►
From YouTube: Building with Codex Fireside Chat - Universe 2022
Description
Join this fireside chat with Oege De Moor, VP of GitHub Next, and OpenAI, on building with Codex, OpenAI's AI system that translates natural language to code. Together they’ll discuss Codex's capabilities, exploring applications from developers across the globe. Drawing on their experience building GitHub Copilot, an AI-powered pair programmer, they'll touch on the unique challenges in working with this powerful technology. They'll also look to the future, and how Codex's new capabilities will empower everyday developers.
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
A
B
C
A
B
So
codex
is
opening
eyes,
AI
system
that
to
translate
natural
as
an
AI
sort
of
generative
model,
it
has
a
lot
of
capabilities,
but
we
find
that
it's
really
particularly
strong
at
tasks
like
program,
synthesis,
co-translation
and
code
documentation.
So
what
really
makes
codex
exciting
is
that
it
actually
expands
programming.
What
I
mean
by
that
is
that
it
sort
of
breaks
down
some
of
the
barriers
to
entry
to
bring
code
and
programming
to
new
Industries
new
problems
and
more
impactfully
new
users.
C
Yeah
and
building
off
of
that
right,
we
knew
that,
like
codex,
would
have
all
of
these
cool
capabilities
when
we
saw
them
in
the
model.
But
what
we've
been
validating
over
the
last
couple
of
years
is
really
the
impact
of
those
capabilities
and
where
they
provide
value
and
yeah
I.
Think
it's
right
where
Katie
said,
which
is
lowering
that
barrier
of
entry
for
people
that
want
to
write
code
but
then
also
making
you
know
the
job
of
season
software
professionals
easier
too.
A
C
So
one
example:
there
is
this
tool
called
flutter
flow,
which
is
a
local
tool
to
build
uis
using
flutter,
which
is
Google's
open
source
software.
So
what
you
can
do
is
you
can
actually
use
like
a
drag
and
drop
UI
to
build
out.
You
know
your
interface,
but
of
course,
then
you
run
into
limitations,
because
there's
only
so
much
that
you
can
do
with
a
UI
and
so
then
what
they
do
is
they
extend.
Then
your
capabilities
with
codecs
to
allow
you
to
add
custom
functions,
so
we
actually
prepped
a
quick
demo.
C
I
can
get
this
to
work
of
it
working
yeah,
so
you
can
see
here.
Actually
you
know
this
is
the
UI
that
you're
using
to
drag
and
drop
typically
to
make
code.
But
then,
when
you
run
into
a
blocker
you
can
come
and
put
in
you
know
what
your
custom
function
should
be.
Maybe
a
couple
of
lines
of
what
you
wanted
to
do
and
then
it
generates
actually
the
code
that
you
can
use
directly
in
your
application.
C
So
this
is
pretty
cool
because
the
makers
of
flutterflow
are
talking
about
when
people
needed
to
use
this
or
needed
a
custom
function,
usually
they
would
have
to
go
in
to
like
communities
to
see.
You
know
what
they
have
built
start
learning
the
language
more
themselves,
and
it
was
just
a
really
slow
process,
whereas
with
codex.
Now
it's
one
of
these
examples
where
they
can
actually
start
to
do
it
with
what's
less
grasp
of
the
language
and
yeah.
So
that's
just
one
example,
but
we
we
see
this
with
like
a
lot
of
Industries
I.
C
C
You
can
give
it,
you
know
your
your
schema,
you
can
give
it.
You
know
the
the
query
you
want.
Like.
Oh
give
me
the
last
10
users
that
signed
up
for
my
application
and
whatnot
it'll
generate
the
SQL.
You
can
run
it
and
then
all
of
a
sudden,
you
have
this
nice,
like
natural
language
interface,
for
us,
for
business
professionals.
To
then
like
query
your
database
like
really
easily.
C
Definitely
like
I
I
think
in
education,
there's
maybe
like
two
ways
that
codex
has
been
particularly
impactful.
One
like
I
mean
you
can
imagine
if
you
have
now
this,
like
interface
from
language
to
code
now,
students
can
kind
of
see,
like
example,
after
example,
of
like
what
their
code
might
look
like
without
having
to
again
like
deal
with.
You
know
small
syntactical
details
or
like
specifics
of
a
new
language
they've
never
seen
before
they
can
kind
of
see.
Okay.
A
So
those
are
all
really
cool
applications
and,
of
course
this
is
not
golden
noticed.
The
open
source
Community
has
started
several
efforts
to
create
open
source
models
very
similar
to
codex,
also
aimed
at
code
iteration,
so
there's
The
Originators
of
this
movements.
How
do
you
think
about
those
efforts.
B
C
Yeah
and
one
of
the
goals
of
our
org
in
terms
of
deploying
these
models
is,
you
know,
first
and
foremost,
to
share
those
benefits,
but
also
then
to
get
feedback
to
learn.
You
know
how
are
these
models
really
useful,
like
we
can
kind
of
see
that
the
model
does
well
on
evals?
You
know
it
does
well
on
demos,
maybe
when
we
try
it
just
things
that
impress
us,
but
really,
where
is
the
value
that's
going
to
last
to
end?
Users
is
something
that
we're
always
trying
to
validate,
and
where
does
it
come
short
like?
C
Where
is
it?
Where
is
it
failing
in
ways
that
are
preventing
it
to
reach
those
end
users,
and
that
kind
of
feedback
is
extremely
valuable?
So
when
we
see
these
open
source
models
and
people
building
on
top
of
them
and
doing
cool
things
with
them,
we
really
love
to
get
that
feedback
and
validation
too.
In
terms
of
what's
working
and
what's
not.
A
So
we
at
GitHub,
we
feel
exactly
the
same.
It's
only
right
a
lot
of
the
open
source
Community,
the
creators
of
the
public
data
sets
Lots
go
into
these
models,
also
build
new
ones
and
give
feedback,
give
feedback
validating
the
basic
assumptions
that
went
into
the
original
codex
and
see
where
all
of
us
can
do
better
and
Advance
the
of
the
science
one
of
these
people
who
leading
this
this
movement.
A
This
setting
right
here
in
the
audience
like
we're
very
keen
on
working
with
you
all
so
we've
talked
about
some
of
the
stuff,
that's
exciting
and
good,
but
it's
never
inevitable
that
with
a
new
technology
like
this
totally
new
applications,
there's
also
are
some
challenges
and
one
of
the
challenges
are
to
being
very
actively
discussed
in
the
community.
Is
this
a
phenomenon
that
sometimes
you
prompt
codecs
and
then
it
will
give
you
back
a
suggestion
of
identical
or
very
similar
to
an
existing
piece
of
our
public
source
code?
A
So
because
of
that
to
someone
to
some
people,
it
gets
the
impression
that
codex
is
just
this
huge
database
or
Snippets
of
code,
and
you
give
it
a
prompt
and
it
will
retrieve
a
snippet
that
may
be
close
to
you
or
what
you
want
to
do.
So,
how
should
we
think
about
this?
Is
it
true?
Is
it
the
database.
B
So
that's
that's
a
great
question.
I
know
that
sometimes
folks
do
think
that
these
models
are
essentially
copying
pasting
from
a
training
data
set,
and
that's
just
not
actually
how
these
models
work.
So
thinking
about
what's
actually
kind
of
going
on
is
that
AI
models
both
for
language
and
code
are
simply
trying
to
predict
the
token
or
character
that
comes
next
so
for
code
in
particular,
these
AI
models
are
actually
generating
entirely
new
code,
using
kind
of
in
a
probabilistic
way.
A
So
from
the
very
beginning,
we
were
aware,
of
course,
that
there
might
be
challenges
of
this
kind.
It's
because
of
that,
when
we
started
building
co-pilot,
we
wrote
that
in
a
in
the
stage
watch
a
stagewise,
it
was
fashion.
We
first
did
the
preview,
specifically
with
the
aim
of
engaging
with
the
community
and
or
talking
about
these
problems
and
the
best
way
to,
on
the
one
hand,
get
the
benefits
of
autotula.
Co-Pilot
brings
you,
but
also
making
sure
that
any
of
the
disadvantages
are
mitigated
so
for
this
particular
problem.
A
One
feature
that
we
built
in
that
you
can
switch
on
optionally
is
to
suppress
any
suggestion
that
matches
a
piece
of
public
source
code.
So
before
the
suggestion
is
shown
to
you,
I
will
actually
look
it
up
in
an
index
of
all
the
current
public
source
code
on
GitHub
and
if
there
is
a
match
it
gets
suppressed.
So
that
has
the
advantage
that
of
you're
totally
sure
you
will
not
inadvertently
be
I'll,
be
quoting
a
piece
of
code
from
some
repository
that
you
don't
know
about.
A
Or
perhaps
it's
actually
an
indication
that
you
shouldn't
be
writing
that
code
at
all.
You
want
to
build
on
what
someone
else
has
done
and
or
simply
take
that
take
that
other
hotels
they
dependency
of
your
own.
This
is
currently
it's
a
commitment
we
are.
Are
we
building
this,
but
it
will
be
a
while
yet
before
we
can
release
it
for
everyone
to,
for
everyone
to
use.
B
A
A
So
the
first
answer
is
that
you
should
always
carefully
that
all
the
suggestions
that
you
get
from
an
AI
like
like
codex,
a
treated
like
a
contribution
from
a
from
Junior
colleague.
It
needs
to
be
tested.
That
needs
to
be
reviewed
and
you
need
to
run
all
the
available
tools.
Do
you
weed
out
potential
law
or
potential
vulnerabilities?
B
A
Models
from
open
AI,
but
it's
also
a
matter
of
giving
the
giving
the
right
problems
if,
in
the
problems,
you're
already
given
instructions
to
write
more
secure
code
chances.
Are
that
that's
what
your,
what
you're
going
to
gather,
what
we're
working
on
on
that
as
well
as
running
very
quick
checks
similar
to
what
they
said
about
the
reference
tool
R2
or
to
just
surprise,
suggestions
about
our
vulnerabilities
in
them
all
right,
so
we've
talked
about
overall
how
code
access
is
is
exciting
and
and
useful.
A
C
Yeah
yeah,
so
I
mean
I,
think
we're
always
working
on
making
the
capabilities
better
and,
like
you
mentioned,
being
able
to
handle
you
know
more
and
more
complex,
prompts
more
and
more
reliability
out
of
the
models.
You
know
we're
also
making
it
a
better
product.
You
know,
making
it
faster,
more
reliable
to
serve,
and
all
those
things
I
think
is
with
the
goal
of
just
again,
unlocking
more
of
that
developer,
workflow
so
being
able
to
have
codex,
be
a
part
of
your
workflow
end
to
end
so
from
starting
with
brainstorming
of
of
design.
C
All
the
way
to
you
know
debugging
or
performance
testing
or
documentation
like
we
want
to
be
there
kind
of
every
step
of
the
way
doing
what
it
does
best,
which
is
again
offloading
those
parts
that
you
don't
want
to
do
so
that
you
can
focus
on
on
the
creative
parts.
So
integrating
more
into
that
workflow
with
you
know,
better
capabilities
and
a
better
way
of
serving
the
models
is
a
big
goal
for
us
and
I
think.
C
Another
goal
for
us
is
again
just
making
sure
that
we
unlock
more
and
more
Industries
for
more
potential
developers
here.
So
I
think
it
would
be
super
cool
if
we
had.
You
know
developers
like
designers
using
these
AI
assisted
tools
to
be
building
mock-ups
musicians.
You
know
like
producing
music
with
code,
because
it's
like
easy
enough
to
do
so.
I
just
think
unlocking.
You
know
that
next
you
know
million
100
million
developers
out
there
to
see
what
they
can
create
is
is
really
like
kind
of
the
end
goal
here.
B
But
these
Technologies
really
help
to
sort
of
alleviate
some
of
the
boring
elements
of
work,
and
instead
programmers
get
to
focus
on
the
creative
aspects
like
deciding
what
a
piece
of
software
should
do
and
so
programming
then
kind
of
shifts
right.
It
becomes
this
design
opportunity
you
get
to
decompose
very
large
problems
into
smaller
ones
and
then
even
into
smaller
blocks
from
there
and
sort
of
when
you
look
at
programming
in
in
that
model,
these
AI
pair
programming
Technologies
really.
B
Are
this
very
beautiful
complement
to
to
kind
of
do
those
little
details
for
for
developers.
You
know,
as
I
kind
of
think
about
where
this
technology
is
going
and
where
really
how
quickly
AI
powered
tools
are
improving.
You
can
pretty
soon
imagine
a
world
where
one
day
anyone
any
human
really
truly
anyone
could
really
just
design
a
kind
of
in
describing
in
either
a
natural
language
or
even
diagrams
what
they
want
to
build
and
see
it
come
to
life,
and
so
that's
kind
of
where
I
see
the
real
sort
of
power
of
this
technology
right.
A
So
you're
actually
not
I'm
glad
that
you
mentioned
Focus.
We.
Actually
we
already
see
that
in
user
studies
of
copilot
today
my
colleague
Irene
did
the
study
of
2000
users
of
co-pilot
and
what
she,
what
he
found
is
that
74
of
people
say
that
it
allows
them
to
focus
on
the
more
interesting
satisfying
part
of
their
work
and
a
whopping
98
said
that
allows
them
to
accomplish
boring
tasks
faster,
and
so
this
is
just
the
beginning.
It's
the
first
attempt
and
we'll
we'll
get
more
much
more
of
that.
C
Yeah,
so
I
think
one
part
is
always
going
to
be
making
the
capabilities
better,
but
we
obviously
serve
these
models
via
an
API,
and
so
one
thing
that
we
want
to
do
is
Empower
developers
to
create
really
cool
applications
on
top
of
the
models,
because
developers
are
really
the
ones
that
are
going
to
understand.
You
know
how
to
integrate
them
in
these
models
into
the
workflows
of
developers.
I
mean
GitHub.
Copa
is
like
the
perfect
example
of
this,
of
taking
the
power
of
codex
and
and
really
able
being
able
to
distribute
it
to
end
users.
C
A
A
A
You
might
ask
for
another
one,
but,
as
you
already
saw
in
a
very
small
way,
with
the
demo
of
hey
get
up
today,
you
can
actually
it
shows
you
a
suggestion,
and
then
you
say
it
change,
no
I
want
you
to
change
the
result
type
to
a
Boolean,
and
it
will
make
a
quick
edit
like
that,
but
in
a
much
more
back-end
back
and
forth
fashion.
It.
C
Makes
sense,
yeah
and
and
I
really
think
that
we
can
get
there
and
to
a
point
where
you
know,
even
in
that,
like
iterative,
State,
you're,
going
back
and
forth
with
codex
and
just
having
more
integration
points
and
I'm
curious.
What
like
you
guys
think
in
terms
of
the
impact
then,
like
broadly
on,
like
the
AI
Workforce
and
the
developer
Workforce,
once
we
kind
of
get
more
and
more
towards
that
future.
Right.
A
I
emphatically
think
not,
and
the
reason
I
think
technology
is
that
if
you
look
at
the
history
of
software
development,
it
comes
in
big
ways
or
the
first
big
way.
I
thought
I
saw
during
my
own
career
is
the
is
that
people
started
programming
in
much
higher
level
programming
languages
and
that
in
itself,
I
mean
we've
forgotten
that
now,
but
that
in
itself
and
people
stopped
our
programming
in
very
low
level
languages
and
went
to
a
higher
level
programming
languages.
A
There
was
also
a
chorus
of
people
who
said:
oh,
surely
that
will
not
as
much
less
work
for
software
for
software
developers.
Well,
the
opposite
happens.
More
people
could
write
software,
and
so
more
software,
more
creative
ways
of
using
software
were
invented,
and
the
second
big
wave,
in
my
view,
was
the
introduction
of
Open
Source
and
once
again
now
you
don't
have
to
build
everything
from
scratch
yourself
anymore.
Instead,
you
can
build
on
the
work
of
others
that
reduces
the
amount
of
work
that
you
need
to
do
as
a
developer
right.
A
Well,
perhaps
both
in
fact
it
enabled
many
more
people
to
build
many
more
interesting
things,
because
the
standard
components
that
everyone
was
implementing
over
and
over
again
for
themselves
could
now
be
built
on
once
and
for
all,
and
so
the
third
wave
where
I
see
AI
are
coming
in.
It
will
do
the
same
thing.
We've
talked
about
it,
but
it
actually
lowers
the
it
lowers
the
barrier.
It
lowers
the
the
floor
for
getting
into
programming,
but
it
lowers
the
4
floor
without
lowering
the
ceiling.
A
Get
started,
but
it's
limited
what
you
can
actually
do
with
them.
You
or
you
reach
a
certain
level,
and
then
you
have
to
go
around
start
programming
again
in
a
normal
or
programming
language
of
other
experts
use
the
promise
of
AI,
and
what
we're
already
seeing
with
a
copilot
today
is
that
you
can
actually
use
precisely
the
same
tools:
the
same
languages
that
the
experts
use.
A
It's
just
easier
to
get
started,
and
so
I
think
that
that's
the
the
message
here,
but
by
making
it
easier,
there
will
just
be
more
programmers
and
the
programmers
will
be
easier.
I
will
be
able
to
focus
on
the
creative
part
of
the
job.
They
will
just
see
many
more
applications
that
we
haven't
even
dreamed
of
today.
C
You
know
the
designers
and
and
the
journalists
and
and
the
musicians
of
the
world
as
examples,
but
I
do
think
that,
like
with
this
potential
impact,
there
are
concerns
because
you
know,
obviously
it's
a
fast-moving
field
anytime.
You
deploy
something
with
this
kind
of
impact.
Then
there
are
broader
impacts
outside
of
what
you
just
deployed
so
definitely
at
open
AI.
This
is
something
that
we
think
about
a
lot,
which
is
the
impact
of
every
deployment
and
what
that
means.
A
If
you
put
this
in
the
hands
of
inexperienced
developers,
it
might
it
make
suggestions,
it's
very
easy
to
hit
that
Tab
Key
accepted
without
knowing
what
you're
actually
doing
so.
We
do
need
a
much
more
conversational
way
of
interacting
with
the
AI
so
that
you
can
ask
it
for
an
explanation.
Ask
it
for
test
cases,
making
sure
that
the
thing
not
is
being
suggested
is
actually
truly
what
you
want
and
need
yeah.
A
B
So
really
we
are
trying
to
build
an
ATI
and
make
sure
that
we
do
that
safely
and
deploy
it
or
its
benefits
in
in
a
responsible
way.
Code
generation
is
actually
a
core
part
of
our
research
roadmap
to
to
get
there
and
largely
that's,
because
code
actually
embodies
a
lot
of
the
kind
of
capabilities
that
we're
interested
in
seeing
from
our
models,
so
code
becomes.
B
Additionally,
code
can
be
verified,
that's
incredibly
helpful
and
practical.
We
can
take
a
model's
outputs
or
suggestions
and
actually
check
if
they're
right
or
we're
on
the
right
track,
simply
by
executing
the
code.
And
so,
basically,
what
that
enables
for
research
is
actually
a
really
helpful
and
effective
learning
Loop,
which
we
then
use
to
kind
of
continue
to
push
the
boundaries
of
our
model's
capabilities
today
and
then
check
it
as
well
as
sort
of
this
Waypoint
along
that
Journey.
A
B
Yeah
so
I
know
we're
almost
out
but
wanted
to
just
quickly
kind
of
cue
in
a
little
bit
of
what's
to
come
and
where
we're
making
our
kind
of
codex
models
and
capabilities
iterative.
So
now
our
latest
codex
prototype
can
actually
explain
its
code
and
debug
its
own
code
and
then
actually
learn
from
its
mistakes
and
correct
them
all
on
the
Fly.
B
B
B
B
Great,
so
we've
given
codex
the
problem-
that's
it
from
here
you'll,
see
in
a
second
that
codex
is
actually
going
to
outline
its
approach
in
natural
language.
From
there
it's
going
to
start
to
move
on
and
step
by
step
sort
of
generate
the
code.
It
needs
review
its
code
and
then,
if
it's
correct,
move
on
to
the
next
step
from
there,
it's
going
to
kind
of
integrate
each
of
these
components
into
the
actual
function
that
it
needs
to
create.
And
then
it's
going
to
try
and
evaluate
itself
again
this
time
by
writing
tests.
B
So
as
codex
goes
and
you'll
see
this
in
a
second,
but
it
will
discover
from
its
a
slight
bug,
and
so
it
tries
to
learn
from
that
mistake
and
actually
iterate
and
correct
itself
and
then
evaluate
to
confirm
that
it's
right
so
here,
basically
what
you're,
seeing
in
this
very
brief
demo
you're,
actually
seeing
all
the
sort
of
capabilities
of
codecs
come
together
right.
It
can
now
not
only
sort
of
explain
its
code
in
natural
language
generate
its
own
code,
but
now
correct
itself,
all
on
the
Fly
and
in
real
time.