
the
actual
the
Real
Results
were
like.
So
we
we
saw
that
the
group
that
used
copilot
was
more
successful
in
completing
the
task
at
all.
So
78
out
of
that
group
finished
the
task
compared
to
70
in
the
group
that
didn't
use
Copilot,
and
that
was
that
was
already
good
news,
but
the
striking
difference
was
that
the
developers
who
used
copilot
completed
the
task
significantly
faster.
B
B
He
leads
the
machine,
learning
embedded
software
and
system
security
lab
there.
That
is
part
of
nyu's
Center
for
cyber
security
and
the
group's
goal,
among
other
things,
is
to
create
better
techniques
to
assess
security
vulnerabilities
and
to
generate
benchmarks.
So,
let's
have
Brendan
join
us
on
stage.
B
Let's
move
hello,
Brendan.
Thank
you
very
much
for
joining
us
I'm,
going
to
jump
right
in.
We
want
to
hear
you,
in
your
own
words,
describe
a
lot
of
the
research
that
that
you
have
been
doing
with
your
group
so
about
a
year
and
a
half
ago
you
did
a
an
evaluation.
You
did
a
study
to
look
into
the
security
of
copilot
suggestions,
and
this
is
a
study
that
has
been
published
and
that
your
team
got
a
distinguished
paper
award
for
which
is
amazing.
Congratulations.
C
C
A
These
are
some
kind
of
fun
results.
That's
awesome.
I
mean
like
it's
not
awesome
that
you
found
all
the
security
vulnerabilities,
but
like
awesome,
research
at
the
same
time,
like
I,
know
that
you
conducted
a
second
study
that
is
now
public,
but
not
yet
published
right.
So
why
did
you
conduct
the
second
study?
What'd
you
find
yeah.
C
So,
with
the
second
study,
we
were
trying
to
address
some
limitations
of
the
original
study.
So
in
the
original
study
we
basically
were
just
letting
copilot
do
its
thing.
There
wasn't
any
human
Loop,
so
no
one
was
sort
of
looking
at
the
code
and
saying,
okay,
that
doesn't
look
right,
I'm
going
to
fix
it.
C
The
other
thing
is
that
we
weren't
directly
comparing
against
human
users,
so
we
didn't
know
how
often
they
would
have
written
security
vulnerabilities
in
that
same
situation,
and
so
this
second
follow-up
study
was
really
designed
to
kind
of
address
those
issues
head
on.
We
actually
recruited
a
group
of
59
users
and
split
them
into
two
groups,
one
of
them
that
had
access
to
a
IDE
plugin
that
used
codex
very
similar
to
Copilot
and
the
other
group.
That
was
writing
code
unassisted.
C
The
way
we
all
used
to
do
before
a
year
ago,
and
then
we
had
them
Implement
a
kind
of
shopping,
cart,
style
API.
That
was
using
a
linked
list
in
C,
and
then
we
went
and
measured
the
two
groups
afterward
on
how
functional
the
code
they
wrote
was
based
on
a
pretty
big
set
of
unit
tests
that
we
had
written
and
then
also
manually
comb
through
all
the
submissions
to
find
all
the
possible
security
vulnerabilities
in
that
code
that
we
could
and
some
sort
of
thing.
That
was
a
little
bit
surprising.
C
Given
the
original
study
was
that
the
two
groups
didn't
seem
to
significantly
differ
in
the
rate
of
security
vulnerabilities
that
they
had,
so
they
were
pretty
close
to
another,
but
on
the
other
hand,
they
actually
were
able
to
write
much
code
that
worked
better.
So
we
had
a
higher
rate
of
people
actually
completing
the
task,
and
the
group
that
had
codex
enabled
were
actually
passing
more
of
the
unit
tests.
B
C
So
I
mean
you
know,
I'm
an
academic,
so
I'm
going
to
add
a
bunch
of
caveats
up
front.
So
you
know
we
looked
at
one
language,
C
and
one
set
of
users,
so
these
are
undergraduates
and
master
students,
so
I'm
not
going
to
try
to
make
strong
claims
about
it,
generalizing
to
all
languages-
or
you
know
all
professional
developers
and
things
like
that.
But
it
seems
like
from
this
initial
study
that
copilot
didn't
seem
to
make
the
security
of
the
code
worse.
C
It
didn't
seem
to
make
it
better,
but
can't
have
everything
so
far
and
I
guess.
The
other
thing
is
also
just
that
you
know
when
you're
working
with
code
you
sort
of
aren't
absolved
from
the
responsibility
of
running
security
tools
and
all
the
standards.
You
know
good
stuff,
like
fuzzers,
static,
analysis
tools
and
things
like
that
that
you
would,
with
any
other
code,
yeah
yeah.
B
A
Need
to
happen,
I
mean
like
even
for
me,
as
I'm
coding,
like
the
Judgment
that
I
make
with
each
suggestion
that
came
up.
I
mean
the
demo
just
there
like
I,
had
10
different
suggestions
to
choose
from
and
that
human
judgment
the
person
thinking
breathing
on
the
other
side
of
the
keyboard
makes
a
lot
of
difference.
So
awesome.
So,
like
that's
a
lot
of
amazing
research
already
and
your
team
I
know
is
doing
a
lot
more.
What
are
you
thinking
kind
of
next
like
what's
what's
next
for
AI?
What's
next
for
security
and.
C
Research
sure
here
so
I
mean
one
thing
that
I'm
really
excited
about.
Looking
at
more
in
my
own
group
and
with
my
collaborators
is
not
just
using
things
like
codex
and
copilot
to
write
code,
but
also
to
understand
code
and
to
potentially
automatically
fix
code
as
well.
So
we've
got
a
follow-up
study,
that's
coming
out
in
May.
You
can
already
read
it
where
we
actually
used
codex
to
try
to
take
security,
vulnerabilities
and
automatically
repair
them
and
then
in
another
line
of
work.
C
We're
looking
at
things
like
if
you're
doing
something
like
malware
analysis,
trying
to
reverse
engineer
some
binary
program.
Can
you
use
something
like
codex
to
explain
to
you
in
natural
language
how
that
code
Works
without
actually
having
to
have
the
original
source
code
available?
So
neither
of
those
things
are
things
I
want
to
like
rush
out
into
production
at
the
moment,
but
they're
pretty
promising
early
results.
That
is
awesome.
That
is
very
cool.
A
B
B
B
A
A
Cool
well
we're
getting
here
near
the
end,
but
I
want
to
take
a
moment
to
celebrate
a
little
bit.
Co-Pilot
has
now
been
available
either
as
a
technical
preview
or
in
GA
for
almost
a
year
and
a
half,
and
in
that
time
we've
had
over
1
million
developers.
Use
copilot,
as
you
heard
earlier
now,
about
half
of
active
developers
are
using
it
for
free
as
either
students
teachers
verified
open
source
maintainers,
but
there
is
one
group
that
we
have
left
out
in
the
cold
a
little
bit.
A
In
fact,
we've
had
some
engineering
managers
literally
see
torches
and
pitchforks
outside
kind
of
their
cubicles
and
windows,
as
developers,
who've
used
copilot
as
individuals
have
been
begging
to
use
it
in
the
workplace,
and
so
that's
why
we're
so
very
excited
to
bring
co-pilot
for
business
to
Market.
Here
this
December
I
do
hope
that
everyone
here
will
kind
of
give
co-pilot
a
try
to
use
it
in
your
work.
Use
it
in
your
non-profit,
use
it
in
your
weekend
project.
A
There
is
a
free
60-day
trial
for
all
individuals
and
co-pilot,
for
business,
of
course,
will
introduce
the
ability
to
manage
licenses.
They
are
pay
as
you
for
what
you
use,
that
is
to
say,
you
do
not
necessarily
need
to
buy
a
big
bulk
package.
There's.
Also,
of
course,
the
ability
to
manage
policy
on
behalf
of
your
users.