SGUCcjHTmGY 1
signs of life got some signs of life
he's going back and yes there we go all
right very nice this is very very nice
indeed it is alive all right great so i
think in order to uh in order to put a
capstone on this game uh we just need to
indeed there's no game if you can't lose
sad sad to say but we do need to
implement that loss condition so first
define what happens when you lose
clear the screen and
show a message
saying you got squashed it should be an
encouraging message ideally um okay well
let's let's so i just kicked this one
off let's make it encouraging so now
modify that function now rewrite that
function
to
also include some words of encouragement
excellent i'm also curious what words of
encouragement the model will choose
so you can do it um so that that's
that's pretty good
and uh so so the only thing is that the
way that this actually was implemented
is implemented as a key down listener
and what we really want is we want a
function that gets called when you lose
so we can get rid of both of these and
we could try this one more time
and so how would you make it different
so you can say
define a function
so make a function
that gets called
okay let's see if it works oh it is a
function it is called you lose
and now rewrite that function to include
words of encouragement all right let's
see what happens here
and sure enough it makes a new
eye try again all right let's let's see
what happens now we actually have to
wire this function up so when the person
and the boulder
overlap at all so constantly check
if the person to boulder up over up at
all and if so you lose
so i'm not even going to say explicitly
call that function
just it's got to figure out that that's
what we want so we'll see if that
happened um yeah do you want to do the
honors definitely oh man all right
moment of truth here moana truth success
you got squashed and a very encouraging
message to try again i think that's very
good life advice from codex right there
okay
i feel like it was a nice game that we
built in a small number of minutes i
think so so we have one more thing to
show you
and that uh with this demo we want to
help expand your mind to the you know to
the possibilities the codex can really
offer and indeed
one of the things that we showed you in
the hello world demo
is that it's very easy
to teach the codex model to use whatever
api you want api doesn't know
and conveniently
all your favorite software comes with an
api in fact i used to work at a company
whose entire
job is to build an api apis are out
there that these days the world is
really programmable and codex is able to
hook into those apis on your behalf and
so that the kind of end-to-end
functionality that i think starts to be
unlocked is that you talk to your
computer and it actually does what you
ask
all right let's let's see how it works
all right so here we have my ipad with
just vanilla microsoft word installed on
it um there's one little one little
secret within it that we'll get to in a
moment um but it turns out that
microsoft word like many pieces of
software has an api in fact it has a
javascript api and hey we built a model
that is pretty good at javascript quite
convenient very convenient so all we did
is that we took this api reference and
we formatted it for codecs and so you
know we kind of trimmed it down it's not
the whole whole implementation of the
whole api um but it's enough to make a
very interesting proof of concept and so
let me show you
the kinds of things you can do so here
is a poem that was actually one of my
favorite poems as a child really oh yeah
yeah it's called the jabberwocky uh it's
very fun um
so i'm gonna paste it into microsoft
word and uh oh shoot let me get rid of
these leading spaces before we start
sorry on this greg this will take
forever hold on hold on
you know what fortunately with the codex
add-in
i don't have to delete them
delete all initial spaces
and it worked it did work the initial
spaces are gone
but all the other spaces are still there
still there and just like before the
instruction at the top was turned into
code which was then run by microsoft
word exactly and so we're just using the
standard microsoft word api here so they
provide a function functionality for you
to get your little sidebar that we show
here and we just basically reuse the
exact same code that we've written for
those other demos and so all that's
going on here is that we use the
built-in speech recognizer so we didn't
write that so if it has transcription
errors
we take no responsibility for it um
but then
we send whatever request is put here to
the api and it generates
actual code in the microsoft word api
and what you see here is a taste of the
future
as the model gets really good as the
neural network gets really good at
turning instructions
to correct api calls it will become
possible to do more and more
sophisticated things with your software
just by telling it what to do and i
think this is the biggest contrast with
gpt-3 like the biggest step on top of
gpd3 in my mind and this wasn't obvious
to us going in but i think it's kind of
emerged from what we've built
gpd3 is a system that you talk to and it
talks back to you so the only impact it
has is in your mind
with codex you talk to it it generates
code which means it can actually
manipulate or you know it can actually
act in the computer world on your behalf
and i think that that's a really
powerful thing that you actually have a
system that can can carry out commands
on your behalf
for example let's do something a little
bit more complicated yep um so uh do you
want to give it a try yes
now make every fifth line bold
okay few i was really worried about the
speech recognition part yes well there
we go oh a success a success indeed so i
think that's pretty good and you know i
think that that
what this kind of demo shows you is what
today's voice assistants have really
been lacking
that i think that what you really need
is you need a system that has the kind
of gpt world understanding so it can
flexibly sort of interpret between
different languages and can really
understand the intent that you're you're
putting forth and while we are very
happy with the neural network that we're
showing you today which is a better code
model than the one we had previously it
is still only just a step
the neural networks the code neural
networks you'll have in the future will
be far better than this so this is only
the beginning of
an exciting future
and so that's the end of our demos uh
we're really excited that you were able
to join us and so just to review uh
today we showed you the latest
generation of the codex model it's
available in open eyes api starting
today so please sign up on on the beta
list
if you want to be able to play with
codex in the context of a pretty awesome
new kind of programming competition that
will be thursday 10 a.m uh we're really
excited for you to get a chance to play
with it so thank you very much for for
tuning in we're excited to see what
you're going to build and thank you for
joining us to experience the magic of
neural networks