November 20, 2025 • 41 mins
Have the basic requirements for building a good user experience actually changed?
In this episode of Futureproof, Xano CEO Prakash Chandran sits down with Tony Casparro, senior staff software engineer at OpenAI, to talk about his current work on ChatGPT. Tony shares some of the past user experience lessons he learned at Netflix, and how they can be translated into an AI world. Together, they cover how to get the most out of AI (and what not to do with it), some of ChatGPT’s critical UX features (deep thinking, pivot points, third-party app support), and the way in which AI is changing how we interact with software.
Topics covered include:
- From streaming to chat: What Netflix taught Tony about usability, accessibility, and anticipating human behavior.
- AI as a design partner, not a design maker: The best results come from working with AI, not outsourcing thought to it.
- The rise of contextual interfaces: ChatGPT’s new third-party apps are changing how users interact with brands and information.
- Cognitive offloading with care: How to delegate tasks to AI while keeping your creative and critical skills sharp.
Subscribe to Futureproof wherever you get your podcasts.
From Xano - The fastest way to create a production-ready backend for any app or agent. Xano unifies AI speed, code control, and visual clarity, so you never trade reliability for velocity. Sign up for free today.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Tony Casparro (00:00):
You know, thinking of AI not as a
replacement for yourself or yourown thoughts, but as a partner
to work with you is a really,really important distinction.
Um, we are working together onthis document.
Uh, it is not just gonna haveit write it and then go off uh
like uh uh some people do.
Um we're gonna work together onthis, we're gonna solve this
problem together.
Uh it's a much better way ofthinking through that.
Prakash Chandran (00:35):
Hi there, I'm Prakash, CEO of Xano.com, and
this is Future Proof.
In this episode, I'm joined byTony Casparo, Senior Staff
Software Engineer at OpenAI,where he's helping shape the
future of Chat GPT.
Tony has spent his career atthe intersection of scale and
design, leading UI and platformengineering at Netflix, founding
multiple startups, and nowbuilding on the front lines of
(00:57):
AI.
Um, Tony, really great to haveyou here.
Thank you so much for takingthe time.
Also a fellow uh slow neighbor.
Uh, it's great to have you alsoin town.
Thanks for so much for agreeingto do this conversation.
I'm happy to be here.
Let's talk.
Awesome.
I love it.
Well, why don't we start withsome background?
Maybe share with us a littlebit uh about uh, you know, what
(01:17):
you've done in the past and whatyou're doing today.
Tony Casparro (01:20):
Yeah, I've kind of had an interesting history as
far as working on the back endfirst.
So doing a lot of databasemanagement, starting at Cisco
Systems, uh, moving over tobiotech at Amgen.
Oh, excuse me, but uh reallykicking off my career when I
moved to Box to do um enterprisesoftware, right?
Really thinking through whatare the needs of you know, B2B
customers, how do we make youknow great software that's easy
(01:41):
to use, good for managementobservability.
Um, and then doing a wild swingto Netflix.
So after enterprise for a fewyears, I really wanted to get
kind of get back in the consumerspace.
And that journey was reallyabout um how do we teach the
world what streaming is?
This is, of course, you know,12 years ago now.
So, what is streaming?
How does it work?
How do you even think abouthaving an entire library at your
(02:03):
disposal and figuring out whatto watch next?
So after 10 years at Netflix,just building Netflix.com, I'm
doing infrastructure work thereas well.
Uh, I am now here at uh OpenAI.
I work on Chat GPT.
Um right now we're building outuh commerce and we're also
doing uh new work with, excuseme, bringing like real to like
um rich responses so that theinformation is not just a wall
(02:26):
of text, but actually hasmeaningful layout, uh
responsiveness to what you'reactually looking for.
And so that's what I work onevery day.
Prakash Chandran (02:34):
Awesome.
So exciting.
And I definitely want to diveinto all that.
But I first maybe wanted tostart with your time at Netflix.
Um, kind of what a uh pivotaland interesting time uh when you
joined Netflix, kind of in thisworld where people didn't
realize what streaming was.
Talk to us about some of likethe um kind of what what the
(02:54):
paradigm was like at the time interms of like the concept of
what it meant to streamsomething and some of the
affordances and things uh thatyou had to think through at
Netflix in order to like usherin this new uh era of streaming
on the internet.
Tony Casparro (03:09):
Yeah, you know, um I built my own servers before
this.
And so it was interesting, youknow, that I had been used to
this kind of new world for awhile, but like showing other
people who were, you know, stillpopping DVDs into the mailbox
and then into their players, um,or a little bit using, you
know, Apple iTunes at the timeto purchase things one-off.
Um, this was a new world, youknow, the same with Spotify when
(03:29):
they kind of introduced peopleto the idea that you have access
to everything.
And you're like, okay, where'smy music?
Well, your music is all themusic.
And there's this concept oflike, okay, but I how do I
actually use this?
Well, how do I actually youknow enable this to work for me?
Um, so same thing with with uhstreaming, right?
Um, we first of all had to showpeople that like um here is
what this movie is about, right?
(03:50):
And doing a lot ofexperimentation on like showing
which kind of images for theparticular users and custom
tailoring it to their needs.
Um, do we show them a wall oftext about what the movie is
about, or do we show them just afew quick words as far as this
is exciting and thrilling andwhatnot?
Um, using um instead of staticassets, like actually moving
assets, showing trailers in lineand things like that, um, and
(04:11):
overcoming a lot of umassumptions people had about how
this was going to work.
I remember going to a lot of uhtests where we would put the
product in front of people andwe'd be watching it behind a
two-way mirror, and they'd beafraid to click on something.
Why are you afraid?
What's going on?
Why are you afraid to click onthis?
Because it's going to chargeme.
And there was this idea that,you know, the world was
piecemeal.
You had to pay per view oneverything you did.
(04:32):
And the idea that it was allavailable.
You could just click on it andit would start playing with one
single cost per month.
We had to educate and overcomea lot of barriers that we take
for granted.
So it was a lot of interestinglearnings.
We did a lot of experimentsthat worked and a lot that did
not work.
Some things where we assumedtoo much about the user's
journey as to why were they atNetflix.com?
Was it to watch?
Was it to actually learn abouta movie?
(04:54):
Was it to figure out somethingto watch later and really
understand uh and make a greatjourney from no matter what
people were doing on the site?
Prakash Chandran (05:01):
So, you know, I think there's so many kind of
parallels in in terms of likewhat you're doing uh at OpenAI
with Chat GPT.
Uh before we touch on that, Iwould love to understand if
there's one key learning thatyou still take with you today
from your days at Netflix, whatwould that be?
Tony Casparro (05:20):
It's too, uh I think it's very naive, but we we
often think like, well,everyone does this the same way
that I do.
Uh and it's simply not true.
Uh you think of something assimple as um scrolling a page.
Uh it turns out there's manydifferent ways to scroll a page.
Some people click go to theright and they click on the
scroll bar and they drag itdown.
Some people use the arrow keys,some people are using voice
commands.
There's millions of different,we're not millions, there's a
(05:41):
lot of different permutationsfor how things people do things.
And when you build an interfaceand you're assuming everyone
thinks the way I do about how tonavigate through this or what
my intentions are, you're goingto miss a lot of people.
And so this is something thatcompanies that you know have a
lot of exposure do really wellis they know, here's how to make
it super accessible.
A great example of this, uh,somebody else told me, is a
(06:02):
steering wheel.
Imagine if a steering wheel wasalways set and you had to grip
it a certain way.
It wasn't a wheel, it was liketwo handles.
It only works for the way thatone person thinks.
But steering wheels, in bytheir design, allow for people
that drive with it on the top ofit, on the bottom of it, on the
side.
It works for all the differentuse cases.
And something as simple as awheel, it can be a beautiful
illustration of how it works foreveryone.
Prakash Chandran (06:22):
That's really a fascinating correlation.
And um, as I kind of alludedto, like, really kind of
poignant to how or the work thatyou are doing at OpenAI,
especially with ChatGPT, becauseit's a similar parallel because
you're also designing in thisnew paradigm shift where people
have access to everything,right?
Maybe not a media library, butnow the world's kind of
(06:45):
information at their fingertipsin a way that had never been
accessible before, like gettingto the answer versus searching
through the sea of blue links.
And there's an input box nowthat serves many different types
of people.
Like I use it, my uh 70 uh orsorry, 82-year-old dad uses it,
uh, my sister uses it, um, mydaughter uses it.
(07:08):
It's crazy that you have somany different people
interrogating and asking uh, youknow, ChatGPT to do things for
them.
So when you think about the thelearnings that you've had at
Netflix and that same paradigm,how have you guys tried to
approach that not problem, but Iguess opportunity of serving so
(07:28):
much of the world through thisinterface uh called Chat GPT?
Tony Casparro (07:33):
It's a really hard problem, right?
As you said, it's a supergeneral processor.
It can do all kinds ofdifferent tasks and you really
have to kind of figure out howto meet people where they're at.
What do they need help with?
Is it relationship advice?
Is it building an essay forsomething?
Is it researching something onthe web and trying to make it
meaningful and accessible tothem?
Um, we've made great strides.
(07:54):
Um, as were before, there was alot of buttons on the bottom of
the toolbar.
We've made great stridestowards you just type whatever
you want, and the model is goingto figure out which engine do I
need to use for this.
Do I need to use a thinkingmodel?
Do I need to use a websearching real-time model,
whatnot, and allowing the uh thebrain of the model to choose
which direction will take you.
There are, of course, you know,hints that you can give it if
you need to say, like, this is adeep learning task I'd like you
(08:16):
to do.
Um, but overall, we wererelying on that model for that
contextualization.
Um, and then in addition withthe response that actually comes
back, giving you meaningfulpivot points, right?
So that if you are looking forreal-time information, you can
pivot to like, okay, great, nowshow me the stock ticker.
Um, or if you're looking for uhwriting a document, okay,
great, open this in a canvas,you know, help let's figure this
out together and work on this.
(08:37):
Um, and so it really is tryingto figure out as soon as
possible what is the intent ofthe user?
What are they trying to do?
And then with the response, bea really meaningful.
Don't just show them a wall oftext, showing them really
contextual information about howthey want to perceive this and
give them those pivot points tothen take the conversation to
the next level.
Um, we've got some great stuffcoming up for this that really
(08:58):
show off these pivot points.
Uh, but this is key to saying,uh, to showing or with hints,
showing the user the capabilityof the product without being
overbearing and showing them awall of text or buttons.
Prakash Chandran (09:10):
Yeah, you know, I think um even in
coordinating this, um, we had tomake sure that you got through
uh a crazy amount of work, uh,obviously culminating in just
the recent developer day thathappened.
It was kind of an amazingmoment, actually, for me to
watch because I really feel likeChatGPT has evolved in uh in
something that was maybeconsidered like a chat interface
(09:33):
and tool to more of a platformand to what you're speaking
about, giving people moretailored contextual experiences
rather than just a wall of textis something that we can just
see the world uh going to.
Tell me, like, as someone thatworks on the interface, how what
are some of the key challengesthat people not may not
necessarily understand or thinkthrough that you have to think
(09:55):
about in terms of presentingthat type of information with
context and tailored andpersonalized to the person
that's searching?
Tony Casparro (10:02):
What's interesting is especially in
this world where the model istrained on a lot of information
and yet there's a lot ofreal-time information that it
can have access to, real-timeinformation in terms of what's
available on the web or toolsout there, all these new um uh
MCP servers and whatnot.
Uh, and so one of thecomplexities is how do you
orchestrate giving the user afast response and at the same
(10:23):
time, basically in parallel, letme go talk to the stock widget,
let me go talk to the weatherwidget, let me go talk to this
MCP server and bring this back.
And so some of the complexitiesof the UI are UI are you know,
how do you show the informationthat you have so far?
How do you show the user thatyou're still bringing in extra
information or that you'reprocessing something really
important and make it not feelslow and janky, right?
Because the power is that itcan build you a great response.
(10:46):
But if it feels like, man,nothing's happening, I don't
even know why I'm waitinganymore, that feels bad.
And so what we do is somethingcalled uh the chain of thought.
So the chain of thought is asit's thinking through whether
that is real-time information ortalking to other servers, we're
showing you little snippets.
Hey, I'm talking to that serverover there right now.
I'm gathering the real-timeweather information.
And that's giving the userinsight into what's happening
(11:07):
behind the scenes in terms of athinking model that's
interacting with the world, um,and then building that useful uh
response in turn.
Prakash Chandran (11:15):
Yeah, that orchestration and kind of giving
a sense to the user around whatuh ChatGPT is doing is just it
when it works really well, itjust feels like really seamless
and elegant, but I can imaginehow much thought goes behind it.
Um, you know, one of the thingsthat we've talked about before
uh is kind of the most uhunderutilized uh tool within
(11:37):
ChatGPT, and we've talked aboutit being deep researcher.
That that was the answer at thetime when we were chatting.
Uh and since you told me that,I actually have been using it a
lot more.
Even though, like, you know, Ithink I'll I'll be in a
conversation and uh it willobviously try to optimize for
quick answers, but I'llexplicitly say, look, for this
one, I want you to think deeplyabout it.
(11:58):
I'll make sure to tell it to gointo deep research, and it
gives me like something really,really thorough.
I'm almost giving it thefreedom to take a little bit
more time to give me like areally thorough answer.
Do you see more people startingto use ChatGPT this way?
Because I think this is whatI'm imagining.
You tell me if this is true.
People, a lot of people, areshifting their behavior from
(12:20):
traditional search to Chat GPT,but they're still using ChatGPT
like search, right?
They're still typing in likemaybe piecemeal words and not
necessarily using it how it ismeant to be and how it can best
be utilized.
Tell us what you're seeing anduh kind of the dynamics around
like deep research as a featureand the best way to utilize
(12:40):
ChatGPT.
Tony Casparro (12:42):
Yeah, you know, um, so the deep research is one
of those hints that I wastalking about.
It's in the toolbar where youcan say, This is how I want you
to answer this question.
And deep research specificallymeans take as long as you want.
I don't want a quick answer.
Um, I want you to be verythorough, go through lots of
sources, take your time, and getback to me with a really deep,
full response.
And so it is one of my favoritefeatures, especially for
(13:03):
shopping for something andcomparing, you know, a bunch of
different review sites or modelsuh for travel advice.
Um, it can be really, reallythoughtful instead of giving you
an answer very quickly.
Um, when you ask about userbehavior, though, it turns out
people still want somethingvery, very quickly.
They don't want to wait, theywant the answer right away.
Uh, and so what we've kind ofcome up with in terms of the UI
is kind of a bifurcated mode.
(13:25):
So it will start to say if itis going to take a while for
something like I do need to dosome research, whatnot, there is
a skip button.
You can say, I don't care, justshow me what you got right now.
And that allows people theflexibility to be like, uh, this
is how I uh want my answer.
Um and so I can't say for surelike what the adoption is like
on some of our uh other featureslike deep deep research, uh,
(13:47):
but people are still very muchin the mindset that they want it
quick, they want it fast, andthey want it correct.
And we are working as hard aswe can to make sure that that is
the path forward.
Prakash Chandran (13:55):
Yeah, I love that bifurcated mode where it's
like, hey, look, and I think itsays something like thinking
deeper for a better answer orsomething.
And then you have the abilityto be like, nope, nope, it's all
good, just get give it to menow.
Um, and I think that's the hardpiece, right?
Kind of balancing that speed,that AI speed that people have
come to expect with thethoroughness that something like
a deep research um can offer.
(14:16):
Um, I want to shift gears alittle bit.
You know, you are a staffengineer, you have been
developing in some capacity foryour entire career.
Uh I think the audience mightbe curious to know how you are
leveraging AI in your day-to-dayto develop software at OpenAI
and uh so you know some of thecore benefits that you've seen
(14:38):
and things that you might haveto like uh supervise or kind of
still be aware of that are stillkind of um evolving in terms of
like the landscape.
Tony Casparro (14:46):
Yeah, absolutely.
Uh, you know, the best way tothink of AI is as an assistant,
right?
It is there to help you withyour tasks.
And so when I get into theoffice in the morning, what do I
do is I look at my plate and Igo, what can I offload to an AI
so that I can work in parallel?
So some requests might come infor some code changes or
connector changes or whatnot.
Um, I can muddle through thoseand go do everything in a serial
(15:08):
manner, but I have the abilitynow to say, let me offload some
of these tasks.
So we have Codex as one of ourum AI coding agents.
It's available to everyone aswell.
Um and it's pretty phenomenalwhen it has access to your code
base to say, you know, this isthe change that it can even
read, of course, now like lineartasks and Slack messages.
And you can say, here's thecontext of the issue, come up
with a solution for it.
Now, uh, like all AA things, itmay not be perfect at start,
(15:31):
but it can actually give youquite a bit of leverage and
thinking through opportunitiesso that when you come back to
the task, you can be like, well,that's pretty clever, but let's
change this, rewire this,perfect.
Now we're done.
And so as opposed to meworking, like I said, serially
in the morning, I'm thinking uhacross many different verticals
at the same time.
I've also got AI, of course,built into my IDE, which is my
uh coding interface.
(15:51):
And now it's like, you know,like, you know, applying some
formatting or markdown or justkind of threading through some
variables and whatnot.
And I can say, just finish thisoff for me.
And so that allows me to have amuch more productive day in
terms of focusing on the biggerpicture tasks uh and then uh
sharing the work with the AI totake care of the more kind of
menial stuff to like just kindof thread through.
(16:11):
Um, and so that's for work andthen also for you know doing
research and things like thatfor other things that we have
going on at work.
Uh it used to be a world likewhere you said, like, okay, I'm
looking for this new um statemanagement.
Let me go to Google.
Okay, I see a bunch of links.
Let me go read through thesearticles and spend half my day
doing that.
Or I can say, hey, here'sanother task for you, not Codex,
(16:33):
but ChatGPT.
Go research the latest statemanagement to make sure the
information is up to date.
Bring me a report, build me aspreadsheet that I can review.
And that is once again, notdoing all the work for me.
It's in a supervised way whereit's helping me and assisting
me, giving me a collection ofinformation that I can then use
to make my next decisions and godeeper as I need to.
Yeah, that's fascinating.
Prakash Chandran (16:53):
So, like there's gonna be developers and
technical builders listening tothis.
And one of the things that yousaid was kind of really um uh
really interesting, just interms of like, you know, at the
beginning of your day, you'rekind of figuring out how can you
leverage this assistant to takecare of some things, to do some
research for you while youfocus on maybe the higher order
bit.
For the people that arelistening, how do you structure
(17:15):
that, that planning processaround what you can delegate to
the assistant and uh to uhleveraging AI for versus what
you need your attention on?
Tony Casparro (17:24):
Yeah, I mean, you know, so as a UI engineer, uh I
do a lot of visual testing.
Um, and it's something where uhour like Codex is extremely
good at like backend codingright now, and it's still
developing its chops in terms ofUI coding and testing those UI
changes.
How do they work?
How do they work foraccessibility?
You know, how does it work on afast connection, solo
collection?
There's a lot of things we needto consider when we're building
user interfaces.
(17:44):
And so some of those tasks, Imight have it like um, for
example, you start building thisfor me.
For example, I was building ananimation the other day, and I
knew it was not gonna hit theanimation correctly, but I was
like, there's gonna be ananimation that occurs when
somebody clicks this button.
I need you to pop open this, dothis, use this variable.
And it did a pretty good job,but at least it was a start for
me to be like a jumping offpoint for my PR that I was gonna
eventually write.
(18:05):
Um, and so when I come in themorning and I'm thinking about
those tasks, uh, I am thinkinguh in terms of like, how would I
approach this problem?
Okay, these are the steps Iwould take.
Let me give those to the to theagent and have the agent take
care of it for me.
If that is too complex, I go, Ican't even fathom my head what
this is gonna look like.
I just have no idea what theoutcome is going to be.
I don't know uh the correct wayto do the performance of this,
(18:26):
whatnot.
I'll take those on and start onthat journey because it's it's
just an unknown.
So to clarify for short, ifit's a known end state that I
want, uh once again, comparingthings uh across different
websites, doing research andanalysis, whatnot, there's a
known outcome.
I want to find the best this.
Uh, that's a great choice forit.
But when you're reallyideating, you're not sure what
(18:46):
it's gonna look like, you'remaybe working with partners and
stakeholders, and you're like,we don't know what that's gonna
look like yet.
That's more the state where Itake ownership and control to
like do the human stuff, talk topeople, get feedback, and
figure that out too.
Prakash Chandran (18:58):
That makes sense.
Um, you know, talk to me alittle bit about where you feel
like uh AI is not quite thereyet, but you're excited to see
it evolve so you can takeadvantage uh of leveraging it
kind of in your day-to-day.
Tony Casparro (19:12):
Yeah, we've made great strides towards um
orchestrating tasks, right?
So this is the idea, like inyou know, a year ago, if you
said, you know, change thispiece of code and structure the
state in this new way, it wouldhave just started spitting out
code, right?
It wouldn't really thinkthrough its tasks.
And so something we've toldthese coding agents now is
develop a set of steps thatyou're going to take, right?
(19:34):
I'm gonna do this, thinkthrough these, think through the
uh side cases and edgeconsiderations here, and then
start coding it.
And you get much better resultswhen you have it think through
a process to follow first.
Um, and so um, in terms of notbeing quite there, the UI piece
is is tricky, right?
Because you need to test acrossdifferent browsers, you need to
test across differentconnections.
(19:54):
Um also AIs don't really knowwhat a beautiful animation looks
like.
They don't know how it feels toa user.
That's something very visceralwhere you go like it should feel
this quick and move this, andyou know, especially when you're
doing multiple uh coordinatingmultiple animations in the same
way, you need that to feelreally good together.
And so it doesn't have a goodfeel for that just yet.
Even in terms of the userinterface, I've experimented
(20:15):
with a lot of um our own toolsand competitors in terms of like
what makes a great UI.
And they guess and they lookokay, but it's kind of using
just random examples.
It's I wouldn't say it's agreat UI designer yet.
Um, there's something reallymagical about the way humans
think through how a userinterface is going to flow and
how that's going to work.
And so I think those are areasfor opportunity.
(20:35):
Backend stuff is getting very,very good at.
Um, you know, if we go to aworld in the future where um
backend code is not even a codebox anymore, it's just a black a
black box where you say, hey,uh, this is the inputs of the
API I want, and these are theoutputs I expect.
I don't care what the codelooks like inside, uh, as long
as it's quick, just go be ablack box.
(20:57):
And we may end up in a worldthere where we have black box,
you know, um inputs and outputsfor Vatican servers.
Prakash Chandran (21:03):
Um, you know, talk to me a little bit about
how uh just in a from adevelopment standpoint, you
think about like security andcontrol over what can sometimes
be the back uh the black box,uh, whether it's front-end or
back end, how you think aboutthat in your day-to-day?
Because I feel like that'sprobably the supervision part
that you're talking about,making sure that you're in the
(21:25):
loop and making sure that it'sdoing the right things.
But you know, in yourday-to-day, how are you kind of
implementing that and thinkingabout it as an engineer?
Tony Casparro (21:34):
Are you thinking, are you asking specifically
like how I use the securitycontrols or how we use the other
thing?
Prakash Chandran (21:38):
How is you you as a company think about those
controls in the develop in yourday-to-day development?
Tony Casparro (21:44):
Absolutely.
So it's person in the loop,right?
You know, uh the AI shouldnever have access to do um what
could be considered um uhdangerous tasks on its own,
right?
You know, executing you know,arbitrary code and things like
that.
It will ask you and prompt you,is it okay if I run this?
Here is the exact command I'mgoing to run, and make sure that
the human is always in controlof those tasks.
Um, as we start to get anexposure into the world of MCPs,
(22:07):
same thing.
We want to make sure that thoseuh checkpoints are there so
there's human in the loop and sothat it's not off and running
on its own.
As our developers, however,build their own products, we're
encouraging them to also putthose same safeguards in place.
However, there might be safelyguarded loops where you go,
like, hey, it's gonna generateimages.
I already have these moderationcontrols in place, I know this
is going to be safe, and you canhave it run that uh with that
(22:29):
oversight, uh, as long as uh adeveloper or somebody has
already made sure that there's asafe path for it to take.
Prakash Chandran (22:36):
So I want to go into just maybe some
practical takeaways for peoplethat are um, you know, might be
building a new project andtrying to figure out the best
way to leverage AI.
You know, you feel like thereis, you've got all these
bypoders and they're you knowbuilding applications very
quickly.
You have people that areleveraging AI alongside of
(22:57):
themselves in a code base, uh,but from a chat GPT perspective,
for a new builder that istrying to think of uh, you know,
developing something new intothe world, or even for a product
owner in a company that'strying to make something new
within their organization, justin terms of a mental model of
how they should be thinkingabout it and leveraging AI, what
(23:18):
recommendations might you havein order to get started in the
most efficient way possible?
Tony Casparro (23:25):
You know, it is interesting because we have
Codex internally, it has accessto our Intari code base.
And so we've had people, uh, Imean, that's always a great
starting point.
If you have an existingproduct, you give it access to
your GitHub repo, it knowseverything about your code.
And when you ask it a questionabout it, um, it can efficiently
use the current styling andpermission sets and uh plans
that you use for your existingcode, and then it can do the
(23:47):
augmentation or adding newfeatures, whatnot.
That turns out to be decentlywell.
Uh, we've had people that areproduct managers, they asked for
those changes, they send it tolike a developer with the change
set.
Uh, the developer can then doslight modifications or whatnot
is needed, but it does help theprocess along because um usually
that process is the developergoing to the person, the person
(24:08):
describes what they want, theygo back and forth, and it's nice
if you know the person can justdescribe everything you want up
front and get that result.
Um, in terms of like startingnew code bases, uh, that is a
tricky one.
You know, zero to one with anew product.
You know, um, I personallywanted to do like, you know,
some more app development andwhatnot.
Um zero to one is always a hardproposition, right?
Because you have to build outsome servers or some uh, you
(24:30):
know, some developer toolkits onyour local machine.
Um I can't speak to anywebsites that currently do like
a great job.
I'm sure they're out there, uh,but it's a great opportunity
because there are so many peoplethat are vibe coding small
applications.
Uh, it'd be fantastic if youcould uh use one of these to
start building up, you know,whether that's a website or an
app that you're gonna release tothe app store uh and get you up
and running quickly.
(24:51):
So uh I guess do some research.
Uh I'm not I can't recommendany particular ones.
Uh Codex is great though forexisting codes.
Prakash Chandran (24:57):
That's awesome.
Um and then, you know, just interms of like uh, you know,
ChatGPT is such a challenge forall of the reasons that we've
been uh speaking about, just interms of the U and I UI UX uh
paradigms that need to beintroduced.
I'm curious to uh if you havethoughts around any UI or UX
(25:18):
patterns that you will seeeither disappear or be
introduced into this new worldof new interfaces uh that we're
entering into.
Tony Casparro (25:28):
So that's a great question.
That's right in my wheelhouse.
So um uh as your users may not,or your listeners may or may
not know, uh we releasedsomething called uh third-party
apps inside of ChatGPT uh justthis Monday.
Um so this is the world nowwhere you can talk to ChatGPT,
and instead of ChatGPT saying,here is a link to Zillow, go
(25:49):
like click out to them and leavethe site.
Uh you can now talk to Zillowor these other uh apps that we
partnered with to start insideof Chat GPT.
So um that means they'recreating custom UI elements.
Uh ChatGPT and these uh thirdparties are now working together
to build visualizations.
Um and that turns out to beextremely powerful.
Another example is Spotify,right?
(26:10):
Um hey, I'm coming up uh for aplaylist, we're having an 80s
party, make me a great playlist.
It could list a bunch of songs,but it'd be great if it could
interface with your Spotifyaccount, build you the playlist,
and then you're done, uh, asopposed to you having to go add
those songs manually or whatnot.
And so bringing these appstogether to be controlled inside
of ChatGPT is gonna be a reallyinteresting new world.
(26:30):
Excuse me.
And then, of course, on theother side of it, we have these
agents that you can build intoyour current sites that go
backwards, right?
So now you are the main app.
Chat GPT is now embedded insideof your application.
Now these two can interface aswell.
Um, it's gonna be a reallyinteresting world in terms of
how people think about you know,branding and um app stores.
(26:51):
I mean, you can see a world nowwhere if these UIs are so like
chat agents are so powerfulenough to build, like integrate
with third parties and evenbuild interfaces on the fly, you
could start to see app storesgo away.
You know, why would I have togo like search my phone and find
this one app if this other appcan just do it all for me,
right?
It can generate the UI,generate my feed, whatever I
(27:13):
need, and build it in a customway.
And so you can see a long tailroad down here where, you know,
uh there is no more searchingfor apps, there's no more
looking for context.
You can just interface withyour device in whatever way you
want, probably through umspeaking and or text, and just
say, show me the weather fortoday.
It doesn't have to open an app,it doesn't have to do anything
else, it can render a rather aweather widget for you.
(27:36):
Uh, you know, tell me how myfinances are doing, you know,
like cross-reference myportfolio and see what I need to
do for my taxes.
And it's generating stuff onthe fly.
Uh, and so there's a reallypowerful new world out there for
those that um are thinkingthrough how does AI and my
application fit together tocreate a single united world as
(27:56):
opposed to I gotta get back tomy site.
Um, and it's gonna be realinteresting to see.
Prakash Chandran (28:01):
I totally agree.
I was actually talking to aprevious guest um about this,
how like the front end uh, youknow, in terms of the way we're
used to digesting, where you Idownload a singular application
or I go to a web application,uh, is changing.
Like because uh the more peopleinterface with tools like a
ChatGPT and get those answers,yeah, the nature of how
(28:21):
companies will express theirbrand through like this new
apps, SDK and otherwise, isgoing to just fundamentally
change.
So we're kind of entering thisera of like personalization.
And uh and another thing thathe he brought up that I thought
was fascinating was a potentialworld of like ephemeral uh
applications.
Because sometimes we build anapp for a purpose, but once that
purpose is done, especiallywith how fast we can create
(28:43):
applications, like it may havebe like purpose built for this
one thing and then it may goaway.
I'm curious, have you have youthought through anything like
that, just even outside of whatuh you're doing at OpenAI, but
just personally, how do youthink about the world of um
digesting information,application development, and how
we're going to utilize orinterface with businesses in the
future?
Tony Casparro (29:05):
Yeah, you know, it's funny.
Netflix did a lot of kind ofone-off applications that were
short-lived.
You know, we were doingexperimentations inside for, you
know, do, you know, does itmatter if we show a different
image to a customer for a moviethan this other image, right?
And we would do theseexperiments where we would have
an idea like that and we wouldneed a UI that was just, you
know, ephemeral.
It was just gonna be there fora while while we evaluated the
(29:26):
efficacy of this product.
Um, and we would spin those upall the time.
Um, and that process was youknow very cumbersome at the
time, and now we have AI to helpgenerate some of those for us,
and it's very, very nice.
Um, in terms of like, you know,um at ChatGPT as well, it is
amazing because when you thinkof like what chat can do now,
(29:46):
um, you know, it more so, like,especially with my team, more
than just a wall of text, right?
We can do tables, we can dointeractive tables, images,
videos, and things like that.
Um, you're getting to a worldwhere, like, yeah, uh, I can
create, you know, I'm I'm I'mlearning something, you know.
Let's say I'm a student, right?
I'm having a little troublewith understanding acute and
obtuse angles.
Uh, you know, build me a quick,you know, uh interface, uh, a
(30:09):
program that I can learn thisfrom.
And it goes, boom, it justgenerated for you within
seconds.
And you go, Oh, I see how theangles work.
Now it's interactive and itshows you, and it was custom to
you.
And how that is better thankind of what we have today is,
you know, today somebody mayhave already generated that tool
and you have to go to thewebsite and to find it, but this
is built inside a chat.
You already has theconversation and context, and it
(30:30):
probably even knows that youlike Minecraft as well.
And it made the tool fun andinteractive, like because you
know Minecraft.
And it's building and creatingthese ephemeral tools that are
unique to you, to your learningstyle.
I think it's going to be such ahuge boon for uh, you know,
children learning in schools oralso even, you know, adults that
are learning new concepts everyday, especially in this tech
world.
Like, okay, I know about this,teach me about that, and it can
(30:52):
do it in a way that's reallyrelevant to you.
And so um, I'm excited aboutthose quick and new interfaces
and tools that it would be ableto generate.
Uh, we're not quite there yet.
Prakash Chandran (31:01):
Yeah, for sure.
I cannot wait for that.
Um, you kind of touched onsomething that I have been
thinking about a lot, just evenwith my own personal usage, um,
you know, and that's literacy,right?
Like I think uh the amazingthing about something like a
Chat GPT is I'm able to deeplykind of think through with or
sharpen my thoughts aroundsomething by just interfacing
(31:24):
with it.
Like I'll be driving and kindof using the voice feature and
having a conversation, and Idon't feel embarrassed to be
like, can you explain thatagain?
Like I'm like eight years old,please.
Um, but it really, really hashelped me clarify um a lot.
On the other side of things,I've found myself like, you
know, ChatGBT is so good at, youknow, generating, for example,
email responses or even thingsaround structuring things for my
(31:47):
board deck that I'm like, Idon't want to sacrifice kind of
that skill that I have of doingthat myself.
I feel like there's like,there's like trying to give it
too much and do too much for youand losing kind of that ability
to shape it yourself.
And then also kind ofleveraging it to like help
sharpen your thoughts and likefrom my my personal point of
(32:07):
view, using it appropriately.
Like that's like the highestand best use for it.
I'm curious how you personallythink about that, right?
That spectrum of like handle itfor me versus using it in a way
that's productive and doesn'ttake away from your uniqueness
and the value that you provideto the world.
Tony Casparro (32:26):
It's an excellent question, um, especially as
we're we're using it more andmore for everyday tasks.
Um, somebody has coined this ascognitive offloading, where
something that was challengingfor you, that's you know,
stressed an ability that youhave, you're now offloading to
the AI.
Um, I have caught myself doingcertain sometimes doing things
where I go, you know, I shouldbe able to figure this out.
You know, maybe it's amultiplication problem or
(32:47):
whatnot.
Let me do this.
Um and it's important to keepin mind that uh as humans, you
know, stretching our mind andkeeping it sharp uh is really,
really important.
Using your literacy tools andusing uh your abilities is
really important.
So, you know, thinking of AInot as a replacement for
yourself or your own thoughts,but as a partner to work with
(33:08):
you is a really, reallyimportant distinction.
Um we are working together onthis document.
Uh, it is not just gonna haveit write it and then go off uh
like uh uh some people do.
Um we're gonna work together onthis, we're gonna solve this
problem together.
Uh it's a much better way ofthinking through that.
And so being careful uh to nottip that too much in the favor
where you're um uh notstretching and you know use
(33:30):
utilizing your brain as much asyou should, and remembering that
that's really important foryou.
Uh the using your brain andstretching it every day is uh
critical to staving off a lot ofdiseases later on.
Prakash Chandran (33:40):
Totally.
And I think um I think just thekind of the key thing that you
said there is making sure you'reworking together and learning
together alongside uh of it.
Because I think for me, atleast that has been kind of the
best use of it.
And I caught myself once whereI was like trying to have it
like answer everything for mebecause like when you have this
kind of magic thing, you're justlike, oh my God, I'm just gonna
(34:01):
use it for everything.
And then I ended up, Iremember, copying a response to
like I had mostly written themessage, but like I just had
relied on it, ChatGPT, to kindof finish it up.
But then I pasted it and itpasted also the um, okay, here's
like an updated response thatmaintains your tone and
everything in the uh the like aSlack message that I had sent.
And it just it caused me topause and be like, okay, did I
(34:25):
really need to do that?
No, probably not.
But like I was, you know, it'sso good that like I I had
started to kind of go in a waythat me personally, I wanted to
kind of co-create and learnalongside with it.
So I feel like I've gotten intoa good balance, but we'll see
as things uh a great example ofthis is a power tool.
Tony Casparro (34:44):
I remember, you know, my dad taught me, you
know, growing up how to use asaw, right?
And like, you know, the painthat it was.
And using power tools is awhole other ball game.
But you have to pay attention.
You have to really make surethat you are in control of that
product.
Um, otherwise it can reallyhurt you.
And so that is the same way Iapproach AI, right?
Like I am, it is a tool that Iuse, uh, but I need to make sure
(35:05):
that I have always control overit.
Prakash Chandran (35:06):
Um, as we start to close, just uh one more
question here, just onsomething that you had
mentioned, which is like, youknow, we talked about kind of
the age of like this newinterface, this personalization
and you know, the AI or ChatGPTbeing able to like tailor things
from a number of different datasources together to give the
user the exact right thing.
If I'm a business today and I'mthinking about the interfaces
(35:30):
that I'm gonna design, I'mstarting an application or I'm
trying to expose my businesslogic uh in a meaningful way,
how would you recommend thinkingabout that?
Like from the interfaces thatthey're creating to um, you
know, their integrations with AIapplications, like where would
you recommend that they start tomake sure that they're
(35:51):
future-proofing themselves?
Tony Casparro (35:55):
Um, you know, so if we're talking about like a
company that's making thingsmaking things for consumers, uh,
first of all, um, most peopleare starting to go to this role
of using voice or text todescribe what they want, right?
So if they come to your site,if you're having them navigate
through a lot of, you know,wizard wizards or menus and
whatnot, um, people are reallycoming to this world to expect,
(36:16):
like, here's what I need to do.
Like, you know, give me acouple options or you know,
allow me to use natural languageto tell you what I want and
give them that approach throughtheir site, right?
Making people learn how yourtool works and the complexity of
that.
Um, maybe there's a place forthat, but most people are
expecting now for the UI to betailored to them instead and not
the other way around.
(36:36):
Uh and when you talk to uhbusiness, business customers
now, uh really making sure thatyou have excellent
documentation, you have theability to search and ask
questions in it.
Um, developers get very, veryfrustrated when they can't see a
lot of good examples orinteractive tutorials or ask
good questions of thedocumentation.
These are kind of now stakes,uh table stakes for people
coming in to work with yourtool.
(36:57):
And so ensuring that um that isreally primed and ready to go.
I'm sure there's third partiesout there that offer this as a
service to set those guides up.
Um, but that is really, reallykey.
Um and so yeah, that is kind ofthe way to win people over
right now.
People are expecting easy,they're expecting low friction.
Um, and if you're like, hey,you have to read all this and
figure it out and muddle throughwhat we we've written, and
(37:19):
we're not gonna give you anyguides or help for on it, it's
like, well, just making itharder, you're reducing that
friction they know.
Prakash Chandran (37:24):
I I heard someone else kind of talk about
there's kind of the userexperience, but also kind of
like an agent experience whereyou have to make sure that
things are really welldocumented and described, just
like you would for a human thatlike needs to understand all the
elements of it.
But when you have also amachine that's working alongside
that human that can digestthings much, much faster, it's
really critical that yourbusiness logic, the way uh the
(37:47):
rules and the processes that youwant people and machines to
understand are really welldocumented.
So that makes a lot of sense.
Tony, you've obviously beenworking on lots of different
things.
There was a dev day.
I'd love to hear about what youhave been working on lately,
what's been keeping youoccupied.
Tony Casparro (38:05):
Yeah, so we uh just released uh Commerce inside
of ChatGPT.
Uh previously, if you werelooking for products like you
were I'm looking for a newexpresso machine or some shoes
or whatnot, we would show you alist of links inside of there.
Um they're really the links arenot ranked in any way, they're
according to what you want.
Um, but there was no ability toactually check out inside of
ChatGPT.
(38:26):
So one of our first kind oflike, you know, these new
third-party products inside ofChat GPT is you can just hit the
buy button and it appears inChat GPT.
So we work with Stripe as apartner to do all the credit
card processing, and then wework with our um merchant
partners, if that's like Etsyand Shopify right now, to do the
transaction.
So this is agentic shoppingwhere we say, here, we'll
(38:46):
collect the credit cardinformation, we'll talk to the
store, we'll orchestrate thetransaction, it's done.
Uh, we were involved in it onlysimply that we have a train, a
copy of the record so that youcan easily access it from that
chat conversation, but it wasall done for you between these
different partners.
So they facilitatedtransactions, they did all the
um in credit card encryption,whatnot, and then the shipping.
Um, and so there's this newworld of once again, inside of
(39:09):
Chat GPT, you have control overthings happening outside to even
do your shopping for you.
Prakash Chandran (39:13):
That's amazing.
So you're really just kind of afacilitation layer, and then
you're like bringing thesepartners together to make sure
that that transaction happensbased on the user's intent.
Tony Casparro (39:22):
Exactly.
And it's great for the user.
We've got the credit cardinformation on file, nice and
secure on our side, well, stripeside really.
Uh, and we're simply making iteasy to hit that buy button
inside of there and do this, doall your shopping right there.
That's awesome.
Prakash Chandran (39:36):
Just as we close, you know, coming off of
all the hecticness uh ofdeveloper day and everything
that you're doing with OpenAI,if you had to leave the audience
um with one thing, justsomething that you would like
them to invest more time doing,whether it be chat in within
ChatGPT or otherwise, what wouldthat be?
Tony Casparro (39:57):
You know, shockingly enough, it's simply
give it a chance.
Um, I have many friends who arein this with me.
They're co-u-engineers andwhatnot, and there's still a lot
of apprehension about, youknow, um its access or controls
or if it'll actually be useful.
Uh, and honestly, give it achance.
Uh, it's the same way it waswith Spotify.
I remember people telling me,like, it's great.
You have these, these, youknow, all the music at your
(40:18):
fingertips.
And I go, like, no, I I likethe way I've always done things
with things with CDs.
I don't want these new tools,and I didn't really give it a
chance.
Um have it do something, youknow, where you know it might
take you a half hour to likesearch through something on the
web or do some piece ofresearch, or you're looking for,
you know, everybody's alwayslooking for, you know, a new
product or, you know, anespresso machine or whatnot.
Have it do some of that taskfor you and see see how well it
(40:40):
can do at that task.
Um, if you're willing to haveit wire up to your code base and
just say, here's a task I wasgonna do today, do today, give
it a try for me and see how itdoes.
Um, these we've made thesethings trivial easy.
Like you don't have to log in,you don't have to make an
account, you can just go to chatgpt.com and just ask it a
question and try it out.
Um, it is uh it couldn't beless friction along that path.
(41:03):
Uh, and it might meaningfullymake a difference into your
day-to-day life.
You might be able to focus onthe higher level stuff that
makes you a better uh umengineer leader in your
workplace.
Uh, you might be able toleverage your time better, uh,
and that makes you a bettereffective uh person in your
field uh and uh all the morepowerful.
Prakash Chandran (41:22):
Tony, what a perfect place to end.
I really appreciate your timetoday.
Thank you so much, buddy.
Thank you very much, Rakash.
Appreciate it.
You know, thinking of AI not as a
replacement for yourself or yourown thoughts, but as a partner
to work with you is a really,really important distinction.
Um, we are working together onthis document.
Uh, it is not just gonna haveit write it and then go off uh
like uh uh some people do.
Um we're gonna work together onthis, we're gonna solve this
problem together.
Uh it's a much better way ofthinking through that.
Prakash Chandran (00:35):
Hi there, I'm Prakash, CEO of Xano.com, and
this is Future Proof.
In this episode, I'm joined byTony Casparo, Senior Staff
Software Engineer at OpenAI,where he's helping shape the
future of Chat GPT.
Tony has spent his career atthe intersection of scale and
design, leading UI and platformengineering at Netflix, founding
multiple startups, and nowbuilding on the front lines of
(00:57):
AI.
Um, Tony, really great to haveyou here.
Thank you so much for takingthe time.
Also a fellow uh slow neighbor.
Uh, it's great to have you alsoin town.
Thanks for so much for agreeingto do this conversation.
I'm happy to be here.
Let's talk.
Awesome.
I love it.
Well, why don't we start withsome background?
Maybe share with us a littlebit uh about uh, you know, what
(01:17):
you've done in the past and whatyou're doing today.
Tony Casparro (01:20):
Yeah, I've kind of had an interesting history as
far as working on the back endfirst.
So doing a lot of databasemanagement, starting at Cisco
Systems, uh, moving over tobiotech at Amgen.
Oh, excuse me, but uh reallykicking off my career when I
moved to Box to do um enterprisesoftware, right?
Really thinking through whatare the needs of you know, B2B
customers, how do we make youknow great software that's easy
(01:41):
to use, good for managementobservability.
Um, and then doing a wild swingto Netflix.
So after enterprise for a fewyears, I really wanted to get
kind of get back in the consumerspace.
And that journey was reallyabout um how do we teach the
world what streaming is?
This is, of course, you know,12 years ago now.
So, what is streaming?
How does it work?
How do you even think abouthaving an entire library at your
(02:03):
disposal and figuring out whatto watch next?
So after 10 years at Netflix,just building Netflix.com, I'm
doing infrastructure work thereas well.
Uh, I am now here at uh OpenAI.
I work on Chat GPT.
Um right now we're building outuh commerce and we're also
doing uh new work with, excuseme, bringing like real to like
um rich responses so that theinformation is not just a wall
(02:26):
of text, but actually hasmeaningful layout, uh
responsiveness to what you'reactually looking for.
And so that's what I work onevery day.
Prakash Chandran (02:34):
Awesome.
So exciting.
And I definitely want to diveinto all that.
But I first maybe wanted tostart with your time at Netflix.
Um, kind of what a uh pivotaland interesting time uh when you
joined Netflix, kind of in thisworld where people didn't
realize what streaming was.
Talk to us about some of likethe um kind of what what the
(02:54):
paradigm was like at the time interms of like the concept of
what it meant to streamsomething and some of the
affordances and things uh thatyou had to think through at
Netflix in order to like usherin this new uh era of streaming
on the internet.
Tony Casparro (03:09):
Yeah, you know, um I built my own servers before
this.
And so it was interesting, youknow, that I had been used to
this kind of new world for awhile, but like showing other
people who were, you know, stillpopping DVDs into the mailbox
and then into their players, um,or a little bit using, you
know, Apple iTunes at the timeto purchase things one-off.
Um, this was a new world, youknow, the same with Spotify when
(03:29):
they kind of introduced peopleto the idea that you have access
to everything.
And you're like, okay, where'smy music?
Well, your music is all themusic.
And there's this concept oflike, okay, but I how do I
actually use this?
Well, how do I actually youknow enable this to work for me?
Um, so same thing with with uhstreaming, right?
Um, we first of all had to showpeople that like um here is
what this movie is about, right?
(03:50):
And doing a lot ofexperimentation on like showing
which kind of images for theparticular users and custom
tailoring it to their needs.
Um, do we show them a wall oftext about what the movie is
about, or do we show them just afew quick words as far as this
is exciting and thrilling andwhatnot?
Um, using um instead of staticassets, like actually moving
assets, showing trailers in lineand things like that, um, and
(04:11):
overcoming a lot of umassumptions people had about how
this was going to work.
I remember going to a lot of uhtests where we would put the
product in front of people andwe'd be watching it behind a
two-way mirror, and they'd beafraid to click on something.
Why are you afraid?
What's going on?
Why are you afraid to click onthis?
Because it's going to chargeme.
And there was this idea that,you know, the world was
piecemeal.
You had to pay per view oneverything you did.
(04:32):
And the idea that it was allavailable.
You could just click on it andit would start playing with one
single cost per month.
We had to educate and overcomea lot of barriers that we take
for granted.
So it was a lot of interestinglearnings.
We did a lot of experimentsthat worked and a lot that did
not work.
Some things where we assumedtoo much about the user's
journey as to why were they atNetflix.com?
Was it to watch?
Was it to actually learn abouta movie?
(04:54):
Was it to figure out somethingto watch later and really
understand uh and make a greatjourney from no matter what
people were doing on the site?
Prakash Chandran (05:01):
So, you know, I think there's so many kind of
parallels in in terms of likewhat you're doing uh at OpenAI
with Chat GPT.
Uh before we touch on that, Iwould love to understand if
there's one key learning thatyou still take with you today
from your days at Netflix, whatwould that be?
Tony Casparro (05:20):
It's too, uh I think it's very naive, but we we
often think like, well,everyone does this the same way
that I do.
Uh and it's simply not true.
Uh you think of something assimple as um scrolling a page.
Uh it turns out there's manydifferent ways to scroll a page.
Some people click go to theright and they click on the
scroll bar and they drag itdown.
Some people use the arrow keys,some people are using voice
commands.
There's millions of different,we're not millions, there's a
(05:41):
lot of different permutationsfor how things people do things.
And when you build an interfaceand you're assuming everyone
thinks the way I do about how tonavigate through this or what
my intentions are, you're goingto miss a lot of people.
And so this is something thatcompanies that you know have a
lot of exposure do really wellis they know, here's how to make
it super accessible.
A great example of this, uh,somebody else told me, is a
(06:02):
steering wheel.
Imagine if a steering wheel wasalways set and you had to grip
it a certain way.
It wasn't a wheel, it was liketwo handles.
It only works for the way thatone person thinks.
But steering wheels, in bytheir design, allow for people
that drive with it on the top ofit, on the bottom of it, on the
side.
It works for all the differentuse cases.
And something as simple as awheel, it can be a beautiful
illustration of how it works foreveryone.
Prakash Chandran (06:22):
That's really a fascinating correlation.
And um, as I kind of alludedto, like, really kind of
poignant to how or the work thatyou are doing at OpenAI,
especially with ChatGPT, becauseit's a similar parallel because
you're also designing in thisnew paradigm shift where people
have access to everything,right?
Maybe not a media library, butnow the world's kind of
(06:45):
information at their fingertipsin a way that had never been
accessible before, like gettingto the answer versus searching
through the sea of blue links.
And there's an input box nowthat serves many different types
of people.
Like I use it, my uh 70 uh orsorry, 82-year-old dad uses it,
uh, my sister uses it, um, mydaughter uses it.
(07:08):
It's crazy that you have somany different people
interrogating and asking uh, youknow, ChatGPT to do things for
them.
So when you think about the thelearnings that you've had at
Netflix and that same paradigm,how have you guys tried to
approach that not problem, but Iguess opportunity of serving so
(07:28):
much of the world through thisinterface uh called Chat GPT?
Tony Casparro (07:33):
It's a really hard problem, right?
As you said, it's a supergeneral processor.
It can do all kinds ofdifferent tasks and you really
have to kind of figure out howto meet people where they're at.
What do they need help with?
Is it relationship advice?
Is it building an essay forsomething?
Is it researching something onthe web and trying to make it
meaningful and accessible tothem?
Um, we've made great strides.
(07:54):
Um, as were before, there was alot of buttons on the bottom of
the toolbar.
We've made great stridestowards you just type whatever
you want, and the model is goingto figure out which engine do I
need to use for this.
Do I need to use a thinkingmodel?
Do I need to use a websearching real-time model,
whatnot, and allowing the uh thebrain of the model to choose
which direction will take you.
There are, of course, you know,hints that you can give it if
you need to say, like, this is adeep learning task I'd like you
(08:16):
to do.
Um, but overall, we wererelying on that model for that
contextualization.
Um, and then in addition withthe response that actually comes
back, giving you meaningfulpivot points, right?
So that if you are looking forreal-time information, you can
pivot to like, okay, great, nowshow me the stock ticker.
Um, or if you're looking for uhwriting a document, okay,
great, open this in a canvas,you know, help let's figure this
out together and work on this.
(08:37):
Um, and so it really is tryingto figure out as soon as
possible what is the intent ofthe user?
What are they trying to do?
And then with the response, bea really meaningful.
Don't just show them a wall oftext, showing them really
contextual information about howthey want to perceive this and
give them those pivot points tothen take the conversation to
the next level.
Um, we've got some great stuffcoming up for this that really
(08:58):
show off these pivot points.
Uh, but this is key to saying,uh, to showing or with hints,
showing the user the capabilityof the product without being
overbearing and showing them awall of text or buttons.
Prakash Chandran (09:10):
Yeah, you know, I think um even in
coordinating this, um, we had tomake sure that you got through
uh a crazy amount of work, uh,obviously culminating in just
the recent developer day thathappened.
It was kind of an amazingmoment, actually, for me to
watch because I really feel likeChatGPT has evolved in uh in
something that was maybeconsidered like a chat interface
(09:33):
and tool to more of a platformand to what you're speaking
about, giving people moretailored contextual experiences
rather than just a wall of textis something that we can just
see the world uh going to.
Tell me, like, as someone thatworks on the interface, how what
are some of the key challengesthat people not may not
necessarily understand or thinkthrough that you have to think
(09:55):
about in terms of presentingthat type of information with
context and tailored andpersonalized to the person
that's searching?
Tony Casparro (10:02):
What's interesting is especially in
this world where the model istrained on a lot of information
and yet there's a lot ofreal-time information that it
can have access to, real-timeinformation in terms of what's
available on the web or toolsout there, all these new um uh
MCP servers and whatnot.
Uh, and so one of thecomplexities is how do you
orchestrate giving the user afast response and at the same
(10:23):
time, basically in parallel, letme go talk to the stock widget,
let me go talk to the weatherwidget, let me go talk to this
MCP server and bring this back.
And so some of the complexitiesof the UI are UI are you know,
how do you show the informationthat you have so far?
How do you show the user thatyou're still bringing in extra
information or that you'reprocessing something really
important and make it not feelslow and janky, right?
Because the power is that itcan build you a great response.
(10:46):
But if it feels like, man,nothing's happening, I don't
even know why I'm waitinganymore, that feels bad.
And so what we do is somethingcalled uh the chain of thought.
So the chain of thought is asit's thinking through whether
that is real-time information ortalking to other servers, we're
showing you little snippets.
Hey, I'm talking to that serverover there right now.
I'm gathering the real-timeweather information.
And that's giving the userinsight into what's happening
(11:07):
behind the scenes in terms of athinking model that's
interacting with the world, um,and then building that useful uh
response in turn.
Prakash Chandran (11:15):
Yeah, that orchestration and kind of giving
a sense to the user around whatuh ChatGPT is doing is just it
when it works really well, itjust feels like really seamless
and elegant, but I can imaginehow much thought goes behind it.
Um, you know, one of the thingsthat we've talked about before
uh is kind of the most uhunderutilized uh tool within
(11:37):
ChatGPT, and we've talked aboutit being deep researcher.
That that was the answer at thetime when we were chatting.
Uh and since you told me that,I actually have been using it a
lot more.
Even though, like, you know, Ithink I'll I'll be in a
conversation and uh it willobviously try to optimize for
quick answers, but I'llexplicitly say, look, for this
one, I want you to think deeplyabout it.
(11:58):
I'll make sure to tell it to gointo deep research, and it
gives me like something really,really thorough.
I'm almost giving it thefreedom to take a little bit
more time to give me like areally thorough answer.
Do you see more people startingto use ChatGPT this way?
Because I think this is whatI'm imagining.
You tell me if this is true.
People, a lot of people, areshifting their behavior from
(12:20):
traditional search to Chat GPT,but they're still using ChatGPT
like search, right?
They're still typing in likemaybe piecemeal words and not
necessarily using it how it ismeant to be and how it can best
be utilized.
Tell us what you're seeing anduh kind of the dynamics around
like deep research as a featureand the best way to utilize
(12:40):
ChatGPT.
Tony Casparro (12:42):
Yeah, you know, um, so the deep research is one
of those hints that I wastalking about.
It's in the toolbar where youcan say, This is how I want you
to answer this question.
And deep research specificallymeans take as long as you want.
I don't want a quick answer.
Um, I want you to be verythorough, go through lots of
sources, take your time, and getback to me with a really deep,
full response.
And so it is one of my favoritefeatures, especially for
(13:03):
shopping for something andcomparing, you know, a bunch of
different review sites or modelsuh for travel advice.
Um, it can be really, reallythoughtful instead of giving you
an answer very quickly.
Um, when you ask about userbehavior, though, it turns out
people still want somethingvery, very quickly.
They don't want to wait, theywant the answer right away.
Uh, and so what we've kind ofcome up with in terms of the UI
is kind of a bifurcated mode.
(13:25):
So it will start to say if itis going to take a while for
something like I do need to dosome research, whatnot, there is
a skip button.
You can say, I don't care, justshow me what you got right now.
And that allows people theflexibility to be like, uh, this
is how I uh want my answer.
Um and so I can't say for surelike what the adoption is like
on some of our uh other featureslike deep deep research, uh,
(13:47):
but people are still very muchin the mindset that they want it
quick, they want it fast, andthey want it correct.
And we are working as hard aswe can to make sure that that is
the path forward.
Prakash Chandran (13:55):
Yeah, I love that bifurcated mode where it's
like, hey, look, and I think itsays something like thinking
deeper for a better answer orsomething.
And then you have the abilityto be like, nope, nope, it's all
good, just get give it to menow.
Um, and I think that's the hardpiece, right?
Kind of balancing that speed,that AI speed that people have
come to expect with thethoroughness that something like
a deep research um can offer.
(14:16):
Um, I want to shift gears alittle bit.
You know, you are a staffengineer, you have been
developing in some capacity foryour entire career.
Uh I think the audience mightbe curious to know how you are
leveraging AI in your day-to-dayto develop software at OpenAI
and uh so you know some of thecore benefits that you've seen
(14:38):
and things that you might haveto like uh supervise or kind of
still be aware of that are stillkind of um evolving in terms of
like the landscape.
Tony Casparro (14:46):
Yeah, absolutely.
Uh, you know, the best way tothink of AI is as an assistant,
right?
It is there to help you withyour tasks.
And so when I get into theoffice in the morning, what do I
do is I look at my plate and Igo, what can I offload to an AI
so that I can work in parallel?
So some requests might come infor some code changes or
connector changes or whatnot.
Um, I can muddle through thoseand go do everything in a serial
(15:08):
manner, but I have the abilitynow to say, let me offload some
of these tasks.
So we have Codex as one of ourum AI coding agents.
It's available to everyone aswell.
Um and it's pretty phenomenalwhen it has access to your code
base to say, you know, this isthe change that it can even
read, of course, now like lineartasks and Slack messages.
And you can say, here's thecontext of the issue, come up
with a solution for it.
Now, uh, like all AA things, itmay not be perfect at start,
(15:31):
but it can actually give youquite a bit of leverage and
thinking through opportunitiesso that when you come back to
the task, you can be like, well,that's pretty clever, but let's
change this, rewire this,perfect.
Now we're done.
And so as opposed to meworking, like I said, serially
in the morning, I'm thinking uhacross many different verticals
at the same time.
I've also got AI, of course,built into my IDE, which is my
uh coding interface.
(15:51):
And now it's like, you know,like, you know, applying some
formatting or markdown or justkind of threading through some
variables and whatnot.
And I can say, just finish thisoff for me.
And so that allows me to have amuch more productive day in
terms of focusing on the biggerpicture tasks uh and then uh
sharing the work with the AI totake care of the more kind of
menial stuff to like just kindof thread through.
(16:11):
Um, and so that's for work andthen also for you know doing
research and things like thatfor other things that we have
going on at work.
Uh it used to be a world likewhere you said, like, okay, I'm
looking for this new um statemanagement.
Let me go to Google.
Okay, I see a bunch of links.
Let me go read through thesearticles and spend half my day
doing that.
Or I can say, hey, here'sanother task for you, not Codex,
(16:33):
but ChatGPT.
Go research the latest statemanagement to make sure the
information is up to date.
Bring me a report, build me aspreadsheet that I can review.
And that is once again, notdoing all the work for me.
It's in a supervised way whereit's helping me and assisting
me, giving me a collection ofinformation that I can then use
to make my next decisions and godeeper as I need to.
Yeah, that's fascinating.
Prakash Chandran (16:53):
So, like there's gonna be developers and
technical builders listening tothis.
And one of the things that yousaid was kind of really um uh
really interesting, just interms of like, you know, at the
beginning of your day, you'rekind of figuring out how can you
leverage this assistant to takecare of some things, to do some
research for you while youfocus on maybe the higher order
bit.
For the people that arelistening, how do you structure
(17:15):
that, that planning processaround what you can delegate to
the assistant and uh to uhleveraging AI for versus what
you need your attention on?
Tony Casparro (17:24):
Yeah, I mean, you know, so as a UI engineer, uh I
do a lot of visual testing.
Um, and it's something where uhour like Codex is extremely
good at like backend codingright now, and it's still
developing its chops in terms ofUI coding and testing those UI
changes.
How do they work?
How do they work foraccessibility?
You know, how does it work on afast connection, solo
collection?
There's a lot of things we needto consider when we're building
user interfaces.
(17:44):
And so some of those tasks, Imight have it like um, for
example, you start building thisfor me.
For example, I was building ananimation the other day, and I
knew it was not gonna hit theanimation correctly, but I was
like, there's gonna be ananimation that occurs when
somebody clicks this button.
I need you to pop open this, dothis, use this variable.
And it did a pretty good job,but at least it was a start for
me to be like a jumping offpoint for my PR that I was gonna
eventually write.
(18:05):
Um, and so when I come in themorning and I'm thinking about
those tasks, uh, I am thinkinguh in terms of like, how would I
approach this problem?
Okay, these are the steps Iwould take.
Let me give those to the to theagent and have the agent take
care of it for me.
If that is too complex, I go, Ican't even fathom my head what
this is gonna look like.
I just have no idea what theoutcome is going to be.
I don't know uh the correct wayto do the performance of this,
(18:26):
whatnot.
I'll take those on and start onthat journey because it's it's
just an unknown.
So to clarify for short, ifit's a known end state that I
want, uh once again, comparingthings uh across different
websites, doing research andanalysis, whatnot, there's a
known outcome.
I want to find the best this.
Uh, that's a great choice forit.
But when you're reallyideating, you're not sure what
(18:46):
it's gonna look like, you'remaybe working with partners and
stakeholders, and you're like,we don't know what that's gonna
look like yet.
That's more the state where Itake ownership and control to
like do the human stuff, talk topeople, get feedback, and
figure that out too.
Prakash Chandran (18:58):
That makes sense.
Um, you know, talk to me alittle bit about where you feel
like uh AI is not quite thereyet, but you're excited to see
it evolve so you can takeadvantage uh of leveraging it
kind of in your day-to-day.
Tony Casparro (19:12):
Yeah, we've made great strides towards um
orchestrating tasks, right?
So this is the idea, like inyou know, a year ago, if you
said, you know, change thispiece of code and structure the
state in this new way, it wouldhave just started spitting out
code, right?
It wouldn't really thinkthrough its tasks.
And so something we've toldthese coding agents now is
develop a set of steps thatyou're going to take, right?
(19:34):
I'm gonna do this, thinkthrough these, think through the
uh side cases and edgeconsiderations here, and then
start coding it.
And you get much better resultswhen you have it think through
a process to follow first.
Um, and so um, in terms of notbeing quite there, the UI piece
is is tricky, right?
Because you need to test acrossdifferent browsers, you need to
test across differentconnections.
(19:54):
Um also AIs don't really knowwhat a beautiful animation looks
like.
They don't know how it feels toa user.
That's something very visceralwhere you go like it should feel
this quick and move this, andyou know, especially when you're
doing multiple uh coordinatingmultiple animations in the same
way, you need that to feelreally good together.
And so it doesn't have a goodfeel for that just yet.
Even in terms of the userinterface, I've experimented
(20:15):
with a lot of um our own toolsand competitors in terms of like
what makes a great UI.
And they guess and they lookokay, but it's kind of using
just random examples.
It's I wouldn't say it's agreat UI designer yet.
Um, there's something reallymagical about the way humans
think through how a userinterface is going to flow and
how that's going to work.
And so I think those are areasfor opportunity.
(20:35):
Backend stuff is getting very,very good at.
Um, you know, if we go to aworld in the future where um
backend code is not even a codebox anymore, it's just a black a
black box where you say, hey,uh, this is the inputs of the
API I want, and these are theoutputs I expect.
I don't care what the codelooks like inside, uh, as long
as it's quick, just go be ablack box.
(20:57):
And we may end up in a worldthere where we have black box,
you know, um inputs and outputsfor Vatican servers.
Prakash Chandran (21:03):
Um, you know, talk to me a little bit about
how uh just in a from adevelopment standpoint, you
think about like security andcontrol over what can sometimes
be the back uh the black box,uh, whether it's front-end or
back end, how you think aboutthat in your day-to-day?
Because I feel like that'sprobably the supervision part
that you're talking about,making sure that you're in the
(21:25):
loop and making sure that it'sdoing the right things.
But you know, in yourday-to-day, how are you kind of
implementing that and thinkingabout it as an engineer?
Tony Casparro (21:34):
Are you thinking, are you asking specifically
like how I use the securitycontrols or how we use the other
thing?
Prakash Chandran (21:38):
How is you you as a company think about those
controls in the develop in yourday-to-day development?
Tony Casparro (21:44):
Absolutely.
So it's person in the loop,right?
You know, uh the AI shouldnever have access to do um what
could be considered um uhdangerous tasks on its own,
right?
You know, executing you know,arbitrary code and things like
that.
It will ask you and prompt you,is it okay if I run this?
Here is the exact command I'mgoing to run, and make sure that
the human is always in controlof those tasks.
Um, as we start to get anexposure into the world of MCPs,
(22:07):
same thing.
We want to make sure that thoseuh checkpoints are there so
there's human in the loop and sothat it's not off and running
on its own.
As our developers, however,build their own products, we're
encouraging them to also putthose same safeguards in place.
However, there might be safelyguarded loops where you go,
like, hey, it's gonna generateimages.
I already have these moderationcontrols in place, I know this
is going to be safe, and you canhave it run that uh with that
(22:29):
oversight, uh, as long as uh adeveloper or somebody has
already made sure that there's asafe path for it to take.
Prakash Chandran (22:36):
So I want to go into just maybe some
practical takeaways for peoplethat are um, you know, might be
building a new project andtrying to figure out the best
way to leverage AI.
You know, you feel like thereis, you've got all these
bypoders and they're you knowbuilding applications very
quickly.
You have people that areleveraging AI alongside of
(22:57):
themselves in a code base, uh,but from a chat GPT perspective,
for a new builder that istrying to think of uh, you know,
developing something new intothe world, or even for a product
owner in a company that'strying to make something new
within their organization, justin terms of a mental model of
how they should be thinkingabout it and leveraging AI, what
(23:18):
recommendations might you havein order to get started in the
most efficient way possible?
Tony Casparro (23:25):
You know, it is interesting because we have
Codex internally, it has accessto our Intari code base.
And so we've had people, uh, Imean, that's always a great
starting point.
If you have an existingproduct, you give it access to
your GitHub repo, it knowseverything about your code.
And when you ask it a questionabout it, um, it can efficiently
use the current styling andpermission sets and uh plans
that you use for your existingcode, and then it can do the
(23:47):
augmentation or adding newfeatures, whatnot.
That turns out to be decentlywell.
Uh, we've had people that areproduct managers, they asked for
those changes, they send it tolike a developer with the change
set.
Uh, the developer can then doslight modifications or whatnot
is needed, but it does help theprocess along because um usually
that process is the developergoing to the person, the person
(24:08):
describes what they want, theygo back and forth, and it's nice
if you know the person can justdescribe everything you want up
front and get that result.
Um, in terms of like startingnew code bases, uh, that is a
tricky one.
You know, zero to one with anew product.
You know, um, I personallywanted to do like, you know,
some more app development andwhatnot.
Um zero to one is always a hardproposition, right?
Because you have to build outsome servers or some uh, you
(24:30):
know, some developer toolkits onyour local machine.
Um I can't speak to anywebsites that currently do like
a great job.
I'm sure they're out there, uh,but it's a great opportunity
because there are so many peoplethat are vibe coding small
applications.
Uh, it'd be fantastic if youcould uh use one of these to
start building up, you know,whether that's a website or an
app that you're gonna release tothe app store uh and get you up
and running quickly.
(24:51):
So uh I guess do some research.
Uh I'm not I can't recommendany particular ones.
Uh Codex is great though forexisting codes.
Prakash Chandran (24:57):
That's awesome.
Um and then, you know, just interms of like uh, you know,
ChatGPT is such a challenge forall of the reasons that we've
been uh speaking about, just interms of the U and I UI UX uh
paradigms that need to beintroduced.
I'm curious to uh if you havethoughts around any UI or UX
(25:18):
patterns that you will seeeither disappear or be
introduced into this new worldof new interfaces uh that we're
entering into.
Tony Casparro (25:28):
So that's a great question.
That's right in my wheelhouse.
So um uh as your users may not,or your listeners may or may
not know, uh we releasedsomething called uh third-party
apps inside of ChatGPT uh justthis Monday.
Um so this is the world nowwhere you can talk to ChatGPT,
and instead of ChatGPT saying,here is a link to Zillow, go
(25:49):
like click out to them and leavethe site.
Uh you can now talk to Zillowor these other uh apps that we
partnered with to start insideof Chat GPT.
So um that means they'recreating custom UI elements.
Uh ChatGPT and these uh thirdparties are now working together
to build visualizations.
Um and that turns out to beextremely powerful.
Another example is Spotify,right?
(26:10):
Um hey, I'm coming up uh for aplaylist, we're having an 80s
party, make me a great playlist.
It could list a bunch of songs,but it'd be great if it could
interface with your Spotifyaccount, build you the playlist,
and then you're done, uh, asopposed to you having to go add
those songs manually or whatnot.
And so bringing these appstogether to be controlled inside
of ChatGPT is gonna be a reallyinteresting new world.
(26:30):
Excuse me.
And then, of course, on theother side of it, we have these
agents that you can build intoyour current sites that go
backwards, right?
So now you are the main app.
Chat GPT is now embedded insideof your application.
Now these two can interface aswell.
Um, it's gonna be a reallyinteresting world in terms of
how people think about you know,branding and um app stores.
(26:51):
I mean, you can see a world nowwhere if these UIs are so like
chat agents are so powerfulenough to build, like integrate
with third parties and evenbuild interfaces on the fly, you
could start to see app storesgo away.
You know, why would I have togo like search my phone and find
this one app if this other appcan just do it all for me,
right?
It can generate the UI,generate my feed, whatever I
(27:13):
need, and build it in a customway.
And so you can see a long tailroad down here where, you know,
uh there is no more searchingfor apps, there's no more
looking for context.
You can just interface withyour device in whatever way you
want, probably through umspeaking and or text, and just
say, show me the weather fortoday.
It doesn't have to open an app,it doesn't have to do anything
else, it can render a rather aweather widget for you.
(27:36):
Uh, you know, tell me how myfinances are doing, you know,
like cross-reference myportfolio and see what I need to
do for my taxes.
And it's generating stuff onthe fly.
Uh, and so there's a reallypowerful new world out there for
those that um are thinkingthrough how does AI and my
application fit together tocreate a single united world as
(27:56):
opposed to I gotta get back tomy site.
Um, and it's gonna be realinteresting to see.
Prakash Chandran (28:01):
I totally agree.
I was actually talking to aprevious guest um about this,
how like the front end uh, youknow, in terms of the way we're
used to digesting, where you Idownload a singular application
or I go to a web application,uh, is changing.
Like because uh the more peopleinterface with tools like a
ChatGPT and get those answers,yeah, the nature of how
(28:21):
companies will express theirbrand through like this new
apps, SDK and otherwise, isgoing to just fundamentally
change.
So we're kind of entering thisera of like personalization.
And uh and another thing thathe he brought up that I thought
was fascinating was a potentialworld of like ephemeral uh
applications.
Because sometimes we build anapp for a purpose, but once that
purpose is done, especiallywith how fast we can create
(28:43):
applications, like it may havebe like purpose built for this
one thing and then it may goaway.
I'm curious, have you have youthought through anything like
that, just even outside of whatuh you're doing at OpenAI, but
just personally, how do youthink about the world of um
digesting information,application development, and how
we're going to utilize orinterface with businesses in the
future?
Tony Casparro (29:05):
Yeah, you know, it's funny.
Netflix did a lot of kind ofone-off applications that were
short-lived.
You know, we were doingexperimentations inside for, you
know, do, you know, does itmatter if we show a different
image to a customer for a moviethan this other image, right?
And we would do theseexperiments where we would have
an idea like that and we wouldneed a UI that was just, you
know, ephemeral.
It was just gonna be there fora while while we evaluated the
(29:26):
efficacy of this product.
Um, and we would spin those upall the time.
Um, and that process was youknow very cumbersome at the
time, and now we have AI to helpgenerate some of those for us,
and it's very, very nice.
Um, in terms of like, you know,um at ChatGPT as well, it is
amazing because when you thinkof like what chat can do now,
(29:46):
um, you know, it more so, like,especially with my team, more
than just a wall of text, right?
We can do tables, we can dointeractive tables, images,
videos, and things like that.
Um, you're getting to a worldwhere, like, yeah, uh, I can
create, you know, I'm I'm I'mlearning something, you know.
Let's say I'm a student, right?
I'm having a little troublewith understanding acute and
obtuse angles.
Uh, you know, build me a quick,you know, uh interface, uh, a
(30:09):
program that I can learn thisfrom.
And it goes, boom, it justgenerated for you within
seconds.
And you go, Oh, I see how theangles work.
Now it's interactive and itshows you, and it was custom to
you.
And how that is better thankind of what we have today is,
you know, today somebody mayhave already generated that tool
and you have to go to thewebsite and to find it, but this
is built inside a chat.
You already has theconversation and context, and it
(30:30):
probably even knows that youlike Minecraft as well.
And it made the tool fun andinteractive, like because you
know Minecraft.
And it's building and creatingthese ephemeral tools that are
unique to you, to your learningstyle.
I think it's going to be such ahuge boon for uh, you know,
children learning in schools oralso even, you know, adults that
are learning new concepts everyday, especially in this tech
world.
Like, okay, I know about this,teach me about that, and it can
(30:52):
do it in a way that's reallyrelevant to you.
And so um, I'm excited aboutthose quick and new interfaces
and tools that it would be ableto generate.
Uh, we're not quite there yet.
Prakash Chandran (31:01):
Yeah, for sure.
I cannot wait for that.
Um, you kind of touched onsomething that I have been
thinking about a lot, just evenwith my own personal usage, um,
you know, and that's literacy,right?
Like I think uh the amazingthing about something like a
Chat GPT is I'm able to deeplykind of think through with or
sharpen my thoughts aroundsomething by just interfacing
(31:24):
with it.
Like I'll be driving and kindof using the voice feature and
having a conversation, and Idon't feel embarrassed to be
like, can you explain thatagain?
Like I'm like eight years old,please.
Um, but it really, really hashelped me clarify um a lot.
On the other side of things,I've found myself like, you
know, ChatGBT is so good at, youknow, generating, for example,
email responses or even thingsaround structuring things for my
(31:47):
board deck that I'm like, Idon't want to sacrifice kind of
that skill that I have of doingthat myself.
I feel like there's like,there's like trying to give it
too much and do too much for youand losing kind of that ability
to shape it yourself.
And then also kind ofleveraging it to like help
sharpen your thoughts and likefrom my my personal point of
(32:07):
view, using it appropriately.
Like that's like the highestand best use for it.
I'm curious how you personallythink about that, right?
That spectrum of like handle itfor me versus using it in a way
that's productive and doesn'ttake away from your uniqueness
and the value that you provideto the world.
Tony Casparro (32:26):
It's an excellent question, um, especially as
we're we're using it more andmore for everyday tasks.
Um, somebody has coined this ascognitive offloading, where
something that was challengingfor you, that's you know,
stressed an ability that youhave, you're now offloading to
the AI.
Um, I have caught myself doingcertain sometimes doing things
where I go, you know, I shouldbe able to figure this out.
You know, maybe it's amultiplication problem or
(32:47):
whatnot.
Let me do this.
Um and it's important to keepin mind that uh as humans, you
know, stretching our mind andkeeping it sharp uh is really,
really important.
Using your literacy tools andusing uh your abilities is
really important.
So, you know, thinking of AInot as a replacement for
yourself or your own thoughts,but as a partner to work with
(33:08):
you is a really, reallyimportant distinction.
Um we are working together onthis document.
Uh, it is not just gonna haveit write it and then go off uh
like uh uh some people do.
Um we're gonna work together onthis, we're gonna solve this
problem together.
Uh it's a much better way ofthinking through that.
And so being careful uh to nottip that too much in the favor
where you're um uh notstretching and you know use
(33:30):
utilizing your brain as much asyou should, and remembering that
that's really important foryou.
Uh the using your brain andstretching it every day is uh
critical to staving off a lot ofdiseases later on.
Prakash Chandran (33:40):
Totally.
And I think um I think just thekind of the key thing that you
said there is making sure you'reworking together and learning
together alongside uh of it.
Because I think for me, atleast that has been kind of the
best use of it.
And I caught myself once whereI was like trying to have it
like answer everything for mebecause like when you have this
kind of magic thing, you're justlike, oh my God, I'm just gonna
(34:01):
use it for everything.
And then I ended up, Iremember, copying a response to
like I had mostly written themessage, but like I just had
relied on it, ChatGPT, to kindof finish it up.
But then I pasted it and itpasted also the um, okay, here's
like an updated response thatmaintains your tone and
everything in the uh the like aSlack message that I had sent.
And it just it caused me topause and be like, okay, did I
(34:25):
really need to do that?
No, probably not.
But like I was, you know, it'sso good that like I I had
started to kind of go in a waythat me personally, I wanted to
kind of co-create and learnalongside with it.
So I feel like I've gotten intoa good balance, but we'll see
as things uh a great example ofthis is a power tool.
Tony Casparro (34:44):
I remember, you know, my dad taught me, you
know, growing up how to use asaw, right?
And like, you know, the painthat it was.
And using power tools is awhole other ball game.
But you have to pay attention.
You have to really make surethat you are in control of that
product.
Um, otherwise it can reallyhurt you.
And so that is the same way Iapproach AI, right?
Like I am, it is a tool that Iuse, uh, but I need to make sure
(35:05):
that I have always control overit.
Prakash Chandran (35:06):
Um, as we start to close, just uh one more
question here, just onsomething that you had
mentioned, which is like, youknow, we talked about kind of
the age of like this newinterface, this personalization
and you know, the AI or ChatGPTbeing able to like tailor things
from a number of different datasources together to give the
user the exact right thing.
If I'm a business today and I'mthinking about the interfaces
(35:30):
that I'm gonna design, I'mstarting an application or I'm
trying to expose my businesslogic uh in a meaningful way,
how would you recommend thinkingabout that?
Like from the interfaces thatthey're creating to um, you
know, their integrations with AIapplications, like where would
you recommend that they start tomake sure that they're
(35:51):
future-proofing themselves?
Tony Casparro (35:55):
Um, you know, so if we're talking about like a
company that's making thingsmaking things for consumers, uh,
first of all, um, most peopleare starting to go to this role
of using voice or text todescribe what they want, right?
So if they come to your site,if you're having them navigate
through a lot of, you know,wizard wizards or menus and
whatnot, um, people are reallycoming to this world to expect,
(36:16):
like, here's what I need to do.
Like, you know, give me acouple options or you know,
allow me to use natural languageto tell you what I want and
give them that approach throughtheir site, right?
Making people learn how yourtool works and the complexity of
that.
Um, maybe there's a place forthat, but most people are
expecting now for the UI to betailored to them instead and not
the other way around.
(36:36):
Uh and when you talk to uhbusiness, business customers
now, uh really making sure thatyou have excellent
documentation, you have theability to search and ask
questions in it.
Um, developers get very, veryfrustrated when they can't see a
lot of good examples orinteractive tutorials or ask
good questions of thedocumentation.
These are kind of now stakes,uh table stakes for people
coming in to work with yourtool.
(36:57):
And so ensuring that um that isreally primed and ready to go.
I'm sure there's third partiesout there that offer this as a
service to set those guides up.
Um, but that is really, reallykey.
Um and so yeah, that is kind ofthe way to win people over
right now.
People are expecting easy,they're expecting low friction.
Um, and if you're like, hey,you have to read all this and
figure it out and muddle throughwhat we we've written, and
(37:19):
we're not gonna give you anyguides or help for on it, it's
like, well, just making itharder, you're reducing that
friction they know.
Prakash Chandran (37:24):
I I heard someone else kind of talk about
there's kind of the userexperience, but also kind of
like an agent experience whereyou have to make sure that
things are really welldocumented and described, just
like you would for a human thatlike needs to understand all the
elements of it.
But when you have also amachine that's working alongside
that human that can digestthings much, much faster, it's
really critical that yourbusiness logic, the way uh the
(37:47):
rules and the processes that youwant people and machines to
understand are really welldocumented.
So that makes a lot of sense.
Tony, you've obviously beenworking on lots of different
things.
There was a dev day.
I'd love to hear about what youhave been working on lately,
what's been keeping youoccupied.
Tony Casparro (38:05):
Yeah, so we uh just released uh Commerce inside
of ChatGPT.
Uh previously, if you werelooking for products like you
were I'm looking for a newexpresso machine or some shoes
or whatnot, we would show you alist of links inside of there.
Um they're really the links arenot ranked in any way, they're
according to what you want.
Um, but there was no ability toactually check out inside of
ChatGPT.
(38:26):
So one of our first kind oflike, you know, these new
third-party products inside ofChat GPT is you can just hit the
buy button and it appears inChat GPT.
So we work with Stripe as apartner to do all the credit
card processing, and then wework with our um merchant
partners, if that's like Etsyand Shopify right now, to do the
transaction.
So this is agentic shoppingwhere we say, here, we'll
(38:46):
collect the credit cardinformation, we'll talk to the
store, we'll orchestrate thetransaction, it's done.
Uh, we were involved in it onlysimply that we have a train, a
copy of the record so that youcan easily access it from that
chat conversation, but it wasall done for you between these
different partners.
So they facilitatedtransactions, they did all the
um in credit card encryption,whatnot, and then the shipping.
Um, and so there's this newworld of once again, inside of
(39:09):
Chat GPT, you have control overthings happening outside to even
do your shopping for you.
Prakash Chandran (39:13):
That's amazing.
So you're really just kind of afacilitation layer, and then
you're like bringing thesepartners together to make sure
that that transaction happensbased on the user's intent.
Tony Casparro (39:22):
Exactly.
And it's great for the user.
We've got the credit cardinformation on file, nice and
secure on our side, well, stripeside really.
Uh, and we're simply making iteasy to hit that buy button
inside of there and do this, doall your shopping right there.
That's awesome.
Prakash Chandran (39:36):
Just as we close, you know, coming off of
all the hecticness uh ofdeveloper day and everything
that you're doing with OpenAI,if you had to leave the audience
um with one thing, justsomething that you would like
them to invest more time doing,whether it be chat in within
ChatGPT or otherwise, what wouldthat be?
Tony Casparro (39:57):
You know, shockingly enough, it's simply
give it a chance.
Um, I have many friends who arein this with me.
They're co-u-engineers andwhatnot, and there's still a lot
of apprehension about, youknow, um its access or controls
or if it'll actually be useful.
Uh, and honestly, give it achance.
Uh, it's the same way it waswith Spotify.
I remember people telling me,like, it's great.
You have these, these, youknow, all the music at your
(40:18):
fingertips.
And I go, like, no, I I likethe way I've always done things
with things with CDs.
I don't want these new tools,and I didn't really give it a
chance.
Um have it do something, youknow, where you know it might
take you a half hour to likesearch through something on the
web or do some piece ofresearch, or you're looking for,
you know, everybody's alwayslooking for, you know, a new
product or, you know, anespresso machine or whatnot.
Have it do some of that taskfor you and see see how well it
(40:40):
can do at that task.
Um, if you're willing to haveit wire up to your code base and
just say, here's a task I wasgonna do today, do today, give
it a try for me and see how itdoes.
Um, these we've made thesethings trivial easy.
Like you don't have to log in,you don't have to make an
account, you can just go to chatgpt.com and just ask it a
question and try it out.
Um, it is uh it couldn't beless friction along that path.
(41:03):
Uh, and it might meaningfullymake a difference into your
day-to-day life.
You might be able to focus onthe higher level stuff that
makes you a better uh umengineer leader in your
workplace.
Uh, you might be able toleverage your time better, uh,
and that makes you a bettereffective uh person in your
field uh and uh all the morepowerful.
Prakash Chandran (41:22):
Tony, what a perfect place to end.
I really appreciate your timetoday.
Thank you so much, buddy.
Thank you very much, Rakash.
Appreciate it.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Burden
The Burden is a documentary series that takes listeners into the hidden places where justice is done (and undone). It dives deep into the lives of heroes and villains. And it focuses a spotlight on those who triumph even when the odds are against them. Season 5 - The Burden: Death & Deceit in Alliance On April Fools Day 1999, 26-year-old Yvonne Layne was found murdered in her Alliance, Ohio home. David Thorne, her ex-boyfriend and father of one of her children, was instantly a suspect. Another young man admitted to the murder, and David breathed a sigh of relief, until the confessed murderer fingered David; “He paid me to do it.” David was sentenced to life without parole. Two decades later, Pulitzer winner and podcast host, Maggie Freleng (Bone Valley Season 3: Graves County, Wrongful Conviction, Suave) launched a “live” investigation into David's conviction alongside Jason Baldwin (himself wrongfully convicted as a member of the West Memphis Three). Maggie had come to believe that the entire investigation of David was botched by the tiny local police department, or worse, covered up the real killer. Was Maggie correct? Was David’s claim of innocence credible? In Death and Deceit in Alliance, Maggie recounts the case that launched her career, and ultimately, “broke” her.” The results will shock the listener and reduce Maggie to tears and self-doubt. This is not your typical wrongful conviction story. In fact, it turns the genre on its head. It asks the question: What if our champions are foolish? Season 4 - The Burden: Get the Money and Run “Trying to murder my father, this was the thing that put me on the path.” That’s Joe Loya and that path was bank robbery. Bank, bank, bank, bank, bank. In season 4 of The Burden: Get the Money and Run, we hear from Joe who was once the most prolific bank robber in Southern California, and beyond. He used disguises, body doubles, proxies. He leaped over counters, grabbed the money and ran. Even as the FBI was closing in. It was a showdown between a daring bank robber, and a patient FBI agent. Joe was no ordinary bank robber. He was bright, articulate, charismatic, and driven by a dark rage that he summoned up at will. In seven episodes, Joe tells all: the what, the how… and the why. Including why he tried to murder his father. Season 3 - The Burden: Avenger Miriam Lewin is one of Argentina’s leading journalists today. At 19 years old, she was kidnapped off the streets of Buenos Aires for her political activism and thrown into a concentration camp. Thousands of her fellow inmates were executed, tossed alive from a cargo plane into the ocean. Miriam, along with a handful of others, will survive the camp. Then as a journalist, she will wage a decades long campaign to bring her tormentors to justice. Avenger is about one woman’s triumphant battle against unbelievable odds to survive torture, claim justice for the crimes done against her and others like her, and change the future of her country. Season 2 - The Burden: Empire on Blood Empire on Blood is set in the Bronx, NY, in the early 90s, when two young drug dealers ruled an intersection known as “The Corner on Blood.” The boss, Calvin Buari, lived large. He and a protege swore they would build an empire on blood. Then the relationship frayed and the protege accused Calvin of a double homicide which he claimed he didn’t do. But did he? Award-winning journalist Steve Fishman spent seven years to answer that question. This is the story of one man’s last chance to overturn his life sentence. He may prevail, but someone’s gotta pay. The Burden: Empire on Blood is the director’s cut of the true crime classic which reached #1 on the charts when it was first released half a dozen years ago. Season 1 - The Burden In the 1990s, Detective Louis N. Scarcella was legendary. In a city overrun by violent crime, he cracked the toughest cases and put away the worst criminals. “The Hulk” was his nickname. Then the story changed. Scarcella ran into a group of convicted murderers who all say they are innocent. They turned themselves into jailhouse-lawyers and in prison founded a lway firm. When they realized Scarcella helped put many of them away, they set their sights on taking him down. And with the help of a NY Times reporter they have a chance. For years, Scarcella insisted he did nothing wrong. But that’s all he’d say. Until we tracked Scarcella to a sauna in a Russian bathhouse, where he started to talk..and talk and talk. “The guilty have gone free,” he whispered. And then agreed to take us into the belly of the beast. Welcome to The Burden.